p

cj

THEORIETICAL AND EMPIRICAL ISSUES IN TOURISM DEMAND

ANALYSIS

by

Maria M. M. Q. de Mello, MSc

6SITY O

Thesis submitted to the University of Nottingham for the degree of Doctor of Philosophy, June 2001

CONTENTS

LIST OF TABLES LIST OF FIGURES ABSTRACT ACKNOWLEDGEMENTS

ix xi xii
I

CHAPTER 1: INTRODUCTION

1.1 1.2 1.3 1.4 1.5

The importance of tourismdemand analysis The objectivesof the study The significanceof the study Methodologyof the study Overviewof the chapters

1 2 3 6 10

CHAPTER 2: UK TOURISM DEMAND FOR FRANCE, SPAIN AND PORTUGAL SINCE 1969:FACTS AND FIGURES 2.1 2.2 2.3 2.4 2.5 Introduction World, European and UK tourism demand The importance for of France,Spainand Portugalasdestinations European tourists for France,Spainand Portugal The UK tourism demand Conclusion 17 22 34 13 13 14

CHAPTER 3: STATIC SINGLE EQUATION MODELLING OF TOURISM DEMAND 3.1 3.2 Introduction The single equation in tourismdemandmodelling approach 3.2.1 The difficultics of tourismdemand modelling 3.2.2 Singleequationmodelsof tourismdemand 36 36 37 37 39

3.3

demand in 3 Variables included tourism of equation models single 3.2.3 . 3.2.4 Static and dynamic single equation modelling of tourism demand Static single equation models of the UK demand for tourism in France, Spain and Portugal 3.3.1 Static single equations of the UK tourism demand for France, Spain and Portugal 3.3.1.1 Assumptions 3.3.1.2 Functional form and variables definition 3.3.2 Analysis of the estimated results 3.3.2.1 Presentationof the estimated results 3.3.2.2 Analysis of the estimated results 3.3.3 Comparison and implications of the estimated results

41 46

49

50 50 51 56 57 69 83 89

3.4

Conclusion OF THE

CHAPTER 4: DYNAMIC SINGLE EQUATION MODELLING UK TOURISM DEMAND

93

4.1 4.2

Introduction in UK tourism time series integration of Orderof andcointegration demandfor France,SpainandPortugal processes 4.2.1 Basicconcepts stochastic of nonstationary integration 4.2.2 Cointegration of time seriesvariables of andorder

91%

98 98 100

4.3

in UK demand the tourism 4.2.3 Orderof integration of variables equations 104 in the UK tourism Cointegration mechanism analysisanderror-correction SpainandPortugal demandfor France, for tourismin France 4.3.1 Tile UK demand for tourismin Spain 4.3.2 Tile UK demand 4.3.3 The UK demand for tourismin Portugal 4.3.4 Analysisof the estimated resultsfor the UK tourismdemandin France, SpainandPortugal 113 116 124 129 13") 141 148

4.4 4.5

Empiricalstudiesusinga comparable methodologyfor the analysisof UK tourismdemand Conclusion

CHAPTER 5: MODELLING THE UK TOURISM DEMAND WITHIN A STATIC SYSTEM OF EQUATIONS FRAMEWORK 5.1 5.2 5.3 5.4 Introduction Theoreticalfeatures of the AIDS model The staticAIDS modelof the UK demandfor tourism Empirical resultsandtheir interpretation 5.4.1 Analysisandinterpretation of the estimationresults 5.4.1.1Interpretation of the cocfflicients'estimates of the system 5.4.1.2Expenditure elasticities 5.4.1.3Own-priceelasticities 5.4.1.4Cross-price elasticities 5.4.2 Comparabilityof the resultswith thoseof other studies 5.5 Conclusion 152 152 157 162 167 167 169 172 173 173 175 176

CHA PTER 6: DYNAMIC SYSTEM OF EQUATIONS IN TOURISM DEMAND MODELLING 182 182 185 188 193 195 199
OF THE 203

6.1 6.2 6.3 6.4

Introduction DynamicAIDS modellingof the UK demandfor tourism Testingthe theoreticalconsistency of alternativcdynamicmodcls Empirical resultsandtheir interpretation 6.4.1 Interpretation of the elasticities'estimates Conclusion
MODELLING

6.5

CHAPTER 7: VECTOR AUTOREGRESSIVE UK DEMAND FOR TOURISM

7.1 7.2 7.3

Introduction Main features of the vector autoregressivemethodology Specification of a VAR model for the UK tourism demand in France, Spain and Portugal 7.3.1 Order of integration of the variables included in the VAR

203 207

215 215

iii

7.3.2 Determination of the order of the VAR 7.3.3 The unrestricted VAR specification of the UK demand for tourism 7.3.4 The forecasting ability ofthe reduced-form VAR I and VAR III models 7.4 7.5 7.6 Tile Johansenprocedure Forecasting with cointegrated VAR specifications Conclusion

219 220

228 232 243 248

CHAPTER 8: THE FORECASTING ABILITY . 8.1 8.2 8.3 ECONOMETRIC

OF ALTERNATIVE 252

MODELS OF THE UK TOURISM DEMAND

Introduction Merits and limitations of quantitative forecasting methods The forecasting performance of econometric models of the UK tourism demand for France, Spain and Portugal 8.3.1 Forecasting procedures:basic concepts 8.3.2 Evaluation criteria of forecasts 8.3.3 Tile forecasting performance of alternative econometric models 8.3.3.1 ARDL models 8.3.3.2 Static AIDS model 8.3.3.3 Dynarnic AIDS model 8.3.3.4 VAR models

252 254

261 262 263 265 265 267 269 272

8.4

Forecasting performance of econometric models for the UK tourism demand: comparative evaluation 274 289 Conclusion

8.5

CHAPTER 9: CONCLUSION

292

REFERENCES

306

iv

LIST OF TABLES

Table 2.1 Table 2.2 Table 2.3 Table 2.4 Table 2.5 Table 2.6 Table 3.1 Table 3.2 Table 3.3 Table 3.4 Table 3.5 Table 3.6 Table 3.7 Table 3.8 Table 3.9 Table 3.10 Table 3.11 Table 3.12 Table 3.13 Table 3.14 Table3.15 Table 3.16 Table 3.17 Table3.18

in Europe, World Tourism Participation of (billion US$) International tourismcurrentexpenditure Europe,France,SpainandPortugal'sshares of world tourism(%) France,SpainandPortugalshareof EU tourism currentreceipts($US million) Arrivals (million) and receipts(billion $US) of France,Spain and Portugal (f million) abroad UK tourismvisits (thousand) and expenditure income Log-lincar functionalform and aggregate and expenditure Log-lincar functionalform and per capitaexpenditure and income Linear functionalform andaggregate and income expenditure Linear functionalform andper capitaexpenditure and income income Log-linearfunctionalform and aggregate and expenditure income Log-linearfunctionalform and per capitaexpenditure and income Linear functionalform andaggregate and expenditure Linear functionalform andper capitaexpenditure and income income Log-linearfunctional form and aggregate and expenditure Log-linearfunctionalform and per capitaexpenditure andincome income Linear functionalform andaggregate and expenditure Linear functionalform andper capitaexpenditure and income 3 for the log-linear and linear forms of regression Elasticitiesestimates for Portugal for the log-lincar and lincar forms of regressions Elasticitiesestimates for Spain for the log-linearand linear forms of regressions Elasticitiesestimates for France Elasticitiesestimates of the demandequationfor Portugal for the demandequationfor France Elasticitiesestimates Elasticitiesestimates for the demandequationfor Spain

is

16 18 20

23 25 57 58 59 60 61 62 63 64 65 66 67 68 73

75
79 80 81 82

Table4.1

ADF(O)andADF(l) unit root testsfor the variablesin levels YF, YS, Y11, PF, PS,PF, I and their first differencesAYF, AYS, AYP, APF APS,APP,Al

Table4.2 Table4.3 Table4.4 Table4.5 Table4.6 Table4.7 Table4.8

Estimationresultsofthe ARDL modelfor the UK demandfor tourismin France Significance PS,. PFI-1.11-1 test for additionalvariablesYFt-1, 1, and Ft., Estimationresultsof the ARDL model for the UK tourism demand in Spain PFt_1, It. Additional variablessignificance test for YSt.1,PSt. i, 1, YPt.1and St., Estimationresultsof the ARDL modelfor tile UK tourism demand in Portugal Joint significanceF-testfor additionalvariablesYPt.i, PSt-i,PPtt_i, It-, and P,. 1 for the long- andshort-rundurnmycoefficients,longEstimates income, and own- andcross-price andshort-runelasticitiesof

118
122

125
128

130

133

Table4.9 Table4.10 Table4.11 Table 5.1 Table 5.2 Table 5.3 Table5.4 Table5.5 Table6.1

velocity of adjustment. in Spainand Portugal ECM for the UK tourismdemand Long-runincome.effectiveown-priceandsubstitute-price for tourismin Spainand Portugal elasticitiesfor the UK demand Estimationof the long-runandshort-runmodelsfor the UK in France tourismdemand andSpain Testsfor the individualandjoint significanceof variables D, SE andT UK Tourismdemand (restricted modelunderhomogeneity andsymmetry) Diagnostic testsfor the equations AIDS models of the unrestricted Expenditure anduncompensated price elasticities(restrictedmodel) Comparison oCclasticitics estimates of'UK tourismdemandin Portugal andSpain TestsI*or tile individualandjoint significanceof thevariables

137 143

144

146

165

168 168 172

175

vi

SEt-1, Dt-1andADt Table 6.2 Table 6.3 Table 6.4 Table 6.5 Table 7.1 Table 7.2 Table 7.3 Testsfor utility theoryrestrictionson the long- and short-run coefficients dynamicmodeland Estimationresultsfor the unrestricted modelsI and11 Diagnostictestsfor the equations of the dynamicAIDS-model Expenditure anduncompensated own- and cross-priccelasticities Unit root DF andADF testsfor the variablesWF, WS, WP, PF, PS,PPand E AIC and SBCcriteria andadjustedLR test for selectingthe orderof the VAR Estimationresultsandstatisticalperformance of the basic VAR I model AIC and SBC selectioncriteria anddiagnostictestsfor the VAR 11 model AIC and SBC selectioncriteria anddiagnostictestsfor the VAR III model LR testsfor restrictivehypothesis on the interceptand variablesof the VAR Forecasting resultsfor the UK expenditure shareof France Forecasting resultsfor the UK expenditure shareof Spain Forecasting Portugal resultsfor the UK expenditure of share Testsof the cointegration rank for VAR I and VAR III models Long-runcoefficients estimates fled the of exactly-identi share equations Testsfor ovcr-identifyingrestrictionsin VAR I and VARIII models Long-runestimates underexact-and of the shareequations over-identifyingrestrictionsof homogeneity, symmetryandnull cross-price effectsfor France and Portugal Table 7.14 Table 7.15 Table 7.16 Expenditure and uncompensated price elasticitiesderivedfrom the VAR III model Forecasting resultsfor the UK expenditure shareof France Forecasting resultsfor the UK expenditure shareof Spain

190

192 194 195 197 217

219 221 225 226

Table 7.4 Table 7.5

Table 7.6

227 229 229 230 236

Table 7.7 Table 7.8 Table 7.9 Table 7.10 Table 7.11 Table 7.12 Table 7.13

238 240

241
242 244 245

vii

Table 7.17 Table 7.18 Table 8.1

Forecasting results for the UK expenditure share of Portugal Forecastaccuracy statistics for different VAR specifications Forecasting results for the log of the UK real per capita expenditure in France

246 247

266

Table 8.2

Forecastingresults for the log of the UK real per capita expenditure in Spain 266. Forecastingresults for the log of the UK real per capita expenditure in Portugal 266 268 268 Forecastingresults for the UK expenditure share of France Forecastingresults for the UK expenditure share of Spain Forecastingresults for the UK expenditure share of Portugal Forecastingresults for the UK expenditure share of France Forecastingresults for the UK expenditure share of Spain Forecastingresults for the UK expenditure share of Portugal Forecastingresults for the UK expenditure share of France Forecastingresults for the UK expenditure share of Spain Forecastingresults for the UK expenditure share of Portugal Forecasting results of the ARDL model for the expenditure shareof France 276

Table 8.3

Table 8.4 Table 8.5 Table 8.6 Table 8.7 Table 8.8 Table 8.9 Table 8.10 Table 8.11 Table 8.12 Table 8.13

269 271 271 272 273 273 274

Table 8.14

Forecastingresults of the ARDL model for the expenditure shareof Spain 276 Forecasting results of the ARDL model for the expenditure shareof Portugal 277 278 279 280 282 Forecastsfor the UK expenditure shareof France Forecastsfor the UK expenditure shareof Spain Forecastsfor the UK expenditure shareof Portugal Absolute and relative forecast error range for the destination shares

Table 8.15

Table 8.16 Table 8.17 Table 8.18 Table 8.19

viii

LIST OF FIGURE S

Figure 2.1 Figure2.2 Figure 2.3 FigureA Figure2.5 Figure2.6

TouristsArrivals Growth Rate 1952-1995 World andEuropean arrivalsandreceipts1987-1997 Europe,France,SpainandPortugal'sshares of world tourism current receipts France,SpainandPortugaltourismreceipts($US million) in the world, Europe,and in France, UK tourismexpenditure SpainandPortugal UK tourist visits to the world, Europeandto France,Spainand Portugal in France(;Cmillion) UK tourismexpenditure UK tourist visits in France(thousands) in Spain(f million) UK tourismexpenditure UK tourist visits in Spain(thousands) UK tourismexpenditure in Portugal(f, million) UK tourist visits in Portugal(thousands) UK visits andexpenditure in Spainand Portugal1969-1979 UK tourismexpenditure (f) in Portugal France, Spain and per visit UK per capitareal expenditure in France,Spainand Portugal Effective pricesof tourism in France,SpainandPortugal UK real per capitaIncome in France First differenceof the UK real expenditure in Spain First differenceof the UK real expenditure in Portugal First Differenceof the UK real expenditure in France tourism First differences the of effectiveprice of in Spain First differences tourism the of price effective of First differences of the effectiveprice of tourism in Portugal First differences of the UK real per capitaincome CUSUM andCUSUMSQstatisticsfor the UK demand equationfor France

14 16 18 21

25 26 27 27 28 28 29 29 31 34 105 105 105 106 106 106 107 107 107 108 134 1X

Figure 2.7 Figure 2.8 Figure2.9 Figure2.10 Figure 2.11 Figure 2.12 Figure 2.13 Figure2.14 Figure4.1 Figure4.2 Figure4.3 Figure4.4 Figure4.5 Figure4.6 Figure4.7 Figure4.8 Figure4.9 Figure4.10 Figure4.11

Figure4.12
Figure 4.13

CUSUM andCUSUMSQstatisticsfor the UK demand

for Spain equation CUSUM and CUSUMSQ statistics for the UK demand
equation for Portugal

135

136 285

Figure 8.1

Models' specification featuresand rank

ABSTRACT

The majority of empirical studies of tourism analysis use a static single equation approach to model the demand for tourism of one origin for one or more destination countries. The interpretation that the and policy economic examination of such studies generally shows implications drawn as conclusions are based on mis-specified models, invalid estimation forecasting inconsistent inference estimates and poor procedures, performance. Static and demand interdependencies tend to tourism neglect equation models of among single destinations, ignore nonstationarity, overlook dynamics and, generally, disregard economic theory. Empirical specifications constrained by these flaws are bound to generate biased and inconsistent estimates upon which no reliable economic analysis or policy implication can be based.

In an anaytica context that fbcueson the UK tourism demandfor France,Spain and Portugalin the period 1969-1997, the main objective of this thesis is to demonstrate that consistentelasticities' estimatesand reliable forecastscan be obtained from empirical modelswhich arebasedon the principlesof economictheory, and specifiedand rigorously tested within the rules of sound econometric methodology. The alternative models 4 to 7 include error-correctionautoregressive distributedlag models estimatedin chapters ,ideal demandsystems(AIDS) (ARDL), static and dynamicalmost and cointegrated vector (VAR). models autoregressive The main findings that emergefrom the study are as follow. The batteryof diagnostictests applied to the dynamic error-correctionARDL models provide sufficient evidence to The classifythem as statisticallyrobust,structurallystableand well-definedspecifications. and evidenceobtainedfor the AIDS and VAR systemsindicatesthem as data-coherent The theoretically-consistent models,complying with the utility maximisationhypotheses. similarity, acrossmodels, of the estimatesof the long-run structuralparameters and the for forecasts further these the they accuracyof models provide supportthe reliability of behaviour,in contrastto the static single explainingandpredictingthe UK tourism demand be 7 in 4' 3. The to can easily equationsestimated chapter specificationsof chapters be loss destinations to to can adapted extended, without of generality, more origins and and alternative contexts such as the demandfor specific regions within a country, specific resortswithin a region or even specific types of tourism productssuchas accommodation or leisurefacilities, within a local area.

xi

ACKNOWLEDGEMENTS

I am deeply grateful to my supervisor, Professor Thea Sinclair, for her insightful comments, advice and constant encouragement while preparing the thesis. More than an excellent supervisor, Thca is a friend from whom I have learned invaluable lessons, which will help me in my future professional duties and researchquest.

My thanksare also due to Mr. Alan Packand Professors Anthony Thirlwall, Alan Carruth and Manuel Mcndes de Olivcira who were always helpful and generousin providing guidance, assistance andadvicc whenever asked.
I would like to thank ProfessorsAlistair Bruce and Paul Fenn for their comments on several chapters of the thesis which helped me to improve its contents. Thanks arc also due to the anonymous refereeswho commented on the model and results of chapter 5. I thank all colleagues in tile Christel Dellann Tourism and Travel Research Institute for their support and help and particularly, Sarah Gerard, whose competence, proficicncy and expertise always makes the most difficult problem appear simple and easy to solve.

I wish to acknowledge the encouragement and supportof my dear friends Cristina, Erica, Ria, Sophia,Jo5o,Jorge,Miguel, Owen,and Roy who madethe life of a PhD studentseem bearable almost and,at times,evenbright and fun.
I am deeply grateful to my parents, brother and sister and to my son, for taking care of me in the darker hours and for giving me constant support during my study years. Without them this thesis would not have been possible. Finally, I am grateful to Kevin for sharing the joys and pains of rcsearch in economics; for the exciting, inspiring, enjoyable and endless discussions about econometric modelling and for lifting it is but (seems in ); theory terms, a contradiction not! economic my spirits high when trouble was in sight, and keeping my feet down when enthusiasm was excessive; for devastating ideas 'original' constructive comments on my and criticisms on my selfindulgent concepts; for brightening my laughs and drying my tears; for listening, reasoning, caring, reassuring, encouraging and comforting; for making my life a wonderful place to live...

xii

CHAPTER 1
INTRODUCTION

1.1. THE IMPORTANCE

OF TOURISM DEMAND ANALYSIS

In the last half century, tourism has becomeone of the world's most important
economic activities. Between 1950 and 1997, international tourism receipts increased at a remarkable annual average rate of 12% and although this growth rate has been slowing down since the mid 1970s, it remained 7% over the last decade, more than double the world's GDP growth' (WTO, Yearbook of Tourism statistics, 1998, p.2-3). During 1997, 617 million tourists travelled the world spending, $US 447 billion, and the World Tourism Organisation (WTO) predicts that by 2020, there will be 1.6 billion tourists spending $US 2,000 billion. According to the same source, tourism is responsible for more than 220 million direct and indirectjobs and, if complementary economic activities linked to tourism are included, tourism accounts for about I I% of the world's GDP.

Thesefigures illustrate the importanceof tourism in the world economy.Indeed, income for jobs, generates revenuefrom foreign tourists createsand sustains additional foreign exchange private andpublic entities,alleviatestradedeficits, increases and reserves financeimports, acts as a catalystto investmentand, overall, contributesto the economic growth of destinationcountries.In addition, tourism often requiresinvestmentin physical and human capital, and. its returns can be realised relatively quickly. Therefore, the activities associated with tourism are generally consideredvital to assistdevelopmentin local andregionalareas of developingand industrialized countries.
As tourism emergesto a centre-stageposition in regional and national economies, it is important to measure and predict tourists' demand patterns and behavioural features.

Whetherand how muchnet economicbenefitstourismbrings to destinations depends on

1The World EconomicOutlook Database that the world GDP growth rate in the 1990sis of the IMF states 3.4%.http://www.imf.org/extemal/pubs/ft/weo/1999/01/data/ngdpLa. csv.
I

the precise form and scale of tourism demand.The key role of tourism demandin the knowledge determinants its successof many economic activities requires of main and forecasts accurate of its future levels.Not surprisingly,in the last threedecades, research on the economicsof tourism has been a growing area of interest for both businessand academic sectors.
1.2. THE OBJECTIVES OF THE STUDY

Evaluationof the magnitudeanddirection of the impactsof tourism expenditure on destinationsrequires estimatesof current and future demand. Hence, tourism demand have forecasting been and studies a growing area in economic literature. modelling However, the majority of studies of tourism demand analysis fail to incorporate the in the constructionof accurateand tools fundamental theoreticalbasisand methodological for models explainingandpredictingtourismphenomena. reliable Empirical studies of tourism demand generally involve the use of econometric models to specify the relationshipsbetween the demand levels and its determinants. Econometricmodelling provides a good basis for forecasting,which is of considerable value for public policy and an important element in public and private investment decisions.However, modelling tourism demandpresentsseveral difficulties, which are First, most of the time seriesusedin the estimation mainly linked with its specific features. Models that overlook this featureof tile of tourism demandare trendedor non-stationary. datamay give rise to spuriousregressions, invalidating statisticalinferenceand forecasting Second, theoryhypothesisshouldbe integratedand testedwithin the procedures. consumer quantitativeframework adoptedto model tourism demand.Models that arc theoretically inconsistentdo not serve well the purposesthey aim to achieve. Third, the inherent dynamicnatureof tourism demandand the possibleexistence feedback effects requires of the explicit incorporation of a time dimension and the consideration of short-run adjustments within a systemof equations structure. The difficulties encounteredin the construction of the models are frequently overcome by means of simplifying assumptions,which permit the specification and estimation of quantitative relationships explaining and predicting tourism demand behaviour.Dependingon the assumptions underlying the constructionof the model, the subsequent specificationwill differ both in its static or dynamicnatureand in the variables

included,functional form adoptedand estimationmethodsused.Different modelsperform differently, being more or less reliable accordingto the theoretical and methodological frameworks within which their specification is fon-nulated.Tourism demand models assumptions, overlooking dynamicsand feedbackeffects, constructed under questionable ignoring the spuriousregression problem, neglectinginterdependencies amongcompeting destinations, and lacking the theoreticalbasiswithin which testablehypothesis of consumer theory can be included, generally give rise to misspecification bias, unreliable and misleading estimation results and overall invalid statistical inference and forecasting procedures. The literatureconcerningtourism demandanalysisshowsthat the large majority of have studies used a static single equationapproachto model an origin's empirical early demandfor tourism in one or more destinations. Thesead hoc modelstend to be basedon lack consistency implausibleassumptions, with consumerbehaviour theory and, as Witt (1995, 458) Witt observe, "the quality of the empirical results obtained is p. and few in A tourism demand behaviour have attempted to studies recent questionable". demandmodelsby with the traditional single-cquation overcomethe problemsassociated However, consideringone or more of the problematicaspects of their modelling strategies. to the bestof our knowledge, theseproblemsin a systematic nonehasyet addressed way. Against this background,the main objectives of this thesis are to discuss and implement alternative methodological approachesto tourism demand analysis which the theoreticalfoundationsof currently usedmodels,and contributeboth to strengthening in econometric to apply andevaluate recentadvances modelling,quality evaluationcriteria, hypothesis in a tourismdemand testingand forecasting procedures context.
1.3. THE SIGNIFICANCE OF THE STUDY

The increasing importance that, in recent years, has been attachedto tourism demandanalysis clearly demonstrates the need to extend the theoretical and empirical the empirical findings contentof the existing literatureand,more important,to substantiate with appropriate and extensivetesting.In this context,this study makesseveralimportant for The demand its geographically UK thesis tourism the contributions. analyses proximate France,Spainand Portugalusing datafor the period 1969-1997. The choiceof neighbours, the countriesinvolved in the empirical analysistook into accountthe fact that the UK is a

Portugal importance France, Spain is to and as tourism origin, which of particular major destinations.Spain and Portugal are interesting cases to their owing consideration -for beginning during At in the the the transition of sample period. economies as position dependence high in displayed 1969, they underdevelopment: symptoms of classic period, income i low low industrial fishing, lack standardsof sation, and of on agriculture and final indicators By 1997, health the welfare. of social yearof education, provision and other European the period understudy,they hadjoined the ranksof more developed economics. France was a high-income country over the whole sample period, allowing for useful The its it between neighbours. concept of neighbourhoodalso poorer and comparison determined the choice of countries permitting analysis of the destinations' dynamic its interdependencies. behaviour and competitive It is well known that tourism demandis responsiveto such variablesas income, is known What is how the responsiveness and exchange rates. not prices relative of demand in thesevariablesaltersduring a country's economictransition and integration to changes into a wider internationalcommunity.Relationships of substitutabilityor complementarity income lower destinations time as over emerge from relative poverty to may change development. information higher levels Little is availableabout whetherlower of achieve
income destinations tend to become more or less competitive over this transition period, either relative to other developing countries or relative to more industrialiscd nations. This

interesting insights theseissues. studyprovidessome concerning The vast majority of studies 6f tourism demandhave relied on single equation 1984; Crompton, (for 1982; Loeb, Uysal models within a static context example, and Gunadhiand Boey, 1986,1ee et al., 1996).Thesemodelsare not derivedfrom consumer demand theory, generally disregard dynamics, the non-stationarityof the timc scries involved, and potential simultaneitybias. As a result, the estimationresultsare unreliable, is the statisticalinferenceinvalid andthe forecasting ability of suchmodels so poor that, as showedin Martin andWitt (1989)andWitt andWitt (1992),eventhe simplestof univariate forecasts. time series accurate the more supply can modelnaYve no-change specification In this study, the specification deficiencies, theoretical flaws and technical inadequacies associatedwith traditional tourism demand modelling are examined and overcomeby the constructionand estimation of alternative econometricspecifications both theoretically consistentand empirically plausible. These specifications which are incorporate dynamics through short-run adjustment mechanisms; consider the inter-

dependences among destinationsand test consumertheory hypothesisusing system of bias Sims' (1980) using vector autoregressive approach, equations;avoid simultaneity and prove the existence of a long-runequilibrium relationshipbetweenthe UK tourism demand (1988) cointegratingvector analysis.The and its determinants accordingto the Johansen in this studyare subjectto rigorousquality scrutinyunderthe rules of the modelsestimated testing,causalityand most recenteconometric suchas structuralconstancy methodologies, Finally the alternativemodels testing,cointegration exogeneity analysisand encompassing. are comparedin their forecastingability and reasonsfor their different performances are analysed. An importantapproach to tourism demandanalysisinvolving systems of equations is the Almost Ideal Demand System (AIDS) of Deaton and Mucllbaucr (1980b). This demandtheory and is particularly valuablefor the principlesof consumer model embodies testing the theoreticalhypothesis of homogeneity and symmetryand for the estimationof cross-price elasticities between competing destinations. Some studies, for example, O'llagan and Harrison (1984) and Syriopoulosand Sinclair (1993), investigatetourism demandusing the orthodox static AIDS approachof Dcaton and Mucilbauer.More recent studiesusing the AIDS approachadd a trend and or other dynamic-like elementsto the orthodox model. In contrastwith previousfindings where homogeneityand symmetryare systematicallyrejected, these 'unorthodox' models seem to supply a 'quasi-dynamic' to supporttheir consistency structurenecessary with the constraintsof consumerdemand 1999; De Mello et al. 2001). Nevertheless, theory, (Papatheodorou the orthodox AIDS model, with or without trend, is derived within a static framework. Specific researchon tourism demand using an explicitly dynamic AIDS system of equations is virtually in issue best knowledge, Lyssiotou (2000) this to the nonexistent of our and, addresses only an empirical study concerningthe dynamicsof adjustmentbehaviourwithin a systemof equationssimilar to the AIDS model. In this thesis,we constructand estimatea dynamic AIDS model which, besidesthe desirablepropertiesalreadypresentin its static version, behaviour. explicitly addsthe fundamental time dimensioninherentto tourismdemand In tourism demandresearch the modellingof dynamicshasgenerallybeenconfined based to the useof error correctionsingleequationspecifications on the Engle and Granger (1987) two-stages approach, (Kulcndran, 1996; Kim and Song, 1998; Vogt and Wittayakorn, 1998;Song et al., 2000). One disadvantage of this methodis that the usual the quality of modelarenot strictly valid andcannotbe usedfor test statisticsto evaluate

doesnot prove that the cointegrating is that this approach inference.Another disadvantage One long-run is is found) (if the alternative equilibrium relationship. unique regression one derivation disadvantages, the these of an error consists of method,which may overcome (ARDL) distributed lag based as model autoregressive on a general correction model in battery Pesaran The (1995,1996). developed by Pesaran Shin tests of et al. proposed and (1996) can then be appliedto confirm or reject the existence of a long-run relationship.In the existing literature, only Song and Witt (2000) supply a condensed example of an demand This in thesiscontributesa chapterto tourism this context. a method applicationof the analysisof the ARDL error correctionmodel. (VAR) approachand the cointegrationvector Although the vector autoregressivc increasingly in have been for used most areasof economicresearch analysis autoregressive the last two decades, on tourism demandanalysishave generallyignored this researchers Kulendran King (1997) Exceptions technique. are and new and Kulendranand Witt (2001). in tourism demandanalysisis To overlookthe importance of the VAR modelling approach toot for the estimationand forecastingof the long-run to leave out a reliable econometric in its determinants. impactson demandinducedby changes Indeed,the VAR approach can identification issues bias as spurious regression, simultaneous overcomesuchproblems and in included from behavioural features the the the theoretical variablesand arising natureof In this thesis we use a the quantitativerelationshipslinking demandto its determinants. for demand UK VAR to the the system estimate structural coefficients of cointegrated tourism. 1.4.METHODOLOGY OF THE STUDY
A major debate in the empirical modelling of consumer preferences is associated

is level individual data issues. data the The at with advantage of modelswith aggregation that they avoid potential aggregation bias. However, reliable longitudinal data sets in following the sameconsumers tourism long time rare and, generally are of over periods is often unavoidable Therefore,someform of aggregation contexts,virtually nonexistent. in been have the this made. case and, of study,no otherchoicecould Throughout the thesis, a great deal of attention is paid to the analysis and included in the models. have interpretation the time the that series path of of events affected The political eventsthat took placein Portugaland Spainduring the 1970s, the changes that

into in EU following integration the in their destinations precedingand occurred these 1986,the openingof the ChannelTunnel in 1994, and events,such as the oil shocksin demand from tourism 1973and 1979,which affectedeconomies a worldwide, are analysed The hypotheses. in integrated the statistical plausible under models perspectiveand relevanceof the variables representingsuch events suggestthat the empirical results historical facts knowledge if of general a robust obtainedare only meaningfuland relevant integrated in is fully involved, the modelling the time and adequately affecting series is fundamental time Therefore, the a of series properties analysis an extensive procedures. demand descriptive The tourism time of the analysis series methodology. of empirical part is carriedout in chapter2. The thesis containsa theoretical,econometricand analytical content.Each chapter issues discussion derivation the theoretical thorough the of a with underlying of proceeds in the modelsandeconometric methodology applied their estimationandtesting. The static single equation models used in chapter 3, serve the purpose of demonstrating the methodologicalflaws of the traditional approachin tourism demand framework dynamic The the theoretical of single equationerror correctionmodels analysis. Pesaran Shin (1995,1996) (1996), follows Pesaran 4 that of and et al. and of chapter incorporating the short-run dynamics missing from the static models, and permitting however, in by basis. A cointegrationanalysis an equation equation systemof equations, destinations, interrelationships efficient method of modelling among can provide a more demand imposition testing the theoretical and of restrictions of consumer and permits behaviour.Moreover, the cointegrationanalysis implementedon a systembasis is both Hence, basis. a system than that performedon a singleequation more efficient and reliable 7. 5,6 is in derive to and the modelsestimated chapters approach used of equations As pointedout in Granger(1981,1990), Harvey (1990,1993),Hendry (1987,1995), Hendry and Mizon (1978), Hendry and Richard (1982,1983), Learner (1983,1987), Phillips (1986), Sims (1987) and in many other studies, the appropriateeconometric modelling of quantitativerelationshipsrelies on a thorough examinationof the economic time seriesincluded,as their statisticalpropertiesreflect the features of the datagenerating The analysisof the by the empirical specifications. processwhich must be approximated 4 and 7 and standard time seriesincludedin the modelsis carriedout in chapters unit root testsof Dickey and Fuller (1979,1981) are performedto establishthe order of integration from low in involved. However, tests these the suffer may power small samples variables of

In breaks. be biased such cases,other the structural tend to contain variables when and (1988) the Peron test, to Phillips variables' the applied establish are and as such methods, orderof integration. The selection of functional form to representconsumerpreferencesis a very important issue in empirical studies of demand behaviour. The ability to model the functional form in a choosing pertinent way relies on preferences' structure an appropriate The being is both tractable restrictive. econometric without excessively and adequate which based in derived this on mainstreameconomic theory specifications. are study models However,they are modified in severalways to apply to a tourism analysiscontext, with between destinations features the the the to relationships origin and attached specific destinations themselves. the among and considered Within a systemof demandequations, thereare many flexible functional forms that indirect functions be the to consumers' utility approximate or cost used andtheseforms can in differ their approximationproperties.Although there are many other substantially may for (see, Lewbel, 1987), flexible functional the example, of possibilities set classifying forms usually adopted in empirical demand analysis can be divided into three major locally flexible functional include 2001): forms, (Fisher the translog which el al., subgroups et al. (1975) and Jorgenson et al. (1980),the AIDS specificationof modelsof Christensen Deaton and Muellbauer (1980) and the generalised Leontief model of Caves and Christensen (1980); globally regular functional forms, which include the minflex Laurcnt in, for (1985,1987), Barnett discussed the generalexponential example, et al. models Banks McLarcn (1996) AIDS Cooper et al. the of model and and quadratic model of globally flexible forms, which include the Fourier flexible model (1997); asymptotically in Chalfantand Gallant (1985)andthe asymptoticallyideal modelof Barnettand discussed Yue (1998). Given a specificdataset, somefunctionalforms will generallyhavemore desirable and well-founded choice of one approximationpropertiesthan others.A knowledgeable definition data for the imply the and set, same the would specificationand estimationof all be form to based functional considered over-perform could of quality criteria, on which one the others.Suchan extensiveanalyticaleffort could be viewed as a subjectof research on in its own right, and is, therefore,beyondthe scopeof this study but shouldbe considered future research.

As Hendry (1995) points out, the existence of a potentially large number of theoretically consistentmodels, which satisfy the required quality criteria makes model is judged by The criteria which a good empirical model are choice a non-trivial problem. failure inadequacy indicate but the since of any may necessary not sufficient requirements, of the model, but the fulfilment of all gives no guaranteeof the model's ongoing in chapters5 applicability. The choiceof the functional form for the systemsof equations in mind, as well as the fact that, with the exceptionof the and 6 had theseconsiderations AIDS model,all the alternativefunctionalforms mentionedabovespecifynon-linearshare equations. in chapter5 is basedon the static AIDS model of Deaton The systemof equations and Muellbauer (1980) and the system of equationsin chapter 6 is a dynamic AIDS specification basedon the models of Anderson and Blundell (1983,1984). The AIDS is spccification with considerable model seenasa particularlyconvenient attractivefeatures and restrictions,can nest a variety of alternative which, with appropriatetransformations functional form In the of the AIDS equations addition, models. gives an arbitrary first-order demand to system,satisfiesthe axioms of preferences any approximation exactly, permits over consumers, and allows for simple linear estimationmethodsand perfect aggregation the impositionof linear restrictionsto testhomogeneity andsymmetry. However, the system approachesof chapters 5 and 6 rests on an a priori division of variablesthat may be questionable, endogenous-exogenous and the time series ' includedin the systems Hence,the estimationresultsobtainedfrom these are nonstationary. models can be deemedspuriousand the statistical inferenceinvalid, if no cointegrated of relationship(s)are found linking the variablesof thesespecifications.In the presence for feedback time series and potential non-stationary effects, an efficient approach for long-run be allows all estimating relationship(s),must a systemof equationswhich variablesto appearas dependent variables,and for appropriatecointegrationanalysis.An by Sims (1980) first these econometric and proposed methodology with was characteristics is usedto specify the models of chapter7. The econometricmethodologyapplied to the VAR modelsof this chapterdraws extensivelyon the conceptsand techniques showedin Engle and Granger(1987,1991), Granger(1988,1997), Harris (1995), Johansen (1988, 1996),Johansen (1998)andPesaran andJuselius(1990),Pesaran et al. (1996). No economicanalysisbasedon quantitativespecificationsis completewithout the forecasting their of performance.The predictive ability of the econometric examination

described in following is the views and procedures modelsused evaluatedand compared Newbold Granger (1986), Granger Fair (198 1), (1987), Yoo Engel and and suchstudiesas (1986), Newbold and Bos (1994) and Clementsand Hendry (1998). All the estimations, Pesaran forecasting inference tests, and were computed with procedures and statistical Pesaran's (1997)Microfit 4.0.

1.5. OVERVIEW OF THE CHAPTERS

Chapter2 explains,by meansof basicstatistics,graphsand tables,the evolution of the UK tourism demandfor its southernneighbouringcountries over the sampleperiod 1969-1997.The analysis of the data in this chapter I's used in subsequent chaptersto basis for the characterisation of relevant variables and help the economic a provide interpretation of the resultsobtained. Chapter3 begins with an overview of early empirical researchin tourism analysis have been the that evaluating econometric critically and models usedto estimate explaining for example,Little, 1980and Witt, 1980have Although someresearchers tourism demand. by including dynamics introduce lagged to attempted variables in their otherwisestatic models,the literature showsthat the large majority of investigatorssuch as, Gray (1966), Artus (1972), Barry and O'llagan (1971), Jud and Joseph(1974), Kliman (1981), Lin and Sung(1983),Papadopoulous and Witt (1985),Gunadhiand Boey (1986),haveuseda static destinations. demand for to tourism more model an origin's singleequationapproach oneor The chapterexamines the estimationresultsobtainedfrom different static single equation for demand UK focusing the' models, on the comparisonof alternativespecificationsof tourism which are estimatedusing different definitions for the variables and different functional forms. The analysis shows that slightly different models can produce inconsistent different thereby results upon which no considerably estimates, providing lack from be the based. disparities The to of a sound reliable conclusions seem emerge can framework theoretical within which plausible empirical methodologyand/or a consistent consumerbehaviourhypothesiscan be fully integratedand tested.Modelling procedures inadequate by deficiencies bound these to empirical specifications constrained produce arc biasedandinconsistent estimationresults. which generate In chapter4, a more reliable approachwithin the single equation framework is derivation integrates by the the specification, of an correction error which considered
10

dynamicdimensionof tourism demandbehaviourabsentfrom the static version.11o... the inter-temporalnature of tourism demandis not the only feature missing from static identified by the utility maximisation singleequationmodels.The behavioural assumptions hypothesis amongcompetingdestinations andthe interdependencies are also neglected and cannotbe fully integrated andtestedwithin a singleequationframework. Chapter 5 examinesthe UK demand for tourism within a system of equations basedon the AIDS model.This modelallows for the integrationandtestingof the approach utility maximisationhypothesisand for the estimationof cross-equation effects in a way functional forms. The AIDS model is formulatedwith not possiblewith other alternative betweenorigin and the introduction of some innovations:the conceptof neighbourhood destinationsand among destinationsthemselves, which is believed to be relevant in the explanation of the competitive behaviour of destinations;the concept of development transition periods,which appearsto affect the destinations'ability to capture increasing foreign tourism receipts in different ways; and the consideration of a non-constant coefficient of the expenditureexplanatoryvariable, which is believed to changedue to factorsthat modify the political andeconomicrelationships between the countriesinvolved. Unlike earlier studiesusing the orthodox static AIDS approach,from which the findings to reject utility theory hypotheses, the "unorthodox" model of chapter5 is well appeared defined,data-coherent andtheory-consistent. Nevertheless, it is possiblethat currentbudgetshares for demand UK tourism the of its southernneighbours dependnot only on currentprices and expenditurelevels, but also desequilibriurn in previousperiods.Tourists' preferences on the extentof consumption may havebeenunstableand the parameters of their utility function may haveshifted over time. In this case,a short-rundynamicmechanism, demand taking accountof the adjustment of towardsits long-run equilibrium, ought to be considered. Theseinter-temporalaspectsof tourists demandbehaviourhavebeenlargely ignoredin the literature,and previousstudies consideringsystemof equations on purely static models approaches generallyconcentrate in order to test the assumptions has More the theory. recent research recognised of utility importance in a of dynamicsin tourismdemandanalysisbut studiesincluding theseaspects frameworkare still rare. The derivation and estimationof a dynamic systemof equations AIDS modelin chapter6 contributes to fill this gap. The dynamicmodel of chapter6 permitsthe estimationof separate long- and shortin pricesand expenditure haveon UK tourismdemand.The longrun effectsthat changes

from from the to those staticversion this are similar obtained model obtained run estimates between the to the existenceof a steadystateequilibrium relationship and seem confirm Nevertheless, UK tourism demandand its determinants. the AIDS systemsof chapters5 in levels 6 their equationsare exogenously the that the of regressors current and assume be in Furthermore, determined, the time the the this used case. series might not although in the variables the equationsare cointegrated, estimationsare nonstationaryand, unless this may give rise to spuriousresults.In chapter7, cointegrationestimationtechniques are both VAR to overcome problems of to equations, permitting of system applied a identiflcation arising from unfoundedassumptions on the endogenous/exogenous natureof the variablesinvolved, and to confirm the existence of structuralrelationshipsbetweenthe for eachdestination. UK tourismdemand andits determinants Chapter8 carriesout a comparative quality evaluationof the econometric modelsof forecasting by 4 7, to their contrasting relative chapters ability over the out-of-sample for 1994-1997 their different performances. The analysis and analysing reasons period forecasters, that the cointegratedVAR model although all models are good establishes Chapter 9 the others. presentsa summary of the main findings and puts over-pcrforms forward some generalconclusions,which might have interesting implications for policy in future purposes research tourismdemandanalysis.' and

12

CHAPTER 2
UK TOURISM DEMAND FOR FRANCE, SPAIN AND PORTUGAL SINCE 1969: FACTS AND FIGURES

2.1. INTRODUCTION

The objectiveof this chapteris to examinethe UK tourism demandfor France,Spain 1969-1997, data in from Portugal the using period collected several sources.The and descriptiveanalysisof the data aims, on the one hand, to give a generalview of the UK tourism demandtrends within the world and Europeancontexts and, on the other hand, to destinations tourism the of suchas France,Spain and Portugalrelativeto examine relevance in world and Europeanterms.The analysisfocuseson the significanceof other destinations for Portugal, the UK tourismdemand France, Spain to that on and other of origins relative and Knowing importancefor UK touristsof thesecountriesrelative to other tourism destinations. their been how have UK time tourists on spending much and money when, where and holidays, helps to provide a context for the quantitative approachadoptedin subsequent definition the of Indeed, the appropriatecharacterizationof relevant variables, chapters. econometricrelationshipsand the interpretationof results arc generallygroundedon a good time behaviour features facts the series economic the trends, of understanding affecting of and for is information therefore, A a sine qua non understudy. thoroughexaminationof available for France, Spain demand UK tourism the and an accurate andreliable quantitativeanalysisof Portugal. The structureof the chapter is as follows. Section 2.2 provides an outline of the behaviourof world, Europeand UK tourism demandover time. It also analyses the relative importanceof UK demandin world and European terms. Section2.3 investigates the world 13

und Europe's most important destinations and considers the position of France, Spain and

Portugal relative to other destinations in world and European terms. Section 2.4 analyses UK tourism demand flor this set ofcountries in visits and expenditure terms. Section 2.5 concludes. 2.2. WORLD, EUROPEAN AND IJKTOURISM

DEMAND

Tourism growth has been impressive during recent years. The number of' tourist ;Arrivals in all countries increased from 69 million in 1960 to 617 million in 1997 (WTO,

1999). In spite ofthc slowdown in the growth rate ofarrivals

since 1960 as depicted in figure

2.1, this variable increased by a factor ol'23 between 1950 and 1997. Figure 2.1: Tourists Arrivals Growth Rate 1952-1995

12 10 8 6 4 2 0 19521960
Source: Yearbook ofTourism

[310.60% 09.10% 05.50% 4.80%

no, 10

04.80% W4.10%

19701980

19901995

Statistics, 1995-1999, W'FO.

Although

econornic and political

factors such as the oil criscs in the 1970s, the in 1990-

economic recession in the 1980s, and the political instability in the 1990s (Gulf'crisis

91 and war in Yugoslavia in 1993-95) may have had a negative influence on tourism flows, world tourism has continued to grow over the last decades and will continue to grow in the next. Indeed, the World Tourism Organisation (WTO, 1999) prcdicts an increase ofits growth

in the years 2000-2010 and that by 2020,1.6 billion tourists visiting countries abroad ratc
antILKIlly will spent around 2,000 billion $I JS. 14

With 15% of the world populationand one-thirdof world GDP, Europeis the largest 60% in tourism, of total arrivals and 52% of total tourism receipts participant world receiving Europehas beenthe world's leading in 1998(Tourism highlights, WTO, 1999).For decades, tourism contributorandits shareof the global tourismmarketis still the largestcompared with that of other regions,althoughits growth rate has beendeclining in recentyears.Until 1980, internationaltourism was typically a North American and Europeanphenomenon but since then, new sources of demand and new competitive destinationshave been playing an importantrole in challengingEurope'sleadingposition.Possiblecausalfactorsfor this change include the lower growth rate of long distanceair travel fares comparedwith short distance fares; competitiveprices and political and economicstability generallyoffered by new long haul destinations; tourist saturation and environmental degradation in some traditional Europeandestinationsand the slowdown of economic growth in main origin countries. However,accordingto the WTO predictionsfor the next decade, Europeremainsthe world's destination losing important 25 percentage tourism although most points of its 1970's share, in Asia, South destination America and Africa will record substantial countries while new period. growth in the sarne Table 2.1 and figure 2.2 show the relative importance of Europe as a tourism in destinationin terms of arrivals and current receiptsmeasured, in respectively, million and (WTO, 1999). thousand million $US for the period 1987-1997
Table 2.1. Participation of Europe, in World Tourism
ARRIVALS (million) RECEIPTS (billion US$)

World Europe Europe/World

11996 1997 1987 1990 1993 1996 1997 1987 1990 1993 1 362 456 519 595 617 175 266 321 425 447 ] s 147 163 220 218 9 226 282 312 351 361 1 1 1 1 55.3 50.8 51.8 48.8 ;j 58.5 8.5 62.4 61.8 60.1 .8 .0

Source: Yearbook of Tourism Statistics (1987-1999), WTO

15

Figure 2.2: World and F,uropcan arrivals and rccclpts 1987-1997

700 600

100

200 100

1997

198S

I i)xt)

1990

1(p) I

pp)

p)')

1'), I

199,)

199(,

pp), Ict-k-ipis (hillioll 1114,1w. 111 WS)

0 Nkorld arrivals (million) Yearbook . sourcc: of Fourism

0 Furopeall ill-l-kill" (Illillioll) Statisitics ( 1987- 1999), WM

N World I-CL-cipts(hillioll

$1 111) aI

In tourism

terms, Furope holds the world", Around 5W/0 of' dic world

leading

position

not onl as a (lest Mat 1011

hut also iis an origin.

CXpendittirc

'I11F M'Opc on fourr"'111 orl g 111,11cs

importaill livc fi the most and mr Out of'

tourism or Igins '11-C'111,0peall countrics. TabIc 2.2 1'. 111-opc mid that of dic likle

international the tourism the expenditure of' \wrld, Current show's most Important origins I'M-the period 1987-1997 (WII

Table 2.2: hilernational tourism CurrentCXPCII(Iittll*C (billion I ISS)

7 World USA Germany IJK France
['111-ope/world ('Vo) IJK/W ()rI (1 IIK/Iti ro pc((6)

1990 266 135 37 34 19 12 14

5 P/o 7 (/o 14"")

1993 321 145 41 41 1 1., -14 45!

199S 99 1 40 -52 24 16 12

1996 42 5 208 51 25 16
60 11

1997 447 217 51 46 '-)8 16

(P .4 6" )

175 97 -)g 25 12 8 5
5W/o 7% 14()/o

/) N) o

6(!/)

Source: Yearbook of tourism statistics ( 1987-1997), WI ()

16

As an origin country generating receipts for other countries, the UK has always been in an important position, both in European and world terms. Although UK tourism expenditure inore than doubled in the last decade, its share of the world expenditure has been slowly decreasing,mostly due to the impressive growth rate of world tourism. The world's increasing trend in tourism spending is particularly notable since 1990. Between 1990 and 1993, the world tourism budget increased by 21% and between 1993 and 1996 by 32%. Although the UK tourism budget increasedby the same rate as the world budget between 1993 and 1996, in the previous three years, UK tourist spending remained the same. Nevertheless, UK tourists contribute around 6% of the world tourism expenditure, securing the third position as the world's most important tourism origin. The UK average share of Europe's tourism expenditure has bcen relatively stable and around 13% since 1985. This share indicates the UK as the secondmost important tourism contributor of Europe. As shown in table 2.2, roughly half of the world tourism budget has been spent

touristssincemid 1980's.Germanyand the UK are the most important annuallyby European European origins and are the secondand third origins, respectively,in world terms.Germany and the UK represent more than one-thirdof Europe'stourism budget.However,since 1995, Germany'stourism spendinghas been decreasingin both relative and absoluteterms. This fact, compounded with the growth rate of UK tourism expenditurebetween1987 and 1997 (133%),which largely surpassed that of Germany(85%), indicatesthat the UK may become the most importantorigin in Europe. For a comprehensive analysis of tourism demandbehaviour it is also important to investigatehow tourists distribute their budgetsaround the world and which countriesare focusing In the next section,we investigatetheseaspects, preferreddestinations. particularly for European tourists. on the importance of France,SpainandPortugalasholiday destinations 2.3. THE IMPORTANCE OF FRANCE, SPAIN AND PORTUGAL AS

DESTINATIONS FOR EUROPEAN TOURISTS In the late 1990s, for 60% of world touristsand European countrieswere destinations recipientsof half their tourism budgets.However,Europe'sshareof world tourism has been declining continuouslyin the last two decades. Since 1980,Europeandestinations have lost 17

do Yet, France, Spain Portugal 12% 7% not seem to ;zArMind and of receipts. of arrivals and
Collow Europe's decreasing trend. On the contrary, these destinations' share of world tourism is fairly stable, particularly in terms of arrivals. France, Spain and Portugal together represent

r0ose to 19% of world arrivals and 13% of world receipts. These destinations' share of' 1-. Lirope's tourism flows is also impressive, with around 33% of' total arrivals and more than 25% ol'total receipts. The relevance ot'these destinations is illustrated in table 2.3 and figure
2.3 (WTO, 1999 and WTO Europe, 1997) which show Furope, France, Spain and Portugal's

tourism tourist receipts. and shares arrivals ol'world . Table 2.3: Furope, France, Spain and Portugal's sharesol'world tourism (%)
ARRIVALS
I Europe/World France/World Spain/World Portugal/World /Worl d (France+Spa in+ Portuga 1)) /L uropc opc (Fratice+Spain+Portugal)/Europc 1

RE. CEIPTS

11996 1980 1990 1995 1996 1997 1980 1990 9 1995 1997 65.1 10.6 13.4 1.8 25.8 38.4 61.8 11.5 7.5 1.8 20.8 33.6 59.6 10.6 7.0 1.7 58.9 10.5 6.9 1.6 58.5 10.8 7.0 1.6 60.3 8.1 6.7 1.1 555.3 '- .3 5 1.5 1 7.6 71 6.9 7.0 7. .0 1.4 6.4 1.1 5 1.8 48.8 6.7 6.5 1.0 6.3 6.0 1.0

1 1 1 1 1 1 1 19.3 19.0 19.4 15.9 16.0 14.4 14.0 13.3 132.3 133.4 1 1 1 1 1 32.4 26.6 28.8 27.6 27.5 27.1

Figure 2.3: Furope, France. Spain and Portugal's sharesot'world tourism current receipts

6050403020. 10. 0. Europe F+S+P France Spain m 1980 01990 Cl 1995 01996 0 1997 mmmmmmo

Ao

Portugal

SOUrce: Destination coujitrytotals according to a single indicator - arrivals of tourists and tourism receipts ( 198797), WTO and Trends ol'Tourisrn Movements and Payments(1980), I'mrope, WTO

19

Around 80% of touristsvisiting European countriesoriginate from Europeitself. Yet, tourism flows are not evenly distributed across Europeanterritory but are instead highly for Union 38% In 1992, European (EU) the twelve members accounted of world concentrated. tourism. With the subsequent of Austria, Finland and Sweden,the EU shareof membership world current receipts increasedto 41%, reinforcing its position as the most concentrated tourist area in the world. According to the OECD (1996), tourism receipts representa substantialshareof total exportsin someEU countries.For examplein 1994,the tourism's in in Greece, 18.4% in 25% Spain 15% Portugal.Howevcr, for total was and export shareof Germanyand the UK, theseshares that tourism doesnot are lessthan 4%. Thus, it is apparent have the samerelevancefor all EU members.Yet, for someEU countries,tourism is a very importantpart of their national economies and in some regions of thesecountries,the main sourceof incomeandemployment. Spain Portugal, impact In the cases the and of of tourism is particularly impressiveat the regional level. For instance,in the early 1970s,regions like the Algarve in Portugaland Costadel Sol in Spainwereunderdeveloped areashighly dependent on agricultureand fishing industrial lacking isation, significant any activities, with precariousnetworks of roads and low income levels and standards and communications, meansof transportation of education, health provision and other indicators of social welfare. A typical tourist who visited these had time, to walk long distances that usually to buy provisionsin the local market, regionsat to get daily bread,and go to the local post office to makea telephone endurelong queues call. This 'environmentallyfriendly' way of touring theseregionsdid not last long. In the last decades, brought about by increasingnumbersof tourists, the two the pressure courseof hand, the one andthe political and economiceventsthat took placein the 1970sand 1980s on (Portugal and Spain attaining democracyafter 50 years of dictatorial regimes, and their for integration in EU), increased the the the the supplied means subsequent on needand other, in the demandof the theseregions' physicaland economiclandscapes to adjustto the changes sole highly rewarding product they could export: tourism. In less than 20 years, the more fishing from these villages to busy, small picturesque populated areasof regions altered millions, from the visitors to accommodating modembig cities, from hostinga few thousand sleepypaceof horsedrawnwagonsto the frantic highwaytraffic, from the quiet beeranddarts high-tech discotheques. in local The speedand to the the crowded and noisy playing pub 19

in in those other this they these are regions, as adjustmentprocessare visible extent of importance. for tourism associated with which are primary countries economicactivities of Tourism receiptsin thesecountriesgive rise to a variety of repercussions, contributingto the Indeed,revenuefrom foreign tourists creates transformationof their economies. and sustains jobs, generates additional income and alters its distribution acrossregions.Tourism receipts business investment to and activity contributing directly and indirectly to act as a catalyst regionaleconomicdevelopment. France, Spain and Portugal are important tourism destinations within the EU, 35% for than of the region's tourismcurrentreceiptsin 1997.Although these accounting more from 29% in 1990to 27% destinations'shareof Europe'stourism currentreceiptsdecreased in 1997,their participationin EU receiptsgrew more than one percentage point in the period, EU total than tourism currentreceipts,as shownin table2.4 one-third of more representing . Table 2.4. France,Spainand Portugalshareof EU tourism currentreceipts($US, million) 1980 EU 1990 1995 1996 1997

50,539 124,075 163,559 168,956 167,070 57,254 60,294 58,937 35.0% 35.7% 35.3% 34.1%

France+Spain+Portugal 16,416 42,332 32.5% (France+Spain+Portugal)/EU 1

Source: Destination country totals according to a single indicator international tourism receipts (1985-1997) and trends of tourism movements and payments (1980), Europe, WTO

France and Spain have always held among the top positions as the world's most destination countries.In terins of arrivals, Franceis the first destinationcountryand preferred Spainis the third, while in terms of current receiptsFranceis the seconddestinationcountry and Spainfluctuatesbetweenthe fourth andthe third position (WTO, 1996).In tourism terms, Portugal is a small country compared with its neighbours France and Spain. However, 1997, in bascd 1985 bctwcen WTO data (1999) tourism that and calculations on show Portugalincreasedfaster (except for Francein terms of arrivals and for the Netherlandsin terms of receipts)than any other Westernor SoutherncontinentalEuropean country. In fact,
tourist arrivals in Portugal increased 60% between 1987 and 1996, only surpassedby France increased increase by Portugal's 275% between 69%, tourism receipts current with an of and 1985 and 1997, only surpassedby the Netherlands with an increase of 288% in the same

20

international has Furthermore, 1.5%since Portugal's tourist of arrivals exceeded period. share 1990,classifyingthe country as one of the twenty most visited countriesin the world (WTO, in termsof receipts,the picture is lessfavourableas Portugaldrops to a 1996).Nevertheless, Figure2.4 showsthe trends position below 25 amongthe most importanttourism destinations. in tourism receiptsfor France,SpainandPortugalbetween1985and 1997.
Figure 2.4: Francc, Spain and Portugal tourism currcnt reccipts ($US million)

30,000 25,000
20,000

15,000
10,000

5,000 0
co CY) co 0) co 0)
00

co

(7)

CC)

CY)

(3) 0)

a) a)

04

0) CY)

or)

4% c

U)

CC)

t*-

G)

CY) 0)

0) 0)

a) CF)

France Spain Portugal] --o-m-&Source: Destination country totals according to a single indicator international tourism receipts (1985-1997), WTO

Total currentreceiptsin FranceSpainand Portugalmorethan tripled between1985and 1997. In this period, France'sinternationaltourism current receiptsincreasedfrom 7,943 to 28,009million $US, implying an averageannualgrowth rate of 19%, Spain's increased from 17% Portugal's increased 8,151to 26,651implying an average from of and growth rate annual 1,137to 4,277 implying an averageannualgrowth rate of 21%. However,if the growth rates by periodsof five years,the picture may appearquite different from that implied are analysed
21

for the whole period.For instance, between1985and 1990,the average by the average annual for for for 26% France, 21% Spain Portugal.Yet, 35% tourism growth rateof receiptswas and for the next five years,thesegrowth ratesdrop sharplyto valuesof 6% for Franceand Spain international for 4% 1997 Portugal. Between 1995 by less tourism receiptsincreased and and 0.5% for Portugal. than 1% for France,around2% for Spainanddecreased Although it is important to know how France, Spain and Portugal's total currcnt interest is behaviour from UK tourists to the time analyse our evolve, over of receipts receipts is This in the next section. total that time their of with receipts. path undertaken andcompare 2.4. UK TOURISM DEMAND FOR FRANCE, SPAIN AND. PORTUGAL The UK hasalwaysbeena key market for tourism in France,Spainand Portugal.Data for tourist arrivalsandtourism receiptscan be usedto illustrate the importance of UK tourism demandfor thesedestinations relativeto that of other origins. For thesedestinations, table 2.5 UK (million) total and arrivals arrivals shows and current receipts from all international touristsandfrom UK tourists(billion $US) in the period 1987-1997 (WTO, 1996-1999). Between1987and 1997in France,total arrivals increased by 81% while UK arrivals increasedby 88%. In spite of some fluctuations in the early 1990's, the UK shareof total in 18%of arrivals Francemore than recoveredits 1987value in the late 1990's,representing by 67%. However, the total in 1997.In the sameperiod, total arrivals in Portugalincreased UK arrivals in Portugalonly increased by 33%. The UK shareof total arrivals in Portugalhas beenrelatively stablebetween15%and 17%.Given the increase in total arrivals,the stability UK indicate in last diversification for Portugal the the tourism share may of some of markets Spainshowsa much slowerincrease decade. of both total and UK arrivals.While total arrivals by a modest9% in in Spainincreased by 54% between1987and 1997,UK arrivals increased from around3 million in 1977to 7.7 million the sameperiod. UK arrivals in Spainincreased in 1988.However,between1989and 1993,this destinationlost around I million UK arrivals its 1987level in 1996.In the last two decades, total arrivals in Spainnever and only recovered decreased. This, compoundedwith the fact that the UK share of total arrivals in Spain decreased from 27% in 1987 to 19% in 1997, seemsto indicate diversification of origins demanding than that implied by the numbersfor tourism in this destination, more pronounced 22

Portugal.Hence,assuming that arrivals may represent a good proxy for tourism demand,the
traditional UK tourists' preference for Spain relative to its neighbours, France and Portugal,

decade. in have decelerated in last Furthermore, the terms of UK arrivals, France to seem it leading for in 1990, in France's first Spain that the time and seems position new surpassed hasnot beenreversed. the UK touristspreferences Portugal Table 2.5: Arrivals (million) andcurrentreceipts(billion $US) of France,Spainard
1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 Totalarrivals 37.0 UK arrivals 6.4 UK/Total. I 7% 11.9 Totalreceipts UK receipts 1.5 13% UK/Total 42.7 49.5 52.5 55.0 59.7 60.6 61.3 60.0 62.4 6.6 7.1 7.0 6.8 8.2 8.2 11.6 11.2 11.5 15% 14% 13% 12% 14% 14% 119% 19% 18% 13.8 16.2 20.2 21.4 25.1 23.6 24.7 27.5 28.4 1.7 2.1 2.6 2.8 3.0 2.9 3.0 3.3 3.1 12% 13% 13% 13% 112% 12% 12% 12% 11% 66.8 12.0 18% 28.0 3.7 13%

France

Spai n

Totalarrivals 27.7 29.8 32.5 34.1 6.3 UK arrivals 7.6 7.7 7A UK/Total 27% 26% 123% 18% Totalreceipts 14.8, 16.7 16.2 18.6 UK receipts 2.7 3.4 2.8 2.7 UK/Total

34.2 36.5 37.3 43.2 39.3 40.5 43.4 6.2 6.5 6.5 7.7 8.2 7.6 8.3 18% 18% 17% 18% 21% 19% 119% 1 19.1 22.2 19.7 21.5 26.7 26.7 1 125.4 2.7 3.1 3.2 3.9 4.5 4.2 4.6 18% 20% 17% 15% 14% 14% 16% 18% 18% 16% 17% 10.2 1.6 17% 15% 16% 4.3 4.3 4.3 9.5 1.6 9.7 1.5 0.8 0.7 19% 16%

Totalarrivals 6.1 6.6 7.1 8.0 8.7 8.9 8.4 9.2 UK arrivals 1.2 1 1.1 1.1 1 1.2 1.3 1 1.4 1.4 1.4 UK/Total 20% 17% 15% 15% 15% 16% 117% 15% l Portuga _ _ 1 Total. 3.7 3.7 4.1 3.8 receipts 2.1 2.4 2.7 3.6 UK receipts 0.4 0.6 0.5 0.5 0.6 0.7 0.6 0.6 19% 25% 19% 14% 16% 19% 15% 16% UK/Total

Source: Yearbook or tourism statistics (1987-1997), WTO and Tourism policy and international tourism in OCDE member countries (1990- 1997), OCDE.

be demand, Tourist arrivals or visits, as a measure tourism misleadingas their can of

levels may not translate into the effective consumption of tourism in the countries visited. Origins' tourism expenditure in a destination or destinations' tourism receipts from an origin,

in tourists to the preferences accordance are generallyconsidered of more accurate measures theoretical micro-foundationsof econometric models currently used to analyse tourism

23

demand.Moreover, tourism receipts have a direct economic impact on the destinations' interesting for therefore, measure policy purposes. and, constitutea more economies In table 2.5, the generaltrendpath of tourism receiptsappears to follow that of arrivals . although in a less obvious way. Between 1987 and 1997, total tourism receiptsmore than is lessthan double.Indeed,total doubledfor Portugaland Francewhile for Spainthe increase for by for 105% for increase France, Portugal 80% Spain. In 133% the same and receipts increased by for 147% for from France, 100% Portugal UK tourists the and receipts period, have UK been 1990, for Spain. Since 70% tourists contributing,on average,12% of the only total tourism receiptsof France,16% of the total tourism receiptsof Spain and 17% of the total tourism receiptsof Portugal.Thesenumbersseemto supportthe view that UK tourists' but from changing progressively slowly mainly favouring Spainto favouring are preferences Franceand Portugalasholiday destinations. In order to understandbetter how UK tourists distribute their budgets among internationalholiday destinations, table 2.6. shows UK tourism visits and expenditurein the Spain Portugal for France, Europe, the period 1970 to 1997 (BusinessMonitor and world, MA6 andMQ6,1967-1993andTravel Trends,1997-1999). For the last threedecades, the UK demandfor tourism abroadhas been growing considerably.Between 1970 and 1998, UK tourist visits abroadincreasedby a factor close to 6 and UK tourism expenditureabroad by a factor of around50. This showsa much fasterincreasein expenditure increased than in tourists numbersand provides an indication of the way in which UK residentshave been changing their patternsof tourism consumptionover time. In 1970, UK tourists' average Even E45 by 1997 taking 408 f. this abroad was spending per visit; per visit. averagewas inflation into account, is still impressive. this increase UK tourists' preferencefor European destinationsis conflirmed in table 2.6 by 90% of total Europe'sshareof UK tourism demand.In 1970,UK visits to Europerepresented UK international 83% in Europe total the tourism abroad and of visits expenditure represented tourism budget. Although these Europeanshareshave been decreasingover the last three important, decades, Europe'sshares 82% and 64%, UK tourism representing are still very of in total respectively,of visits and expenditure 1997. Figures 2.5 and 2.6 show the relative in (E (thousands) UK million) world, Europe and in magnitudesof visits and expenditure France,SpainandPortugal(FSP)asa region. 24

Table 2.6. IJK tourism visits (thousand) and current expenditure (f million) abroad
World Visits 1970 8482 Europe France Spain Portugal Expend 8 11 67 177 306 483 452 492 468

Expend Visits 382 917 7662

Expend Visits 317 704 1942 6831 10422 10260 10879 11700 1059 2149 3844 4_523 6865 9645 9834 11149 11518

Expend Vi ct ,i 33 1583 HI 375 642 1482 2107 2015 2256 2663 2521 ::2: 61: 7 4175 5096 8239 7545 8281 9650

Expend Visits 73 207 428 939 1528 2877 2704 2825 _ 3236 164 94 364 709 982 1211 1102 1304 1299

1975 11992 1980 17507 1985 21610 1990 31150 1995 41345 1996 42050 q 5; 1997 4 957 8 4 8800 19) (9)
Source: Business

10283 114676 27384871 9886 15386 16310 17136 19900 26268 34418 34213 37745 39500

81 3687

Monitor

MAO ( 19 /u- 199-1) ana i ravel

. ( 1994- 1997) I rends

Figure 2.5: UK tourism Current cxpencliture in the world, l''Uropc, and in France, Spain and Portugal
20000 18000 16000 14000 12000
10000 i

8000 6000
4000

2000 0
1970 1975 1980 1985 1990 1995 1996 1997 1998

M World 0 Europe 0 FSP

Source: Business Monitor MA6 (1970-1993) and Travel Trends (1994-1997)

25

Figure 2.6: UK tourist visits to the world, Furope and to France, Spain and Portugal
50000 45000 40000 35000 30000 25000 20000 15000 10000 5000

0 1970 1975 1980 1985 1990 1995 1996

1997 1998

MWorld MEurope OFSP

Source: Business Monitor MA6 (1970-1993) and'i Iravel Trends ( 1994-1997)

Among European destinations, France, Spain and Portugal seem to be preferred by many UK tourists. Indeed, these three countries' share ol'UK tourism demand in the world and FUrope has been increasing since 1970. These destinations' share of, total t JK visits abroad from in 33% 1970 to 46% in 1997, and their share ol'UK visits to F, increased increased urope from 37% in 1970 to around 55% in 1997. In terms ol'UK tourism cxpciiditurc, France, Spain and Portugal have been major recipients with an average share of' more than one-third of' the total UK tourism budget and more than halt'ol'UK tourism expenditure in EuropcThe evolution over time of tourism demand for France. Spain and Portugal as individual holiday destinations for UK tourists, may be better analysed using figures 2.7 to 2.12. These figures show the UK tourism current expenditure (f million) and UK tourism 1969-1999. France, Spain (thousand) Portugal the in period visits and in

26

in France(E million) Figure 2.7: UK tourismexpenditure

3000 2500 2000 1500 1000 500

(3) (0 C)

9-

ci rC)

LO

frC)

(D P(D

qCo a)

(1 Co 0)

to Co (3)

tCo C

( Co (

%( 0

(1

Lin

C)

)enditure In Francej

Figure 2.8: UK tourist visits in France (thousands) 1400012000 100008000 6000 4000

2000
i . II II II .i II II II II II II I. II If II I

(3) (D C) I-

VPCD T-

CY) C) I-

U) ro) T-

rtC) T"

0) r0) T-

T00 a) V-

(Y) CC) CD V-

U) CC) 0) V-

rco 0) T-

(7) co (3) V-

T0) 0) T-

CV) CD 0) V-

LO 0) 0) V-

f(7) 0) V-

CY) G) CY) V-

I-

I in France UK visits

Source: Business Monitor MA6 (1969-1993) andTravelTrends(1994-1999)

27

Figure 2.9: UK tourism expenditurc in Spain (f million) '500 30001

2500 2000 1500 1000 500

rz rG) 0)

r0

Co (

CY) CO (

CD (M

Co 0)

Co C)

(3) C)

( cn

( C)

( (

0) C)

I-

UK expenditure In S

Figure 2.10: UK tourist visits in Spain(thousands)

12000 10000 8000 6000 4000 2000
0iIIIIIIIIIIIIIIIII CY) (D Cl) Vrmm CV) rto rCD Ir-_ CD CD I0 Vco CD 0) co 0) LO OD CD
F00 C) Co v0) m (3) Lin C) r_ (3) ch 0)

UK visits in Spai
Source: BusinessMonitor MA6 (1969-1993) and Travel Trends (1994-1999)

28

Figure 2.11: UK tourism expenditure in Portugal (f million) 50045040035030025020015010050VT 11---11------111,, 11j),,,,,,

C) (0 (D

9r0

m rc3)

V) r0)

rz (Y)

(3) r(D

%%-

) CX) (3) 9-

te) Co 0 4-

PCo a)

( Co C

%C) (M

tin C C

rC) (D

(Y) C) C)

UK expenditurein Portu

Figure 2.12: UK tourist visits in Portugal (thousands) 1500 -1

1250 1000750500 250-

v-

m rl_ (D

LO r,cm

r_ t(

(7) r_ (

IrCo 0) v-

M Co C) v-

Lt) 00 C 9-

rCo C Ir-

(Y) Co CI) 1-

Ir( 0)

v) CY) C

r(Y) (3)

CY) C) - cn

UK visits in Portugal
Source: BusinessMonitor MA6 (1969-1993) and Travel Trends (1994-1999)

29

Figures2.7 to 2.12 presentincreasingtrendsfor both UK expenditureand visits. UK in in 1969 1999. In increased from I 12 France to the same to million around million visitors from 1.5 increased Portugal UK Spain II to to to around million and million, visitors period, from little more than 100 thousandto around 1.5 million. A similar trend exists for UK by a factor of The level of UK tourism expenditurein Franceincreased tourism expenditure. 106 between1969(f.25 million) and 1998(E2663million), by a factor of 52 in Spain (from L62 to 0236 million) andby a factor of 94 in Portugal(from L5 to E468million). Assumingthat France,Spain and Portugal'sproximity to the UK gives this region a destinations in UK it is likely to tourists' preferences, other relative comparativeadvantage in flows tourism that exogenous one country havea significant impact on the shocksaffecting demandfor the other ricighbouringcountries.In fact, this appears to be the casefor the 1974 Franco's dictatorial in Portugal, the substitution of regime in Spain,and Spainand revolution Portugal'smembership of the EU in 1986.The 1974revolution in Portugaland the political in led Spain, 1976-1977 dictatorial to the substitution of of events regimesby parliamentary in both thesecountries.Initially, between 1974 and 1979, these eventshad a democracies negativeeffect on UK tourism flows to these destinations.However, the consolidationof democracyin both countriesand their simultaneous have into in EU 1986 to the entry seem had a positive impact on UK tourism demandin the following decade.Indeed,in 1980-1989, both destinations increasinglevelsof UK expenditure. present almostuninterrupted Nevertheless, the 1974revolution had a major negativeimpact on the UK demandfor 1974-1976 Portugal,also affecting Spain. The Portuguese instability of political and social in UK visits and expenditure in Portugal.In this period,UK demandseems a decrease caused to have been diverted mainly from Portugal to neighbouringSpain. Hence,initially, Spain to benefit from the instability in Portugal.However,the subsequent political upheaval appears in Portugal, in Spain,compounded (communist PREC the process) with revolutionary ongoing for from Peninsula diverted a substantial Iberian UK the rest of the the tourists part of away levels 1973. UK Given the decade. Only by 1979,did both destinations their of visit recover for Portugal,the plots in figures low levelsof expenditure 1970s, the particularly and visits of Therefore,figure 2.13 is 2.9 to 2.12 do not clearly show theseUK tourism demandchanges. demonstrate in data first decade in UK to the the decrease to the order constructed magnify of following 1974. in Spain Portugal, visits andexpenditure and 30

in SpainandPortugal1969-1979 Figure 2.13: UK visits andexpenditure

3000250020001500 1000 1969 1971 J-UK 1973 1975 1977 I Spain visits to 1979 0,0, . 10000000

3002001000 1969 1973 1971 1-UK_visits 1975 to Portuga 1977 1979

500 400 300 200 100

0
1969 1971 I1973 1975 1977 1979 UK expenditure in S

50 40 30 20 10 0 1969 1971 J-UK 1973 1975 1977 1979 expenditure in Po, u7gal

Source: Business Monitor MA6 (1969-1979)

31

Sincethe mid 1980's,the numberof UK visitors in Francehasgenerallyexceeded that in Spainby amounts that canreachthreemillion visits. However,the UK tourism expenditure in Spainis generallylarger than that in Franceby amountsrangingbetweenL200 million and LIOOO million. This fact illustratesthe problemsof using tourist numbersas a measureof tourism demand.Although UK tourists visiting Spain are less in numberthan those visiting France,UK tourism expenditurein Spain generally exceedsthat in France,sometimesby feature is detected for Portugal Spanish This same when comparing and amounts. substantial UK visitors and Spanishand UK tourism expenditurein this destination.Spanishtourists but little; in in huge UK Portugal tourists to spend very numbers arrive smallernumbers come but their spending part of Portugal'stotal tourismearnings. a substantial represents Smaller numbersof tourists associated with relatively larger expenditurelevels, are length longer different linked to average of stay a and/or a spendingpropensity generally displayedby tourists from specific origins in different destinations.Different behavioursin tourism demandindicatedby suchmeasures as the averagelength of stay,may be due to both individual characteristics of the destinations countriesand to specific featuresof the tourists' demandbehaviour.For instance,in the last three decades, the averagelength of stay in all decrease. destinations Given the generalincrease leisuretime in average continuous a presents in industrialised the of across world and,particularly, countrieswhich constitutemajor sources tourism demand,the decreasein averagelength of stay may indicate greater mobility of in demand This greatermobility may indicate changes tourists amongpreferreddestinations. behaviour favouring 'complementary' destinations more than competing ones. Distance betweendestinationsmay play a major role in defining these complementari ties, from the tourists' point of view. If this is true, then neighbouringcountries are likely to attract the touristswho, having little time to waste,want to spenda part of their attentionof cross-border OHowever,from the destinations'point of holiday in one country andthe rest of it in other(s). long for important is keep in the territories to tourists their as as possible. strategy view, Therefore,the higher mobility of modem tourists may trigger competitivebehavioursamong between destinations, than those countriesfar apart. observed neighbouring moreaggressive
0Thesearehypothetical by (or be However,given specificresearch. supported otherwise) which should assertions length of stay in individual the increasingaverageleisure time availableto UK tourists and their decreasing budgets tourism,it is plausiblethat UK touristshave to international countries allocated associated with increasing duringtheir yearlyholidays. beenvisiting morethanonedestination

32

For France,Spain and Portugal,it is interestingto observethat UK tourists' average in in accordance length of stay, althoughdecreasing the trend all other observed with general UK destinations,presentsrelatively stable values when compared,for example, with the in this measurefor Italy or Germany.In fact, UK tourists' averagelength of stay decrease in Francefrom 8 daysin 1975to 6 daysin 1997,in Spainfrom 13to 12 days,and in decreased Portugalfrom 7 to 5 days. However,in Germanythis measurefell from 12 to 6 and in Italy Monitor MQ6,1969-1998). from 14 to 9 in the sameperiod(Business In general,the longerthe average of stay,the smallerthe numberof visits as the same hand, longer On imply is the time. other stays usually more spending. over extended visit
Fewer visits and more spending imply higher expenditure per visit. Therefore, the expenditure destinations higher for be longer is length to with average of stay. 11iis may expected per visit France UK larger in to visits are number than those to Spain, UK explain why although that in France. tourism expenditure in Spain generally surpasses

However,the expenditure per visit in a destinationcountry may reflect more thanjust fact, length In of stay. as shown in figure 2.14, the UK expenditureper average an extended in is higher Spain in Portugal than that or France.Yet, given that UK tourists' average visit length of stay in Portugalis smallerthan in Franceand Spain,and has decreased most since 1969., with travelling to any of thesedestinations and given that costsassociated are expected to be similar for UK tourists,the levelsof expenditure per visit in Portugalmustbe linked with length than the other of stay. aspects The main reasonsthat can explain this apparentlyodd feature of UK demandfor Either Portugalmay be linked with tourism pricesand the spending the visitors. propcnsityof tourism price levels in Portugalare substantiallyhigher than thoseof its neighbours and have Portugal beenso for the last three decades, UK tourists visiting or are more prodigal when than when visiting Spain or France.Probably,a combinationof theseand other factors may Nevertheless, if destinations. in three these explain the differencesin UK expenditure per visit interesting for is do differences, these then there a very scope and prices alone not explain innovative field in' tourism demand analysis, focusing on the determinantsof different destination in different from tourists countries. propensities origin spending of a given

33

in France,SpainandPortugal(f) Figure 2.14: UK tourismexpenditure visit per

450 400 350 300 250 200 150 100 50
rz 0) 9-

(3) (D 0) v-

vrC w-

ei r0) v-

to r0) v-

(3) r(7) 1-

9Co 0) 1-

cn Co c3) 9-

Lii Co m N-

rCo 0) 1-

(7) Co C)
Ir-

1C) C
le-

ri C) (D
T-

ti') C (

Pv-

I--&- France --*-Spain -0-- Portugal

Source: Business Monitor MA6 (19,69-1993) andTravel Trends(1994-1997)

2.5. CONCLUSION In the courseof the last two decades, broughtaboutby increasing the pressure numbers of tourists and the potential economicbcnefits and costs that tourism brings to national and regional economies,intensified the need to understandand measurethe magnitude and direction of changesin tourism demandbehaviour.Whether, how much and for how long tourism receipts contribute net benefits to a region depends,among other things, on the demand for form tourism the the structureof the local economies and on and scaleof precise
that region. Hence, as tourism emerges as an important factor of development and source of reproducible wealth, it becomes imperative to implement a comprehensive quantitative demand behaviour. tourism examination of

Quantitative empirical studies of tourism demand generally involve the use of between the variable measuringdemandand to the econometric relationships models specify 34

basis for its determinants. Econometric economicanalysis generally provides a good modelling and policy evaluation.However, appropriateconceptionand implementationof quantitative methodological analysesrequirestwo basic things: plentiful and reliable data, and adequate tools to build, estimateandtest modelsthat can validly describethe pastand predict the future data investigators, in levels. demand Plentiful to tourism and reliable are rarely available of in has been improvement Although there tourism contexts. some research recently, particular, the more reliable suppliersof tourism information generallymake it availableon yearly basis, for levels, relatively short periodsof time. Yet, thesedifficulties should not and at aggregate deterresearch needbig eggs', there is nothing wrong with smaller and although'big omelettes long asthey arc well cooked. as omelettes Whatever the quality and quantity of the information available, the appropriate building of an econometricmodel and interpretationof its estimatesrequires a thorough The data time the series. relevant must be inspectedin detail, in order to examinationof determine how the series behave over time. Data specific features, compoundedwith knowledgeof eventswhich could havehad a significant impact in the evolution of the series, in building, be integrated the then model can permitting a more precisespecificationof the interrelationships, within an adequate variables quantitative framework. Although 'torturing the data until they confess' is not the generalidea, every aspect,even minute, can help to better be based. modelson which reliable inferenceand forecastingprocedures construct can The analysis of the data in this chapter complies, with these requirementsand helps to characterizethe relevant variables and define the quantitative framework within which be relationships can plausible modelledand interpreted. Using the information provided in this chapter,the remainderof the thesisfocuseson different methodologicalapproaches to the modelling, estimationand forecastingof tourism demand.Basedon thesedifferent methodologies, severaleconometricmodelsarc constructed to analysethe behaviourof UK demandfor tourism in France,Spain and Portugal over the in 1969 is 1997: to a static single equationapproach considered chapter3; a dynamic period dynamic is in 4; almost ideal demand and equation single model estimated chapter static 6 (AIDS) in 5 (VAR) systems of equationsare used chapters and and a vector autoregrcssive is exploredin chapter7. Chapter8 evaluates the ability of thesemodelsto predict the approach levelsin thesedestinations. Chapter9 concludes. UK tourism demand 35

CHAPTER 3
STATIC SINGLE EQUATION MODELLING TOURISM DEMAND OF

3.1. INTRODUCTION Econometric includethe modellingof tourism demandis not an easytask. The reasons complexity of the motivational structure underlying the decision-mak-ingprocess; the heterogeneity of the products and services supplied; the existence of and multiplicity fact the that transportationplays a role in tourism consumption;the regressors; endogenous intcr-temporaldependence of current demand;the existenceof qualitative factors,often nonaffecting tourists' travelling decisionsand the limited availability of reliable data, measurable, Theseconceptualand practical difficulties are frequently overcomeby meansof simplifying assumptionsmade to permit the specification of empirical models explaining demand behaviour.Examplesof such models are static single equation specifications,which most have in adopted, empirical studiesof tourism demand. researchers Following Johnsonand Ashworth's (1990) appeal "comparisonsacross studies are
(p. 150), still largely ignored in tourism literature, this chapter analyses the needed' urgently modelling of tourism demand within a static single equation framework, and provides a systematic examination of estimation results obtained from different studies using this focuses first, on the demand. The UK the analysis approach and a set of observations on different UK from the models single equation of estimation of comparison results obtained demand for tourism in France, Spain and Portugal. The differences between models relate to included, in definition the the variables changes and measurementcriteria of and to the use of different sets of regressorsand linear and log-linear functional forms. Then, we proceed with

36

the comparisonof these estimation results with those obtained in other empirical studies involving the UK as an origin country,using the samemethodological approachand a similar sampleperiod. follows. Section3.2. providesa literaturereview of tourism The chapteris structured as demandstudieswhich have used the single equationapproach.In section 3.3, static single equationmodelsof the UK tourism demandfor France,Spainand Portugalare estimatedwith different functional different for 1969-1997, forms, definitions the using period observations for different The estimation some variables, and criteria setsof rcgressors. and measurement 3.4, In then these are models presented alternative and analysed. section other studies resultsof for demand UK tourism which use similar cmpirical methodologiesand analogous the of sampleperiodsare analysedand comparedwith the results obtainedin the previous section. Section3.5. summariscs the main findings anddiscusses their implications. 3.2. THE SINGLE EQUATION APPROACH IN TOURISM DEMAND MODELLING
3.2.1. TIIE DIFFICULTIES OF TOURISM DEMAND MODELLING

Studiesof tourism demandbehaviourare subjectto specific problems.The complexity of the motivational structure underlying the decision-making process, and tile limited data appearto be two of the main difficulties which have had a availability of relevant significant influence on the way most empirical analysishas beenconducted.A quantitative in tourism demandanalysisrequiresa formal statisticalframeworkwithin which tile approach impacts on the demandvariable causedby changesin its determinants can be validly and for basis Appropriate reliable accuratelymeasured. econometricmodelling provides a good estimation and forecasting which is of considerablevalue for public policy and private investment decisions. However, appropriate modelling, estimation and forecasting of theoretically consistentand empirically plausible econometricspeciricationsare not easy to in demand tourism achieve contexts. Ratherthan constituting a single economicactivity, tourism embodiesa large set of
production and supply activities including transports, communications, catering, entertainment,

travel agencies,tour operators,advertising firms, and the production and saleof numerous

37

items.An additionalcomplexity,which distinguishes tourism from other economicsectors,is in the locationof its production. that tourism productsmustbe consumed Tourism demand can be analysed from many different perspectives. From a tourists' flows can be analysedas domesticor internationaldemand geographical perspective, for groupsof countries,one sole country,specificregionswithin a country,specific local areas facilities local From within areas. a motivational point of view within a region, specific for destinations, influencingthe decisionto travel and the preference the demandfor tourism (personal interest be travelling), as a part of the production can viewed as private consumption (business travelling) or evenas a part of the policy-making processaffecting this and process (diplomatic Demand for be tourism the missions). economy can also of analysed other sectors by different tourists, them of grouping categories such critcria as age, gender, considering income; different level types of tourism productssuch as sport, considering of nationality or cultural, religious or scientific events,and consideringdifferent componentsof the tourism product such as accommodation, catering or transportation. Irrespective, of which is characterisation adopted, tourism analysis remains complex, since it is a demand for involving, heterogeneous services and goods on the supply side, many different industries, and cost structuresand, on the demand side, the considerationof a production processes influencing tourists' behaviour. In addition, and as a result of the diversity of determinants dislocation processthat tourists undertaketo consumethe goodsand servicesof unavoidable their choice, tourism demandis much more sensitive to non-economicinfluences,such as political instability, natural disastersand other special factors, than the majority of other behaviour. demand Theseconceptual and practicaldifficulties, inherentto the research within this context, in their attemptsto justify someof the simplifying assumptions that investigatorsundcrtak-c specify econometric modelsexplaining tourism behaviour.However,when such assumptions inadequate or questionable,they lead to the construction of models which provide arc determinants. in its instance, For for demand the to changes unreliableestimates responses of inter-temporal infinitely the and separability of assumptionsconcerningan elastic supply demand detennined inadequate if simultaneously are consumption may prove pricesand and if decisionsto purchasetourism are interrelatedover time. Econometric models constructed ' do not take into accountthe presence of endogenous undertheseassumptions rcgrcssors and/or 38

that of a dynamic structure,leadingto identification, simultaneityand serial correlationbias. Another exampleof inappropriate specificationis to ignorethe possibility that the decisionto consumespecific types of tourism may be related to the consumptionof other goods and include Other tourism. those of types of potential problems services, as well as other inappropriate structureand/orover the components of tourism aggregation over the preference display different bchaviours, individuals different tastes and groups may or and productssince in the samedestination. different typesof tourismmay be consumed In addition to misspccirications resulting from the lack of a theoreticalbasis within be included, hypotheses testable the econometric can and methodologyused which reasonable in the estimationof tourism demandmodelscan also add problemsto the estimates'accuracy. do not take into consideration This is generallythe casewhen the estimationprocedures the from data, the presence arising regressing relationships nonstationary of spurious problem of inaccuracies serial correlation residual problem, associatedwith small severecollinearity, feedback the existence possible and of effectsandlaggedstructures. sampleestimations The difficulties in modelling tourism demand arc broadenedby the existenceof influencing demand, factors as well as the inaccuracyor unavailabilyof dataon unquantifiable the objectively measurable ones.The scarcity of reliable statisticson tourism, addedto the data inadequacies the much of required of mainly compiledfrom secondary or through sources into important be limitations taken sampling that surveys, are non-representative must is it demand difficulties, Given tourism not these specifying considerationwhen equations. frequently demand have been that tourism studiesof surprising econometric reportedmore not for basis in the economics literature.Nevertheless, thosethat havebeenundertaken a constitute in differences further research, for the allowing comparisonof resultsand critical analysisof obtainedand statisticalinferenceprovided. estimates
3.2.2. SINGLE EQUATION MODELS OF TOURISM DEMAND

The majority of studiesin tourism demandanalysisuse a conccPtof demandwhich includes the entire bundle of goods and servicesthat tourists usually purchase(transport, Most of thesestudiesuse time services). accommodation, related and catering,entertainment determinants focusing data the to on of tourist flows at series regresssingle equationmodels 39

the national level. Single equationmodels in studiesof tourism demandgenerally concern different origin and destination countries, different time periods, different sample sizes, in different They different measurement relation to many variables. also vary criteria and For the underlyingthe models,the example,the assumptions of models' specification. aspects form functional definition of the dependent that the the regressors, adoptedand and of variable have for Despite differences, these the the econometric studies a estimation. used methodology its in demand is to the tourism to sensitivity changes of evaluate common aim, which depend firms and governments Consumers, determinants. on the reliability of suchinformation for decision and policy making. However, the precision of the estimatesobtained and the forecasting inference depends the and procedures, crucially on the statistical validity of framework theoretical the within which the modelsare specificd,and on the use of robustness for the estimation of the relationshipsbetween the of a sound econometricmethodology determinants. its dependent variableand Quantitativeformulationsof tourism demandcan roughly be groupedinto two kinds of ' both in dynamic models contexts. static and models:singleequation,and systemof equations Reviews of tourism demandstudies, focusing on the modelling, estimation methodologies in Johnson are given and Ashworth (1990), Sheldon(1990), Tanscl and/oranalysisof results (1991)and Lim (1997). The majority of earlier quantitative studiesused a static single equationapproachto demand demand behaviour. A typical tourism explain single equation model of a static function canbe described as follows: E0,EikITRU9TRik 9 1, Pik i; PO is demand by Dij i is income for destination j; Yj are tourism and of origin where origin Eik k; EU destinations in i destination j to to and prices origin relative and relative competitor Do
9 9 f(yi PO Pik

Z)

k. destinations TRij between i i j between and competitor rates origin and areexchange and and TRik arc transport costs between i and j and between i and competitor destinationsk. Z demand. tourism the represents a setof qualitativevariablesaffecting origin's

1Recently, studies of statistical simulation of tourism demand regarding habit persistencemodelling contribute to (200 1). Johnson for Darnell the this and example, enlarging scope of classification, see,

40

The functionalform adoptedin the large majority of single equationstudiesin tourism demandresearch,is the log-linear form. Although the linear form has also been used, the its direct is both form log-linear to the of supplying convenient related property popularity of its Lee (Witt Witt, 1992; 1996) to the et claimed and al., and of relevantelasticities estimates form. involve linear Most time series to the studies good empirical performancerelative analysison a yearly basis,althoughthere are also examplesof the use of pooled and cross Mak 1977; Kliman, 1974; 1981; Lin Sung, 1983; data (Jud Joseph, et al., and and sectional Trembley,1989;YavasandBilgin, 1996;Romily et al., 1998). The empirical specificationof single equationmodel is defined within the framework Some of the most common underlying this categoryof models arc: of a set of assumptions. between independence tourism consumption and the consumptionof other goodsand services; between decisions to purchasetourism, in destinationj and to purchaseit independence invariant intertemporal tastes;perfectly elasticsupply; of separability consumption; elsewhere; and acrosstypes of tourism; and predetermined acrossconsumers or cxogcnous aggregation Although someof theseassumptions or, at can be deemedreasonable explanatoryvariables. least, unavoidable,others may be consideredcontroversial.Additional assumptions can be involved found in different contexts,depending on specific featuresof the origins/destinations in that researchers to of particularaspects andon the relevance want enphasizc their analysis.
3.2.3. VARIABLES INCLUDED IN SINGLE EQUATION MODELS Or, TOURISM DEMAND

The definition and measurement of variablesincluded in single equationmodelsalso for demand For dependent tourism, the the vary acrossstudies. example, variable,representing Lin (1981), by in Kliman be total and tourists, numberof visits, arrivals or as can measured Sung(1983)and Gunhadiand Boey (1986); visits per headof the origin countrypopulation,as in Witt (1980a, 1980b)and Witt and Martin (1985); total real expenditureor receipts,as in Gray (1966), Barry and O'llagan (1971), Jud and Joseph(1974), Loeb (1982), Lin and Sung (1983), Uysal and Crompton (1984); and per capita real expenditure,as in Artus (1972) and Loeb (1982). However,total or per capita tourism expenditureis generallyconsideredas an it because, tourist tourism visits or arrivals, unlike numbers, consumption appropriate proxy of for changes depictsfluctuationsin tourism spendingpatterns,revealingthe factorsresponsible 41

in expenditure levels. In some circumstances,increasing numbers of tourists can be due to higher inflation rates,tourists' lower spending by decreasing accompanied expenditure is be likely lower Hence, length to tourism of stay. expenditure a more average propensityor accurate measure for tourism demand and, given its direct effects in the destinations' for interesting variable policy purposes. a more economics, different definitions The independent to are also subject and measurement variables denjand The in income the tourism variables representing studies. origin's are usually criteria (disposable) income few by total capita real with or per exceptions, notably Lin and measured Sung(1983) and Gunadhiand Boey (1986), who use the real per capita national income,Jud Domestic Gross Product, (1974), Little (1980) Joseph uses and who who uses real per and disposal income The is the real of origin consumption. country viewed as an adequate capita is final ' Whether is income tourism a total considcred good. measure since or per capita real depends on the equationspccificationand restrictionsimposed. moreappropriate Several considerationsrelated to the price determinantsof demand must also be for demand First, tourism concerns the a bundle of goodsandservicesrangingfrom addressed. A for tourism-speciric tickets. travel to theatre price index would be an appropriate air measure this determinantbut the difficulties involved in its construction usually preclude its use. for tourismin onedestinationcanbe sensitiveto changes demand Second, in the relativeprices destinations and,therefore,thesevariablesshouldbe included.Finally, touristsdecide of other tourism in one destinationby, amongother things, comparingtheir own domestic to purchase different foreign in As domestic the foreign prices. prices with and prices are expressed idea be the into translated these of to must currencies clearer a common currency provide a relative purchasingpower. Hence,both prices and exchange ratesof the origin relative to the demand. its destination determinants to of competitors and are generallyconsidered relevant Due to the unavailabilityof tourism-spccificprice indexes, the majority of investigators acceptthe consumerprice index as a good proxy, althougha few, for exampleGunadhiand Boey (1986), constructspecialtourism price indexesfor the demandanalysisthey conduct. Except for Grey (1966), who excludes relative Prices from his model on the basis that
2 As notedin Songand Witt (2000), if the research focuses tourism demand, on business a moregeneralincome variablesuchasGDPshouldbe used.

42

in th e exchangerate are "changes in foreign prices, other than those caused by changes investigators (p. 86), demand... " influence to most use relative prices as relevant unlikely determinants of tourism demand. As shownby Martin and Witt (1987),the relative consumer be for index, the exchange of rates, can a reasonable proxy or consideration with without price in Hence, the cost of tourism, but exchange are acceptable. prices are present alone not rates tourism demandstudieseither in the form of relative prices betweenorigin and destination (Jud Joseph, 1974, Witt, 1980;Kliman, 1981; prices and with or without competitors'relative Witt and Martin, 1985; Uysal and Crompton, 1984), in the form of relative prices and (Artus, 1972; Little, in 1980; Loeb, 1982) form tile regressors, or as separate of rates exchange (Kliman, by 1981; Gunadhi Bocy, 1986; Martin rates exchange adjusted and and prices relative Witt, 1987),usuallydenominated effective (or adjusted)prices.Whetherit is more appropriate to considerrelative prices and exchange rates as separateregrcssors or effective prices, has in the literatureby a few authors(071aganand Harrison, 1984; Syriopoulos, beendiscussed 1995; Sinclair and Stabler, 1997).The argumentsfor and againstare generallybasedon the in long-run that theseexplanatoryvariablesmay haveon tourism changes cffects short-runand demand. Modelsspecifiedwithin a static single equationapproach generallyintendto describe a long-run equilibrium relationship.In this case, it seemstheoretically more appropriateto includeeffectivepricesratherthanrelativepricesand exchange determinants. ratesas separate Sincetransportation costsmay represent an importantpart of the total price of visiting a
destination, some studies (e.g. Jud and Joseph, 1974; Kliman, 1981), consider these costs as a be determinant demand. tourism The can of separate price of tourism consumption disaggregated into three essential elements: the price of travelling, the price of commodities The destination's the the generally services purchased after arrival, currency. and and price of in demand included definition demand implies tourism the that a of adopted price variables inclusion first the the two these three of a equation embody of elements, making unnecessary in for However, if the right-hand side of transportation. the price variables separate price demand regressions do not account for transportation cost, this cost should be included as a be is in this the cost can contexts, considered various separate regressor unless, as case irrelevant or not a major influence in the decision-making process. Ruling out a theoretically its difficult inclusion data to the that obtain and are grounds relevant variable on would decrease the degrees of freedom (Uysal and Crompton, 1984), that there are complexities

43

involved in the constructionof such variables (Gunadhi and Bocy, 1986), or that other for studieshad found it insignificant(Locb, 1982),do not seemreasonable criteria omitting When a transportcost variable is found to be, on a theoreticalbasis,a relevant-variables. separaterelevant determinant of tourism demand for a particular set of origins and destinations,it should be included in the model. For instance,if researchconcernstile far demand for destinations UK apart suchas Ireland and Australia,a transport analysisof is determinant likely be I lowevcr, tile appropriate to a of preferences. cost variable relevant form of its inclusion is neither evident nor simple. The reasonconcernsthe diff"icultics and accuratemeasurefor the cost of attachedto the constructionof a comprehensive travel cost variable can be constructedon an average travelling. Generally,an acceptable basis,that is, transportcosts can be measured as the weighted meanprice of all types of Therefore,a meaningful transportusedto move tourists from an origin to the destinations. is difficult to obtain due to the complexity of tile fare structure, transportation cost variable frequency in the seasonal network, route of departures changes and, generally,tile lack of reliable information about tourism traffic. In addition, if transport cost is an important variable in tourists' decision-making process, tourists' choice among alternative destinations, should take accountof this docn-ninant.Hence,if the cost of travelling is a demand determinant for tourism of relevant a given destination,the cost of travelling to its ought to be includedin the demandequation.However,alongsidethe problems competitors it is above, not alwaysclearwhich competitorsshouldbe considered. stated Given these difficulties, most studies which include a separatetransport cost it variable,measure as the economyclass air fare of a return trip betweenthe origin and destination capitalcities. Yet, this is not alwaysadvisablesince,in manycases, a significant Moreover,a largeshareof shareof the incomingtouristsusesothermeansof transportation. the air traffic is covered by charter flights with complex seasonalcost and transfer structures,which depart from locations other than the origin's capital city and arrive at in the airports areas otherthanthat proximateto the destinations'capitalcity. In thesecases, inaccuraciesof the measureused to proxy the travel cost cause the estimatesof the respectiveelasticity to be biasedupward or downward,dependingon the under- or overthat the transportcost statement of the transportvariable.In addition, it hasbeenobserved income high tile with variable,measured negative correlation as airfarcs,showsa variable by the insignificanceof one or both causingcollinearity problems.Theseare oflen revealed of the correlatedvariablesand by wrongly-signedcoefficientsattachedto those variables. 44

However, the omission of a relevanttransportcost variable on thesegroundsis hardly a solution, since collinearity can be overcomeby such meansas additional data containing by relevant more variation in the independent variables.In contrast,estimationbias caused by suchmeans. variablesomissioncannotbe overcome Unavailability of data is not an acceptable for reason omitting a relevanttransport cost variable,or any other pertinentvariable for that matter. However,when that must be the case,researchers shouldbe awareof, and explicitly refer to, the estimationbias inscrtcd in their misspecificdmodels. For instance,it can be shown that the effect of omitting a relevantvariablenegativelycorrelatedwith a variablethat is includedcan causean upward ' Moreover, the omission of any bias in the estimatedcoefficient of the included variable. leads to (false) autocorrclationin the estimated relevantvariablealways model, with all its for known adverse consequences the validity of the estimationand inferenceprocedures. it is not surprisingthat severaltourism demandstudies, Given theseconsiderations, found included it insignificant (Gray, 1966; Little, have transport cost a variable which 1980).However,thereare also examples of a significant negativerelationshipbeing found between the transportcost variableand the demandfor tourism, althoughthe validity of the due to the models' misspecifications. For example,in Kliman estimates may be challenged (1981), the estimates of the transportcost elasticitiesof the Canadiandemandfor tourism ranged between-0.94 (Italy) and -3.09 (Portugal), and in Jud and Joseph(1974) the Latin estimatedtravel cost elasticity of the world-wide demandfor tourism in seventeen American countrieswas -0.665, while that of the US tourists' demandwas -2.022. It be noted that, even if theseestimates by should,nevertheless, were sanctioned appropriate inference the coefficients of travel cost variablesmeasured modelling and as procedures, airfares,represent partial elasticitiesasthey can only indicatethe effect of a specificair fare changeon the tourism demandfor a given country or region,celerisparlbus. The effect of changesin transport costs other than that specific airfarc cannot be inrerred from the estimates obtained. in tastescan havea significant effect on tourism demand.However,due to Changes in aggregate the difficulties attachedto this variable measurement terms, most empirical studiesuse trendsto proxy its behaviour.Due to the link betweenindividual preferences tastes,the explicit modelling of this variablein a tourism context structureand consumers' level of analysis. Nevertheless,the appropriate should take place at a disaggregated

45

integrationof sucha capriciousvariable in tourism demandmodelsstill hasa long way to in go empiricalresearch. Tourism demandcan be responsiveto advertisingcampaignsdesignedto attract is by When destinations. this the to expenditures case, relative marketing visitors specific the destinationsshould be included as an explanatoryvariable. For example,Uysal and Crompton (1984) found that this variable had a significant positive effect on the German for between 0.094 demand Turkey, (Germany) 0.596 Spanish tourism ranging and and (Spain).A review of demandmodelsthat include marketing variablesis provided in Witt (1987a). Martin and Tourism demandcan also be sensitiveto specialfactors not included in any of the disturbances, disasters, discussed political natural above: sportingevents,religious variables fairs When factors defined international these they can arc examples. are clearly meetings, demand in included have be the their equations and effects a straightforward easily interpretationthrough their estimatedcocfficicnts. Some examplesof earlier empirical are studiescoveringseveralof thesespccialeffectswithin a static single equationapproach Little (1980),Locb (1982),Uysal andCrompton(1984),GunadhiandBocy (1986).
3.2.4. STATIC AND DYNAMIC SINGLE EQUATION MODELLING OF TOURISM DEMAND

With few exceptions,a specification with an explicitly dynamic structure of the demandfor tourism, distinguishing betweenshort- and long-run cffccts, has rarely been in estimated the single equationcontextsof early researchin tourism analysis.Generally, earlier empirical studies are charactcrisedby the absenceof an explicit theory of is, behaviour, the by its intcr-tcmpoml that consumers' mostly nature; overlooking formation of expectationsand habit persistencefeatures.These studies assumethat the ignore level demand depends its determinants levels the and of current of only on current the possibility of past demandpatternsaffecting current consumption.They provide no explanationof the processby which tourism demandoccurs over time, and neglect tile level demand tourism to the of current possibility of an adjustmentmechanism propelling its long-runequilibrium value.Therefore,thesestudiescanonly be viewedas specific cases and not as a generalcomprehensive within the wide rangeof plausibledemandbehaviours, analysisof tourists' conduct.
3 SeeGujarati,1995,pp. 457-8,for example.

46

As a result, important aspectsof consumertheory have been disregarded and the differencesbetweenshort- and long-run responses of tourism demandto changesin its determinants havegenerallybeenignored.For example,the theory of inter-temporalchoice decidehow to allocatepresentand future consumptionaccordingto statesthat consumers their rate of time preference,meaning that current consumption can depend on any Consumptionmay be a combinationof current, past or future values of its determinants. function of the discountedpresentvalue of expectedfuture income,and uncertaintyabout future incomecan deterconsumption or reduceits currentvalue by inducing precautionary decisions in imperfect Moreover, are often made consumption a context of savings. information, uncertaintyand liquidity constraintsthat restrict current demand.Hence,the level, symmetryand accuracyof information, as well as the degreeof imperfection in tile ingredients important in formation Consequently, the of expectations. relevantmarkets,are the explanation of current demand often requires the formulation of a theory of if into taking account consumers are "forward-looking" or "backwardexpectations, looking". If consumers are"backward-looking",the demandequationshouldincludelagged if income they arc "forward-looking", a suitableexpectations the variable; valuesof process into be taken account(Sinclair and Stabler,1997). should Tourism demandspecifications modclled as appropriatedynamic fonns also permit the examinationof important theoreticalaspectsconcerningthe sensitivity of demandto that, in the variations in prices and exchangerates. For instance,it is generally assumed short-run, tourists are more aware of exchange rates than relative prices. Ilowever, favourablevariations in exchangerates can be offset by higher inflation rates and these compensatorymovementsof prices take time before they are fully acknowledgcdby in Hence, tourists. the short-run, the investigation of exchangerate cffccts potential separatelyfrom price effects assumesparticular importancewhile in the long-run, the cffective price variable (relative prices adjustedfor exchange rates) are more likely to be relevant for tourism demandthan the separatevariables of prices and exchangerates. Further issues concerning the dynamic specification of tourism demand responsesto changesin price variablesare related to the fact that tourism purchases can be made in implies This their that lagged consumption. advance, as well as simultaneously, with actual be in to their may values, appropriate'in certain price-change current variables, addition in future On hand, the cases. changes relative pricesand/or exchange expected other

47

rates are less likely to be relevant determinantsof tourism demand due to the lack of information(unpredictability)generallyattached to fluctuationsin thesevariables'values. The few early studies that introduce dynamic elements in their single equation specifications,mostly do so by including a lag structure in the variables,without a clear justificd be (e. the theoretical that within which structure principles can of explanation g. Witt, 1980 and Martin and Witt, 1988).More recent studiesdo considera thcory-bascddynamic in approach the formulationof their modelsand explicitly show how the empirical modelsare derived from the principles of economic theory (e.g. Syriopoulos, 1995, Vogt and Wittayakorn,1998,Kim and Song,1998andSonget al., 2000). The above discussionsuggeststhat most empirical researchusing single equation
demand fails important to tourism to study theoretical approaches consider quantitative several and methodological matters related to the modelling, estimation and inference procedures. Empirical inadequacies such as omission of relevant variables, incorrect functional form, or ignoring the problems attached to regressing nonstationary data, are bound to result in misinconsistent produce which estimates, invalid inference and unreliable specified models forecasts.These models' limitations are summariscd in Syriopoulos' (1995, pp. 318-9) remark that "the most critical weaknessof these studies has been the gencralfailure to pay attention to the 'dynamics' of tourism demand, as these works rest upon a static (and, therefore, incomplete) theoretical and empiricalframework. " An appropriate approach within the single equation framework would have to consider a flexible dynamic econometric model, involving all possible relevant determinants as well as an equilibrium correction mechanism. Such a model would allow for the separateexamination of short- and long-run impacts of the independent variables', changes on tourism demand levels. The construction of a dynamic model could be implemented according to the "general to specific" methodology discussed in, for example, Hendry and Mizon (1978), Hendry and Richard (1982) and Hendry (1987,1995). Unlike the traditional "specific to general" approach applied in most tourism demand studies, where the manipulations carried out on the 'simple' in model order to attain a more general one arc both theoretically unfounded and econornetrically inappropriate, the general to specific approach consists of modelling a theorybased demand equation in the most general terms possible, and attains afinal parsimonious

48

specification,which is legitimately derived from the general form by testing plausible imposed restrictions on the variables'coefficients. 3.3. STATIC SINGLE EQUATION MODELS OF THE UK TOURISM DEMAND IN FRANCE, SPAIN AND PORTUGAL We demonstrate the limitations of the static single equation approachin modelling by estimatingthe UK demandfor tourism in France,Spainand Portugalusing tourismdemand, and comparingthe resultsobtained.The specifications altcmativespecifications, arc definedby forms, different functional different aggregationlevels for the expenditureand considering income variablesand different sets of rcgrcssors. The specificationsusedare summariscdin the following diagram:
Dcstination i

Log-LinearFunctionalForm

Linear FunctionalForm

Aggregate variables

Percapitavariables

Aggregate variables

Percapitavariables

I RelativePr

[ -Exchange

Rates

[-Relative

Rates PricesandExchange

I EffectivePrices

Hence, for each destination, sixteen different models are estimated.The analysis of the data-cohercnce, statisticalperformance, of thesemodelsallows for and theoreticalconsistency a systcrnatic cvaluationof the mcthodologicaltools applicd in thcir dcrivation, and providcs a 49

legitimate basis for comparisonof the estimatesobtained.In the next section, we start by for framework demand discussing UK the the the assumptions tourism modelling of underlying in France,Spain and Portugalover the period 1969-1997. A full descriptionof the variables included is given. The static single equations resulting from different specifications are interpretation OLS the an of the estimationresults is provided. estimatedwith method and Finally, we compare these results with those of other similar studies and analysc the implicationsof this comparison exercise.
3.3.1. STATIC SINGLE EQUATIONSOF THE UK TOURISM DEMAND I-'OR FRANCE9 SPAIN AND
PORTUGAL

3.3.1.1. Assumptions The models are based on a number of assumptions. Separability between tourism

is labour in supply assumed, that the decisions sinceit is postulated to engage and consumption decision in to the to IIcncc, the former tourism consumption. prior made engage work arc paid are not related to thoseunderlying the spendingof the earnedincome on tourism products. SeparabilitybetweenUK consumers'expenditurein tourism and their expenditureon other This implies the separation into goodsand servicesis also assumed. expenditure of consumers is independent groups of commoditieswhere each group of preferences of the preferences underlying other groups.Therefore, it is assumedthat the expenditureallocatedto France, Spain and Portugalby UK tourists is independent from the expenditureon other goodsand including tourismbudgetsassigned to otherdestinations. services, UK tourism in France,Spainand Portugalcan be viewed as a clearly defincd product, with relatively constantmotivations underlying the tourists' decisionto visit thesecountries, over the period under study. Indeed, not only is the UK demand for these destinations (business travellerscompriseonly predominantly a demandfor holidaysby private households a small proportion of total flows), but also this group of destinationshas specific common ' in from in Europe the them or world. attributeswhich separate others,either

4 The most important are climatic factors, relatively small cultural difTerenccsand their proximity to the origin country rendering transportation costs relatively unimportant in this context.

50

We also assumeseparate demandequationsfor UK tourism expenditurein France, ' SpainandPortugal. In the long-run, France,Spain and Portugal's supply of goods and servicesto UK to be perfectly elastic, since the level of UK tourists' tourists may reasonablybe assumed be can viewed as relatively small comparedwith that of national consumptionat any period, in fact, In investmentin the tourism sectoris generally tourists most areas. and other residents future both demand, to current and and tourism-rclatedsupplyingactivities undertaken satisfy in destinations such as the onesunder consideration, are particularly vigorous where tourism both important is Furthermore, of regionalandnationaleconomics. part revenue an a shift from facilities in hotels has been in to self-catcring observed these countries, accommodation Therefore,"the spectreof constraints. contributingto overcomeany potentialaccommodation full capacity can reasonablybe ignored' (Gray, 1966, p.86). As a result, prices may be least, ' at predetermined or, explanatoryvariables. exogenous assumed Since the static equationsof the UK tourism demandfor France,Spain and Portugal included in the section nonstationarytime series, the regressionsmay be next specified spuriousunlessthe variablesare cointegratcd,that is, unlessa genuinelong-run equilibrium "exists involved. the time Consequently,for the time being, we among series relationship implying that the that the static equations a such relationship exists, assume specificdrepresent ' between UK its determinants. demand for the state relationship equilibrium steady tourismand 3.3.1.2.Functional form and variables definition The modelling of a structuralrelationshipbetweenthe UK demandfor tourism and its detenninants forra further functional issues the requiresthe discussion adopted and of suchas

3 Although thesedestinations destinations, from the assumption other can be viewed as a separate' group of Spainand Portugalcan be regarded separate as dubious,given their equations of UK tourismdemandfor France, interdependent in the first place.This previouslyassumed commoncharacteristics which makethesedestinations is inherentin the singleequationframeworkand,therefore, inevitablein this context.However,it will assumption be relaxedin chapters 5,6 and7 wheresystems areconsidered. of equations ' Tourismpricesare frequentlydetermined (e.g. package in advance holidaysand preof the actualconsumption booking conditions)and do not ordinarily respondto the level of demandin the short-run.Nevertheless, the assumption of exogenous pricesis relaxedin chapter7. 7 This assumptionis given empirical support in chapter4, where the cointcgrationanalysis the of relevant variables within a dynamicsingleequationframeworkis carriedout.

51

the definition of the variablesincluded in the single equationsto be estimated.Since theory doesnot provide a clear indication about which functional form shouldbe used,there is no a functional form the should be linear or log-linear. Although in priori certainty about whether the linear form, the coefficientsare interpreted as the absolutechangein the dependent variable its determinants, in while in the log-linear form the coefflicients per unit absolutechange represent the relative change of the dependentvariable per unit relative change in its both forms give important information concerningthe impactsthat changesin determinants, have on tourism demandlevels. However, no direct comparisoncan be made the regressors betweentheseforms using the usualeconometric criteria since in the first case,the dependent in in form, log is in linear form. Moreover,since the variables' the the the second and variable different things (absolutechanges in the linear case,and relative changes coefficientsrepresent direct in log-linear be between the them as well. case), no comparison can made or elasticities be if indirect Nevertheless, the relevantelasticitiesare computedftom can made comparison an Y X, formula (X*/Y*), form (DY/DX). linear the say, on using regressing, model c= wherec a Y X, to the with respect of and Y* and X* are the averagevaluesof the elasticity represents 8 independent Y X. dependent variable variable andthe Although some empirical studies have shown that a number of tourism demand by a linear functional form (Smcralel al., 1992), approximated can be reasonably relationships the large majority of studies applying a static single equation approachto tourism demand analysisusea log-lincar functionalform for their models.This choiceis mainlyjustified on the groundsof the sameprocedurebeing usedpreviously, or infcrior resultsbeing obtainedwith the linear form. However, in the particular modelling exercisewe are involved with here, no decision be could clear-cut made concerning the appropriatefunctional form. Indeed, the data, functional form tests and goodnessof fit measures, indicate both as the analysis of equivalent.Hence,the estimationresults of both the log-linear and linear specificationsarc to permit a thoroughcomparison reported, of statisticalpcrfonnanccs. The dependent the.UK demandfor tourism in France, Spain and variable expressing Portugalis the total expenditure UK touristsallocateto eaclidestination.Data for the UK total

1 The elasticities are computed for the average values of Y and X. In this way, estimates of the average elasticities are obtained. Hence, in the linear form the elasticities are not constant as their magnitudes depend on the chosen values of Y and X, while the log-linear form provides constant elasticities.

52

by destinationsand measuredin million poundssterling, tourism expenditure,disaggregated Afonitor (1970-1993) AL46 The Business from continuedas onecommonsource, were obtained The values of the Travel Trends (1995-1998),so reducing potential data inconsistencies. budget by dcflating UK in the tourism dependent terms allocatedto computed are variable real index. by its destination price consumer each demand for UK the tourism arc The variablesconsidered to be relevantdeterminants of both income (in disposal aggregate and per capita terms) and the price of the origin's real its by is two separable the origin's consumerprice components: tourism which represented index, both relative to that of the destinationand to that of its competitors,and the exchange both destination its for Data that the between competitors. the origin's currencyand of and rate from indexes International the UK and exchange rates were obtained the population, price Financial Statistics(IMF, Yearbook1980-98). Aggregate and per capita real expenditureand income were chosen to represent,
its in determinants dependent the and one variable of alternative specifications. respectively, However, since the UK population did not vary much over the sample period (from 55 million in 1968 to 58 million in 1997), any measurement errors are not likely to be serious and the have income to a expected per capita expenditure or and not are aggregate of consideration substantial effect on the estimated results. Given that differences in the distance between the UK and any of the destinations under be is it to be little variability argued that, over time, consideration are relatively small, can is to in Hence, travelling. the transport expected not relative costs of a cost variable expected add much explanatory power to the model. Other special events, which appear to be relevant to the explanation of the UK demand

for tourism in France,SpainandPortugal,are takeninto accountby dummyvariablesincluded in the demandequations.Such eventsarc the political upheavaland economicinstability in Portugal following the revolution of April 1974, and the political changesthat occurred in Spainfollowing the deathof Francoin 1976.It is assumed that theseeventsaffectedthe three destinationsin different periods. For instance, the Portugueserevolution influenced UK tourism demandin the region more intensivelyand for a longer period in the caseof Portugal (1974-1979)than in the casesof its neighbouringcountries.The UK demandfor tourism in Spain's Spain was affected by both the Portuguese and revolution own processof political

53

dummy in destination. durable less Hence but the this their variable effects were change This sameperiod is also includedin the equationfor Spainis definedfor the period 1974-1977. in Spain for France, in that the the assuming political change rather than equation considered that in Portugal,had greatereffectson the UK demandfor France.However its pertinencein ' be is destination to this secondary. equation expected In the static version of the single equation model, no intcrtcmporal dcpcndcnccof destinations interdependencies among competing tourism consumptionand no explicit are is it that for. Hence, the the separately arc estimated and assumed only equations allowed have demand for determinants the the current relevant effects on contemporaryvalues of from demand features is likely to the tourism The modelling of such of tourism. omission in Hence, bias the evidence of serial models' estimation results. cause misspccification in be the is these estimation to of static singleequationmodels. expected correlation the static form of a log-lincar single equationmodel of the Given theseconsiderations, i is:` UK demandfor tourismin destination LREXi = ai + PiLRDI+ + yoLRPj 86LERj+ iliDi + u, (3.1)

UK P (Portugal); S (Spain) LREXj is logarithm tourism i=F (France), the real of and where LRPJ i; is income; logarithm LRDI in destination disposable UK the and of real expenditure LERj are, respectively, the logarithms of tourism relative price in destination i and the UK destination between i These the and currencies. variableswere constructed rate exchange

asfollows:
LREXi = In(REX,) = In( ) EX, cpli

9Otherevents, for thesedestinations, which arebelievedto havehada relevantimpactin the UK tourismdemand are the oil crisesof 1973and 1979,Portugaland Spainjoining the EU in 1986and the openingof the Channel Tunnel in 1994.However,the dummy variablescorresponding to theseeventswere not statistically significant thesevariables wereomitted from thesemodels.Yet, context.Therefore, whenincludedin a staticsingleequation the significanceof theseeventsmay be capturedby dummy variables as will be shown in following chapters, derinedin differentwayswithin a dynamicsingleequation of equations'specifications. andsystem 10In tables 3.1 to 3.12 showing the estimationresults below, the superscript"a" and "c" over the relevant denotethe representation of the aggregate variables andper capitaformsof the expenditure and incomevariables The log-linear form is indicatedby adding the letter "U' addedto the namesof all variablesin respectively,. by theomissionof this letter. (3.1),andthe linearform of the modelis denoted equation

54

is is EXi UK REXj UK the the and nominal expenditureallocatedto where real expenditure i, andCPli is the consumer destination price index of destinationi;
LRDI = In(RDI)

income; is disposable RDI UK the where real

(CP'UK)

LRPj = In(RPj)= In

cpli

destination i is in CPIUK UK index; is RPI the the and consumer price price relative whcrc (CURuK CURi LERj = In(ERj) = In between destination CURI is UK i is ERi the the national the rate and currencies, exchange where is i CURUK UK destination the and currency. currencyof Di is a dummy variable which takes the value of unity for observationsin the period 1974-1979in the equationfor Portugal,and 1974-1977in the equationsfor Franceand Spain, is ith disturbance in the term a wcll-bchaved stochastic equation. ui and zerootherwise;

Spainand for tourismin France, Hence, the singleequation models of the UK demand
Portugalare:
LREX F
CtF + +YFFLRPF PFLRDI+ YFpLRPp + YFsLRPs

+8FpLERp

+SFsLERs

+8FFLERF

+'IFDF

+uF

LRPF + 8spLERp + 8ssLERs + 8sFLERF+ iisDs + us LREXs = as + PsLRDI + yspLRPp+ yssLRPs+ YSF
LREXp = ap + PpLRDI + yppLRPp + ypsLRPs + ypFLRPp + SppLERP + 8psLERs +

8pFLERF

+ ilpDp

+ up

disturbances be the to arc assumed where white noisestochastic processes. The estimationof these equationsis pcrfon-ncdusing OLS and four different sets of in four different to variables,giving rise explanatory equations: regression1, only the regression in included; in 2, the alone rates arc considered; exchange prices are relative regression in both 3, 4, only effective relative prices present; regression regression and exchange ratesare " included. pricesare

11The cffective price of tourism in destinationi is defined, in the log form, ) LP, as: = ln(P, = ln(RPj/ERj), whereLP is the logarithmof the effectiveprice (P) and RP and ER are definedas above.The linear form of the modelomitsthe letterIV in this variable'sname.
55

3.3.2. ANALVSIS OF Til E ESTIMATED RESULTS

in tables 3.1 to 3.12 labelled Portugal,Spain and The estimatedresultsare presented France.There are four tables for each destination.Each of the four tablespertaining to one destinationidentifies a different specificationof the models: the first table, specifies a loglinear functional form with aggregate values for the expenditureand income variables; the form for functional the expenditureand the the with per capita values second,specifies same incomevariables;the third, specificsa linear functional form with aggregate expenditureand incomevariables;the fourth specifiesa linear functional form with per capitaexpenditureand income. In some specifications, one or more variablesarc deleteddue to their insignificancc. The resultingestimationresultsarc labelledwith capital lettersto distinguishthem from those included. For for in Portugal, the the the variables example, equation obtained with all is in 4 includes different 4(A) two tile carried out ways: regression estimationof regression Spain France 4(B) Portugal, and and regression excludesthe effectiveprice of cffcctivc prices in for France, 4 3 the France. In the are performed estimation of regressions and equation of includes 3(A) in Regression different three all ways. relevantvariables eachparticularmodel; dummy Portugal; 3(B) the the of and rate omits relative price and exchange regression includes Spain. 3(C) Regression 4(A) the all relevant exchange omits rate of regression in 4(B) the dummy the and model; particular each regression variable variables excludes 4(C) omits the effectiveprice of Spain. effectiveprice of Portugal;regression The tables display the estimates(t values in brackets)of the intercept (INT) and coefficientsof the variablesincludedin eachmodel. The last four rows of eachtable show the for for R fit, F the overall significancc the adjusted statistic eachmodel's goodness statistic of test, and the Durbin-Watson(DW) and LagrangeMultiplier (LM) statistics for the test of distribution The LM follows and for this residual serial correlation. statistic a chi-square *, in The brackets. the symbols "and I represent, statistic correspondent p values are shown the 1%,5% and 10%significancelevels. respectively,

56

3.3.2.1.Presentationof the estimatedresults PORTUGAL

Table 3.1: Log-linear functional form, and aggregateexpenditure and income

3 RP and ER

Relative Prices Exchange Rates (ER) (RP)

4 Effective Prices A B

INT LRDla LRPp LRPs LRPF LERp LERs LERF LPP I-Ps LPF Dp
K2

-9.415 (-1.29) 1.318 (2.07)0 -1.420 (-3.29)* 3.541 (3.32)o -0.213 (-0.45)

-0.894 (-0.11) 0.568 (0.88)

1.610 (0.18) 1.222 (2.08)0 -1.696 (-4.19)9 8.051 (4.10)9 -1.981 (-2.31)"

-1.924 (-0.27) 0.579 (1.61)

-1.176 (-0.21) 0.538 (1.92)1

0.200 (0.76) -0.345 (-0.53) 0.417 (0.75)

-0.758 (-1.226) -2.132 (-3.68)* 2.399 (2.92)o -0.999 (-2.29)* 0.774 (1.22) 0.133 (0.19) -0.942 (-3.06)o 0.824 (1.46)

-0.323 (-3.05)o 0.694 13.67 0.863 4.2(0.04)

-0.254 (-1.48) 0.557 8.05 0.846 5.6(0.02)

-0.457 (-3.66)9 0.804 15.39 1.517 0.4(0.52)

-0.296 (-2.16)0

-0.297 (-2.22)0

F DW LM

0.665 0.679 12.12 15.78 1.013 1.005 3.6(0.06) 3.6(0.06)

57

Table 3.2: Log-linearfunctionalform andper capitaexpenditure andincome

1

3 RP and ER 2.457 (0.35) 1.234 (1.98). -1.684 (-4.33)* 8.047 (4.09)9 -1.989 (-2.34)0

Relative Prices Exclange Rates (RP) (ER) INT LRDIO LRPp LRPs LRPF . LERP LERs LERF I-Pp I-Ps LPF Dp
K2

4 Effective Prices A B -3.559 (462) 0.554 (1.43) -2.886 (-0.61) 0.504 (1.66)

-8.759 (-1.71) 1.403 (2.03)0 -1.430 (-3.41)o 3.519 (3.34)o -0.196 (-0.42)

-3.122 (-0.53) 0.628 (0.87)

0.177 (0.65) -0.339 (-0.52) 0.474 (0.86)

-0.768 (-1.26) -2.129 (-3.67)o 2.401 (2.92)o -1.021 (-2.37)0 0.767 (1.20) 0.150 (0.22) -0.958 (-3.10)o 0.824 (1.45)

-0.326 (-3.08)o 0.672 12.46 0.863 4.23 (0.04)

-0.246 (-1.43) 0.521 7.08 0.838 5.82(0.02)

-0.460 (-3.72)* 0.790 14.14 1.517 0.40 (0.53)

-0.291 (-2.15)" 0.639 10.93 1.013 3.6(0.06)

-0.292 (-2.20)" 0.654 14.21 1.005 3.6(0.06)

DW LM

58

Table 3.3: Linear functionalfonn andaggregate expenditure and income I 2 Relative Prices Exchange Rates (ER) (RP) INT RDIa RPp RPs RPF ERP ERs ERF Pp -287.3 (-1.07) 0.004 (1.88)' -421.7 (-3.06)o 670.9 (3.24)o -34.1 (-0.37) 0.406 (0.94) 0.060 (0.08) -8.704 (-0.65) 359.7 (1.49) -0.0005 (-0.27) 3 RP and ER B A -482.0 (-1.66) 0.005 (2.64)0 -601.2 (-4.05)o 1234.7 (3.48)o -85.6 (-0.70) 0.05 (0.08) -2.24 (-2.87)o 17.91 (1.04) -60687.0 -55459.4 (-2.24)0 (-2.97)o 0.49 (1.01) -1.74 (.2.89)9 -306.2 (-1.95) ' 0.005 (2.86)e -625.1 (-4.67)* 1011.2 (3.70)* 4 Effective Prices A B 149.5 (1.18) 0.003 (1.35) 174.0 ' (2.02) 0.001 (1.49)

Ps
PF

50231.2
(1.68)

53062.3
(1.93)'

427.7 (0.27) -87.4 (-3.34)o 0.690 13.49 1.008 4.29(0.04) -88.4 (-2.08)0 0.525 7.20 0.934 4.81(0.03) -121.6 (-3.79)o -134.0 (-4.73)* -84.1 (-2.47)0 -83.9 (-2.51)"

Dp
it- 2

F DW LM

0.643 0.765 0.657 0.755 16.17 11.09 14.41 11.76 1.406 1.049 1.447 1.044 1.39(0.24) 1 1.16(0.28)1 3.83(0.05) 13.98(0.05)1

59

Table 3.4: Linear functionalform andper capitaexpenditure and income

1

3 RP and ER A B -8.85 (-1.70) 0.006 (2.60)0 -10.44 (-4.22)o 21.65 (3.47)o -1.54 (-0.71) -5.69 (-1.99), 0.005 (2.79)9 -10.92 (-4.87)o 17.69 (3.66)*

Relative Prices Exchange Rates (RP) (ER) INT RDIO RPp RPs RPF ERP ERs ERF Pp -5.62 (-1.18) 0.005 (1.95)' "7.70 (3.33)o 11.71 (3.26)o -0.49 (-0.30) 0.006 (0.75) 0.0013 (0.10) -0.114 (-0.47) 5.86 (1.27) -0.0004 (-0.17)

4 Effective Prices A B 2.88 (1.26) 0.001 (1.15) 3.42 (2.29)0 0.001 (1.20)

-0.0002 (-0.01) -0.04 (-2,84)9 0.32 (1.05)

0,008 (0.91) -0.030 (-2,83)*

-1110.1 (-2.34)"

-1004.5 (-3.02)9 921.8 (1.86)'

Ps
PF

861.9
(1.60)

8.90 (0.32) -1.58 (-3.41)9 0.670 12.38 1.005 4.26(0.04) -1.49 (-1.95), 0.470 5.96 0.895 5.36(0.02) -2.18 (-3.83)9 -2.40 (-4.76)o -1.45 (-2.42)" -1.44 (-2,45)"

Dp
K2

F DW LM

0.747 0.611 0.736 0,626 14.76 10.76 9.80 12.70 1.394 1,435 1.039 1.033 1.46 (0.23) 1 1.20 (0.27) 3.9 (0.05) 14.1 (0.04)

60

SPAIN Table 3.5: Log-linearfunctionalfonn andaggregate expenditure andincome I I 1

3,

RelativePrices Exchange Rates

(RP) INT LRDI" LRPp LRPs LRPF LERP LERs LERF I-Pp I-Ps LPF Dp
it- 2

RP and ER
-12.434 (-1.41) 1.640 (2.72)0 -0.872 (1.84) ' 2.307 (1.27) 0.941 (1.12)

(ER) -7.602 (-1.41) 1.123 (2.37)0

EffectivePrices
-19.990 (-3.47)o 1.944 (6.57)e

-10.733 (-1.96), 1.581 (3.31)o -0.805 (-2.25)" 2.865 (2.99)0.634

0.025 (0.12) 0.429 (0.92) -0.085 (-0.22)

0.225 (0.36) 0.103 (0.20) -0.328 (-0.40) -0.870 (.2.02)' -0.624 (-1.37) 1.280 (2.09)0

0.248 (2.55)* 0.859 34.98 1.743 0.06(0.80)

0.134 (1.14) 0.798 23.15 1.244 3.73(0.05)

0.257 (2.08)0 0.839 19.19 1.69 0.15(0.70)

0.195 (1.54) 0.819 26.35 1.238 3.45(0. 61

F DW LM

Table 3.6: Log-linearfunctionalform andper capitaexpenditure andincome I I 1 2 3 Relative Prices Exchange Rates RP and ER

(RP)
INT LRDI' LRPp LRPs LRPF LERP LERs LERF I-Pp I-Ps LPF Dp
K2

(ER)
-7484 (-1.88) 1.176 (2.24)0 -9.902 (-1.45) 1.657 (2.58)" -0.823 (1.82) 1 2.270 (1.24) 0.915 (1.09) 0.012 (0.06) 0.432 (0.93) -0.068 (-0.18) 0.191 (0.31) 0.121 (0.24) -0.319 (-0.39)

_ E ct've Prices

-8.618 (-2.24)0 1.614 (3.12)9 -0778 (-2.23)0 2.799 (2.94)* 0.633

-16.648 (-3.49)o 2.022 (6.31)9

-0.822 , (-1.95) -0.644 (-1.39) 1.252 (2.07)" 0.246 (2.53)" 0.843 31.17 1.740 0.07(0.79) 0.134 (1.14) 0.778 20.60 1.248 3.69(0.06) 0.252 (2.05)' 0.821 17.08 1.68 0.17(0.68) 0.190 (1.51) 0.800 23.45 1.234 3.49(0

F DW LM

62

Table 3.7: Linear functionalform andaggregate expenditure and income I I 1 2 3 Relative Prices (RP) INT RDI' RPP RPs RPF ERP ERs ERF Pp PS
PF

4 Effective Prices

Exchange Rates (ER) -0.644 (-0.00) 0.012 (1.49)

RP and ER

-4162.3 (-3.67)o 0.027 (3.07)9 -1409.3 (-2.43)0 3817.7 (4.16)* 897.7 (2.52)0

-3739.3 (-2.54)0 0.023 (2.51)0 -1305.9 (1,96) ' 3529.1 (1.98) . 1316.3 (2.31)0

-1081.2 (-2.04)0 0.033 (6.82)o

1.102 (0.64) 4.419 (1.43) -47.241 (-1.00)

0.110 (0.04) 0.811 (0.52) -75.7 (-0.94) -337383.5 (-2.18)0

-177229.5
(-1.47) 18810.8
(2.47)0

Dp
it- 2

297.4 (2.54)"
0.871 38.73

125.1 (0.82)
0.800 23.43

226.9 (1.66)
0.864 23.28

237.2 (1.30)
0.810 24.9

DW
LM

1.768
0.19(0.67)

1.176
5.46(0.02) j

1.603
0.65(0.42)

1.080
5.64(0.

63

Table 3.8: Linear functionalform andper capitaexpenditure and income I

1

Relative Prices (RP) INT RDI" RPp RPs RPF ERP ERs ERF Pp -72.4 3.58)* 0.027 (2.84)* -23.6 (-2.44)" 65.7 (4.14)o 15.5 (2.46)0

2 Exchange Rates (ER) -0.481 (-0.03) '0.011 (1.23)

3 RP and ER

4 Effective Prices

-65.6 (-2.48)0 0.023 (2.35)0 -22.2 (2.03) ' 61.3 (1.97) ' 22.8 (2.27)"

-20.81 (-2.17)0 0.035 (6.54)9

0.020 (0.63) 0.076 (1.40) -0.804 (-0.95)

0.001 (0.02) 0.032 (0.52) -1.3 (-0.92) -5588.2 (-2.11)0

PS
PF Dp
jZ2

-3246.9 (-1.51)
322.5 (2.43)0 5.2 (2.52)" 0.856 34.26 1.771 0.18(0.67) 2.2 (0.83) 0.778 20.58 1.182 5.36(0.02) 4.0 (1.66) 0.848 20.56 1.607 0.62(0.43) 4.1 (1.28) 0.789 22.00 1.083 5.59(0.02)

F DW LM

64

FRANCE Table 3.9: Log-linearfunctionalform andaggregate expenditure and income

2 RelativePrices ExchangeRates
(RP) (ER)

3 RP and ER
A B

4 Effectiv e Prices
A B

INT LRDI' LRPp LRPs LRPF LERP LERs LERF LPP I-Ps LPF Dp
it- 2

-3.983 (-0.65) 0.979 (1.83)' 0.475 (1.19) 0.079 (0.07) 2.349 . (-6.33)4o

-10.340 (-1.50) 1.480 (2.46)0

-10.319 (-1.16) 1.465 (2.40)0 0.089 (0.19) 0.259 (0.14) -2.564 (-3.01)o

-9.021 (-1.58) 1.506 (3.48)o

-27.762 (.3.75)o 3.125 (8.22)e

-20.885 (-4.40)e 2.803 (110.39)o

1.495 (2.46)0 -2-951 (-6.90)o

0.796 (3.18) -0.335 (-0.57) -0.879

0.320 (0.50) -0.452 (-0.88) 0.622 (0.75) -0.637 (-1.45) 1.033 (2.41)0 -0.661 (-1.20) 1.458 (2.49)0 -1.430 ' (-1.82) 1.660 (3.12)e -2.100 (-4.03)e

-0.062 (-0.58) 0.971 189.04 0.752 9.41(0.00)

0.080 (0.54)

0.034 (0.27)

0.161 1 (0.99)

F DW LM

0.976 0.947 0.973 0.951 0.952 100.37 127.12 226.25 110.25 186.51 0.698 1.236 1.297 1.342 1.220 9.40(0.00) 13.37(0.07) 2.01(0.16) 2.43(0.12) 1 13.51(0.06) 1 65

Table 3.10: Log-linearfunctionalform andper capitaexpenditure and income I II 1 3 4 2 RP and ER Effective Prices Relative Prices Exchange Rates (ER) (RP) B A B A INT LRDI' LRPp LRPs LRPF LERP LERs LERF I-Pp I-Ps LPF Dp
it- 2

-4.082 (495) 0.981 (1.70) 0.473 (1.21) 0.083 (0.08) -2.348 (-6.30)9

-8.637 (-1.70) 1.509 (2.25)0

-8.066 (-1.16) 1.434 (2.21)" 0.143 (0.31) 0.260 (0.14) -2.608 (-3.07)o

-7.057 (-1.58) 1.519 (3-13)o

-19.808 (-3.19)* 3.272 (7.81)9

-14.954 (-3.77)o 2.978 (9.94)o

1.516 (2.42)* -2.954 (-6.83)o

0.797 (3.11)o -0.325 (455) -0.891

0.280 (0.44) -0.444 (-0.85) 0.640 (0.77) -0.632 (-1.43) 1.019 (2.37)0 -0.530 (496) 1.447 (2.40)" -1.536 (-1.94) ' 1.629 (3.01)* -2-065 (-3.90)9

-0.063 (458)

0.081 (0.54)

0,027 (0.22)

0.143 (0.87)

0.970 F 180.97 DW 0.752 LM 9.43(0.00) L 1

0.944 0.974 0.972 0.947 0.949 95.74 214.55 120.90 101.74 175.67 0.703 1.293 1.229 1.346 1.235 9.32(0.00) 13.48(0.06) 12.07(0.15) 2.43(0.12)_ 1 13.37(0.07)1 ----

66

Table 3.11: Linear functionalform andaggregate and income expenditure 2 Relative Exchang e Prices Rates 1 (RP) INT 1771.7 (3.49)o -1513.4 (-2.87)o 0.032 (6.42)*
1

3 RP and ER A 1569.2 (2.55)* 0.013 (3.40)o 1946.4 (7.00)* -2411.3 (3.24)* -1030.7 (-4.33)9 -115.7 (-0.19) -593.4 (-1.76)' B -1280.9 (-2-94)9 0.034 (7.74)o C 1305.9 (2.37)* 0.014 (3.71)9 1883.1 (6.99)9 -1943.1 (-3.45)o -1117.0 (-5.08)* -1.94 (-1.56) -1.381 (-0.60) 47.562 (1.16) 48.32 (2.05)" 277328.2 (3.39)o A -1188.9 (-4.23)o 0.026 (10.02)o

4 Effective Prices B -1506.8 (-5.17)o 0.032 (15.00)* C -1190.1 (-4.20)o 0.028 (14.00)o

0.013 RID18 (3.37)t RPP RPs RPF ERP ERs ERF Pp PS PF Dp

2

1716.0 (6.61)o -2464.2 (-5.99)o -791.7 (-4.97)* 0.385 (0.35) -1.034 (-0.52) -3.227 (-0.11)

-1.597 (-1.23) 1.408 (0.96) 25.452 (0.76)

262890.0 (3.21)o
(0.66)

79175 4 46797.6
. (1.24) -13625.9 (-3.79)o -3403.6 (-1.12) -11183.2 (-3.14)o

F DW 1 LM

143.1 -138.70 -211.7 -191.2 -62.2 -174.1 (2.51)0 (-2.44)0 (-2.19)" 1 (-3.64)o (-0.64)__ , (-1.98) , 0.986 0.958 0.953 0.987 0.985 0.970 0.958 0.969 129.24 294.01 115.51 256.40 180.22 374.63 214.82 219.94 0.931 1.705 1.792 1.467 0.869 1.492 0.904 1.226 12.1(0.15) 19.7(0.00)10.6(0.42) 1 7.1(0.01) 1 0.3(0.60) 1 1.6(0.21) 1 8.3(0.04) 14.0(0.05) 1

67

Table 3.12: Linear functional form and per capita expenditure and income

3 RP and ER

4 Effective Prices

Relative Exchange Prices Rates

(RP) I
INT RDI" 31.9 (3.55)e 0.012 (2.87)o

(ER)
-27.5 (-2.60)* 0.033 (5.49)o

AI

BI

CI

AI

BI

23.6 27.7 -24.7 -22.2 -28.8 -22.5 (2.50)" (-2.99)9 (2.36)" (-4.30)9 (-5.27)o (-4.33)o 0.035 0.013 (3.06)o (6.70)* 33.6 (7.34)o 0.013 0.027 0.033 0.029 (3.29)o (9.42)9 (13.69)o (14.09)9 32.8 (7.36)o

30.3 RPP (7.02)o RPs RPF ERp ERs ERF Pp PS PF

1

-42.0 -5.94)o -14.5 5.20)9 0.010 (0.47) -0.017 (-0.46) -0.132 (-0.24)

-33.9 -41.3 -0.547 (3.17)e (-0.05) (-3.44)* -18.5 -10.5 ' (-4.41)o (-1.70) -0.026 (-1.11) -0.022 (0.88) 0.467 (0.79) -0.024 (-0.56) 0.795 (1.06) 0.836 (2.00) ' 5066.1 (3.57)9 4839.0 (3.40)o
(0.67)

-20.0 (-5.20)o -0.031 (-1.43)

1433.3 875.3
(1.24)

-249.7 -206.2 -62.6 (-3.51)o (-1.12) (-3.29)o -3.3 (.3.62)o -1.05 (-0.59) -2.5 (-2.44)0

-2.4 -3.8 -3.2 (-2.40)01 (-2.24)1 (-1.93)' it- 2 0.985 0.947 0.952 0.984 0.986 0.967 0.952 0.966 F 352.42 100.64 237.33 111.3i 274.10 163.10 187.13 199.05 DW 0.847 1.688 0.900 1.458 1.770 1.508 0.904 1.244 0.7(0.39)1 7.7(0.01)1 10.2(0.00)1 0.4(0.55)1 1.5(0.23)1 8.4(0.04)1 3.7 1 LM 12.2(0.14)1 Dp

68

3.3.2.2. Analysis of the estimated results

1,2 and 3 of all equations, In regressions collinearity problemsarc likely to arisedue to the high correlationbetweenthe incomevariableand the relative pricesof Portugaland Spain (respectively,0.96 and 0.94 in the log form and 0.97 and 0.93 in the linear form), betweenthe 0.97 Portugal Spain (respectively, Portuguese the of and and prices relative rate and exchange 0.99 in the log form and 0.96 and 0.98 in the linear form) and betweenthe relative prices of Portugaland Spain(0.99 in the log form and0.97 in the linear form). Theseproblemsmight be individual the through non-significance of estimated and signs unexpected expressed high le the significance of model and values. relatively overall with coeflicients, associated 2 and 3 for Spain.Indeed,regressions in regressions 2 and 3 for Collincarity is ratherapparent Spain are overall significant and have high explanatory power; however, most of the individually Yet, their and some of signs as are not expected. non-significant are coefficients instead disappear to effective when prices, of relative prices and these symptoms seem included asrcgressors. rates,are exchange large In the presence tend to the coefficients regression show standard of collinearity, be Yet, implying they that with great prccision. as collinearity violates cannot estimated crrors So, be it that unbiased consistent occur. estimators still even can shown assumption, no classic in the presenceof severecollincarity, the OLS estimatorsstill remain BLUE (bcst linear is induce large direct The to effect of multicollincarity standard only unbiasedestimators). independent little is But, the this and effect of small samplesizes variableswith also crrors. data depends The the the the and variability of regrcssors on size sample variability. Therefore, Achcn (1982, 83) the choice. as on researchers' points out, the not p. availability and is be done ' 'what 'what to the multicollinearity? about cquivalent question should questionof " is for be ' "no if be done the can size small? which statistical given". sample answer should Sincemulticollinearity is a featureof the sampleand not of the population"we do not 'testfor multicollinearity' but can, if we wish, measureits degreeIn any particular saml)le" (Kmenta,1986,p.431). Thereareno methods of detectingmulticollincarity; "what we haveare formal, but (Gujarati, informal thumb the thumb, and some rules of same" of some all rules

12 (1983,p. 300-1),high collinearity is "afact offifie', not aproblcm; and ad hoc solutions Accordingto Learner "can be disastrously Inappropriate'. regression andridge regression that havebeenusedsuchas stepwise

69

include 335). 'symptoms' These describedabove (individual the the of rules recognition p. insignificanceof regressors with high R2 values and overall significanceof the associated high the and wrong signs of coefficients sensitivity of the estimatesto small regression, in the sampledata), as well as somemeasurement changes entities such as the coefficient of 2), factor between (VIF) defined VIF=I/(l-rij the variance-inflating regressors and correlation as 2 2 between is i J. As the the coefficient of correlation explanatoryvariables and where rij ro infinite VIF to the giving a measure unity, approaches of the speedwith which the approaches increase include the covariance and of estimators with collincarity. Other measures variance the high pair-wisecorrelation,the conditionindex andthe toleranceindex (seeGujarati, 1995). The Durbin-Watson(DW) statisticand/or the Lagrangemultiplier (LM) test for serial
level detect 5% in regressions 1,2 and 4 for this the at significance problem correlation, Portugal, and in the log form of regression I and all regressions 2 for France. Therefore, no derived be from the estimates of these spccifications. The DW test is can conclusions reliable inconclusive in all regressionsfor Spain. However, the LM test detects scrial correlation at the 5% level in regressions2 and 4 in the linear form and in regression 2 in the log form for Spain. No scrial correlation is detected in all regressions3, and I and 4 of the log form for Portugal; in regressions I and 3 of the linear form for Spain or in regressions3 and 4 of the log form, and in regressions 1,3(A), 3(C) and 4(A) of the linear fon-n for France. Hence, only regressions 3 for Portugal in both functional forms, I and 3 in the linear form and 1,3 and 4 in the log form for Spain, and regressions 1,3 and 4 in the linear form and 3 and 4 in the log form for France best the as providing results, given the reservation that a static single equation considered arc ideal be the meansto conduct a reliable analysesof the UK tourism demand. may not The following analysis is carried out at several levels: first, each destination country is considered separately; then, only the reliable regressions,according to the econometric quality criteria described above, are considered; finally, the interpretation of the coeff Merits' estimates

is given separatelyfor the log-linear and for the linear functional forms. Although for both functionalforms,the estimates do not obtainedwith the per capitaand aggregate specifications differ much, the goodness of fit and overall significancearc slightly superior for the models with the aggregate valuesof the expenditure and incomevariables.Hence,the analysisfocuses form of the models. on the aggregate

70

Demand Equation for Portugal The log form of regression 3 for Portugal is significant overall, with an F statistic of 15.39. The model explaining 80% of the UK demand variations. Except for the Portuguese level less individually 5% the or significant at exchange rate, all other explanatory variables are for income The the elasticity estimated coefficients' estimatesare as expected. and the signs of UK tourism demand in Portugal is 1.2, meaning that if the UK real disposable income increases by 1%, the demand for tourism in Portugal increases by 1.2%, celerls paribus. The UK demand for Portugal is expected to decreaseby 1.7% for each 1% increase in its relative Spain France have The of and a significant although relative prices price, celeris paribus. in for increase While demand Portugal. 1% the UK tourism the relative a on opposite effect in demand UK for Portugal, increase 8% induces tourism the Spain the of same an of price in This Spain decrease, in 2% France, tourism increase celerisparibus. classiries causes a price in 11 Portugal. in France, tourism as a complement of as a substitute and, The Portuguese currency exchange rate is not significant and has the 'wrong' sign If its the this coefficient would to negative sign on significant, estimate. coefficient attached be less is UK tourists, the to they to the willing the visit would that escudo cheaper mean Portugal. If tourism in Portugal is a normal good, this result is incongruent. Yet, the nonUK to Portuguese that tourists the sensitive rate means not are exchange significance of been is has in the the the which plausible since always escudo, escudo relative value of changes for Nevertheless, the the to coefficients of exchange rates sterling. a weak currency relative Spain and France are significant at the 1% level, and their signs confirm the substitutability and if between French Hence, Spanish, tourism, the the ratio and respectively. of complementarity Spanish (French) currency and the pound increasesby 1%, meaning that the national currency decrease demand for UK Portugal be the to would sterling, weakening relative would (increase) by 2.1% (2.4%), ceteris paribus. The interpretation of the coefficient estimate of the dummy variable is linked to that of the intercept and must take into account the fact that there is no logarithm attached to the interpretation has it in itself, intercept the The no sensible economic represents, as variable.

13 In addition,the estimated Theseresultsappear to be inconsistent with empiricalevidence. sensitivityof the UK in the relative price of Spainseems for Portugalto changes tourismdemand excessive and can only be explained by the mis-specification above. stated problems

71

form cl.61=5.00, the averageUK tourism demandfor Portugal in the year base(1990), when
UK real disposal income is f1million and the value of sterling equals the values of all other is intercept downward in 1974-1979, this the However, the shift shifts and period currencies. (1.61-0.46)=3.16 be intercept from 5.00, Hence, the value the must subtracted value e signiricant. in be interpreted decrease UK This the fl. 84 the can as average value million. which gives 1974-1979. for demand Portugal the tourism over period

is detccted. for Portugal, Both In the linear form of regression no autocorrclation -3 demand 76% UK B, tile A variations. of versions and are significantoverall andexplain around " France the cffcct Considering the versionwherethe price and exchange are omitted, rate of income is (EI UK in UK 0.005, increase that the the million) meaning of an absoluteunit by an average demandfor Portugalwould increase amountof E5000,celerisparibus. The ownlevel. its 1% Ilowevcr, is' impact tile and significant at negative,as expected, relative price interpretationis not simple given the relativeprice variable's definition. 711c relative price is a increases from, in if I indexes. base Hence, this the to between two year ratio say, price ratio in increased destination by I following in the that this 1.01 the mean prices would year, impact in UK In (pp), the the this that constant. case, of remained prices given point percentage demand for decrease L6.3 Portugal be UK the would a of this relative own-price changeon for be interpretation Similar to the can given relative price coctTicicnt million, celerisparibus. by I pp, assumingthat prices in the UK do not change,tile Spain: if prices in Spainincrease by E10.1million. demandfor tourismin Portugalincreases The interpretationof the exchange also has to take into consideration rate coefficients. for does UK demand Portugal The definition. not respond significantly to the variable's
but it does in to to the the the significantly pound relative escudo, of respond value changes if Hence, in the the to the the the value peseta. of relative value of pound relative changes in decrease, is demand for Portugal UK by increases 1, to tourism the on expected peseta Given UK by fl. 7 the tourists' spending every magnitude of paribus. million, celerls average, less difference destinations, in these more or one peseta per pound makes considerable a year for the choice of the country to be visited.

14In versionA, the relative price and exchange rate of Franceare individually non-significant.Their omission B. The F test confirmsthe omitted variablesas improves the goodness of regression of fit andoverall significance irrelevant.

72

In the caseof the linear functional form, the interpretationof the dummy variable's demand for 1974-1979 UK in Portugal is direct. Its the that the period value means coefficient high diffcrs This E134 decrease and million. value seems extremely of an average experienced log-linear form. information from the the obtained with equivalent sharply Comparisonbetweenestimatesof the log-linear and linear forms of regression3 for Portugalcan be carried out if averageelasticitiesare calculatedfor the linear form using the from in 3 formulaegiven above.Table 3.13 showsthe elasticitiesestimates regression obtained the linear and log-linearforms.
Table 3.13: Elasticities estimatesof regression3 log-linear and linear founs for Portugal Regression3 Log-linear Linear RDI 1.22 1.40 RPp -1.70 1 -1.45 1 RPs 8.05 3.09 RPF -1.98 ElIp -0.76 0.28 ERS -2.13 1.09 , ERF 2.40

depending for the UK demandfor Portugalvary considerably The elasticitiesestimates is in for Spain 8.1 fonn is The functional the of estimate relative elasticity price used. on which in linear for form. The less half fonn, log-linear that the the than value elasticity and the be form, is log-lincar Spain to the around with and around -1 estimated rate of -2 exchange form, log-linear France form. In linear the the of are relative price and exchange rate the with indicated as relevant explanatory variables of the UK demand for Portugal and their in linear form, destination Portugal. Yet, France the of a as complement coefficientsqualify is indicating in is UK Portugal that tourism these significant, not coefficients neither of in France. the to price and/or exchange rate of relative sensitive changes
Demand Equation for Spain

Regression3 in the log-linear form shows several symptoms of collinearity: the insignificant the and/or presentunexpectedsigns, the value of the variablesare majority of W is F is high (0.84) the significant overall and regression with statistic value of an adjusted is 1.64 income level. The 5% 19.2.17he the elasticity and at significant relative price estimated level its 10% indicates is Portugal the that, celeris and at coefficient estimate significant of 73

0.87% decrease in leads for increase in demand Spain. This 1% UK to that a price a paribus, indicatestourism in Portugalas a complement to, ratherthan a substituteof, tourism in Spain. is irrelevant. The dummy's coefficientis significantat the 5% level but its magnitude 4, disappear log-linear form the to the symptoms of collinearity seem of regression -In insignificant, has the price coefficient still own-effective now the expectedsign. and, although Yet, it should be mentionedthat the LM test would detectserial correlationat the 6% level. The model explains82% of the dependent variablevariability and it is significant overall with France in have 26.4. Changes F the price of effective of a significant and positive statistic an influenceon the UK demandfor Spain,indicatingtourism in Franceas a substitutefor tourism in Spain.Tourism in Portugalappears, onceagain,as a complementof tourism in Spain,as a 1% increasein the effective price of Portugalinducesa decrease of 0.87% in the demandfor tourism in Spain,celerisparibus. The incomeelasticity is 1.94and significant at the 1% level, but the dummy variableis now insignificant. As mentionedabove,relative prices and exchangerates are supposed to be relevant demand for in UK I lowever, Spain, tourism. the the variables of case of exchange explanatory have demand for UK do to the tourism. relevance explanation of appear any not rates _to'thc RegressionI in 'the log-linear form- does not include the exchangerates as rcgrcssorsbut, fit: it features 86% dependent tile good of a explains presents of variable nevertheless, is F 35, does overall with an statistic of not presentevidenceof scrial variations, significant for individually the variables price, all except own-relative arc significant and and, correlation income is The level, 1% the the signs. elasticity significant expected at positive and present in increase A 1% Portugal induces (France) decrease the relative of price unity. a above in the UK demandfor Spainof 0.81%(0.63%),celerisparibus. The dummyvariable (increase) but is level its is 5% Tile the significant at magnitude not relevant. estimateof the coefficient for is Spain level, 1% the elasticity significant at positive and around2.9. Yet, it relativeprice in the relative price of Spain induce increases UK demand for indicatesthat increases the ill Spain,which is an odd result. Collinearity and/ormisspccificationbias may be responsible for the unexpected sign, significanceandmagnitudeof this cocfflcicnt. The linear form regression1, shows the income cocfficicnt as significant at the 1% level. Its magnitudeindicatesthe averageincreasein UK demand for Spain, measuredin increase in fl. fmillion (E27000), income,ceterisparlbus. The the UK disposable million per 74

coefficients',estimates of the relative price of Portugal andTrance are significant at the 5% level, and their magnitudesindicate that aI pp increasein the relative price of Portugal inducesan approximate decrease (increase) (France),assuming pricesin the UK unchanged, of f 14 (0) million in the UK demandfor Spain,ceterisparibus. Thesevaluesindicatetourism in Portugaland in Franceto be, respectively, to and substituteof tourism in Spain. a complement The estimateof the dummy variable's coefficient is significant and indicatesthat the UK dcmand.for Spain increasedby, E297 million in the, period 1974-1977.The coefficient's is indicating Spain the samepeculiar the significant and positive, of relative price of cstimate

before. resultas
3, symptomsof multicollincarity are detected, In the linear form of regression probably due to the inclusion of the three (irrelevant)exchange rate variables.The interpretationof the is for 1. The for the logto that similar regression given elasticitiesestimates estimationresults in the table 3.14. 1,3 and4 arepresented linear and linear forms of regressions for the log and linear fonns of regressions for Spain Table 3.14: Elasticitiesestimates
Regression RDI 1 Log-linear 3 4 1 Linear 3 1.58 1.64 1.94 1.51 1.29 I (R)Pp -0.81 -0.87 -0.87 -0.66 -0.61 (R)Ps' 2.87 2.31 -0.62 (R)PFI 0.63 0.94 1.28 0.23 0.10 -0.33 ERP ERs ERF

2.35 0.71 T t2.17 1.04

0.01 1

0.10

-0.55

indicates relative prices tor regressions I and 3 and c1lative prices for regression4.

Important differencescan be observedin the estimatesof the own-price elasticities, between in 2.87, France, those the and and which range of price elasticity of which range -0.62 between 0.63 and 1.28.The otherelasticitiesestimates seemto be fairly similar. DemandEquation for France In the log-linear functional form, only regressions 3 and 4 presentacceptable quality In to the criteria. regression3(A) the dummy variable and the relative price according usual 75

insignificant from Portugal the variables omitted and were arc manifestly rate of and exchange 98% dependent 3 (B) Version the the of variations variable explains of of regression model. Except 226. for Spain, F is the of, of statistic rate all exchange and significant overall with an have levels, 5% 1% their the and or estimatedcoefficients variables are significant at -the. is indicating increase by in 1.5% income 1.5, The that the estimate an elasticity expectedsigns. UK demandfor Francefollows a 1% increasein the M real income, celerls paribus. The is indicating 2.95, for demand UK France that the estimate price elasticity own-relative by around3% for each 1% increasein the own relative price, celerlsparibus. The decreases demandfor tourism in Francealso respondssigniricantly to changesin the relative price of Spain as a 1% increasein this price induces,celerisparibus, an increaseof 1.5% in the UK demand.Hence,Spain is a substitutedestinationof Francefor UK tourists. A close to unity in for if is France the that, exchange rate suggesting with changes elasticity estimate associated the French franc becomes1% cheaperrelative to the pound, the UK demandfor tourism in by aroundI%, ceterlsparibus. Franceincreases 4(B), all coeflicientsare significant at the 1% level. The model explains In regression
95% of the variability of UK demand for France, and is significant overall with an F statistic of 186.5. However, it should be noted that the LM test for scrial correlation would detect this level, be 6% Tile the that the omittcd variables suggesting may rcgrcssors. at relevant problem income elasticity estimate is 2.8 meaning that, celeris paribus, the UK demand for tourism in France increases by around 3% per 1% increase in the UK real income. The own cffectivc indicates increase is 1% in 2.1 induces France that the which a estimate of price price elasticity demand in for UK France, for in decrease The demand UK 2% the tourism celeris paribus. a France increases by 1.7% if prices in Spain increase by 1%. This result indicates Spain as a destination France. of substitute The linear form of regressions 1,3 and 4, present acceptable quality according to the detected, 1, In the model explains 98.5% of no signs of correlation are usual criteria. regression the variations of the dependentvariable, it is significant overall with an F statistic of 374 and individually level, 1% the the significant at coefficients are presenting the expected signs. A all in income UK disposable (El has the million) unity change positive an estimated positive impact on demand of E13,000. The coefficient for the relative price of Portugal is highly if in Portugal increase indicates by I pp, while prices in the UK that prices signiricant and

76

in increases France by for E17 UK'demand tourism the million, ceterts remain constant, is'a for Portugal France. indicates destination The that substitute coefficient sign paribus. destination for indicates Spain Conversely, the this the estimated relative as price of coefficient in increase in France. Indeed, tourism the to, than aI pp a substituteof, a complement rather for E25 in demand induces decrease UK France, Spain the of around million a relative price of increase by decrease in if in France I However, the tile pp, estimated prices ceterisparlbus. be E7.9 UK demand million, celerisparibus. will only Regression3(B) excludesthe dummy, relative price and exchangerate of Portugal.
Ilowever, both the DW and LM tests detect scrial correlation in this model indicating the Version being (C) omitted. variables of this model omits only the relevant possibility of increased Its Spain. explanatory power, overall significance and absence of of exchange rate for Except Portugal, the this correct. choice exchange of reveals as rate all the scrial correlation levels. individually 1% 5% The the at or significant estimated cocfficicnt of other variables arc increase in indicates for France that the value of the pound by I franc an the exchange rate induces an increase in the UK demand for France by E48 million. Given the relatively strong UK in France franc (e. E1958 the tourists' the the of expenditure and magnitude g. value of in increase franc is UK likely in 1997), the the per pound relative value currency aI of million difference decision interpretation France. 17he to the to visit of economic to make a significant for is to that regression 1. given the other coefficients similar In the linear form of regression 4, the omission of explanatory variables appears to

detected by DW LM (A) IIcncc, the tests. the as and/or version correlation only causeserial impact Celeris increase be in LI UK the to paribus, estimated of acceptable. a million seems incomeis a E26 million increasein the demandfor France.The cffcctive price of disposable Spainis not a significantexplanatoryvariablein this model but the effective pricesof Portugal is interpretation due The France to the way these of estimates complicated are. economic and in which the cffective price variable is defined.11crice, to be appropriateto an exampleseems explaintheseestimates. Considerthe following valuesextractedfrom the data, for tile relative price, exchange in France 1: the period 1990-199 rateandeffcctiveprice of

77

CP'F 2-RPF CPIUK

CURF ERF= CURUK 1 9.72 9.98

PF =

RPF
ERF

1990 1991

1.000 0.975

0.1029 0.0980

The relative price of France(RPF)is defined as the ratio betweenthe price index of is d6lined by the France (ERF) Franceand the price index of the UK. The exchange rate of buy. The is francs France defined (PF) by I that effective can of sterling price of pound amount the relativeprice andthe exchange the ratio between rateof France.Considerthe exchange rate in 1991 franc francs incrcascd fl), (9.72 1990 that the the relativc valuc pcr and assumc of of francs as a Ipp increasc by I pp. Thcn, in 1991, Elwould bc 9.6228 (9.72-0.0972=9.6228) in in incrcasc franc's 0.0972 If the of valuc. rclative changc no occurs the rclativc: mcansan in France, in in 1991 be 0.10392.1 lencc, the tourism tile effective change price would of price the cffcctivc price of Francein 1991relative to its value in 1990is 0.10392-0.1029=0.00102. by This change multiplied the coefficient's estimateof the effective price of France(-13625.9) is 413.9 million. Hence,-13.9 represents the impact on the UK tourism demandfor France franc inercascs by I pp, given that the own-rclativeprice did not the the value of relative when hand, On is the that the exchange other assume change. rate constantbut the relative price of by I pp. Then, its value increases from I in 1990,to 1.01in 1991.The cffcct Franceincreases in is 1.01/9.72=0.10391. the Then, the cffcctivc price would this effective price variation of is by 0.00101from 1990to 1991.This increase increase by cocfi"icicnt multiplied the estimated in the 413.9 million, representing the impact on the UK demandfor Franceof a Ipp increase France, in Therefore, the tile exchange rate and of given price exchange rate. changes relative in relative prices causechangesin the effective price which impact on demandin the same

way.
Considernow the following valuesof the samevariablesfor Portugal:
RPP = CPIP
CPIUK

ERp =

CURp

CURUK

pp =

RPp ERp

1990 1991

1.000 1.052

254.41 255.64

0.0039 0.0041

78

by I pp from 1990 Assumethat the value of the escudo relative to the pound increases relative price to 1991.Hence,the value in 1991is 251.87escudos per pound.Givcn a constant is in 1/251.87=0.00397. The in 1991 its Portugal, now change the effective effective price of by be, This 0.00007. the estimatedcoefficient of the then change multiplied price would impact is E19.4 This (277328.2) Portugal the the on million. value represents of price cffcctive by I pp, provided for Francewhenthe exchange UK tourism demand rate of Portugalincreases hand, is if On does the the other rate constantbut exchange. that the relative price not change. its in by be, 1990 in I I 1.01 increases Portugal value pp, will say, and the relative price of Then, 1991.The effect of this variation in the effective price would be 1.01/254.41=0.00397. before. This by incrcases by 0.00007 as change, multiplied the codficient's the cffcctivc,price impact demand for is UK France is L19.4 the tourism the same on which million estimate, Portugal. Table 3.15 in increase I by the the of shows rate clasticitics exchange of pp an caused linear functional fonns for log-linear France. the the of regressions and with obtained estimates for France for the log and lincar forms of rcgressions Table 3.15: Elasticitiesestimates 1 ElIs ERF (R)Ppl (R)Ps' (R)PFI ElIp Regression RDI 3 (A) 3 (B) Log-linear 4 (A) 4 (B) 1 3 (A) Linear 3 (C) 4 (A) 1.47 1.51 3.13 2.80 1.18 1.18 1.27 2.36 1.30 1.48 1.43 1.30 -0.66 0.09 0.26 1.50 1.46 1.66 -2.46 -2.40 -1.94 0.44 -2.56 -2.95 -1.43 -2.10 -1.02 -1.32 -1.44 -0.29 -0.36 1 I 0.29 1 I 0.30 0.57 0.32 -0.45 -0.64 0.62 1.03

-1.77 III ___ ' Indicates relative prices tor regressions I and J -lJorregression4. and ellcuive __ prices

log linear forms in for Francepresentimportant The estimates the and of regressions discrepancies. For example,the incomeelasticity rangesbetween1.18 and 3.13. Ilic price of Portugalvariable is not significant in the log form but is significant in tile linear form. The Spain between 1.66. Ilic the of range price and estimates of elasticity own price elasticity -2.46 79

The for 2.95. between 1.02 the exchange estimate and rate of Franceelasticity cstimateranges 1.03 (significant). between 0.30 (non-significant), and ranges models' An overall view of the estimationresults,is providedin tables3.16,3.17 and Table 3.16 presents 3.18, showingthe elasticities'estimates the obtainedwith eachregression. for for 3.18 for 3.17 France Spain. Portugal, table table and results

demand for Portugal Table 3.16: Elasticitiesestimates the equation of

1 PORTUGAL Log-linear 1 Linear Log-lincar 2 Linear Log-linea, Aggregate Percapita Aggregate Percapita Aggregate Percapita Aggregate Percapita Aggregate Percapita Aggregate (A) (B) Aggregate Percapita(A) Percapita(B) (A) Aggregate (B) Aggregate Percapita(A) Percapita(B) (A) Aggregate L i near (B) Aggregate Percapita(A) Percapita l(B) RDI 1.32 1.40 1.12 1.39 0.57 0.63 -0.14. -0.11 1.22 1.23 1.39 1.39 1.67 1.39 0.58 0.54 0.55 0.50 0.83 0.28 0.82 0.28 -1.70 -1.68 -1.39 -1.45 -1.38 -1.44 -1.00 -0.94 -1.02 -0.96 -0.86 -0.79 -0.90 -0.82 8.05 8.05 3.78 3.09 3.77 3.08 0.77 0.82 0.77 0.82 0.86 0.91 0.84 0.90 0.19 0.17 0.13 0.15 -0.35 -1-98 -1-99 -0.34 (R)Pp -1.42 -1.43 (R)Ps 3.54 3.52
( 11) PF

ERp

Ells

ERF

-0.21 -0.20 -0.13 -0.11 0.20 0.18 0.19 -0.76 -0.77 0.03 0.27 -0.01 0.26 -0.35 -0.34 0.23 1 0.04 0.05 2.13 . 2.13 . -1.40 -1.09 -1.43 -1.07 0.65 0.42 0.47 -0.31 -0.23 2.40 2.40 0.64 1

2.05 -0.98 1 2.04 -1.02

Linear

Log-linear 4

80

for the demand Table 3.17: Elasticitiesestimates equationfor France

FRA NCE 1 1 Aggregate Log-linew Per capita Aggregate Linear Log-l i new 2 Linear Per capita Aggregate Per capita Aggregate Per capita Aggregate Log-lincai Per capita Aggregate (A) RDI 0.98 0.98 1.18 1.10 1.48 1.51 2.91 3.02 1.47 1.51 1.43 0.09 0.26 1.50 0.14 1 0.26 -2.56 -2.95 -2.61 (R)Pp 0.48 0.47 (R)Ps 0.08 0.08 (R)PF -2.35 -2.35 -1.02 -1.07 0.80 0.80 0.07 0.11 0.32 -0.34 -0.33 -0.21 -0.20 0.45 . -0.64 0.28 1 -0.44 -0.88 0.89 . -0.04 0.09 . 0.62 1.03 0. 1 ERp ERs ERF

1.30 1 -2.46 1 1.32 -2.41

3 Linear

Aggregate (B) Percapita(A) Percapita(B) (A) Aggregate Aggregate (B) Percapita(A) Percapita(B) Aggregate (A) Aggregate (B) Percapita(A) Percapita(B) Aggregate Percapita

1.52 1.18 1.27 1.19 1.19 3.13 2.80 3.27 2.98 2.36 2.55 2.48 2.66 1.30 1.24 1.36 1.30 1.48 1.43 1.46 1.43 -0.66 -0.53

1.52 40 -2-. -1.94 -2.37 -1.94 1.46 1.66

2.95 . -1.32 -1.44 -1.37 -1.48 -1.43 -2.10 -0.29 -0.36 -0.28 0.33 .

-0.63 0.29 0.26 .

1.02 0.30 0.57 0.31 0.56

Log-lineai 4 L inear

1.45 1 -1.54 1.63 . 2.07 0.44 0.46 -1.77 -1.45 -1.80 -1.49

81

for the demand Table 3.18: Elasticitiesestimates equationfor Spain

SPA IN Log-lincar Aggregate Per capita Aggregate Per capita Aggregate Log-linear 2 Linear Per capita Aggregate Per capita Aggregate Per capita Aggregate Per capita Aggregate Per capita Aggregate_ Per capita RD1 1.58 1.61 (R)Pp -0.81 -0.78 (R)Ps 2.87 2.80 2.35 2.30 (R)PF 0.63 0.63 0.71 0.70 0.03 0.01 0.13 0.13 -0.87 -0.82 -0.61 2.31 2.27 2.17 0.94 0.92 1.04 1.03 1.28 1.25 1.48 1.43 0.23 0.19 0.01 0.01 0.43 0.43 0.56 0.55 0.10 0.12 0.10 0.23 -0.09 -0.07 -0.34 -0.33 -0.33 -0.32 0.55 . -0.54 1 Ellp ERs ERF

Linear

1.51 1 -0.66 1.52 -0.63 1.12 1.18 0.67 0.62 1.64 1.66 1.29 1.29 1.94 2.02

Log-linca,

Linear

2.15 -0.59 1 -0.87 -0.82 -0.62 -0.64

Log-linear 4 Linear

1.85 1 -0.98 -0.60 1 1.97 0.64 -0.92 1

The differences in the magnitude of the elasticities estimated from models with in in dcrined defined terms, and with models variables aggregate pcr capita terms variables in 0.3. This is that, terrns there suggests of elasticities estimates, practically no never exceed differencebetween modelsusing aggregate or per capitavariables,either in the log-lincar or in the linear forms. Then, at least from this point of view, the results are consistent.However, there are striking disparities in the estimateswhen different sets of rcgrcssorsor different functional forms are considered.For example, in the equation for Portugal, the elasticity between for Spain in 0.77 8.05; the the equationfor Spain,tile of ranges price and estimate between in 2.87 for France, the the ranges elasticity estimate and and equation own-price -0.64 income for Spain the and of price range,respectively,betwccn -2.46 and 1.66 and estimates bctwcen1.10and3.27. Given that all regressions were estimatedusing the samesampleperiod, tile samedata destination the countries, within tile samc static single equation set, same origin and 82

framework, the discrepanciesfound in the estimation results do not pennit strong or demand UK behaviour in destinations. inferences tourism the these about convincing in involved Furthermore, the regressions time the estimation the arc nonstationary, as series inference invalid. Hence, any policy directives be the statistical spuriousand results may undertaken on the basisof theseresultswould be controversial. Nonstationarityaside,the resultssuggest that the functional form adopted,the set of definition included in the the the of price and variables play a crucial role rcgrcssors Consequently, the of coefficients' significance estimates. comparison signs and magnitude, betweentheseresultsand thoseobtainedin other studiesis likely to be qualified as a futile if for Nevertheless, only even exercise. pedagogicalreasons,there arc and meaningless important lessons to be learned from the comparison of the methodologiesused, the inferred from interpretation the conclusions studiesconcerned of resultsofferedand with the investigationof tourism demand.Our next task is to analyscseveralof thesestudiesand, when possible, to comparc methodologies,results and conclusions, as well as their implications.
3.3.3. COMPARISONAND IMPLICATIONS OF THE ESTIMATE. D RESULTS

for analysismust havesimilarities both betweenthemselves The modelsselected and in different this the study, as consideration used of completely models methodologies with focuses fruitful Hence, UK demand for the tile selection on purpose. studies of no serves tourism within a static single equation approach,using comparablefunctional forms and include which and similar explanatory variables relative or effective prices sampleperiods, Under thesesetsof conditions,the studiesselected rates. exchange are Witt's (1980b) and/or for demand in UK 16 destination tourism the of main countries; Locb's (1982) analysis demand in UK US; Uysal tourism the the of and Crompton's(1984) study of tile examination UK tourism demandfor Turkey; Papadopoulos and Witt's (1985) estimation of the UK demandfor tourism in Greece;Gunadhiand Bocy's (1986) analysisof tile UK demandfor tourismin Singapore and Witt andMartin (1985) andMartin and Witt's (1987) studiesof tile for tourismin several UK demand major destinations. Witt (1980b) examinesthe UK demandfor tourism in sixteen destinations,using log-linear 1965-1972, data for functionalform and the OLS methodof the the pooled period

83
0

demand by by (air travel Witt's tourism and mode of sca/land) and study separates estimation. variable is type of holiday (independent and inclusive),using dummy variables.The dependent income is in The by terms. the the variable measured capita per and numberof visits measured in for three of the specifiedmodels the variables explanatory cocfficicnts' estimates someof following in displayed table: the arc
Relative Per Costof Relative Dep!ndent MODELS costof capita travel price variable income tourism Model I Model 2 Model 31 Visits Visits A Visits 0.52 1.39* 1.45 0.24 . -0.05 -0.69* -0.74* -0.20 0.36 Lagged Travel dependent mode variable 1.32* 0.15 0.91

Travel time -1.00* 1 -0.17* -0.27

Holiday type 1.39* 0.16

* indicates significant at the 5% level

In Model 2, travel time, per capitaincomeand the laggeddependent variable seemto provide for is No demand UK tourism the needs. other cxplanatory variable all the explanation
holiday have travel Hence, type travel, the mode, and which a significant cost of significant. impact on the demand for tourism in Model 1, vanish as determinants of that same demand behaviour in Model 2. Moreover, the impact of the travel time variable in Modcl 2 is less than

in 1. is lcvcl in It Model that the 20% of its estimated possible variables uscd Models I value In Model differences 3, first 2 them the trends related. make spuriously which of share and data. in In to the avoid problems of seems an attempt nonstationary what variablesarc used, between long-run UK demand its is determinants however, the the relationship this case, and difficult to analyse, sincemodelsin differencesare intendedto dcpict short-runbehaviour.The first differencespecificationindicatesa significant role for the tourism cost variable, but no is othervariable statisticallyrelevant. Loeb (1982)studiesthe UK tourism demandfor the USA using time seriesdata for the functional form log-linear OLS. The dependent is 1961-1978, and variable a real period Loeb specifiestwo alternativemodels: using nggrcgate tourism expenditure. expenditureand income;and using per capitaterms for the dependent variable and the incomevariable.Some in following the table: the arc reported of estimated elasticities 84

I Models Dependent variable

Percapita Aggregate Relative real real price income income 1.04* 0.87 -6.36* 5.25*

Exchange rate 4.07* 2.64

Per capita Model I LAg expenditure en Model 2 gg gregate g expenditure xpen .tu

0.95 0.86

* indicates significant at the 5% level

The income variable is significant when measuredin per capita terms but not so when for in The terms. estimates the exchangerate elasticity arc considerably measured aggregate different in both models. If the R is the main criterion for measuringthe quality of a income is I lowevcr, the the capita expenditure specification with per and superior. regression, R2, by itself, is not a dependable criterion of statistical quality. In contrastwith the results Loeb's from in the that previously, results show models estimated variables measured obtained
differences in terms the cocfficicnts estimates. substantial capita can produce or per aggregate Given that per capita variables can be obtained by simple division of the variables' aggregate levels by the UK population and that the UK population did not vary much over the sample fit differences indicate the of estimates and goodness reported may period considered, misspccificd regressions.

Papadopoulos and Witt (1987) estimatethe UK demandfor tourism in Greecefor the log-lincar functional form, Cochranc-Orcutt 1972-1982 (CO) estimation the using a period UK dependent the the of as number of visits per capita population variable.Their and method income, per capita relative cost of tourism, travel cost, advertising explanatoryvariablesarc: dummy Their in following the variable. table: a estimation and results are reported expenditure
Relative costof tourism

Dependent Percapita variable real income

Advcrtisin71 Travel cost expenditure

Dummy

DW .. stapstic

Number of 6.67* visits per -1.67 capita II * indicates signif icantat the 5% level

-0.28 I I

0.26 .I

-0.40 I

2.49 I

0.92 _j

85

is per capita incomeand this fact alonespeaks The only relevantvariablein this regression for the quality of the model. The CO method is used to deal with serial correlation problems detectedpreviously. However,if the serial correlation detectedis due to misspccificationsin the model, the CO methoddoesnot solve the problem. Furthcnnorc,the DW statistic is not a
is detection CO With the estimation of performed. method whenever such a specification valid for few degrees freedom is R2 the the the estimation of procedure, magnitude of and WW R2 degrees 1000 freedom is 0.30 to of with always preferable a meaningless, as an or

0.99 with 10degrees of freedom. Uysal and Crompton (1984) estimatethe UK demandfor tourism in Turkey for the functional form. log-linear dependent The is 1960-1980 using a variable derinedin two period Turkey" "number "expenditure in Turkey by tourists". tourists of visiting and ways: alternative The model with the first dependent variable(Model 1) is cstimatcdwith OLS while the model is dependent (Model 2) the variable estimated with CO. The independent second variables with in both models are per capita income, relative price index, exchangerate and promotional in following The the table: estimation results are presented expenditure.
Dependent variable Number of tourists Total expenditure Percapita income Exchange rate 1,68* 1 2.09 -1.49 2.84 1 0.28 Promotional expenditure 0.28* j 0.93

Models

Relativeprice 1.57*

R2

DW

Model I I Model 2 L.

-0.064

0.87 I

1.64 I 1.59

* indicates significant at the 5% level

According to the authors"wheneverserial correlation was detectedby the DIV statistic, a Cochrane-Orcutt procedure was usedin an attempt to alleviate the problem" (p. 293). Yct, be bias Strongcollincarity is cannot "alleviated" by the useof CO procedure. misspccification likely to be presentin Model 2 where all coefficicrits are individually insignificant, In an attemptto improve the model, the authorsomit the variable 'promotional expenditure'(PE). I lowcvcr, omitting this variabledoesnot seemto be the solution in spite of the authors' claim "the removal of variable PE increasedthe coefficientsof the Incomeand the exchange rate in "increase" (p. 296). The the text refersto the magnitudeof the cocMicients' variables" word 86

in increase An the but their omitted magnitude can occur when even significance. not estimates it instead it is the that, of solving problem, makes worse variable relevantwhich would mean by the omissionof a relevantvariable.The DW statistic in Model 2 is meaningless since the CO procedureis used. In Model 1, the income elasticity estimateis insigniricant with the inconclusive for detection indicates DW test the the of serial an statistic wrong sign, and "that Hence, doubt, Inconclusive the authors' mostof remark-, not certaintyl means correlation.
dismiss is (p. 296), to the not enough the estimated equations werefreefrom autocorrelailon" in for 20 DW in 15 tile the was equations which statistic of out correlation presenceof scrial inconclusive zone.

Gunadhiand Bocy's (1986) study the UK tourism demandfor Singaporein the period in OLS. dependent is form The functional log-linear the 1965-1981, variable and model with a
income, the capita arc real relative shopping per tourist arrivals and explanatory variables in following Their hotel tile estimates arc showed prices and exchange rate. prices, relative table:
Relative shopping prices -0.41

Mod'71s

Dependent variable Tourists arrivals Tourists aff ivals

Percapita income 3.74*

Relative hotel prices

Exchange rate

R2

DW

Model I Model 2

-0.20

-0.01

0.94 I I 0.91

1.27

7.30*

1.91

* indicates significant at the 5% level

In Model 1, only the income elasticity estimateis significant which, consideringthe high F0 for However, indicate the tile the solution of multicollincarity. authors' presence value, may is (Model 2) insignificance individual to model using a stcpwisc a second estimate problemof but income, in the variables one, are per capita estimationprocedure which all explanatory for DIV "the Indonesia The that then statistic claim proved conclusive only authors omitted. for UK, implying thesetwo equationsneither a misspecification that the nor an omission and four lines below, has been this contradictory they add, and made" of significant variables UK DWstatisliefor Indicates "Although the the rejection offirst-order the equation statement: IoW R2 income the the that value and significance of alone,suggests relatively autocorrelation, 87

may he other explanatoryvariablescapableof increasingthe explanatory power of the model. being has led by the incomevariable thus Inducing to their their picked effects rip omission ... (pp. 245-46). Yet, one could say, instead,that tlicir income elasticity" an over-estimateof omission may have led to misspecificationbias, renderingthe model invalid and uselessfor inferenceproposes. No valid conclusions canbe retrievedfrom suchmodels. Witt and Martin's (1985) estimatethe UK tourism demandfor severaldestinationsin log-lincar functional form the pcriod 1965/8-1983, and the OLS and CO mctliods.As using a in Witt (1980),the authorsseparate the UK tourismdemandby modeof travel (air and surface) inclusive). holiday The dcpcndcnt (independent is by type and of variable visits per capita and income, the capita real variables arc per explanatory relative cost of tourism,cost of travel, and lagged dependent Their trend. and a variable rate, estimationresultsfor destination exchange in France, Italy following Spain the table. arc prcscntcd and countries
Dependent Destination Percapita Relative Costof income Travel cost variable Visits/head (V/11) independent, by air France OLS Italy OLS Spain OLS France CO Italy 1.43* 2.69* 1.46 I 3.81 5.55* 0.87 I 1.35* 0.94 0.73 2.54* 2.12* 2.88 -0.21 -0.32 -0.35 -0.12 -0.11 -0.75* -0.44* -0.51 -0.46 -0.91 -0.54 ' -0.46 -1 '-0.08 I 1 -0.36 -1.60* -0.02 -0.11 -0.03 I 1.16 . Exchange Lagged (V/I 1) rate 0.75 1.08* 0.72* 1.23 0.77* 0.52* I 0.72* 0.46* 0.31 0.67* 0.82* 3.68* 2.08 I I 0.31 -0.05* 0.62*

Trend

DW 2.06 2.00 2.00

le 0.82 0.93 0.97 0.74

Visits/hcad (V/I 1) inclusive, _CO by air Spain CO

I 0.11*

0.88 0.91 0.97 0.55 0.66 0.94 0.64 0.74

Visits/head
(V/11) independent, by surface

France CO Italy CO Spain CO France CO Italy OLS Spain CO

I -0.04

Visits/hcad (V/11) 1 inclusive, by surface

1 - inuicates signiticantat tne :i,/o ievei

88

The inclusion of an explanatory variablein one equationand its omissionfrom another similar one, is not explainedby the authorsand can only be viewed as relatedto the statistical basis is for the the useof different setsof variables. a sound not of models,which performance Suchcriteria for excludingor including variables"leave a puzzleover economicinterpretation that requires resolution if the studiesare to provide insights into economicbehaviour,and if (Johnson into a trawling of the data simply to obtain the best Y11"' they are not to degenerate CO 1990, 149). Again be Ashworth, the to estimation method seems p. usedto 'correct' and linked be to misspecificdregressions. serialcorrelationwhich may The resultsreportedin the studiesselectedfor comparisonand in many othcrs, differ considerablywhen the author(s)eliminate,add, and/or changethe definition of one or more included justify However, to comments arc no clarifying satisfactorily such variables. differences. In addition, when scrial correlation is detectedthere is no discussionof the isspcci ri its being by The detected is caused cation m effors. scrial of correlation possibility inherent be disturbances 'corrected' by to to the therefore, the Cochraneand, assumed of1cn Orcutt estimationmethod.In some studies,the dependent variable is mcnsurcdin aggrcgnte in per capitaterms. For the sakeof consistency, termswhile the incomevariableis considered in income both be In terms. the should measured same and addition, someauthors expenditure interpret insignificant variables as if they were significant determinantsof the dependent draw implications from conclusions and others policy and modelsthat would not pass variable, a morerigorousquality examination. 3.4. CONCLUSION Research on tourism analysishas largely beenfocusedon the demandside, attempting to establishits determinants and quantify the effects of changesin them on the dependent find determinants The the studies generally main of tourismdemandto be the origin's variable. income, rates,relative prices,and a numberof other qualitative and quantitative exchange real factors,dependingon specific circumstances of the countriesanalysed.The literature shows that the majority of investigatorsuse a static single equation approachto model tourism , demand behaviour.Thesemodelsgenerallyinclude different origins and destinationcountries, different sample sizes, different measurementcriteria and dclInitions for the variables 89

involved, and different estimation methods. In spite of these differences the researchers' commonaim is to estimatethe sensitivity of tourism demandto changesin its determinants. dependon the accuracyof such information for decision firms and governments Consumers, inference However, the the tile of precision estimates and validity of and policy making. and forecasting proceduresdependscrucially on the robustnessof the theoretical framework for the the tile of a on use sound econometric specifications'and methodology underlying specificd. modelling,estimationandevaluationof the quantitativerelationships The main contribution of this chapter is to show how small differences in the demand tourism within a static single equation context can affect the of specifications inconsistent signs and significancc, estimates magnitudes, providing resultsupon coefficients' implication be based. Taking UK tile conclusions can or policy reliable as an origin no which destinations, demand Spain Portugal for France, tile tourism we as modelled and origin's and log-lincar lincar functional forms forms), (the different dcrinitions different for the and using income dependent variable (aggregate the variable and and per capita values),and diircrcnt (exchange variables ratesalone, relative prices alone, exchange ratesand sets of explanatory A, together prices and effective alone). widc rangeof elasticitiesestimates prices was relative functional form by by the the changing simply of models or changing the set of obtained in the as caseof using relative pricesand exchange variables, ratesseparately or, explanatory instead,using them combinedto form effective prices. This instability of tile estimateswas further confirmed when earlier studies of tourism demandwere analysed.While in some income has demand, in is a positive and significant real effect on others,this regressor studies, insignificant and/or has a negativeimpact on the dependent variable-, while someestimations for this as significant rates alone rcgressois, exchange claim others present role relative prices flind both be included that others should still and as relevantexplanatoryvariables.Even alone do dctcn-ninants' the the on significance agree and sign of when studies effects for similar nation pairs, their magnitudescan differ so sharply that neither a consistent origin/desti judgement be nor a sound can established can be madefrom the estimationresults. comparison Moreover, either evidence of collinearitY and autocorrelationis ignored and insigniricant rication bias, such as omission of relevant results are interpretedas significant, or misspcci variables,arc 'blindly' tackled as true scrial correlation and 'solved' by the application of suchasthe Cochrane-Orcutt. alternative methods estimation 90

Which variablesshould be included?Which functional forms should be used?Which formal be Econometric trusted? models are quantitativerelationshipswhich estimates should link theory and data to allow for the understanding of economicbehaviour.The building of is econometricmodels generally associatedwith the objective of providing a consistent descriptionof the data generatingprocessand a reliable meansof predicting its behaviour. However,by nature,modelsarc simplificationsof "the real thing" and as such,they can differ is, in That they are not equally useful or reliable to portray their empirical relevance. radically they intend to explain. Indeed,as Learner(1987, p. 1-2) points out, "models, the phenomena Afore from Importantly, metaphors. othenvise, are merely a practical standpoint, stochasticor depend does inferences indicate that substantiallyon the choiceOfthe ivhensensitivityanalysis be by doubt inferences the can relived only eliminating altogether the about metaphor, problem's metaphor". From the analysisof resultsprovidedby the modelling and estimationexercises of this is by those the that the estimated other studies, conclusion of numerous main chapter and the of variables' cocfficicnts and the general statistical signiricance signs and magnitudes, form dcrinition functional depend the tile the on crucially adopted, models, on ofthe of quality included, In the the on set of regrcssors considered. and other words, within a static variables inference depend "on to the seems choiceof tile metaphor". approach, equation single Static single equation models of tourism demandtend to neglect interdependencies ignore from destinations, data, problems arising nonstationary overlook among competing dynamicsand lack an explicit theoreticalbasiswithin which consumcrsil preference structure Hence, be be these modelled. models can only appropriately viewed as specific cases can demand bchaviours, but the of plausible wide range not as comprehensive or reliable within descriptions of tourist's general conduct. Modelling procedures constrained by these bound inadequate faults to produce empirical specifications arc which generate methodological 'elimination altogether'of the static single biasedand inconsistent estimationresults.Perhaps, if is indispensable is in directions. 'metaphor' to research proceed more valuable equation The featuresof consumerbehaviour generally omittcd from static single equation modelling are, on the one hand, the specification of a dynamic structure which clearly in demand functions long-run the effects separates short- and and, on the other )land, tile derivationof demandfunctionsfrom an explicit economictheoreticalbasiswhich permitsboth 91

influences the formal testingof utility theoryhypotheses and the consideration of cross among destinations.In chapter 4, we addressthe first omission integrating the dynamics of UK tourism demandin a single equationerror-correctionmodel. In chapters5, we deal with the second omissionby meansof a static systemof equationsapproach.In ch3ptcr6 both features model. arc includedin a dynamicsystemof equations

92

CHAPTER 4
DYNAMIC SINGLE EQUATION MODELLING OF
THE UK TOURISM DEMAND

4.1. INTRODUCTION The first objectiveof econometric modelling is to provide a coherentexplanationor the observed behaviour of economic variables. Understanding relationships between involves ideas data theoretical a process which relates empirical phenomena and economic framework. decisions have be In this to process a number of made a quantitative within idea, is be to tested the among altcrnative sensible choice which ones, of going startingwith inrorniation data Reliable is models. empirical of using quantitative statistical set againsta information, however by it in Yet, itself, this this exact process. cannot explain crucial Nor however theoretical can a causal structure, phenomena. clever and creative, economic be 'realistic' without the supportof empirical evidence.Quantitativeanalysisof economic for based the theory a systematic search matching of comprises observation and phenomena born However "all models. models not econometric are equal and we seekfor on empirical thosewhich are usefulin practice" (Hendry, 1995,p. 3). Among the investigatorsconcernedwith understanding the how's and why's of
be behaviour, to there appears some consensusabout tile abstract concept which economic def"ineswhat a model is, and which attributes are inherent in a 'good' model. Economic large that too to economics agree arc and too complex for the seem researchers development of 'true' models. A model is "a free creation of the mind", "a metaphor", "a creative process", "a simplified representation", "an art". These qualifications establish the separation between the "Iruth of realiV' its and caricature. That models are "Incvitably false" seem to be uncontroversial. Indeed, even if the construction of true models were be doubtful. For example, a road map is a simplificd handling their would possible useful

93

lanes, motorways, countryside rivers, and of railroads, of a complexnetwork representation from interpreted, A If take to one correctly a road map can enoughand mountains. accurate B. The "real road", however,would not evenfit inside the car. it is importantto distinguish'betwcen Nonetheless, theoreticalandempiricalmodcls. Despitehavingthe commonobjectiveof seekingfor the "best" approximationof reality and in practice (theory underlines the structure of empirical despite their interdependence findings can change theoretical postulates), theoretical models models and empirical latent theoretical variables,while empirical models establish among relationships assume latent linking to observed Georrcspondencc andmeasured variables. relationships' In researchwork, it is generally assumed that a true data generatingmechanism
cxists within the complexity of the economy, and that the objective, of tile modelling Tile data is this to mechanism accurately about as possible. make as statements process determine data in determines the turn, tile the of sct which, propcrtics generating process from licnce, "the delta model. an empirical of observed estimating use results obtained hetiveen distinction fundamental theory and empirical models since empirical creates a be by default, simply a recombination of ivhatever process generated the models must, data" (Hendry and Richard, 1982, p. 6).

There also appears to exist a consensusamong researchersin setting the 'good' For empirical of models. and attributes example, relevance, characteristics simplicity, theoretical plausibility, explanatory ability, accuracy of cocfl'icicnts and forecastingability are desirablepropertiesof empirical models that Christ (1966,1975) investigators Charcmza In of work many other recent more such and as underlines. Deadman(1997), Granger (1990), Hendry (1987,1995), Hendry and Wallis (1984), Learner(1985) and Mizon and Richard (1986), we can find similar main and auxiliary for It model as useful and an empirical reliable. seems generallyaccepted qualifying criteria that a goodempiricalmodelshouldexhibit a good fit, absence of residualautocorrelation or (data coherency),valid exogencity assumptions,parameterconstancy, hetcroscedasticity ' data theoryconsistency, admissibility and encompassing.According to Hendryand Wallis (1984),an adequate empiricalmodelhasthe ability to "describehistorical data without

'The generalto specificmodel-buildingapproach, data in a parsimonious seeksnot only to characterize way within a generaltheoreticalframework,but alsoto provide a statisticalbasisagainstwhich other modelscan is usually seen as the quality of econometric models which allows the be evaluated.Encompassing investigator to seehow well a given model accountsfor the findings of rival studies.I lence,encompassing requires anygiven 'good' modelto explainthe resultsobtainedby othermodels.

94

fiture, fit be to the to to well equally consistentwith the producing systematicmisfit, systemand to encompass alternative explanationsof underlyingtheory and measurement 6). (p. In Hendry (1995), a broader,lesstechnical the sameset of endogenous variables" for description, that the states good empirical models allow and perhapsmore ambitious for interpretation the evaluationof tile explanatory context and simplified a within of reality power of competing theories, permit the accumulation'and consolidation of empirical knowledge and give a scientific approach to the understandingof human conduct. Summing-up,an adequate model seemsto rest upon its ability to portray tile past,explain judge future, the the present, models rival and encompass competingtheories. predict How to devclop such cconometricmodcls from a givcn thcorctical structurcand a how desirable to the and recognisc evaluate system, and properties of a measurement 'good' model, are neither obvious nor settled among investigators. A unique and does On deep there the to modelling not exist. empirical path good contrary, arc consensual differences of opinion among researchers concerning the building, interpretation and
fact, different In models. several of empirical processes modelling strategies coevaluation literaturc for (sce in 1990, Granger, for economics example, applied contemporary a exist be linked issucs in flic These to to spcciric seem methodological cngagcd questions survey).

building, estimationandtestingof empiricalmodels. discussion Contemporary of economicproblemsis heavily influencedby the results 1940's the produced analysis since econometric early within the so-called of empirical its According to critics, traditional econometricmethodologyappears traditional approach. to modelswith poor forecastingability, shakyinferenceprocedures, to lead,in manycases, dissociationbetweentheory and empirical evidence,questionable assumptions and overall Frequently, fit to theoretical models to economic results. attempts estimation unreliable time seriesled to a numberof statisticalproblemssuchas auto-correlatcd residuals(despite distributed independently disturbances), 'wrong' signs,insignificanceor the assumption of doubtful magnitudesof coefficients and high collincarity among explanatoryvariables. Moreover, vital parametersin some models seem to be very unstable as the model specification changes.These problems are diagnosedand 'eliminated' without further for in the first place.An exampleof consideration aboutpossiblereasons their appearance this practiceis the detectionof residualauto-correlation with Durbin-Watson(DW) statistic by its the application of the Cochranc-Orcuttestimation elimination' and subsequent disease by 'removing' the symptomsseemsan "camouflaging However, the method.

95

1983, 197). The searchfor the best fit basedon such " (Hendry, to p. unlikely route success is far highest from being 'significant' R t-values the or process criteria as an acceptable of Newbold instance, Granger building. For (1974) show that spuriousregressions and model R2 involving independent high tend to present randomwalks and 'signiricant' t-valucs. Unstructured "specification-search",as Lcamcr (1978) adequatelyrcnamcd tile harshdesignation of "data-mining",may produceapparentlygood resultsbut inappropriate is fit inference. A the typical of a polynomial in t of order N-I to a time example statistical Thereare,of course,more subtleways of specification-scarch but seriesof N observations. be R2 t-valucs and must similar practical viewed with caution and used as with results. discussion Further the than of stating quality rather ways of a precise model, guidelines of the use and performance of modelsbuilt within the traditional econometricmethodology in, for (1980), Hendry 1 Icndry found be example, and Morgan (1995), Lcamcr (1983), can Lovell (1983),Sims(1987)andWallis (1989). According to the critics of this methodology,the ongoing researchseemedmore best how the to estimatea model than with the processof its with question of concerned However, is While sonic the there among critics, even some disagreement. specification. is to that the credibility ascribed any reported econometric out estimate relatedto tile point 1983b), is by it (Leamer, obtained otherspoint out that a model's credibility process which does not dependenton its 'modc of discovery' but on whether it will survive latter 1987) "the (Ilendry, that and validity of any outcomeIs Intrinsic to theproduct, evaluation discovery its best to or construction, method of niethodologv can at reveal the beneJ71s not drawbacks ofalternative researchstrategies"(I Icndry, 1995,p. 10). and However, most critics seem to agree that the practical problems of model ignored in traditional literature,ought to be addressed. selection, generally and specification 11crice,confronted with increasing specification uncertainty, modern methodological from the change emphasis estimationto modelling. This changeis apparentin approaches recent applied work, where considerableattention is paid to the processof modelling cconomic time series using alternative strategies,and to the qualitative cvaluation of is Little attention currently paid to the issue of estimation methods econometricmodels. in literature for half the which predominated a century. The failure of traditional econometrics to producesatisfactoryforecastsor resolve divergences between to be linked not to the use,but to competingeconomictheoriesseems the abuseof its methodologicalprinciples and to a widespreadgrowth of senseless data-

96

for 'best' it is Yet, the to a structured and purposeful search conduct mining. possible data is Acceptable the the mining. specification-scarch worst aspects of model avoiding better from to a one, using sensible evaluation criteria. model one process of moving However, if changes associated with this process arc based on some extremes of datalead from the they than the 'best' - model's path. search away not mining, more often Nevertheless, it has to be recogniscd that, in practice, some spcciflcation-scarch is is in Charemza Dcadman (1997), As the and matter not whether pointed out unavoidable. data-mining is involved in the modelling process, but how its sensible use may contribute to achieve satisfactory specifications. While model-scarch processes which deliberately ignore or conceal conflicting results arc unacceptable, the purist's case of a judicious just leading is to tested model, well-spcciried estimated a once, argument and a economic lies In the these appropriate econometric modelling. extremes middle of nonsense. The origins of modem methodology in model building processesarc believed to be

linked with the work of Davidson,Hendry,Srbaand Yco (1978), known in the literatureas the DIISY paper.Sinceits publication,this work has beengenerallyseenas an important influenceon the way cconomctricians usetime seriesdatato model economicrelationships. in the DIISY paper have received considerable Indeed, some of the issues addressed for development building leading the to the of new methodological approaches attention, include gcncral-to-spcciric models, which of econometric modelling, errorand evaluating integration specifications, vector autoregressivc and and cointcgrationanalysis. correction According to Hendry (1995), a 'data-basedrevision strategy' in model-building is Given the empirical success, claim some gcncral-to-speciric can approach. a which framework, theoretical and a system a generalmodel can be spccificd taking measurement into considerationthe sample size (which, a prIori, constrainsits generality), previous findings (for features data (for example, nested special cases) and special empirical Ilic estimationof the general example,laggedreactionsor rapid adjustmentprocesses). be testedagainstvariouspreparameters' unrestricted estimates which can modelprovides Specific theoreticalmodels are usually nestedcaseswithin tile dcf*incdnull hypotheses. be it. Alternatively, the generalmodel can be tested can against and more generalmodel form consistent simplified, until a parsimonious results.This should then be testedfor all the desirable attributesof a goodmodel. large The existence of a potentially numberof theoreticallyplausiblemodelswhich alsosatisfythe quality criteria rendersmodcl-choicca non-trivial problemas the criteria by

97

judged features desirable but the model are a good which of are necessary not sufficient In addition,the fact that one model encompasses requirements. anotherundersomecriteria imply that the latter shouldbe retainedand for example)doesnot necessarily (forecasting, Nevertheless, if the next rival the former discarded,as they may servedifferent purposes. does model and still satisfy the mentioned encompass model an existingadequate criteria it better for be being, the time asa model. Yet, as Ilendry (1983,p. 199)stresses can qualified, has been hazardous it "unill tested to against new evidence a model rigorously would seem how pleasing'lhcse its implications, no matter place muchweighton seem". Although there are not sufficicnt conditions to ensurethe finding of the 'perfect' be there of necessary conditions which model, arc a number can usedto rule out empirical inadequate Given on the best remainingcandidates. models.This allows us to concentrate the unsatisfactoryempirical results obtained in chapter 3, a change of model-building is investigates IIcncc, this the UK demand practically self-imposed. chapter methodology for tourism using someof the morerecenteconometric methodsin the building, testingand evaluatingof empiricalmodels. The structureof the chapteris as follows. Section2, presentsa succinct literature
in The integration time cointegration analysis series economic of models. order of review of the variables included in the models is also addressed. In section 3, a gcncral-to-speciric implemented build is to econometric models with crror-correction mechanisms approach for the UK tourism demand in France, Spain and Portugal. The estimation results obtained Section 4 these arc also provided. models presentsa critical examination of estimation with in different studies using the same methodology. Section 5 concludes. obtained results

4.2. ORDER OF INTEGRATION AND COINTEGRATION OF TIME SERIES IN UK TOURISM DEMAND FOR FRANCE, SPAIN AND PORTUGAL
4.2.1. BASIC CONCEPTS ON NONSTATIONARY STOCHASTIC PROCESSES2

The basic ideas of applied cointegrationanalysisarc simple to understand and to Thus, it may be useful to start use,althoughthe underlyingtheory is not so straightforward. of stochastic with somebasicconcepts processes and time seriesanalysis.
2 For a clear basic the of concepts of cointegration analysis see, for example, I laff is explanation and simple (1995) or Charemzaand Deadman(1997).

98

time, is a family of real A stochasticprocessdenotedby {Yt), where t represents Yt. A let For (Yt) be Yt. Y2s YI, as simplicity, referred valued random variables ..., joint if (in be is the to and a, strict or strong sense) stationary stochasticprocess said in if displaced In time. distributions the process arc unchanged of conditionalprobability is the of stationary used, restricting concept process wcaker a situations practical most definition the the to of process. the variances and covarianccs means, stronger of scope if be (wcak-ly) is Yt to Hence,a stochastic stationary said process E(Yt)=t
E(yrt)2=(12 Yk-Kyt-Ayt+klt)l

k=0 Consequently, between Yt Yt+k If k lag is thcn7k the and 2 a2. covariance at WhereYk . if a stochastic process is stationary, its means, variances and covarianccs (at various lags) is fulfillcd, is the If these time. process conditions not one or more of remain constant over

be to nonstationary. said is An importantspccialcascof a nonstationary stochastic proccss a randomivalk. A following dcscribcd by bc Yt, the cquation: proccss can randomwalk stochastic
Y,

Yt-I

Zt

(i)

identical independent 'Another Zt random variables. and representsa series of where important specialcaseis the nonstationarystochasticprocessdenominatedrandoin is-alk described be drift as which can ivith a

[L*O
is p where a constant.

(ii)

In the literature,the conceptof a time seriesis often usedalongsidethe conceptof a be being by Any time series can viewed as a stochastic generated stochasticprocess. be data this a set of can regarded as particular rcalisation of specific a and process, process. stochastic To makethe notationcompatiblewith that of most econometric textbooks,let us use instead Zt denote denote instead Yt time to to series stochastic and process, ct of a a of yj independently distributed identically continuousrandomerror variableswith and seriesof The is no serial and correlation. stochastic variance process ct called a zeromean,constant while noise. With this new notation the equations y, = y,-, + c, and y, = It + yt-1+ c,, drift, describe a random and walk with walk a respectively. a random

99

The nonstationarityof an economictime seriesmay be apparentwhcn it is plotted in downward This direction. its by tendency to time propensity move an upward or against is calleda trend. A time seriescantrend up"or downwardsas a result of randomshocks.In this case,the seriesis calleda time serieswith a stochastictrend. However,a trend in a nonstationary time seriescan occur if its meanis a specific function (linear, for example) of time. Supposethat a time series can be described Po Plt In is P, PO that time t, the this + + + yt + so c, case series where = gt = asy, = ji, ct is A deterministic have trend trend. to stochastic-detcrministic mixed process also a said PO pit be described + yi-I + ct a3 + y, as = possibleand can be discussed The processes abovecan viewcd as specialcasesof a broaderclassof described be as can which processes nonstationary

+ct yt =oyt-I
processes.
4.2.2. COINTEGRATION AND ORDEROF INTEGRATION OFTINir, SERIESVARIABLES

(iii)

have been discussing Therefore, 0=1. the we are callcd unit root stochastic processes where

Nonstationarity of timc scrics has always bccn regardcd as a problun in it flic levels to spurious can relationships since give rise of among studies econometric been by, for has Granger It Newbold (1974), shown cxanipic, variables. and economic
Dickey and Fullcr (1979) and, more recently, Phillips (1987), that tile statistical properties dubious. Yet, in implicit time series are an assumption of regressions using nonstationary literature is data time the that concerning regression analysis of series such are much of

inference is forecasting If the the this case, statistical and not procedures may stationary. fact, be In regressingone nonstationarytime scrics on another may present valid. not An are meaningless. apparentlysatisfactoryestimationresults even when the regressions illustrative examplewould be the ratherpointlessregression of a linear trendon a quadratic trend. Another such example would be the case where the variables are subject to a deterministic These trend. than a regressionsoften provide apparently stochasticrather high R and 'signiricant' t-statistics, although they such as good estimation results,
3 It is assumed that the stochasticprocessc, is a white noise (zero mean, constant variance and no However,theseconditionsmaYbe relaxedto permit, for example,serial correlation.If the autocorrelation). (i) and (ii) can no longer be called random walks but yj is still the processes effors are autocoffelated, nonstationary.
100

interpreting 'danger' This the shows of regression results represent spuriousrelationships. involving deterministicor stochastictrendedvariables: the apparentlyrobust but invalid determine it difflicult to whether an economic rclationship estimation results can make from data. in fact, Deadman has, Cliarcmza by As the the theory any support and suggested for data (1997) point out, regression sense makes only analysis not subjectto a trend, that is, only if the variablesinvolved areeitherindividually stationaryor a linear combinationof them is stationary.Since most economic data series contain trends (are nonstationary), thesehaveto be purgedbeforeany sensible regression analysiscanbe performed. from trend a an economic time series is by convenient way of purging differentiation of a nonstationary differentiation.In other words,the successive time series 4 The later, series. a stationary number of times that a time series sooner or will produce, integration in bc differentiated to tile to stationarity achieve gives of order order of needs d timcs before time seriesyt needsto be difTcrcntiatcd that series.Hence,if a nonstationary it achievesstationarityis called an integratedprocessof order d and denotedby yt - l(d). For example,the first differenceof a randomwalk, with or without a drift, is a stationary integrated l(l), first Therefore, their series walks are of order random one, or and series. for a seriesto differencesare integrated seriesof order zero,or 1(0).Still, it is not necessary be a randomwalk for achievingstationarityby differentiation.Stationarityin a time series its if by differentiation, In be this case, even crrors are autocorrclatcd. achieved can however,the nonstationary seriesis not a randomwalk and its errorsare not white noise.A is integrated such as ct a stationary process or an seriesof order zero,that white noiseseries is, a 1(0)variable. it scemsthat nonstationary From the preccdingdiscussion, timc scricsvariabicscan be a major problem for applied economics.Nonstationary or trended variables, either deterministic, inrcrencc invalid to spurious may give regressions, or rise and stochastic forecasting proceduresand, generally, make regression results difficult to interpret. Unfortunatelyin economics, most time seriesarc subjectto somekind of trend, that is, are be if this problem must addressed meaningful relationshipsarc to be and nonstationary, regressions. obtained with econometric

4 Provided,of course,that the nonstationaryseries can be transformedInto a stationary series through differentiation.It may be that the seriesis not integratedat all, so that no matter how many times It Is into a stationaryseries. differentiated it will neverbe transformed
101

The remedyof differentiatingnonstationary seriesuntil stationarityis attained,may long-run loss leads be ideal to the the this usually model's of procedure solution as not an forms difference first l(l) instead levels, Indeed, their of variables of regressing properties. long-run is information imply losing the on variables' which relationships valuable may first differences. in first diiTcrcnccs Models by levels by their their generally and not given is both behaviour, needed what are models can reproduce whereas which rcflect short-run dynamics the the structuralrelationshipand short-run underlyingthe adjustmentprocessto equilibrium. Most of the theoriesunderlying the relationshipsbetweeneconomicvariablesare diffcrenccs. Thosc lcvcls thcir thcir arc static, stcady-state and not cstablishcdconsidcring fully levels that the long-run theories to their variables' adjust which assume equilibrium or long-run equilibrium in the currentperiod. However,if this assumption doesnot hold, the behaviour the the should variables' relationship of of the reflect econometricspecification
importance for inference, The such matters of statistical economic adjustment process. implied the rcconsideration of the modelling problem using analysis and policy evaluation, information. long-run in levels to shortand obtain variables

Cointegration. theory statesthat if there is a stablelong-run relationshipamongthe

levels of economic variables, they cannot diverge much frorn cach other over time, implying that the variables are cointegratcd. Cointegration means that one or more linear individually is In these they stationary, variables although arc of not. other combinations be linearly into if two scrics nonstationary can combined or more a single time words, 5 itself is be stationary, the original variables are said to cointegratcd. If these series which far from from they the cannot cointegratcd move apart are each other and variables 66attractor"which is their long-run equilibrium relationship. In contrast, if they are not is it (but have long-run that not necessarily so) such possible variables no cointegrated drift from "The can arbitrarily apart each other. and poiver of economic relationship different shouldforce an attractor variables to move together it? the longequilibrium as if indivitlually in if (Engle the they even short-run and are nonstalionar)P and even not run Granger 1991, p. 8). Therefore, "synchrony" of movements between time series variablcs is the intuitive idea underlying their cointegratcd long-run relationship.
5The stationary from linear such combinationof non-stationary seriesresulting variablesmay be the residual A regressionin which residualsarc stationaryor 1(0) is a cointegrating regression. seriesof an estimated regressionand the vector of parameterslinking the variables within this regressionIs known as the cointegrating vector;

102

Cointegrationanalysisis concernedwith the developmentof ideas, conceptsand long-run investigation for the meaningful of relationshipsamongeconomictime methods in detection, involved The the estimation and testing of cointcgrating series. problems been have focus the of, much recent research work. regressionsare complex and Fortunately,there are somegenerallyaccepted simple rules that can guide us through tile of modelling,estimatingandtestingcointcgratcdrelationships. process Let us assume that economictheory postulates an equilibrium relationshipbetween (yt, Px,. I(I) If yt follows an equilibrium that such series xt) y, pair of nonstationary a = instant in time, then we can rewrite the equilibrium relationship as any at path to an 'attractor' in Engleand Granger's y= Px or y- Px = 0. The line y= Px corresponds (1991) nomenclature.However, the attractor equation is not expected to hold at all levels disturbances drive instances the variables' can as away from tile equilibrium path. Therefore,out of equilibrium, their relationshipis betterdescribedby y, - Px, = Cc,where is be the to the extent which relationship of out equilibrium represents ct and may called an 0) variablewith zero mean"there will be a tendencyforthe "equilibrium error". As ct is a 1( be line, line Ot, to act as an attractor" (Engel and to the thusfor the Xd and around points Granger,op. cit.). Since the linear combinationof nonstationaryvariablesgives rise to a Therefore,"cointegration Is single stationaryseriesct -1(0), the variablesarc cointegrated. (op. ). The reverseis also true, as the existence ofan conditionfor altractor" cit. a suJf1cient the existenceof a long-run relationshipbetweena set of variables implies that they are is Furthermore, betweencointegratcdsystems there a correspondence cointcgrated. and tile error correction mechanism (ECM). An ECM constitutes one case of a systematic through which cointegratcdvariablesarc preventedfrom drifting apart process adjustment from the 'attractor' line. Therefore,cointegratcdvariablescan always be viewed as being by crror-corrcctionequations. The converse is also true. generated From the previous discussion we conclude that any meaningful econometric analysisof an equilibrium relationshipbetweeneconomictime serieslevels must include involved. Cointegrationamong a set of economic the tests for cointcgration. of variables variables depends,among other things, on the order of integration of those variables. Although there is a similarity between cointcgration and order of integration tests (commonly denoted'unit root' tests), the latter are performedon univariate time series while the former deal with the relationships among a group of variables where

103

(unconditional IY) each has a unit root. Consequently,before any scrisible regression identify is be it to the order of integrationof eachofthc analysiscan perfon-ned, essential relevant variables.In the next section, all the relevant time series variablesused in the modelling of the UK tourism demandfor France,Spain and Portugalare examinedand a setof teststo determinetheir orderof integrationis performed.
4.2.3. ORDER OF INTEGRATION OF VARIABLES IN UK TOURISM DEMAND MODELS.

The variablesincluded in the econometricmodels of this chapterare the sameas thoseusedin the previouschapter.Hence,we begin this sectionby recalling the definition of the variablcsdcscribcdin chaptcr3. Ncxt, wc prcscntplots of the variablcs' lcvcls and first differences,for an easierinspectionof their trend features.Finally, we test for their integration order of using the Dickey-Fullcr (DF) and AugmentedDickcy-Fullcr (ADF) unit root tests. The UK demandfor tourism in eachdestinationi, [i=F (France),S (Spain) and 11 (Portugal)], is measuredby the logarithm of the per capita UK tourism expenditure destination i, deflated by its to consumerprice index (CPli). YF, YS and YP allocated denote the UK tourism demand for France, Spain and Portugal, respectively.Tourism in destination i j arc measured by the logarithm of tli and competitivedestinations prices indexes in destination i 0) and in the UK, adjustedby the relevant price consumer of ratio exchangerate. Tourism effective prices in France, Spain and Portugal arc denotedby, PF, PS PP. ne UK real per capita income is measured by the logarithm and respectively, disposable income, dcflated by the UK consumer denoted the index capita of per and price by 1. Figures4.1 to 4.10 Presenta set of graphsshowing how the variables' levels and their first differencesevolvedover the sampleperiod. Figures4.1 to 4.3 relateto the levels of the variables,and Figures4.4 to 4.10 relate to their first differcnces.Figure 4.1 presents the UK real per capita expenditurein France, Spain and Portugal. Figure 4.2 presents tourism effective prices in thesethree destinationsand Figure 4.3 the UK per capita real income.Figures4.4 to 4.10 present the individual plots of all the first differenccdvariables.

104

Figure 4.1: UK per capita real expenditurc in France, Spain and Porttigal
2.0--

0.5-

0.0 1969

1974

1979 Years

1984

1909

1994

19197

Figure 4.2: Effective prices of tourism in l, rancc, Spain and PortUgal

0.15 0.10-

0.050.00 -0.05 -0.10

19 69 1974 1979 Years 1904 1909 1994

pp ......... --------...................... -------...... ............. ------PF

PS

1997

Figure 4.3: UK real per capita Income

3.40 3 34 3-28 3.22 3.16 3.10 3.04 2.98 2.92 286 2.00 1969 1974 1979 Years 1984 1989 1994

105

First difference of the UK real expenditure in France Figure 4.41:

0.150.10 0.05 0.001 -----0.0 -0,10 1969 1974 1979 Years 1984 1989 1994 1997

--

----

----------

- -------------------

DYF

--

--

------

Figure 4.5: First difference of the IJK real expenditUre In Spain

0.15-0.10 0.05 0.001 ----------0.05 n -i nf

------------

DYS oO' .

1969

1974

1979 Years

1984

1989

1994

1997

Figure 4.6: First Difference of the UK real expenditure in Portugal

0.20 OAS

--------0.10 ojool ---------------llXK

0 JO'S

DYP

Mar

Years

106

Figure 4.7: First differences of the effective price oftourism in France

WOO -OJ02
-OiDd 0 jola 0 DO

ll

DPF

Years

Figure 4.8: First differences of the effective price oftourism in Spain

0 JOB
OJDA 0 J012 ojoof -ojw
0 JOK

---

-------

--

---

-----

-----

---

-----

DPS

0 JOS -ojolal

I gerg

19914

1997

Years

Figure 4.9: First differences of the efflective price oftourism

in Portugal

OJDO -OJ02-

------------

-----------

DPP

-OJDIL-

1974

lcjn

I CPU

1997

Years

107

Figure 4.10: First differences of the UK real per capita income

ID C) 4 0,03-0 001 000 02

.oof
------------------

F-I,

-001 0,0 2 1969

1974

1 117 1; Years

Some of the variables' fleaturescan readily be spotted from the graphs above. For
level both difference first the the that the and plots show instance, peculiarly variables behave

in a sub-period of' the sample that can roughly be placed at 1973-1999. This

for is behaviour Portugal Spain. Several the to variables related more obvious and peculiar events, which canjustify this behaviour, took place within this period: tile oil crises of' 1973 of' 1974; the substitution of dictatorial regimes in

Portuguese 1979; the revolution and

Portugal and Spain t1orparliamentary democracies in the mid 1970s; the Portugal and Spain in likely have 1986. These breaks in EU in to the events are produced structural integration the data which may have repercussions on the conventional unit-root tests performed

below. For some series, these breaks may not have a strong enough impact to alter tile for However, Spain, the tourism tests. the price of series ol'Portugal and a clear conclusions dill'crent be is have be to tests to expected and may not perflornled to establish conclusion their order of integration. Apart from this potential problem, a fairly 'normal' for all series, as the oscillating observed behaviour is

movements of' their first dilTerclices seem to

indicate stationarity and, hence, the presence ofa unit root in the variables' levels.

Besides the potential problem of existing structural breaks, it should he noted that how is, That tests to the tests are also sensitive unit-root performed. conventional are some the outcome may vary whether the test regressionsinclude a non-zero mean or a time trend. is data to an the or autoregressive process average moving a assumed underlie whether generating process, and whether the tests are perl'ormed using classical or Bayesiall Jansen Thornton, inferences 1991 (Dickey, ). This sensitivity can be partially and statistical JUStiliedby the lack of power ofthese tests against an alternative hypothesis ofa stationary but larger than unit root. Consequently, the DF and ADF tests are not necessarily tile best in for to way search stationarity some tirne series and whenever these tests present dubious have be to may results,alternative approaches implemented.
IN

is integrated hypothesis that time Suppose the of a seriesvariableyt we want to test by is Assume that yt generated orderone.
Yt :., --pyt-I
+
ct

(iv)

is (white identically distributed is and p noise), stationaryserieswith zero mean wherect an I&I. 11o: If be hypothesis test the to A null would procedure a real number. straightforward is (iv) is the is said to 110 not rejectedthen yt nonstationaryand equation autoregressivc 6 is integrated have a unit root. If JpJ<1 then the series yt of order zero. stationary or However, as Newbold and Davis (1978) show, tests based on OLS estimations of Therefore, in tests root should unit significance. spurious seriesmight result nonstationary be performedusing a stationaryseriesas dependent variable. An appropriatemethod for in Dickey is Dickey-Fuller (DF) integration test the and proposed of yt testingthe order of (iv) to the suchthat7 Fuller (1979)which is based estimationof an equivalentregression on
Ayt = 8yt-l + ct

JpJ=I) is 8=0 (which is 110: hypothesis to the the against equivalent null the null where is integrated implies favour in Rejection that W. the the yt of alternative null of alternative in is [is 1(0), dependent (v) Since the the stationary variable null, under of order zero. I(l) 1(0) implying that yt is I(I)], equation(v) is a regression variable variable on an of an In for these the the t-statistic in the cases, arc not valid. values critical usual and suchcases for found in, for be is known the the which values can critical as r statistic t statistic In (1997). (1991) Charemza Deadman MacKinnon (1976), Fuller and and example, is for band (1997) Deadman's tables, Charemza a given numberof provided a critical and in (V) for 8 If level the the t cocfficient statistic regression significance. of and observations is smaller(morenegative)than the lower value of the critical band,the null hypothesis of a for in If favour be has the the to computed of stationarity yl. of alternative rejected root unit hypothesis be band, is the the than the cannot null critical uppervalue of t statistic greater is inconclusive. band falls test If the the the critical within statistic computed rejected.

6 it is conventionally implying in implausible "explosive" that >1 processes p are economics and, in assumed be integrated for to it is thantwo. of an order greater series economic ractice, unusual The DF test can also be used for testing the order of integrationof a variable generated as a stochastic testsfor the process with drift. Additionally, the Dickey-Fullerequation(v) canbe modified to accommodate deterministic trends. and of stochastic simultaneous presence

109

it into does is the take that DF A weakness possibility test consideration not the of OLS If the disturbances terms the white noise, the error are not a et. of serialcorrelationof
by Fuller Dickey A (v) and solution suggested are not cf(icicnt. estimates of equation (1981) consists of using lagged dependent variables as additional explanatory variables to 8 'whiten' the error terms.

be is to tested the In the presence regression of autocoffelation

Ayt = 8yt-I + m PiAyt-i + ct

(vi)

The test is now called AugmentedDickey-Fuller test and denotedby ADF(p); p is the (vi). in The dependent lags the side of equation right-hand variable of the number of before, is implement the the test alternative the null and to with same same as procedure hypothesisand the samedecisioncriteria The ADF test is generallyregardedas the most it is "Infinite Yet, integration tests. samples the not very powerful cff icient among simpler for alternatives Hi: 6=&<], when 8* is near unity. There is a size-power trade-off dependingon the order of augmentation usedto deal with the residual serial correlation. large ADF 217). 1997, Pesaran, tests (Pesaran with a p relative to the samplesize p. and havevery low power. The order of the ADF regressionis selectedby using the two step procedurefor is based First, the the criteria such as order selected, on augmentation selection. model Akaike Information Criterion (AIC) and the Schwarz BayesianCriterion (SBC). T'hen, is length, ADF lag test the performed. usingthe selected in table 4.1. The variables' The resultsobtainedfor the ADF(p) testsarc presented
levels and their first differences (indicated by the operator A) arc written in the first column; the following columns present the t-statistic for the ADF(O) and ADF(I) regressions with intercept and no deterministic trend for all variables, the AIC and SBC criteria for tag level. These 5% for the ther significance critical statistic at selection and the critical values in MacKinnon (1991, 1). Table the surface estimates response using values are computed The MacKinnon critical value at the 5% level for the ADF(O) test (27 observations) is 2.975, and for the ADF(I) test (26 observations) is -2.980. As these critical values do not

1The lengthof the laggedvariables is the minimum which secures to be included(the orderof augmentation) a with noisepropertyfor the error terms.
110

differ much we use-2.980 as the uniqueMacKinnon critical value for both tests.The bold for indicate the comparisonwith the critical value, accordingto the valid statistic values selectioncriteria. Table 4.1: ADF(O)and ADF(l) unit root testsfor the variablesin levels YF, YS, YP, PF,
PS, PF, I and their first differences AYF, AYS, AYP, APF APS, APP, AT. Variables YF AYF YS AYS yp AYP PF APF
PS

Test ADF(O) ADF(I) ADF(O) ADF(I) ADF(O) ADF(I) ADF(O) ADF(l) ADF(O) ADF(I) ADF(O) ADF(l) ADF(O) ADF(I) ADF(O) ADF(I) ADF(O) ADF(l) ADF(O)

t-statistic -1.9706 -1.9518 -4.9874 -3.4481 -1.9431 -1.8724 -5.5582 -3.8840 -2.1911 -2.6103 -4.1809 -3.9364 -1.8686 -2.3113 -3.4354 -3.4165 -2.0832 -2.4279 -3.7881

AIC criterion 42.63* 41.70 39.12* 38.13 36.21 * 35.22 33.33* 32.38 29.19 29.82* 26.00* 25.54 56.53* 54.76 53.05* 50.38 56.85* 54.26 52.29* 48.92 55.12 56.67* 54.34* 50.85 79.88* 76.37 80.70* 77.50 1

SBC criterion 41.33* 39.75 37.86* 36.24 34.91* 33.28 32.07* 30.49 27.89* 27.88 24.74* 23.66 54.99* 52.82 51.75* 48.49 55.52* 52.32 51.00* 47.04 579 54.73* 53.04* 48.96 78.55* 74.43 79.35* 75.56

5% critical value

2 98 - .

2.98 .

-2.98

-2.98

2.98 .

AP9 pp APP

ADF(l) 1 -2.9683 ADF(O) -1.4130 ADF(I) -2.5996 ADF(O) ADF(l) ADF(O) ADF(I) ADF(O) ADF(I) -2.4806 -2.2523 -0.1979 -0.0670 -3.9281 4.0379

-2.98

-2.98

Al

III

The tests clearly indicate all variablesin levels as nonstationaryand, except for APP, all implies This in first differences that all level variables,except for variables as stationary. the price of Portugal(PP),canbe considered as integratedof order one or I(l) accordingto the ADF test. Indeed, given the ADF test results of nonstationarityfor both the level variable PP and its first difference(APP),we cannot concludethat PP is an l(l) variable underthe rulesof this particularunit root test. Theseresults were expected,given the economicand political eventsof the mid 1970'swhich are likely to have affectedthe structureof all time seriesin general,and of the series'price of Portugal' in particular. Theseeventsaffected Portugal in a particular in in due 1974. Tile effectsof this revolution are to this the country political revolution way believed to have lasted until, at least, 1979. Indeed, during the period 1974-1979,tile Portuguesecurrency devalued sharply, foreign and domestic investment attained their lowest valuesever, inflation roseto its highestvaluesever, and generalpolitical and social instability adversely affectedall economicsectorsin the country. I lencc, it is quite possible that a structuralbreakexistsin the seriesPP. The weakness of the ADF testsrelatedto the presence of structuralbreak in a series in Perron(1989,1990).Perron(1989) suggests is addressed teststo examinestationarityin the presence of varioustypes of shocksand shows,amongother things, that the effect of a "pulse" variable (a dummy taking the value of unity when the shock is 'active' and zero is I(I) on an variable permanent while on an 1(0)variableis not. otherwise) We believethat we can safely assume variable PPto be I(l) and its first difference to be 1(0),given the lack of power of the ADF test and given that the variable for the price in be Portugal I(l) MacKinnon tourism would accepted as with of critical valuesat the 10% level, and is accepted (1997) critical valuesat the as I(l) with the Charcmza and Deadman 5% level.9 Moreover,the graphof the variable's first diffcrcncc, exceptfor the pcriod 1974. 1979, shows the typical oscillatory movementof a stationary series, Consequently, we concludethat all relevant time seriesincluded in the regressions of the UK demandfor tourism in France,Spain and Portugal are integratedof order one. The ncxt step is to

9 In Chapter (1988)testto investigate 7, we usethe Phillips-Perron the stationarityof time seriesPP.Basedon this test results,we concludethat the variable in first differencesis stationary,and the variable in levels is to the Phillips-Perron Therefore, test,variablePP is integrated non-stationary. according of orderoneor I(I).

112

in built investigate the these cointegrated are using the empirical models whether variables The following sectionaddresses 4general this issue. to specific' modellingapproach. 4.3. COINTEGRATION ANALYSIS AND ERROR-CORRECTION MECHANISM IN THE UK TOURISM DEMAND FOR FRANCE, SPAIN AND PORTUGAL The needto build, estimate, test and evaluatemodelswhich combineboth long-run in linear the time, the at same which, arc and stationary and short-run properties has led to the reconsideration they the assemble, of regression variables combinationof been focusing in levels. Researchers' has involving attention recently variables on analysis the modelling of economictime serieswhich, although individually nonstationary, can be linearly combinedinto a single serieswhich is itself stationary.Serieswith such propriety is Hence, known cointegration analysis concernedwith the as cointegratedseries. are investigationof long-run relationshipsamonga set of variables in levels, each of which (unconditionally)hasa unit root. In section4.2.2 we established the link betweencointegrationand meaningfullongin levels, I(I) as using an example a among variables pair of nonstationary run relationships designated illustrates The "attractor" tile equilibrium path that y-px=O, variables,yt andxt,. forcesthe variablesto movetogetherin the long-run,evenif not in the short-runor evenif they are not individually stationary. If such an "attractoe' exists then, in the out of be the tile yt-Pxj=ct, expected value of residuals relationship ct should zero. equilibrium Furthermore, the residuals' scries should be stationary,so that the variables' levels may deviatefrom their equilibrium valuesin the short-run,but shouldconvergeto it in tile longfind (meaning Failure to the that they arc not a stationary relationship among variables run. implies that the long-run relationshipdoesnot exist in any meaningfulsense. cointegrated) Summarising,cointegration is a sufficient condition for the existence of a long-run in levels and the existence of a long-run relationshipamonga relationshipamongvariables implies that they arecointegrated. setof variables Most economictheoriesstipulatelong-run relationshipsamonga group of variables the existenceof a cointegratingvector combining the group of seriesinto a and assume a test for cointegrationcould be performedby univariateseriesct. Given this assumption, unit root test to the seriesof residuals applyingany conventional of the long-runregression. However, this procedurerequires that the hypothesised cointegrating vector is known.

113

Sincein most casesthe cointegratingvector is unknown, some linear combinationof the be be first to stationaryand the unknown cointegrating assumed relevantvariablesmust we first addressthe building of empirical vector must then be estimated.Consequently, modelsdescribingthe relationshipbetweenthe UK demandfor tourism in France,Spain determinants. Then we estimatethe underlyingcointcgrating and Portugalwith its assumed vectors.Finally we test for stationarityof the resultingseriesof residuals. Demandtheory establishes the existenceof a steadystaterelationshipbetweenthe its demanded income and and commodity own-price,consumers'disposable of a quantity prices of the commodity and its substituteand/or complementcommodities.Theoretical framework, static models are generally establishedwithin a which assumesthat the adjustmentprocessof the economicvariablesto their long-run equilibrium values occurs Hence,in general,theory doesnot provide inflannationon the dynamicsof instantaneously. the short-runadjustmentprocess.However, it has been rccognisedthat empirical models information both longthe can provide on and short-run aspectsof economic which behaviourareimportant. An approachto modelling short-run dynamics was proposedin Davidson et al. (1978) and has beendevelopedby a number of researchers ever since. In Davidsonel al. (1978), the short run dynamicsis modelledthrough an ECM which allows tile dependent its long-run following disturbance to to return equilibrium path variable a shock. Cointegratedvariables can always be viewed as being generatedby an crror-correction Therefore, a sound model-building strategy to analysc tile existenceof a mechanism. between UK demand its the tourism (as well as relationship cointegrated and determinants, its short-run dynamic adjustmentprocess),would be to estimate the underlying errorThis from may model. model result to correction applying a 'general to specific' approach distributed lag model (ARDL). Through successive hypothesistesting, an autoregressive form which is datathe ARDL model may be reducedto an interpretable and parsimonious ARDL modelcan thenbe usedto derive andstatisticallyrobust.The parsimonious coherent the short-run error-correctionmechanismembeddedin its specification and to test for cointegration amongthe variablesincluded. Assumethe following ARDL modelof the UK tourismdemandfor destinationi. 10

10 Themaximumlag-length by the AIC andSBC criteria. of the ARDL modelis one,as suggested

114

PS, + cioPr-t Yi, = cti + ajHt + ai,PPt-I+ bioPSt + b, + c,,Pr-, + -, -, + PioIt+ pilIt-I + 8ivt + oiyit-I + Uit

(4.1)

is logarithm P (Portugal); Yi UK i=F S (Spain) (France), the the and of per capita where i; destination PP, PS logarithms in PF the tourism and arc of the effective expenditure real Portugal,Spainand France;I is the logarithmof the UK real tourism pricesin, respectively, is dummy V disturbing income disposable a variable representing a and event per capita for the a sequence of observations unity and zerootherwise. value of which assumes If a steady-state equilibrium relationshipexists,the variablesin equation(4.1) may is, its from time, that their any x assumes at variable any values; equilibrium not move long-runequilibrium value suchthat, xt.j= xt = x*. Thcreforc,equation(4.1) can be written long-run the cquilibrium rclationshipsuchthat, as reprcscnting stcady-statc +b, +ail l io PS*+c"+c"PF*+P'O+P"I*+ pp*+bio yi* = + I-0i I-0i I-Oj I oi I-Oi 'i
or * +(Xi4l * +CCiSV) Yi * -(ccio + (xjjPP * +cci2PS* +a, =0 3PF kjo + kil

81 V I-Oj
(4.2)

where the coefficients,

a, =

I-Oj

for all k=a, b, c, P and j=O, 4 and cc,, =-81 , ..., I- Oj

impacts determinants' long-run dependent the the the of changes on variableand represent Cc! is the intercept. (Xio I -oj in logarithms,tile cccoefficientsmay be Sinceall variables(exceptV) are measured interpretedas long-run elasticities. On the other hand, the coefficients of the variables'
(except V) in 4.1 equation represent short-run elasticities, and 1-0 represcnts values current

the adjustment velocity. If the long-run equilibrium given in (4.2) is disturbedby a shock, its equilibrium by an effor-correctionmechanism form canbe restored which can be described as
) V, Yi, - (ajo + cci,PPt + a, + + + ai3PF, a,, (XIA = C't 2pS, (4.3)

dcnotcd hcrcafter. EQE is 'cquilibrium the crror', as whcrecit The ECM modelrepresenting the short-rundynamicadjustmcntprocess to the longrun equilibrium relationship,canbe derivedfrom equation(4.1) in the following way: first, from both sidesof equation(4.1) the laggeddependent subtract variableYi,. I obtaining
115

+ b,,PSt-I+ ciopEt + c,,Pr-, AYit = cci+ ai,PPt+ a,,PP, + bioPSt -, -,

8ivt (Oi I)Yit-I polt P. + + + Uit Ilt-I

0,01, Rearranging bioPS, 8, V, and cioPF, and then, sum and subtract ajoPP, -,, -,, -, -,. -,, collectingtermswe obtain
AYit = ajoAPPt+ bioAPS,+ cioAPFt+ PioAll + 8,AVt -

I 0i) Yit-I -

(1-0i a,

bio bil + + Co Cio + ail ai(, PS + pp, + + _, I-oi I-oi I-Oj

Plo+ oil I + ---ivt-l + I- oi t-I I-Oj (4.4)

)l +Uit

The expressionin bracketsin equation (4.4) rcpresentsthe equilibrium crror of be (4.4) Therefore, (4.3). can simplificd suchthat, equation equation AYit = ajoAPPt + PioAll+ 8iAVt - (I - Oj)EQE+ ui, + cioAPF, + bioAPSt (4.5)
Equation (4.5) represents the effor-correction model associated with the long-run dynamic derived from both ARDL (4.2), the models are and general model equilibrium

in (4.1). model In the next sectionswe define for each destinationi, a generaldynamic ARDL derive the associated error-corrcctionmodels. and model
4.3.1. THE UK DEMAND FOR TOURISM IN FRANCE

The ARDL generalspecificationof the UK tourism demandfor Franceis assumed to be given by the following equation:
YFI = (xF+
aFOppt+ aF]Ppt-l+ bFOPSt+ bF1PSt-1 + CFOPF1 + CFlpr., -, +

dFOYP, +

dFlypt-l

+ eFOYSt+eF1YSt-1

+PFOlt +PF11t-1

SF3F9497 OFYr + +8FIF7581t +8F21787891 +

+ Ur, t-1

(4.6) Hence,we assume that the currentvalue of the UK tourism demandfor France(YFt) lagged fft. lagged its depends the on value and current and valuesof own- and on own i), forj=F, S, income P), UK PJt. (PJt (It destinations' and real capita and per prices competing 1 It.,) and real per capita expenditureallocatedto France's ncighbouringcompetitors(YSt that the eventstaking place during the pcriods 1974YSt.1and YPt YPt.1).We also assume in 1981(dueto the oil price shocksof 1973and 1979,Portuguese revolution 1974,political during Portugal in Spain the period 1975-1979 and and Spain'sjoining the regimechanges

116

OECD in 1980), 1987-1989 (due to Spain and Portugal's simultaneous joining the EU in 1986)and 1994-1997 (dueto the openingof the ChannelTunnel), had a significant impact in France. demand Hence, three dummy variables are added in UK for the tourism on during (4.6) theseperiodsand zero otherwise. taking the of unity value equation Table4.2 presents the estimationresults(t-valuesin brackets)for the generalARDL model (denotedby 'regressionI'), and for the intermediateand final parsimoniousforms (denotedby 'regression2' and 'regressionY) which were derived from the former by the testing of plausiblerestrictions.The last eight rows of table 4.2 presentseveral systematic goodness of fit (adjustedR2, overall significanceF criteria for evaluatingthe regressions test and the residual sum squares)and statistical information based on the Lagrange Multiplier (LM) version of severaldiagnostictests:LM test for residualserial correlation; Ramsey'sRESETtestusingthe square of the fitted valuesfor the functionalform; residuals based kurtosis test the skewness and on of residuals, and residuals normality heteroscedasticity fitted values. test based on the regression of squared residualson squared For all theseteststhe respective p valuesarc shownin brackets. The t-statistic for individual significanceand the LM and F tests for additional variables joint signiflcancc,indicate the variablesfor tourism price in Portugal,UK expenditurein Portugaland UK expenditurein Spainas irrelevantvariables.The samestatisticsindicate tourism pricesin Spainand France,UK real incomeand the one-lagdependent variableas in the parsimonious model 'regression 3'. The three variables relevant explanatory I and 2 have beenmergedinto one single dummy (F) dummiesappearingin regressions in which appears regression3. The construction of the 'merging' dummy F takes into considerationthe relationship betweenthe cocfficients' estimatesof F7581, F8789 and 2 estimationresults,the cocfficient estimates F9497.As indicatedby regression for F8789 different F9497 from eachother and are not significantly different are not significantly and from 80% of the coefficient estimatefor F7S81. In fact, the null IIYPOtllCSiS 110:82=83 A82=0.881 is strongly supportedby the Wald test statistic which is -X(2)=0.006(0.997). Therefore, the new dummyvariableF assumes the value of unity in the period 197S-8 1, tile value 0.8 in the periods 1987-89and 1994-97and zero otherwise.Moreover,an F-tcst for additionalvariableswas performedon the joint significanceof PPt, Plit.1, PPt, YPt.1, YSt and YSt.1 using regressionI againstregression2, and the null of no significancewas not rejected at the 1% level with the value of F(6,13)=1.82.

117

Table 4.2: Estimationresultsof the ARDL model for the UK demandfor tourismin France
Regression 1
Intercept PPt PPt., Pst Pst-1 PFt PFt., it It-1 ypt ypt. 1 Yst Yst. 1 F7581 F8789 F9497 YFt. j Adjusted R2 Residual Sum Square F-statistic Serial Correlation Functional Form Normality fieteroscedasticity 4.9185 (-5.03) -0.5394 (-1.78) 0.0455 (0.12) 0.0154 (0.05) 0.8235 (1.71) 0.8283 (2.38) -2.0355 (-6.44) 0.3103 (0.62) 1.6040 (2.85) -0,0415 (-0.55) -0.0444 (-0.66) -0.0833 (-0.61) -0.1532 (-1.67) 0.0641 (3.35) 0.0921 (3.91) 0.0947 (2.91) 0.4269 (6.08) 0.9971 0.0350 574.18 17.41 (0.00) 0.001 (0.97) 0.718 (0.70) 0.835 (0.36)

Regression 2
4.6021 (4.50)

Regression 3
4.6108 (-7.50)

-0.1563 (-0.55) 1.0450 (3.06) 0.2145 (0.83) -2.127 (-6.09) 0.9654 (2.50) 0.7498 (1.65)

-0.1672 (-0.79) 1.0441 (4.19) 0.2233 (1.10) -2.1189 (-8.31) 0.9633 (2.80) 0.7539 (1.83)

0.0956 (5.20) 0.0758 (3.89) 0.0776 (2.09) 0.3551 (5.25) 0.9957 0.0080 623.26 0.02(0.90) 0.14(0.71) 1.50(0.47) 0.49(0.48)

0.09S2 (6.75)

0.3568 (5.97) 0.9961 0.0080 870.42 0.01(0.92) 0.12(0.73) 1.53(0.47) 0.52(0.47)

118

All statistics and diagnostictests presentedin Table 4.2 show regression3 as a of residualserial correlationor beteroscedasticity, robustmodelwith no evidence with high
functional fit, form acceptable power, explanatory good and evidence of a normal distribution of the residuals. Therefore, regression 3 is acceptedas a good model. "

From the regression 3 estimation results, we can now derive the long-run long-run the and compute cffects of the determinants'changes equilibrium relationship on derived long-run from the ARDL The the dependent estimates variable. for France the UK tourism demand are: a -4.6108 Intercept:a^ -x =-= -7.1688 0= 1-0 0.35682
A

model of

Priceof Spain: a,
A

bo+b, -0.16721+1.0441 X ---1.3634 1-0.35682 1-0

= 1-0
A

Priceof France:a2

0.22328 2.1189 -2.9473 1-0.35682

PO P, 0.96329 0.7539 + + UK Income: a3 =. ==2.6699 1-0.35682 1-0

A

DummyF:

OC4 =80.095218

1-0

1-0.35682

0.14804 0.64318 =

Velocity of adjustment:I-6=I-0.35682

Therefore, the estimated long-run equilibrium relationship for the UK tourism

in France demand canbewrittenas(t ratiosin brackets):

YF, =-7.1688+1.3634PS, (-13.46) (4.60) It 14804F, +26699 +0 +U^, -2.9473PF, '(7.08) (-10.34) (i 6.25) (4.7)

The residuals of regression(4.7) constitute the EQE series which is used to ECM the with the long-run relationship dcrivcd from the ARDL associated construct is by The EQE series given model.

EQE = IN ttt

7.1688+1.3634PS, +0.14804r-, -2.9473PFt+2.66991

(4.8)

for demand ECM for UK tourism in Franceis the the and AYFt = boAPSt + PoAlt+ doAPSt + 8AFt- (I - O)EQEt-I+ v, + coAPFt (4.9)

"We address the issueof coefficients'stability in section4.4. The models' forecasting ability is examinedin 8. chapter

119

for ARDL the model, The estimationresultsof the error-correction selected representation in brackets): (t-ratios following the are
AYF, = -0.1672APS, +0.2233APFt +0.9633AI t +0.09522AFt -0.64318EQE (6.75) (-10.75) (2.80) (1.10) (-0.79)

W= 0.839 Adjusted

RSS=0.008

F(S,22)=29.79

DW=1.95

for t-ratios be the all variables that It should coefficients and respective noted t-ratios (exceptthe error correctionterm) arc the sameas the cocfficientsand corresponding 3 from impacts) (the estimation. rcgrcssion obtaincd short-run of the currentvaluevariables

(0.059809) is its deviation is (1-0), the EQE as same As the although standard coefficient is by its in 3, dependent lagged t-valuc now givcn variable regression thecoefficient of the (1-0.35682)/0.059809-10.75. The next step is to apply the conventionalunit root tests to the EQE series.If the is in (4.7) long-run is equation the then expressed equilibrium relationship series stationary for The involved tests root the unit results of are cointegratcd. variables meaningful and 12 follow: EQE using an interceptandno trendareas MacKinnon criticalvalucs" 5% 10%

Tcst ADF(O) = DF ADF(l) ADF(2) ADF(3) ADF(4) ADF(5)

EQEserics -3.1898 -3.0332 -2.5659 -3.5544 -3.6796 -4.2366

4 315 .1

3 86 72 - .

The results indicate cointegrationat the 5% level for a lag-Icngthof 5 in the test in be it tests the Yet, that situations, conventional unit some root noted should regression. may not be powerful enoughto enablea clear conclusionaboutthe stationarityof tile series between UK in the the Hence, the analysis of relationship cointegration under analysis.
12 The AIC and SBC criteria indicate five lags as the appropriate lag-length to eliminate autocorrclation in the EQE series.Therefore, the number of observations available is now 23. 13The MacKinnon critical values are computed for 23 observations and three l(l) regressors for which the lag-length The is dimension likely to weaken the power of the test. is being tested. null of non-cointegration

120

tourism demandand its determinants we apply, alongsidethe conventionalunit root tests, advancedin Pcsaran the estimationand testing procedure and Shin (1995) and Pesaran et (1997) it is hereafter. In Pesaran Pesaran PP denoted (1996), the procedure and as al. lies in "the thefact that this testing that estimation of and strategT main advantage claimed 308). 1(0) (p. I(l)" the it can be appliedirrespective regressors are or ofthether In the first stage,tile existence involvestwo stages. The PP procedure of a long-run for is by joint F-statistic between tested the the the computing variables relationship Pcsaran in The lagged the test the error correction model. and variables of signiricance Pesaran(1997) error correction version of the ARDL model correspondingto the UK demandfor tourism in Franceis: AYFt = 70 + YIAPSt + 72APFt + y3AIt+74AFt
+ 'IpSt-l + 'YFt-, + '3, t-l + *'4Ft-j

(4.10)

YE, + w, + %, -, between is long-run be hypothesis the to tested the a relationship exists whether and is long-run F. Hence, I YF, PS, PF, the relationship null of non-existcncc of a and variables %s= for i=Ipo.. 0 againstthe altemativc I 11:%1; 0, any & s 95 The relevantstatisticfor this test is the F-statistic.However,the (asymptotic)distributionof definedby Ilo:
integration irrespective 111c is F-statistic the the order of of regressors. this non-standard, 14 (1996) in 's included If Pesaran the tables. computed et al. are values critical appropriate F-statistic exceedsthe upper value of the critical band then we can reject the null of non's between long-run the relationship variabics. of a existence
Xff %3 %4 1%22--

Another feature of this procedureis the possibility of establishingwhether the 16 forcing" is follows. This Once "long-run the variables. procedure as rcgrcssorsare between long-run been (statistically) has the variables relationship a of existence in by the the test the of null, we repeat exactly the sameterms,using as rejection recognized in first difference, by In the case the the dependent variables explanatory one. one variable hypotlicsis for France, demand UK test the the same null considcring,successively, we of

" The tablesprovide a 'band' coveringall possibleclassificationsof the rcgressors into 1(0), 1(l) or even fractionallyintegrationorders.If the computedF-statisticfalls out of the band,a conclusivedecision(either be If the computedF-statisticfalls within the band,the result is the can made. null) or rejecting not rejecting inconclusive. 15 Sincethe EQE term can be viewed as expressing the relationshipbctweenthe variables' laggedvaluesin by Pesaran (1997) (4.8), test the et may be seenas equivalentto a test for individual al proposed equation in itself, (4.9). This procedurefor testing for cointcgrationis EQE regression the variable significanceof referredin Benerjqe el al. (1998). 16In Pesaran (1997), the term 'long-run forcing' variablesis used interchangeably and Pesaran with the Therefore, a long-runforcing variablecanbe interpreted concept of exogeneity. asan exogenous variable.

121

dependent AYF AYS the as Al now appears APS, APF, while variable, as the variable and for tested, the If, the of non-existence null regressions additional the all regressors. of one longis be that there only one of a long-runrelationshipcannot rejected,we may conclude dependent indicates 'true' AYF the linking variable, the as which variables relationship run " The differenced first obtained results explanatory variables. as the variables other all and in Table 4.3. PP the for this setof testswithin the second are presented procedure stageof Ft. YFt. PSt. Ht. It., for 4.3: Significance test and Table additionalvariables 1 1, 1, j,
Dependent variable AYF APS APF AT L Explanatory variables APS,APF,AT andAF AYF, APF,AT andAF APS,AYF, AT andAF APS,APF,AYF andAF F-statistic (5,18) 28.2135 1.2927 2.3165 2.6287 Critical band at the I% level Test result

Rejection of the null Non-rejectionof the null (3.219; 4.378) Non-rcjcctionof the null I Non-rejcctionof the null I

between UK long-run the The null hypothesisof non-existence relationship a of data, by is determinants its in the France demand reinforcing strongly rejected and tourism I'lic P11 tests the above. procedurealso conventional unit root with the conclusionobtained indicatesthat there is no other significant long-run relationshipamongthe variables.This indicatesthat the explanatoryvariables in the UK dcmand for France equation can be forcing' 'long-run variables. viewedas
Regressors'exogeneity is one relevant aspectof empirical models' evaluation which

in many studies, for example, Richard (1980), Engle, lIcndry and has been addressed (1988) and Johansen Richard (1983), Hendry (1983), Johansen and Juselius(1990). The depends, the on the among other requisites, static models of results estimation validity of inference Hence, to the sanction statistical that exogcnous. are all regressors assumption has tile based the models, variables exogcneityof explanatory estimationresultsof such on is The be exogcncity a complex one, proved. concept of or statistically assumed either to its levels involve definitions The 'super' from 'weak' to various exogeneity. of ranging
17 In somecircumstances, the resultsof theseadditionaltestsfor the existence of 'long-run forcing' variables is for A be to test the appropriate procedure exogencity more of a set of variables may not very powerful. (1988) and in Johansen "fully maximisedlikelihood systems procedure"proposedin Johansen and Juselius is usedin chapter8 within a systemof equations'approach. (1990).This procedure

122

derivations which are outsidethe scopeof this and statisticalrepresentations mathematical in issue Hendry's (1995, based Therefore, this surnmarised a way, on we address chapter. "weak 181-191) exogencityandunit roots". of exposition pp. Let the long-run cointegratingrelationship betweenthe UK tourism demand for be i andits detenninants destination
yit + alp't = GCO + (X2pjt + a3lt + Clt

(4.11)

in destination i destination P Pi tourism the and competitive of price where and are defined before. is disturbance the term the and variables as other are el respectively, by Suppose that the PP procedureindicates,alongsidethe long-run relationshipexpressed (4.11), another long-run relationship bctwccn Pi and all the other variables; that is, it indicatesPi as a potential endogenous variable. The relationshipbetweenPi and all the be can written as variables other (ao )]+ Pi, pit-, (4.12) APit = (pAPit-I+ p[Yit-I + a, + a, + a3l, c2, -, -, lagged in brackets (4.11), term the the represents error of equation cl. expression where According to Hendry (1997), there arc tlircc confligurations of paranictervalues holds: whereexogencity

is is, between When there that contemporary correlation cl and c2, no 9=p=0 and a)
then (4.11) is a valid regression equationbetweenl(l) variablesand cov(clt, c2t)=O, for the parameters Pit is both weakly and stronglyexogenous of interestaj. b) When9=p=0 but cov(clt, c2t)#-0, from 'simultaneity then equation(4.11) sufTers if be defined bias' but a valid regression the variable APit is included equationcan for tile in the long-run relationshipand Pit is both weakly and strongly exogenous impact interest. The the variable Allit 'corrects' for the addition of of parameters between correlation el and c2 and restoresvalid single-cquation contemporaneous inference. but 0, be Pig When c2t)=O cov(elt, q* cannot strongly cxogcnousfor the c) p=0 and be interest, but In this case,equation(4.12) could weak-lycxogenous. of parameters is uninformative about the cc parametersof interest, allowing single equation inference to be valid without lossof information. The ARDL model is a dynamic specification which includes current as well as
lagged values of the explanatory variables. Moreover, if an ARDL model includes the

appropriatelag length for all variables,that is, if the model is correctly specified, the

123

between is the terms error already accountedfor. correlation possiblecontemporaneous Hence,there is no needto include any 'impact variables' to correct for contemporaneous fact, have In Pcsaran (1995) in Shin ARDL models. and shown correlation well-spccified that "valid asymptoticinferences on the short-run and long-run parameterscan be made, ARDL ARDL least the the the the of model, once squaresestimates order of model is using for between to the possible contemporaneous correlation appropriatelyaugmented, allow error terms. Therefore, the ARDL method continues to be applicable even If the and irrespectiveof whether they are I(]) or not", explanatoryvariables are endogenous (Pesaranand Pesaran 1997, pp.183-4). Consequently,if the existence of a long-run dependent its between be the variable cannot and regrcssors relationship cointegrated rcgrcssors of the appropriately augmented of potentially endogenous rejected,the existence ARDL model should not affect the validity of the statistical inference based on its estimationresults. Given theseconsiderations and sincethe testsperformcdindicatethe existcnceof a
long-run relationship between the UK demand for France and its determinants, we conclude that this relationship comprises a steady-state equilibrium path acting as an 'attractor'

from the through variabIcs wandcringtoo an crror-correctionmcchanism,prcvcnts which, far apart,evenif disturbedby shocksin the short-run. We now turn to the cointcgrationanalysisof the relationshipbctwcenthe relevant for UK demand in Spain. the tourism the equation within variables
4.3.2 TilE UK DEMAND FOR TOURISM IN SPAIN

The starting point for explaining the UK demand for tourism in Spain is the dcfinition.of an ARDL generalspecificationsimilar to the one usedpreviously.I Icncc, it is demand in UK be described Spain by tourism that the the following cquation: can assumed
YSt =cc+ aOPP, +a, PP, +bOPSt+bPS, +cOPFt+c, Pr,, +dOY131 +d, YP, + -, -, _, -, (4.15) 62S8689t POT, PIT, 8IS7576t YF, + + + + + e, + eOYF, + OYSI-I + u, -, -,

dummy S8689 S7576 variables.I'lle former accountsfor the events represent where and following Spain"s tourism the Portuguese which affected revolution of 1974; the latter for the reflects consequences Spainof the Portugaland Spainjoining the EU in 1986. Theestimationresultsfor equation(4.15) arc presented in table4.4,

124

in Spain Table 4.4: Estimationresultsof the ARDL model for the UK tourism demand
Regression I Intercept PPt PPt-1 Pst Pst-1 PFt PFw it It.1 ypt YPt.1 YFt YFw S7576 S8689 YSt.1 Adjusted
R2

Regression 2 -2.7561 (4.78)

-2.1633 0.25) -0.5765 (-1.28) 0.6273 (1.11) -0.0573 (-0.12) -1.0731 (. 1.73) 0.8412 (1.63) 0.0294 (0.06) 1.3607 (1.72) 0.4067 . (-0.48) 0.0883 (0.63) 0.2970 (3.28) 0.0060 (0.02) 0.1544 (0.92) 0.1202 (2.54) 0.0609 (2.48) 0.0129 (0.10) 0.9552 0.009 39.37 3.54(0.06) 2.60(0.11) 0.54(0.76) I 1.64(0.20) - T

-0.7219 (-2.90)

1.0444 (3.56) -0.5367 (-2.23) 1.1209 (5.11)

0.2687 (2.88)

0.1117 (3.95) 0.2713 (2.33) 0.9294 0.023 51.78 1.14(0.29) 0.03(0.87) 0.24(0.89) 0.54(0.46)

ResidualSumSquare F-statistic SerialCorrelation FunctionalForm Normality Ileteroscedasticity

125

The structureof table 4.4 is similar to that of table 4.2. The procedureused to Considering 2 follows. 1, to the the specific regression was as generalregression reduce for dummy I tested, the we variables' cocfi"icients,the estimationresults of regression
hypothesis 110:81=28f--8 which was not rejected by the Wald test with the statistic value dummy into S7576 S8689 Therefore )=0.0004(0.99). the variables are mcrgcd and a X2(1

in I 1975-1976,0.5 in 1986-1989 dummy (S) the takes the value period period which new between dummy YSt. YPt., S High tile variable collinearity and new and zero otherwise. 1, S YSt., indicating detected, to the explanation of the that the contribution of and was dependent this variable variable almost overlapsthat of the variable YPt.l. Consequently, for joint F-test We the significanceof additionalvariables an also performed was omitted. 2 spccirication.Tile null of no added It.1,YFt and YFt.I againstregression PSt-1, PPt,PPt-1, F 0.886(0.53). statistic with an value of was not rqJected significance in table 4.4 indicate regression2 as a All statisticsand diagnostictests presented is licterosccdasticity There no statistical evidence of or residual scrial good specification. distributed, it high the are normally presents a residuals explanatorypower, correlation, functional form. Therefore, fit 2 take we and acceptable as a good regression good in between its 'UK demand Spain tourism the the of relationship and representation 2 estimationresults,we can derive the long-run equilibrium From regression determinants. in demand Spain long-run UK tourism the compute and cocfricicnts' estimates. model of Theseestimates suppliedby the ARDL (1,0,1,0,0)arc: -2.7561 ,=&. Intercept:ao = 1-0.27126 -3.7820 -= 1-0 bo Priceof Spain: 6c, 1-0
Price of France: a, =.. 1-0

0.72186 1-0.27126

-0.99055

-0.5367 =1.0444 = 0.69659 1-0.27126

UK Income: CE3 =

POX 1.1209_ 1.5381 = 1-0.27126 1-0 0.36876

do. 0.26873 in Portugal:&4 = UK Expenditure 1-0 _1-0.27126 0.11171 8 0.15329 Dummy S: &5 = = 1-0.27126 = 1-0
Velocity of adjustment: I-6=I-0.27126 = 0.72874

126

Iberefore, the long-runequilibrium relationshipfor the UK demandfor Spainis (t-ratios in brackets):

A 5 st +ul 0.991 PSt+O. 697PFt+I. 53811 +O. 369YP, + 0.1 5 YS, =-3.782(7.47) (1.85) (2.75) (4.161 (5.96) (-2.58)

The residuals of regression(4.16) constitute -the EQE series which is used to from long-run derived ARDL ECM the the the relationship with associated construct is by The EQE series given model. ) EQEt = YSt -(- 3.782-0.991PS,+0.697Pr-,+1.5381t+0.369YI),+0.153S, (4.17)

The crror-correction model associatedwith the long-run relationship is given by the following equation(t-ratiosin brackets): 121 044APFt Alt +O.269AYPt-,+0.112,S, - 0.729EQE-1-1 +1 A-ht =-0.7219APS, +I. (3.11) (-6.25) (2.88) (3.56) (3.95) (-2.90) AdjustcdR=0.712 RSS=0.0231 F(6,21)=12.28 DW=2.30

The next step is to apply the conventional unit root tests to the EQE. series. If the long-run in (4.16) is the then equilibrium cquation relationship expressed stationary series

The resultsof unit root testsfor is meaningfuland the variablesinvolved arc cointegratcd. ' 8 following EQE produced the results: MacKinnon 19 critical values 5% 10%

EQEtseries ADF(O) ADF(I) ADF(2) ADF(3) 4.1948 -4.4975 -4.9520 5.1480

4.5643

-4.1510

The resultsshow that at the 5% significancelevel the EQE seriesis stationaryand, hence, the variablesin the long-runrelationshiparecointegratcd. As before, we apply the PP procedureto confinn the existenceof tile long-run by indicated the conventional unit root tests and to establish whether the relationship
11 lag length for the test regression The SBC andAIC criteria indicate3 lags as the appropriate to eliminate is estimated The regression in the EQE series. with intercept, autocorrelation no trendand25 observations. 19 25 observations, The MacKinnoncritical valueswerecomputed considering an intercept,no trend and four is beingtested. l(l) regressors for which cointegration

127

ARDL The PP forcing" the "long-run of effor-correction vcrsion variables. are regressors in is Spain for demand for UK tourism the model
AySt
+ 734611 + 74AYpt + 75ASt + 72APFt + 70 YIAPSI --'

+'IPSt-l + '2"t-l

+ '31t-l

+ '4ypl-l

+ 'SSI-l

+ kySl-l

(4.18)
+ Ut

between long-run is be hypothesis tile tested to the a relationship exists whether and is hypothesis The S. YP PS, PF, I, YS, test the null of non-cxistence of under and variables by defined 11o: long-run relationship, a
kj-"2 %2' X3 %42-%6 0

againstthe alternative

III: Xj*0, for any i=1, .. J. As before,the relevantstatistic for this test is the F-statistic The included in (1996). Pcsaran results et al. and the appropriatecritical values are in Table 4.5. this presented are procedure obtainedusing Ypt., 11.1, SI., Table 4.5: Additional variablessignificancetest for YS,.,, PS,.,,PFI. and I,
I Dependent variable AYS APS APF AT AYP Explanatory variables AYS, AYF, AT, AP andAS AYS, APF,AT, AYP and AS APS,AYS, AT, AYP and AS APS,APF,AYS, AYP and AS APS,APF,AYS, AYP and AS F-statistic (6,16) 5.0172 1.1216 1.6770 2.1567 1.5938 (2.850; 4.049) Critical band at the 5% level Testresult Rqjcction of the null Non-rcjcctionor the nulI Non-rejcctionof tile null Non-rcjcctionof the null Non-rcjcctionof the null

The null hypothesisof non-cxistcnceof a long-run relationshipbetweenthe UK is rejectedby the data, supportingthc conclusion demandfor Spain and its determinants indicates The PP test the above. conventional unit also root with procedure alreadyobtained that there is no statisticalevidenceof other long-run relationsamongthe variablesexcept its (YS) for that linking the dependent variable and explanatoryvariables.IIcncc, we can infer that the UK tourism demandfor Spain comprisesa meaningful long-run equilibrium if from drifting 'attractor' the which prevents variables apart, an even path acting as disturbedin the short-run.We now turn to the cointegrationanalysisof the UK demandfor tourismin Portugal.

128

4.3.3 TiIE UK DEMAND FOR TOURISM IN PORTUGAL

To explain the UK demandfor tourism in Portugalwe start with an ARDL general it is Hence, UK demand that the to the previously. used ones assumed spccificationsimilar by the following equation: for tourism in Portugalcanbe described YP, = cc+ a,PP,+ a,PP, + bOPS, + b,PSt-I+ c, PF,+ c,Pr-, + Polt+ P111-1 + -, -, + c,YF, + 81117479, + u, + 82118391 + OYPt-, + dOYSt + d,YSt-I + eOYFt -, (4.19)

The dummy variable P7479 corresponds to the Portuguese revolution period. The dummy P8391 relatedto the period 1983-1991deservesfurther explanation.This period (1983-85) (1986-91) the to the and post-joining of tile pcriods pre-joining corresponds in from integration EU. Just the Portuguese the ordealsof a revolution emerging economy's by imposed "straightjacket" IMF, the the of restrictive economic policies still enduring and the traditionally closed,weak and unstablePortuguese economy had to undertakemajor its full integration flexibility for the to openness and required reforms necessary achieve The lagged Portuguese behind EU the other the economics. economy stronger among Europeaneconomiesto such a high degree that the Europeanauthorities allowed an lasted 31 1989 but December which, officially, until period adaptation extended beyondthat date.Therefore,tile interceptdummy variableP8391is 'unofficially' extended for (4.19) to the to account theseevents. equation added in table4.6. The estimated resultsof equation(4.19)arepresented

129

in Portugal Table 4.6: Estimationresultsof the ARDL model for the UK tourism demand
Regression 1
Intercept ppt ppt. 1 Pst Pst-I PFt PFt. 1 it It-I Yst Yst. 1 YFt YFt. 1 P7479 P8391 ypt. 1
Adjusted R2

Regression 2
0.9416 (1.37) -0.8424 (-3.57)

Regression 3
0.9407 (1.45) -0.8430 (4.20)

-2.5674 (-0.64) -1.6415 (-1.70) -0.0600 (-0.04) 0.6904 (0.56) 0.8940 (0.61) 1.1783 (1.05) -1.1255 (-0.97) 1.5390 (1.08) -0.4445 (-0.26) 0.1351 (0.37) 0.0439 (0.15) -0.8014 (-1.45) 0.4158 (1.47) -0.1557 (-1.90) -0.0821 (-1.51) 0.2667 0.01) 0.7797 0.035 7.37 2.06(0.15) 1.34(0.25) 0.146 (0.93) 0.00(0.97)

1.0381 (2.74)

1.0381 (2.81)

1.6300 (2.35) -1.7633 (-2.43)

1.6305 (2.43) -1.7636 (-2.50)

-0.1624 (4.27) -0.0534 (-2.10) 0.3992 (3.74) 0.8360 0.044 20.66 1.68(0.20) 0.26(0.61) 1.67(0.43) 1.16(0.28)

-0.1623 (4.71) 0.3994 (4.02) 0.8438 0.044 25.31 1.58(0.21) 0.24(0.63) 1.68(0.43) 1.16(0.28)

Residual Sum Square F-statistic Serial Correlation Functional Form Normality rile teroscedasticity

The searchfor the model expressed by regression3 obeys the same rules used before; that is, we start from the general ARDL model including all variables and, by testedrestrictions,reduceit to the speciflc form of rcgrcssion3. We means of successive

130

for PFt. F-test PPt. PFt, YSt, PSt. joint the additional variables significance performed 1, 1, 1, 2 specification.This test doesnot reject the null YSt-1, YFt and YFt_Iagainstthe regression I F 0.357(0.925). the variables, with an statistic additional of significance of value of no by the F test, we established After imposingthe restriction suggested that the coefficient estimate of variable P7479 was approximately three times the coefficient estimate of for dummy hypothesis Therefore, the P8391. the tested, variables' cocfficicnts, we variable 81=382: 110: =8 which was not rejected by the Wald test with the statistic value of dummy P7479 into Therefore, P8391 1(0.92). (I)=0.01 the variables and were merged a )? in in 1974-1979,0.33 dummy P the the the period of value unity period which assumes new is includedin the regression P 3 equation. This 1983-1991 new variable andzerootherwise. From the regression3 estimationresults we can derive the long-run relationship betweenthe UK dcmand for Portugal and its determinants,and compute the long-run by long-run The ARDL (1,0,0,1) the estimates given modelare: estimates. coefficients' 0.94069 Intcrccpt: cco = -. = 1-0.60065 1-0 a 1.5661

50 -0.84304 1.4035 Priceof Portugal:c3c, == ^= 1-0.60065 1-0 Priceof Spain: CC2 UK Income: a3 Dummy P:
&4

bo
A

1-0

1.0381 1-0.60065

1.7283

PO + P, 1.6305 -1.7636 A= 1-0.60065 1-0 -0.16232 == 1-0.60065 0 -A

-0.2215

-0.27024

Velocity of adjustment:I-0=

1- 0.39935= 0.60065

11encc, the equilibrium relationship for the UK demand for tourism in Portugal is given by: YPt = 1.566- 1.404 PPt+ 1.728PSt - 0.222 It - 0.270 P, + Ot (2.53) (-0.60) (-3.78) (1.29) (-4.14) (4.20)

The residuals of regression(4.20) constitute the EQE series which is uscd to long-run ARDL model.The EQE scricsis given by ECM the the with associated construct [1.566 I PPt 1.728PSt 0.222 It EQEt = YPt 0.270 Pt + -1.404 The estimated for the selectedARDL resultsof the error-correctionrepresentation in brackets): follow (t-rStios modelareas

131

AYPt =-0.843 APPt+1.038APSt +1.631 All - 0.162APt - 0.601 EQEt-, (-4.71) (-6.05) (2.43) (2.81) (-4.20)

AdjustedR2=0.730

RSS=0.044

F(S,22)=15.84

DW=2.31

20 following for EQE this test The conventional rcsults: providesthe unit root MacKinnon 21 critical values
5% 10%

Test ADF(O)=DF

EQEtseries

-4.9110

-4.0583

-3.6785

The resultsshowthat, at the 5% significancelevel, the EQEseriesis stationaryand

the variables in the long-run relationship are cointegrated.

As before, we apply the PP procedureto conrirm the existenceof the long-run

by indicated tile tests, to the whether root establish unit and conventional relationship The PP ARDL forcing" "long-run tile of crror-corrcction version variables. rcgressors are in is demand Portugal UK for tourism the model

AYpt'- + 73Alt+ 74APt + 72APSt 70+ 71APPt (4.21) '4pt-l+ SYpt-l '2A-l + '311-1 + + Wt + + 'IPP t-I is long-run between be the hypothesis tested to whether a relationship exists the and is long-run P. Hence, I PP, YP, PS, the of a relationship null of non-existcncc and variables X3 4"' ks=O Xl %2' for i=1, 5. Xi* ""2 I 11: 0, the definedby Ho: any alternative against ..., As before, the relevant statistic for this test is the F-statistic and the appropriatecritical 's (1996) The in Pesaran found this tables. results obtained using procedurc et al. are values in 4.7. table arepresented

20 lag-lengthfor the test regression The SBC andAIC criteria indicatethat the appropriate with interccptand 28 ADF(O) is test the Therefore, trend with observations we use available. no zero. 21 28 observations, TheMacKinnoncritical valueswerecomputed considering an intercept, no trendandthree is beingtested. for which the null of non-cointegration l(l) regressors

132

Pt., It. PPtt. F-testfor additionalvariablesYPt.1,PSt. Table 4.7: Joint significance and 1, 1 1,
Dependent variable AYP APP APS I API Explanatory variables APS,APP,AT, andAP APS,AYP, AT, andAP APP,AYP, AT, andAP I APS,APP,AYP, andAP F-statistic (5,18) 6.0411 6.1219 6.3743 2.1197 Critical band at the 5% level Test result

Rejection of the null Rcjcction of the null [3.219;4.378] Rejection of the null I I 1 Non-rcjcction ofthe null _

The above results suggestthat a long-run relationship exists between the UK its determinants, indication for demand Portugal the supporting obtainedwith and tourism that Spain and Portugal the conventionalunit root testsabove.The PP test also suggests is, forcing' for YP 'long-run be that the the treated of explanation as cannot variables price 22 being PP PS endogcnous. and test indicatesthe possibility of
4.3.4. ANALYSIS OF THE ESTIMATED RESULTSFORTHE UK TOURISM DEMAND IN FRANCE,
SPAIN AND PORTUGAL

In this section, we start by establishingwhether the regressions'coefficients are interpretation focus Then, the the economic of estimationresults we on stable. structurally is The in the the cocfflcicnts structuralstability of regressions' previoussection. obtained testedby using the cumulative sum of residuals(CUSUM) and the.cumulative sum of is CUSUM by (1975). The Brown (CUSUMSQ) tests test proposed et al. residuals squared in detecting for systematic changes the regression cocfficicnts, Whereas particularly useful detecting departures from is for CUSUMSQ the constancy test or sudden useful random the below CUSUM The the tile the show of and graphs plot coefficients. regression of CUSUMSQ statistics for each of the three ARDL equations.The pair of straight lines
bordering the plot of the statistics representsthe conridencc interval at the 5% signiricancc level. If either of the lines is crossed, the null hypothesis of correct specirication for the

level. be 5% the at must rejected regression equation

22See section 4.3.1. on the subject of endogenous regressorswithin a long-run relationship dcrived from an appropriately specified ARDL model.

133

Figure 4.11: CUSUM and CUSUMSQ statistics lor thc I JK domind equation Ilor Francc

Ir

-15, 1970

1975

1980

1985

1990

1995

1997

Plot of Cumulative Sum of Recursive Residuals

0.

0
' -0.5 1970

1975

1980

1985

1990

1995

1997

Sum Plotof Cumulative Residuals of Squares of Recursive

134

Figure 4.12: CUSUM and CUSUMSQ statistics tor the I JK demand cquation flor Spain

15
10

...................................................

-5
-10

-15' 1970

1975

19so

1985

1990

1995

1997

Sumof RecursiwResiduals Plotof CumulatW

Ir

0
' -0.5 1970

1975

1980

1985

1990

1995

1997

Sum Residuals Plotof Cumulative of Squares of Recursive

135

Figure 4.13: CUSUM and CUSUMSQ statistics flor the I IK dcniand cquallon I'or PortUg, 11

1 rl

-15, 1970

1975

1980

1985

1990

1995

1997

Sum Residuals Plot ofRecursive ofCumulative

1.0 0.5
0.0

...................................................

w, .............................................................

' -0.5 1970

1975

1980

1985

1990

1995

1997

Sum Residuals Plotof Cumulative of Squares af Recursive

136

be hypothesis indicate The tests cannot that the null of correctly specifiedequations, 24 Therefore, the level that 5% regressions' can we accept the significance of rejectedat . coefficientsare structurallystableandadd one more positive point to the quality evaluation of the models. Table 4.8 We turn now to the economicinterpretationof the cocfticients estimates. for longthe t values and short-run coefficicrits of the estimatesand respective shows income, own- and cross-priceelasticities,dummy variables and adjustmentvelocities of UK demandfor France,Spain and Portugalequationsobtainedpreviously. Symbols is, 11 level. 10% 5% 1%, I and significance respectively, and represent, Table 4.8 Estimatesfor the long- and short-rundummy coefficients, long. and short-run income, andvelocity of adjustment. own- andcross-price of elasticities I SPAIN PORTUGAL FRANCE I Short-run Long-run I Short-run Long-run Short-run Long-run pp
1 PS PF 1 Yp Dummy variables Velocity of Adjustment 0.148 (7.08)o 0.095 (6.75)* 1.363 (4.60)o -2.917 I (. 10.34)e 2.670 (16.25)* -0.167 (-0.79) 0.223 (1.10) 0.963 (2.80)s -0.991 (-2.58)e 0.697 (1.85)' 1.538 (7.47)o 0.369 (2.75)0 0.153 (4.16)o 0.722 . (-2.90)* 1.044 (3.56)o 1.121 (5.11)* 0.269 (2.88)o 0.112 (3.95)o -0.270 (-3.78)o -0.162 (4.71)o -0.222 (-0.60) 1.631 (2.43)

-1.404 (-4.14)*
1.728 (2.53)"

-0.843 (4.20)o
1.038 (2.81)o

0.643 (10.75)9

0.729 (6.25)o

0.601 (6.05)o

to The statisticalevidenceobtainedfrom the estimationof the ARDL modelsseems We countries. about neighbouring postulatedthat the conjecture support our previous likely it is find link between the to more evidence of a are, countries closer geographically
24We also appliedthe recursiveleastsquares and the rolling leastsquares regression regression estimation Thesetests confirm the cociTicients'stability methodsfor testing the structuralstability of the equations. indicated by the CUSUMandCUSUMSQtests.

137

destinations indicates as competitors rather than them, which close neighbouring France link bctwcen fact, In and the suggest significant results a estimation complements. Spainthroughthe price of Spainin the equationfor Franceand the price of Francein the in Spain Portugal Spain between for Spain, the through the of price and and equation for in in Spain. UK Portugal for the Portugal the through equation expenditure and equation for FranceandPortugal. the equations However,no significantlink is foundbetween In general,the elasticitiesestimateshave plausible magnitudesand the expected income the When and own-priceelasticitiesarc, respectively, significant, statistically signs. from tile and cross-priccelasticities commodities, normal expected as and negative positive destinations from than complements. are which competitors rathcr as expected positive are TIc velocities of adjustment are reasonably high and of similar magnitude across destinations, suggesting that UK tourists adjust fast to changes in their demand UK tourists need more than one period (15 to 16 Yet, for all destinations, determinants. its long-run demand fully to their to equilibrium path. months) adjust
In the equation for France, the estimate for the long-run own-price elasticity is

in in fall increase in France implying tourism that causes an prices a negative,as expected, is, in Tourism France for demand thcrcforcg, UK this tourism a non-nal country. the in for France long-run UK demand in tourism However, the the responds while commodity. its in in highly the to this country, short-run price changes negatively significantly and for The is these statistically signiflcant. reason not resultsmay although positive response indicates UK (0.64), tourists that the need estimate which adjustment of velocity on rest in determinants fully for 16 (around the to than adjusting changes months) oneperiod more immediately be for Therefore, France. demand their short-run changes will not perceived of (within one period) and plans alreadymade for crossingthe Channelwill remain so, no 25 in occur prices. matterwhat changes The positive sign of the long-run estimatefor the Spain's price clasticity in the by destination If indicates Spain France. for France, prices as a competitor of rise equation 1% in Spain, the UK tourism demand for France increasesby 1.4%. Ilowcvcr, this in indicating UK demand is the tourism that the short-run, elasticityestimate not significant in Spain.On the other hand,the UK for Franceis not affectedby short-termprice changes
23 by increases in expcnditure in demand Notethat increases aremeasured andalthoughthe numberof tourists (or diminish) in increase in prices,thosethat the Channel to same even the remain response an may crossing havealreadydecided (demand) to do so are likely to maintaintheir decision,which will increase expenditure dueto the increase in pricesbut not necessarily the numberof tourists.

138

in Franceboth in the long- and in demandfor tourism in Spainis sensitiveto price changes in The demand for is income France in the the elastic the short-run,and more so short-run. long-run and inelastic in the short-run,while that for Spain is income elastic both in the long- and in the short-run.A possiblereasonmight be the fact that Spain is a "primary" budgets increases in (a UK Thus, its destination tourists' to short-lived ncighbours. relative lucky bet in a wining horse,for example),if directedto tourism, will preferablybe spentin Spainratherthan France.The demandfor Portugalis incomeelastic only in the short-run, in long-run budget do Portugal in UK increases the tourism the significantly. not affect as The long- and short-run estimatesfor the dummy variable cocfricient indicate in for demand France 1975-1981,1987-1989 impacts UK the the periods and on positive 197S-1976 demand for in UK Spain impacts the the 1994-1997, on periods and positive demand UK for impacts Portugalin the periods 1974-1979 the 1986-1989 on and negative location, because is hypothesis It 1983-1991. that, their these of geographical our and The fact that, for all destinationcountriessharea relatedfaith in UK tourists' preferences. destinations'equations, the dummyvariablesincludedare definedover similar periodsmay be interpretedas an extra link among these ncighbouring countries concerningthe UK demand for tourism. We believe that the change of political regimes to democracyin joining the EU on the Portugaland Spainin the mid 1970's,their subsequent simultaneous I` of January1986,the substantialappreciationof the Spanishpesetabetween1986 and in Tunnel brought in Channel 1994, the the 1991,26 of opening about, although and different magnitudesand signs, consequences which affect the UK tourism demandfor thesedestinationsboth in the short- and long-run. Franceseemsto have profitcd in tile (1974) from Portuguese the the of upheavals revolution and political regime short-run, in from long-run, democratic in Spain (1976) the the the and, stability of regimes change joining both EU from in their these the the opening of the countries, and established ChannelTunnel, compounded with the decreasing popularity of sometraditional tourism 1990's. longThe in dummy Spain the tile and of mid short-run estimates since resorts indicate for Spain based in the similar effects equation on reasonssimilar to coefficient to the equationfor France,meaningthat UK tourism in Spain may thosecorresponding
26 boomcdandthe real valueof the Spanish FollowingSpain'sentry into the EU, consumption and investment increased the currencies For example, tradingpartners. against of Spain'smajor European peseta substantially the peseta/deutsche rate appreciated almost 30% by 1991.1lowcver. the processwas mark real exchange fell investment in 1992: A similar accountof sharplyandthe real exchange reversed ratedepreciated. starting factscanbe given for the caseof Portugal,bearingin mind the differences the of size and strengthbetween two economies.

139

havebenefitedin the short-runfrom the ordealsof the Portuguese revolution,particularly in the period 1975-76,and in the long-run from Spain and Portugal's changeof political in The EU. the political and social upheavalsin Portugalduring the and entrance regimes its during its 1980's, 1970's, then the the economy of and vulnerability and slow and mid have 'long-memory' to seem effects on UK tourists' uneven recovery ever since, impacts for dummy the the negative which of coefficientson the UK account preferences, demand for tourism in Portugal. Furthermore,,these cffccts seem to bear some long-run income for the the of non-significance elasticity estimate. responsibility The positive signsof the elasticity estimates of the UK demandfor Spainrelating to in indicate destination France Spain this country as a competitor of and more changes price long-run. increases in in Price France have in than the the shorta positive, elastic and so in level) UK demand (1% for Spain in long-run, tourism the the effect while in significant the short-run,this effect is alsopositivebut inelasticandonly significantat the 10%level. In the equationfor Spain, the results indicate the own-price elasticity of the UK
demand as negative and below unity both in the long. and short-run. Therefore, tourism in Spain is a normal good. The estimates for the sensitivity of UK tourism demand to price in in in impact Spain Spain has that suggest a permanent rise prices a negative on changes the UK demand for Spain, but a positive effect on the UK demand for France and Portugal. This seemsto support further the hypothesiscd existence of a strong link between the UK tourism demand for Spain and France on the one hand, and that for Spain and Portugal on the other. The association between the UK tourism demand for Spain and France is by between the these two countries, qualifying them as significant price effects supported in UK between The tourists' the UK tourism demand preferences. association competitors for Spain and Portugal is supported, in the equation for Spain, by the significant and in increases UK that the tourism expenditure in Portugal have on tourism effects positive demand for Spain and, in the equation for Portugal by the positive and significant effects that price changesin Spain have in the UK demand for Portugal. Price changes in Portugal do not affect UK tourism in Spain, while price changes in Spain do affect, positively and significantly, UK tourism in Portugal, both in the short- and in the long-run. As for France, UK income increasesalso have positive effects on the UK tourism demand in Spain, both in the short- and in the long-run. However, these effects are greater for Spain in the shortlong-run. in Therefore, France tends to bencfit more than Spain for France the run, and

140

in UK tourists' wealth, while Spain tendsto benefit more than from permanent increases Francefrom short-runpositive impactson the UK per capitaincome. In the equation for Portugal, the estimatesindicate the own-pricc:elasticity as in long-run. They inelastic in the the and negative short-run and elastic also negativeand in indicatethat the sensitivityof UK demandfor tourism in Portugaltowardsprice changes Spainis significant,positiveandelasticboth in the short-and in the long-run. neighbouring This showsPortugalasa destination competingwith Spainfor the UK tourists' preferences. in Portugalnffcct the UK demand Apparently,in the short-run,own-priccincreases
for this destination more negatively than own-price increases in Spain or France. In the long-run, the negative effects of own-pricc increases are greater for France and Portugal than for Spain. The high relative sensitivity of the UK demand for Portugal towards ownin in Spain, Portugal changes and compounded with the positive cffcct price changes price that increasesin the UK demand for Portugal have in that for Spain, should make prices a destination key for Spain in the targeting tourism policies competitive priority and implemented by Portugueseauthorities. The UK income elasticity and adjustment velocity further longthat tourism the policy changes suggest should concentrate on rather estimates income The the than on elasticity estimate indicates Portugal as a top short-run short-run. destinations for UK in longtourists three the the neighbouring among whilst, preference income in do have UK impact increases demand for tourists' their not significant a on run, Portugal. In contrast, positive, significant and high long-run effects on tourism demand are for from income for France UK Spain Therefore, particularly, and, increases. expected tourism policy in Portugal should focus on reverting some of these effects in its favour. The above considerations seem to be consistent with the underlying economic theory, permitting the specifications and the econometric methodology applied in their derivation to be seenas reliable means to analyse the UK tourism demand for France, Spain interesting from Other Portugal. aspects may arisc comparisons between these results and and those of other studies using similar methodologies and identical origi n-dcstinations. 4.4. EMPIRICAL STUDIES USING A COMPARABLE

METHODOLOGY

FOR

THE ANALYSIS OF UK TOURISM DEMAND

Studies of tourism demand using single equation models which are dcrivcd from economic theory and apply appropriate methodological tools for their specification,

141

estimationand testing are few in number.These include Syriopoulos(1995), Kulcndran (1996), Kim and Song (1998), Vogt and Wittayakorn (1998) and Song, Romilly and Liu (2000).The studiesby Syriopoulos(1995) and Song et al. (2000) are particularly suitable for the comparisonpurpose we have set, becausethey apply the general to specific to constructempirical error-corrcctionmodelsof the UK demandfor tourism in approach the samedestinations with. Comparedwith most previous literature in we are concerned tourism demandanalysis,thesetwo studiesshow considerableadvancesin aspectsboth Syriopoulos(1995) includesa dynamicdimensionthat was theoreticalandmethodological. lacking from Song the empirical study almost every on subject. el al. (2000) previously for long-run techniques the the analysis most recent cointegration of evaluation apply demand its determinants between in models derived with the tourism and relationships (1987)two-steps Engel-Granger method. Syriopoulos' (1995) study analyses the UK demand for tourism in several Mediterraneandestinationsincluding Portugal and Spain, during tile period 1960-1987. This study considers a generalto specific approachfor identifying tile long-run and shortdeterminants' dependent the changes on of variable using an crror-corrcction cffccts run longThe information model contains, simultaneously, and short-run about the model.
impacts of the regressors' changes on the UK demand for tourism. I lowevcr, the errorin is imposed this specified model mechanism as a restricted or partial form of correction

by, for example, the 'normal' ECM that would emergefrom an ARDL model as suggested from (1997), Pesaran Pesaran or and an assumedlong-run equilibrium relationship, as Granger by Engel (1987) two-stcpsmethod.However, Syriopoulos, study and suggested does not test for cointegrationof the long-run relationship which is derived from his final Therefore, do know if the long-runcocfficients' estimates model. we not parsimonious hold as steady-state equilibrium valuesor not. The UK demandfor tourism in destinationi 27 by following is given the equation:
lt + P3Arp, + PAP, AYt :-- PlAlt + P26? +P, 02cet +8Ay, -, +OY-Pt-2 + P, Ac, + P62e, + P7Acrp, + p, 2crp, + p9ACCI +7111-2+72CPt-2 +Y3CCPI-2+PIIDV+u,

where 'y' is the UK demandfor tourism measured as the origin's tourism expenditurein the relevantdestination,T is UK per capitareal income, Irp' and V are,respectively,the relativeprice and exchange rate of a destinationrelative to the origin, 'crp' and 'cc' arc,
27 We changed thenotationto facilitateinterpretation andcomparison,

142

rate of a destinationrelative to coMpetitors,'ep' is respectively, relativeprice andexchange DV the own-effectiveprice of a destination and 'cep' areeffective pricesof its competitors. is a dummyvariable.The variables' first and seconddifferentialsarc denotedby A and A2, in logs The so their cocfficients represent are measured variables elasticities. respectively. The short-runeffectsare given by the cocfficicnts attached (p to the variablesin difTerenccs cocfficients), while the long-run effects are given by the coefficicrits attached to the variables in levels (y coefficients) when appropriately weighted by the velocity of . 4.10 Syriopoulos' Tables 4.9 summarise and estimation results for the adjustment Spain of and Portugal. rcgrcssions in Spainand Portugal. Table4.9: ECM for the UK tourismdemand
SPAIN Alt Arpt A2rpt Aet &2e , t Acrpt Ayt.1 (Y-I)t-2 lt-2 CA-2 cept-2 DV
Source: Syriopoulos (1995) Table I a, p.324.

PORTUGAL 1.72 (3.67)

0.89 -(3.67) -0.98 (-2.22) 0.93 (1.24) 0.94 (1.86) -1.47 (-3.83)

2.98 (3.54) 0.94 (1.94) -0.27 (-1.73) -3.71 (-2.43)

-0.18 (-2.94) -0.60 (-2.93)

-0.51 (-3.43) -0.25 (.2.99) 0.45 (5.52) -0.46 (-2.96) -0.74, (.2.90) -0.38 (-3.59)

-1.26 (-3.68)

143

Table 4.10: Long-run income,effective own-price and substitute-price elasticitiesfor the for tourismin SpainandPortugal UK demand
SPAIN Income elasticity Own-price elasticity Substitute-price elasticity
Source: Syriopoulos (1995), tables b, 7 and 5, p. JJU-332

PORTUGAL 2.80 -1.84 -2.9628

1.00 -2.10

Someof the estimationresultspresented are peculiar and seemdifficult to interpret in a coherentway both in theoreticaland prior empirical knowledgeterms.For instance,it demand for does UK justify Portugal difficult to the to that significantly not react seems but in Portugal, to the positively reacts significantly and changes own-price short-run for fact, In A2rp the cstimate cocfficicnt changes. short-run own-price of rate growth is in Portugal, 2.98. This the rate of own-price changes growth represents variable,which by in increases 1%, UK Portugal if tile the that changes own-price growth rate of means 3%. The magnitudes demandfor Portugalrisesby approximately and signsof the estimates in the equation in competitors'prices(CCPt-2) for the short-and long-runeffectsof changes for Portugalpresentthis country as a strong complementof all the other Mediterranean in for Spain, the the absence destinations equation of significanccof this while considered, has Mediterranean Spain indicates the that competitors no or complements among variable Anotherpuzzlingmatteris the following: how can onerelatethe negativesign destinations. first in difference dependent for the the the the of variable estimate equation coefficient of for Portugal(Ayt.1),which can be interpretedas a negativegrowth rate of the UK demand for tourism in Portugal, with the highly positive long-run income clasticity of the UK In addition, consideringthe origins UK, Ocn-nany USA, demandfor this destination? and Portugalas the most incomeelastic destination(2.80,2.85 and Syriopoulos'studypresents 3.32,respcctivcly)amongall Meditcffancancountrics.This is hardly crcdibIc, particularly in the caseof the USA, given the relative obscurity of Portugalas a tourism destinationin the 1960'sand the political, social and economicinstability of this country in the 1970's Syriopoulos' data two-thirds than of sampic. which coversmore

29 (1995)work this elasticityestimateis incorrectlywritten as2.56. In table8 of Syriopoulos

144

The resultsfor Spainseemto be more plausiblethan thosefor Portugal,particularly for the relative and effective own-price elasticitiesin the in the short-run.The estimates be to elasticities of comparablemagnitudesand signs to corresponding short-run, seem found in other studies.However,the estimatefor the long-run elasticity of UK demandfor Spain to own-price changes(-2.10) seemsto be overvalued. For the long-run income be for Spain, tested could not rcjcctcd. was and unity a value of elasticity Some of the above questions could be partially answered if we take into in (1995) This Syriopoulos' definition term the the model. correction error of consideration definition implies that deviations of the UK tourism expenditure from its long-run in into by takes 'corrected' that adjustments account a mechanism only equilibrium path are adjustingthe short-runvalues the incomevariable.Hence,the effor-correctionmechanism leading be long-run to the undcr-estimation of tile the mis-reprcscntcd, to equilibrium may long-run hence, to the the of cffccts. ovcr-estimation velocitiesand adjustment Songet al.'s (2000) study focuseson the examinationof the UK tourism demand including FranceandSpain,during the period 1965-1994, for twelve destinations usingpreidentify integration to the time analysis and gencral-to-spcciric approach series modelling short-run error-corrcetion mechanismwith tile the long-run equilibrium and associated (1987) two-stagemethod.Besidesthe usual right-handside variables,this Engel-Grangcr includes "destination additional explanatory variable, preference an and study constructs "non-economic toivardsthe destinations". index", which measures preferences The dependentvariable in these models is total holiday visits per capita which in formally from the these those with obtained comparing estimation results us prevents increased 's Song Nevertheless, the provides analysis of et al. study previous section. knowledgeaboutboth the UK demandfor tourism behaviourand the use of cointegration demand Furthermore, to tourism techniques models. can somecomparison applied analysis if, althoughbcaringin mind that Songel al.'s elasticity estimates be established are related to the number of tourists while our models' elasticity estimatesare related to tourism 29 for tourism cxpcnditurc. we considernumberof touristsas a rough proxy cxpenditure,

29 is generallyconsidered Tourismexpenditure, of ratherthantourist numbers, asa moreappropriate measure tourismdemand or decrease at a different pacefrom that of expenditure or may sincearrivalsmay increase decreases (increases) due to lower (higher) effective (decrease) while real tourism expenditure evenincrease lengthof stayof the incomingtourists. propensity,or lower (higher) average prices,lower (higher)spending Therefore, the interpretation of tourist arrivalsasa proxy for tourismexpenditure maybe misleading.

145

According to the authors, the ADF test suggeststhat all timc-scries variables
included in the UK demand for tourism equations are l(I). The cointegration tests indicate in fail but 12 to do so in the them cointegration eleven of equations perfon-nedon the equation for France. This means that the statistical inference based on tile estimation be long-run between the tile UK this valid since relationship equation may not results of demand for tourism in France and its determinants may be spurious. For a group of six (including Spain) out of the twelve equations, both long- and short-run models pass all tile

functional diagnostictests for residualsnormality, hctcrosccdasticity and autocorrclation, form and structuralstability. Not surprisingly,the equationfor Francefails severalof thcm. In table4.11 we present the estimationresultsobtainedby Songcl al. (2000) for the long- and short-runmodelsof the UK demandfor Franceand Spain. Table 4.11: Estimationof the long-runandshort-runmodelsfor the UK tourismdemandin FranceandSpain
FRANCE SPAIN

Short-run Short-run Long-run Long-run ECM Cm F. PFt(A) PSt (A) PIRI(A) PITt (A) it (A) PREF (A) 2.123 (11.23) 0.818 (8.80) 1.665 (2.12) 0.904 (10.89) 0.951 (4.13) 0.785 (2.39) -1.079 (-3.14) -0.778 (-2.32) 0.496 . (-2.54) 1.372 (4.73) 0.67 . (-3.16) 2.199 (19.81) -0.491 (-2.44) 1.047 (3.39) -0.329 (-1.25) 2.770 (3.68)

0.791 1 0.919 (11.63) (10.00)

EQEt_j
k2

0.985 1.08 -3.03

-0.286 (-1.26) d2 0.881 8 18

-0.899 (4.16) 0.981 1.86 -4.87 0.855 1.86

DW ADF(O)

Source: Song et al. (20UU),appendix A, pp. 621-24

146

The variables' notation is modified to facilitate interpretationand comparisonof PIT in PIR, PS, Therefore, PF, the tourism represent, and respectively, price of results. France,Spain, Ireland and Italy; I standsfor UK real per capita income, PREF is the from OLS index EQE the the residuals series obtained estimation represents and preference first In 4.11 long-run the table the namesof the column of we present equilibrium model. of the explanatoryvariableswhich, when in levels, representthe relevant variables for the first differences in (A), represent long-runequations the relevantvariablesfor the and,when k2 The last (ECM) three the table equations. presentthe rows of short-run/cffor-correction for DW the cointegrationtest run as the the t-statistic values statistics as well as and between brackets The ADF(O) unit root test for the EQEt. series of residuals. values 1 30 ioS. beneath t-rat the respective the coefficients"estimates represent The authors claim that the results of the ADF(O) test indicate absence of for France, in long-run the that tile which equation means relationship cointegration determinants betweenthe UK demandfor tourism and its assumed may not exist in any inference forecasting be invalid for this and and statistical way procedures may meaningful focus for Spain. Therefore, the the we will on estimation results equation of equation. being diagnostic longBesides tests as well as cointegrated, a postulated passingall be be data-coherent theoretically-consistcnt to a reliable and must run specification long-run between dependent the the of underlying relationship variable and representation determinants. However,someof the estimationresultsreportedin Songel al. its assumed interpretation. is instance, it For difficult to a with plausible economic associate seem for tourism in Spain,both hardly believablethat Irelandshouldbe sucha 'strong' substitute in the short- and in the long-run, whereasItaly is a 'weak' complementand only in the long-run. Moreover, the estimation results reported indicate Ireland as a significant destinations for 12 is Spain considered, while all presentedas a significant substitute destinations Germany, Austria, Nederland in UK USA the as of such and complement Otherwise, the estimationresultsfor Spainseemconsistent tourists' preferences. with those high in to the relating speedof adjustment obtained our study,particularly of UK touriststo inelastic demand, in the their own-priceelasticity and the elastic incomeelasticity, changes both in the long- andin the short-run.
30As pointed out in Engel and Granger (1987), the t-ratio of individual significancc, as any other testing procedure for the first-step equation coefficients obtained with the Engel-Granger method, are not valid means for statistical inference. Indeed. becausethe long-run (first-step) equation is assumed to be a steady. stateequilibrium, a dynamic mis-spccirication may exist, invalidating inference.

147

4.5. CONCLUSION
In chapter 3, we concluded that the static single equation approach for modelling tourism demand was not the appropriate means for undertaking UK tourism demand basis behavioural both lacks This the theoretical upon which plausible approach analysis. features can be explicitly modelled and interpreted, and the empirical methodology with be testing and valid estimation and specification, procedures can which a sound implemented. In static single tourism demand equations, the essential features usually dynamic incorporating hand, the process by the explicitly a structure one omitted are, on its long-run behaviour to adjusts equilibrium and, on the other which short-run consumer hand, the formal. testing of utility theory restrictions. Furthermore, within a single equation framework, the existence of interdependencies among competing destinations may be indicated but not formally identified or tested, since such interdependencies need to be Hence, following inconsistent the configuration. a system of equations within specified in 3 (a introducing by wide range of elasticities chapter estimates, obtained obtained results in implausible the with unexpected models, associated signs and magnitudes minor changes for some of the coefficients' estimates and poor statistical performance of the models), our task for the remainder of this study was the construction of altcmativc cnipirical models inherent flaws in the omissions and overcome static single equation models. could which In this chapter we addressed the modelling of the UK tourism demand within a

dynamic single equationapproach,basedon the principles of economictheory and using for the specification,estimation and evaluation of the appropriateeconometricMethods UK demand The in followed tourism this construction of models adopted. chapter a models "data-basedrevision" strategy known as the gcncml-to-spcciricapproach.Tile general into defined ARDL dynamic taking the an structure as consideration was natureof model throughthe insertionof an crror-correctionmechanism. the UK tourismdemand The dynamicspecificationof the ARDL modelsassumed that, in order to maintain in the steadystaterelationship,touristsadjustthe currentvaluesof their expenditure partly in to response current changes the explanatoryvariables,and partly in responseto tile dcsequilibriumobservedin the previous period. A parsimoniousform of tile model was then derived from the general ARDL spccification through the systematic testing of plausiblerestrictionsand comparativequality evaluationof alternativemodels. From the derived for the long-runequilibrium relationship form, final the equations we parsimonious 148

information the about the short- and and short-runeffor-correctionmodel which provide in its determinants. long-runimpactson the UK tourismdemand originating from changes Prior to the estimationof the parsimoniousARDL models we testedthe order of The unit root testsindicatedall integrationof all the time seriesincludedin the equations. the variablesincludedto be I(l). Nonstationarityof time serieshasalwaysbeenregarded as it in since can generate studies spuriousrelationships amongthe empirical a major problem levels of economicvariables,giving rise to invalid estimation,inferenceand forecasting if long-run However, that theory a suggests stable relationshipexistsamongthe procedures. levelsof economicvariables, they cannotdivergetoo far apart from eachother and from an dattractor'which definestheir steady-state equilibrium path. In this casethe variablesare linear Cointegration is be that to means a of variables combination cointegrated. said individually If two or more nonstationary time seriescan they arc not. stationaryalthough be linearly combinedinto a single serieswhich is itself stationary,the original variablesare long-run be In this to a case, genuine cointegrated. relationship exists among the said longdefincs levels, this the run equilibrium pathor attractor. and relationship variables From the parsimonious ARDL models we derived the long-run equilibrium in for tested stationarity their residualsseriesusing conventionalunit root and regressions tests. These tests indicate stationarity of the residuals and hence, tile cxistcncc of a in However, a single equation specification, tile cointegratcd cointcgratcdrelationship. be found The long-run unique. may not uniqueness of a equilibrium relationship be by Pesaran Shin (1995) tested the using can methodology suggested and and relationship implemented We (1996). Pcsaran thesetesting procedures and tile results obtained el al. long-run between the the of a genuine and meaningful existence relationship support dependentvariable (the UK demand for tourism) and its determinants.Tile statistical from both the the long-run and the short-run estimation results obtained validity of by battery is further diagnostic testswhich providedsufficient confirmed a of specifications ARDL to the specificationsas robust, structurally stable and classify statisticalevidence well-definedmodels. The resultsindicatea significant link betweenSpain and France,on the one hand, hand, but Portugal Spain the other on no signiricatit link between'Franceand and and in Spainaffect the UK tourism demandfor Francein the long-run Portugal.Price changes in Franceaffect tile UK demandfor Spainboth but not in the short-run, while price changes in the long- and in the short-run.The magnitudes and signsof this interrelationship qualify

149

Franceas a strongerlong-runprice-competitor to be a relative to Spain,while Spainseems France. 'flicrefore, we can conclude that to relative strongershort-run price-competitor demand UK Franceand Spain sharea stronglink of dependence tourism concerning and, is bi-directional. further,that this dependence In the long-run, the UK demand for tourism is more sensitive to own-price fact, increases Portugal. In in in France Spain in France the than or own-price of variations in the UK tourism demandfor this country three and two times greater, induce decreases long-run, for Spain Portugal. In from France's the than those and gains respectively, in UK per capita income are larger than those expectedfor Spain or Portugal increases from increases. benefits in Portugal these the more short-run, while in Portugaldo not significantly afTectthe UK demandfor tourism in Price changes Spain while price changesin Spain do have a significant impact on the UK demandfor Portugal. This meansthat Portugal is not a significant pricc-competitorof Spain While in the UK tourism Spainis a significantprice-competitor of Portugal.In addition, increases demand UK for in Portugal Spain the tourism affect significantly and expenditure both in in longin This the tile a modest magnitude, short-run. although and positively, in in Portugal UK Spain. the they tend to tourists the that spend spend more more means Therefore,the UK tourists' preferencefor Spain relative to Portugal, is apparentand it destination, depends happens UK Portugal,. tourism to tile that more as a on what seems demand for Spain, than Spain depends on Portugal. This reinforces our previous link between Spain further, Portugal the of a existence strong and about and, expectations indicatesits unidirectionalnature. The aboveconsiderations permit us to qualify the empirical specificationsused in this chapterand the methodologiesemployed in their derivation, as sound and reliable UK demand for in France, for Spain and Portugal. The the tourism analysing means for obtained allow a theoretically consistent and data-plausible estimation results interpretation of reality within a simplified context, pen-nitting the consolidation of UK behaviour both in knowledge tourists' the the short- and in the longabout empirical in fulfil The this the quality critcria of good empirical chapter estimated run. models models: they are consistentwith the underlying economic theory, compatible with the empiricaldata,simple,relevant,plausible,accurate,reliable, and useful. Their forecasting 'ability will be addressed in chapter8.

ISO

Nevertheless, the equation-by-equation natureof the modelsderivedin this chapter disallows the formal testing of utility theory restrictions involving cross-equation coefficients.Indeed,although they suggestthe nature and direction of the destinations' do for ARDL the the formal testing of these specifications not allow competingconduct, hypotheses. Therefore,the structureof the model must changeto embracea system of be included. features these can within which equations

151

CHAPTER 5
MODELLING THE UK TOURISM DEMAND WITHIN A

STATIC SYSTEM OF EQUATIONS FRAMEWORK

5.1 INTRODUCTION In chapter3, we followed the traditional approachof tourism demandanalysisusing a 4, framework In to specify empirical models. we adopteda more chapter static single equation dynamic in ARDL the of crror-correction estimation modelsandassociated methodology recent in Spain demand for UK France, These have Portugal. tourism the the and models equations dynamics demand for improved behaviour, treatment the of and allow of of explicit advantage in However, testing the a specification and of single equations. econometric methods demand behaviour theoretical tourists' tile of analysis requires consideration of comprehensive demand functions be Indeed, tourism which within should specified. as we principles economic becomes directly theory to complete systems of consider equations, relevant at all on move As inference. the analysis: specification, estimation, and statistical econometric stagesof Deaton(1986, p. 1768)points out "it is not possible to sludy applied demandanalysisivilhout keepingstatistics and economictheory simultaneouslyIn vleiv". This view is rcflcctcd in a base derivation the studies which explicitly empirical of of their models on the number behaviour, has become functional known 'flexible' using what consumers' of as principles formsto specifya theory-consistcnt systemof demand equations. Since Stone's (1954) first estimation of a demand system explicitly derived from been has there theory, a continuingsearchfor alternativespccificationswhich can be consumer been Many have to consumers' preference structure. considered approximations models good but theoreticallyconsistentmodelswhich allow for generalincomeandprice effects proposed,

152

few. Indeed, functional the are choice an appropriate of and also nest simpler specifications form to mimic the true data generating processhas beenthe ccntre of intensedebateamong for the last threedecades. investigators One of the earlier specificationsfrequently uscd to test constimcrthcory hypothesis, (1968,1969) is known Barten by (1965) Theil. f and and as the Rotterdam was irst proposed but instead Stone diffcrcnced is This to that of uses variables of their model similar model. levels. As mentionedin chapter4, it is generallyrecognisedthat models using variables in differencesarenot the bestway to estimatelong-runequilibrium relationships, which is one of the main interestsof this study. In addition,applicationsof tile Itottcrdam-typcmodel tend to basic behaviour due, the theory assumptions of consumer which contradict supply results being based "an (Dcaton to on unhappy approximation or choice offunctionalfortn" perhaps, 1980b, 72). Muclibauer, p. and In more recent studies many other 'flexible' functional forms have been used to indircct functions. Although (19870 Lewbel 1990,199 1, or cost utility consumers' approximate full rank demandsystems 1995)groupsthesefortris in termsof fractionaldemandsystci-ns, and Engel curve approximations, thereare other classificationpossibilitiesthat appearto describe their different propertiesin a much clearerway. For instance,Fislicr el al. (2001) divides the locally flexible functional forms, which flexible functionalforms into threemajor sub-groups: includethe translogmodelsof Christensen ct al. (1975),andJorgensen ct al. (1982),tile AIDS Muellbaucr (1980a, 1980b) Dcaton Lconticf the and and of gencraliscd systemof specification (1980); globally regular functional forms, which include the miniflcx Cavesand Christensen Laurent models discussedin Barnett et al. (1985,1987), the generalexponentialmodel of Cooper and McLaren (1996) and the quadratic AIDS model of Banks ct al. (1997); and forms, include flexible Fourier in flexible discussed the which globally model asymptotically ChalfantandGallant(1985)andthe asymptoticallyideal model of Barnot and Yuc (1998). The locally flexible functional forms have been rccogniscdto have some limitations instance, define. For Caves and they the to regiQns regularity are able associated with Christensen (1980) show that when preferences are not homotlictic and substitutionincreases Lcontief (GL) has the system amonggoods, generalised a rather small regularity region, and the Monte Carlo analysisof Guilkey and Lovell (1980) shows that the GL and the translog (TL) modelsfail to provide a satisfactoryapproximationof the true data generatingprocess.

153

Anotherproblemwith the TL modelis that it can classify goodsas complements when they are The 2001). is AIDS flexible form (Fisher model al. a widely used et substitutes actually logarithm) expcnditurcfunction which derivedfrom a PIGLOG (price independent gcneralised However, Chalfant (1987) Fourier have using seriesand Ramajo region. a small regularity may (1994) using Laurent series show that the regularity region of the AIDS model can be increased. considerably A partial solution to the problemof small regularity regionshasbeenprovided by the higher larger forms flexible development theoretical regularity and rank models of with regions that can approximatebetter more generalEngel curves.Theseare globally regular functional for example,in Cooperand McLaren (1996).The rank of a demandsystem forms as discussed, hasimplicationsfor non-linearEngel curves.SomeEngel curvescan be more non-lincar than higher rank models,suchas the rank 2 AIDS and translogmodelsaccountfor. In thesecases, the quadraticAIDS (QUAIDS) of Banks ct al. (1997) which is a rank 3 demandsystemby better AIDS Engel the these can model, approximate of non-lincar curves.I lowcvcr, extension fail (2001), in Fisher these to provide an cffectivc approximation models may et al. of asnoted the curvature(derivatives)of the true utility or cost functions. Asymptotic globally flexible forms such as the Fourier flexible form (FFF) and the asymptotic ideal model (AIM), also functions, can provide asymptotically global approximations known as scmi-non-parametric for morecomplexdatagenerating processes aswell as for their partial derivatives. It shouldbe noted,however,that no matterwhat degreeof flexibility is adopted,there is alwaysthe possibility of regularity conditionsbeing rejected,and more flexible form than the locally flexible are not immune to this problem. This is so not only becausethe data load independent bear "heavy but technique the of assumptions" a of used, also generally be due factors bias and mis-spccirlcd because to can of regularity such as aggregation violation by itself. In addition, given a specific dataset, dynamicswhich no functionalform overcomes the flexible functionalforms may differ substantiallyin their approximationproperties.Indeed, underlyingspecific economicrelationshipsis unknown,we process asthe true datagenerating in (1980b) Muellbauer Deaton that "there Is no reason to believethat utility agreewith and functions are exactly translogarithmic (or any other prc-derincd functional form, for that matter) either at the individual or aggregate levels". At most, functional forms are Therefore,it should approximations can be more or lessaccurate. and, like all approximations,

154

forms may havemore interestingasymptotic becmphasised that while the scmi-non-parametric it does functions, follow that one is to be preferredto the than the not parametric properties fic datasets. othergiven speci Recent findings in the econometricsliterature, state that estimation and statistical inferencecrucially dependon the integrationand cointegrationpropertiesof the time series involved. For example,Attfield (1997) showsthat with the linear shareequationsof an AIDS be homogeneity rejectedoncethe cointcgratingpropertiesof the time seriesarc cannot system, imposedin the estimation.This implies that, in the spirit of Engel and Granger(1987), testing for linear cointegration and deriving linear forms for the underlying crror-corrcction for forms is AIDS flexible functional Except the model, all referred mechanism essential. by Granger (1995), have tile non-lincar and, as share equations pointed out non-lincar above is largely is There tentative new, complex a and subject. variables modellingof non-stationary linear demand if data between the trade-off with or non-lincar estimating systems: generating a linear be biased; is the with estimates obtained a approximation may non-linear, if the process from be imprecise the obtained results non-lincar may cointegrated, models are variables (Fisher ct al., 2001, p.67). As will be shown in chapter 7, the variables involved in the demand for France, Spain UK Portugal tourism the and are cointcgratcd. of estimation Therefore,the AIDS model appearsto be a good choice to approximatetile UK tourist's structure. preference The AIDS model is seen as a particularly convenient specification with attractive features transformations and restrictions,can be madeto ncst a variety which, with appropriate for Anderson Blundell 1983,1984 (see, 6). The AIDS example, and chapter and models of first-order demand to approximation arbitrary any system,satisfiesthe axioms model gives an for permits perfect aggregation over exactly, consumers, allows simple linear of preferences imposition linear the of restrictionsto test homogeneity methodsand and symmetry. estimation Furthermore, early results obtained with the AIDS specification showing scrial correlation hypothesis demand have been the theory, of qualified as possible problemsand rejecting dueto the staticnatureof the orthodoxmodel,non-stationarity shortcomings of the variablesor the result of income effects arising from aggregation.Indeed, the results of more recent models,capturingpreviously omitted characteristics empiricalapplicationswith aggregate of

155

dynamic introduction by behaviour the of adjustments or trend tenns, tend to consumers' demand consumer postulates. all with resultsconsistent produce The demand theory conceivesan individual consumer endowed with a fixed incomeand facing a marketof commodities with given prices.The consumer's objectiveis to 'a specific bundle of goodsin order to maximiscthe to allocatehis/hertotal expenditure hypothesis his/her The function that preferences' order. utility maximisation rcflects utility functions demand derivation for to given prices the relating quantitiespurchased of allows andexpenditure. Deriving demand equations systems from consumer behaviour theory has inclusion it In the of a numberof theoreticalconstraints on advantages. particular, permits the parameters of the equations,leadingto parsimoniousand cfficicnt modelsas more a imposition incorporated be It knowledge the via can of such also assures restrictions. priori between be total the that and expenditure equation providing models can each consistency feature dcnaand In in theory- the terms-' an addition, essential of consumer used aggregate interdependence of related commodities- can also be embodied within these systems restrictions.Indeed,"the essence throughthe imposition of cross-equation of the consumer demand systemsapproach consists in providing empirical demand analysis with a demand deal framework for interdependencies to the with of various conceptual 57). 1977, (Barten, p. commodities" Although most investigators concerned with tourism demandanalysishavechosen for White their specifying empirical approach models others, equation such as a single (1982), O'llagan and Harrison (1984), Syriopoulos and Sinclair (1993) Papatlicodorou (1999) and De Mello et al (2001), linked their empirical specificationsto theoretical in integrated based theory, utility consumer using system of equations on the principles AIDS model. Rather than the ad hoc reasoningthat tends to underlie the static single first-order AIDS the model a gives approximationto any demand equationapproach, the axiomsof choice,permitsperfectaggregation satisfies over consumers without system, imposingparallel Engel curves,is data-consistent, simple to specify and easyto estimate. Therefore, the AIDS systempermitsnot only the estimationof the cornpIcteset of relevant

1Aggregate datahave relatively less incomevariation than disaggregated data. Low Incomevariationsand found in data,may favourfunctionswith simpleEngelcurves large aggregate often relatively price variations, and/orlow rank suchas the AIDS model. In contrast,functionshaving more income flexibility relative to data. priceflexibility suchasthe QUAIDS model,maybe preferredwhenusinghouschold-level

156

interdependencies information the of competing on elasticities which supply crucial but alsothe formal testingof the consumer theoryassumptions. products, In this chapter,a version of the 'orthodox' static AIDS model is applied to the in France, introduction Spain for Portugal demand UK tourism the the and with analysisof between destinations innovations: the the the two neighbourhood and origin conceptof of during inclusion the tile the of real expenditure coefficient non-constancy a possible of and is believed be former factor in The to the explanationof a relevant under analysis. period the competitive behaviour of the destination countries, as cornplementarity or is by the the cross-price clasticitics, signs of of particular relevance shown substitutability, in this context.The latter is basedon the belief that tourists' allocationof expendituremay factors due that modify the political and economic relationships time to changeover This is betweenthe origin and the destinationsand among tile destinationsthemselves. , in Spain, interesting Portugal the and cases of as thesecountriesexperienced a particularly at the beginningof tile sampleperiod (1969), to a transition from 'developing' economies 'developed' status at its end (1997). The considerationof France as a neighbouring destinationallows for the comparisonof tourism demandbehaviourbetweena developed behaviour its the the to this examination of which and extent countryand poorerncighbours, in its dcten-ninants. less becomes time, given changes similar over moreor derivation of This chapterproceeds as follows. Section5.2 providesa summarised its AIDS and main features.Section5.3 prcsentsthe static AIDS the model, assumptions for in demand France, description UK Spain Portugal tourism the and and a of the of model 5.4 included. Section reports the estimation results obtained and their variables interpretation.A comparisonbetweenthese results and those obtained in other studies demand is AIDS 5.5 Section tourism to the model analysis also addressed. applying concludes.
5.2. THEORETICAL FEATURES OF THE AIDS MODEO The AIDS model can be viewed as an extension (including price cffccts) of tlic

Working-Lesermodel which relates the budget shareswi- to tile logarithmof the total

Thedcrivationof the AIDS modelin this sectionfollows that of DeatonandMuclibaucr(I 980a,1980b)

157

The Pilogx. + that model restsupon a specificclassof preferences ai wi = expenditure x, such the PIGLOG class- which are represented via a cost or expenditurefunction dcrining the level As to attain a specific at given of utility prices. necessary expenditure minimum Muellbauer(1976) shows,the PIGLOG class permits an exact aggregationover consumers imposing identical preferences. without budgetor total expenditurewhich is to be spentwithin a given Let x be the exogenous These be bought in products can non-negativcquantities period on someor all of n products. (qlgq2g ) be H, Let the the of n. q= quantities vector pig n products qi at given prices ...tq,, ..., (pj9p2, the vector. prices and p= purchased, -IPn) is Fn, Defining The budgetconstraintof the representative the utility consumer q, p, =x. iftl functionas u(q), the consumer's aim is to maximiscthe utility subjectto tile budgetconstraint:
pqi =x max u(q) subject to Fn, i-I

(5.1)

The solution of this maximisationproblem leadsto the Marshallian(uncompensated) (p, functions demand qj = gi x). defined be Alternatively,the consumer's problemcan as the minimum total expenditure
level to of utility u*, at given prices: a specific necessary attain min 1: p,qi =x subject to u(q) = u* W
n

(5.2)

dcmand The solutionof this minimisationproblemleadsto the I licksian (compcnsatcd) functionsqj = hi(p, u). Therefore,a cost function canbe definedas
C(p, u) n Y.pi hi (p, u) =x W

(5.3)

Given the total expenditurex and prices p, the utility level u* is derived from the solutionof the problemstatedin equation(5.1). Solving (5.3) for u, an indirect utility function is obtained suchthat:
U=

Therefore, the AIDS modelspecifiesa cost function which is usedto derivethe demand functionsfor the commoditiesunderanalysis.The processof derivationcan be summariscd in the following threesteps:

x) V(P,

(5.4)

158

First,

a Pi

Ilicksian demand functions. hi(p, is derived the establishing u) =

function is indirect for (5.3) Second, the utility obtained,suchthat u= v(p, x). u, solving Third, hj, v(p, x)]=gj(p, x) is retrieved stating the Ilicksian and the Marshallian demand
functions as equivalent.

The Ilicksian andMarshalliandemandfunctionshavethe following properties: I. Adding-up: Ypihi(p, u)= Zpjgj(p, x)= x, meaningthat all budgetsharcs sum to unity; ii 2. Homogeneity: hi(p, u)=hi(Op, u)=gi(p, x)=gi(Op, Ox) VO>O, mcaning a proportional has in the no effect on quantitiespurchased; and expenditure prices all change

3. Symmetry:

0 Pi

(p, clh, u) Vi j, that meaning consumer's choices are consistent; = a j Pi

cl pj is negative semidcrinitc, that is, for

4. Negativity: The n-by-n matrix of elements

form 4, the quadratic any n vector

pj

in This that a rise prices means :50.

in demand fall in asrequiredwhenthe commoditiesunderanalysisarenormal goods. results a The AIDS model specifythe following cost function: In C(p,u) = a(p)+ ub(p) where a(p) = ao + In ai pi +IF, F,y, In p, In p, 2111 and b(P)=Prip, 1, (5.5)

Thederivativeof (5.5) with respect to In pi is: clIn C(p,u) Inp, porlpOil Fy,, + + up, oc, = In ii cl pi
As C(p, u) =x 4* In C(p, u) = In x, thcn

(5.6)

In x= a(p)+u bp) Solving(5.7) for u we obtain

(5.7)

159

In x- a(p) U= bp) Substituting (5.8) in (5.6) we have 0 Inpi C(P) pi pi p1q. (0) hi -w = =a C(O) a Pi C(O) x' =d, +Fy, Inp, +Pj! nx-a(p)j

(5.8)

If we define a price index P such that In P= a(p), then 0 InC(p, u) 0 Inpi or Inp, Pi[lnx InP] Fy, + + oci = -

In wi = ai +Yyo Inp, + Pi

jP

(X)
F.y*kt In Pk In p,

(5.9) (5.10)

In P'= o:O+ where

7- (1k kkt

In Pk +

1E

basic (5.10) AIDS (5.9) the the equations of are model. and equations i is a destination In a tourismanalysis context, countryamong a groupof n alternative by touristsof a sPecific demanded destinations origin. The dependent variablewi, represents destinations. budget i share This destination tourism to the the allocated set of n origin's of j by in i destinations is (p) in tourism the priceof and alternative variability explained share's deflated (x) destinations, by to by the the allocated group of capita expenditure n price per and indexP.Themodelis based on thefollowingassumptions: 1.All budget sumto unity,theadding-up restriction, requiring: shares ZPj=O, j: yU=O, fbrallj; Zaj=I, iii has no effect on the quantities 2. A proportional changein all prices and expenditure homogeneity the restriction, requiring: purchased, Z 0, for all i; j 3. Consumers' thesymmetry areconsistent, restriction, choices requiring: j; i, for all yo= 7ji, 4. A rise in pricesresultsin a fall in demand the negativityrestriction, which rcquiresthe for destinations. elasticities all condition own-price of negative

160

The restrictionsimposedon a and y comply with theseassumptions and cnsurethat

homogeneous function If individual linear defines P (5.10) of prices are prices. as a equation defined be "P to then approximately appropriately proportiOnal any will relatively collinear, InPk by logarithm F'Wk Stone, Is for index, the the ofwhich one used =Inl)*" example, price (Deaton & Muellbauer, 1980a, p.76)3. Hence, the deflator P in equation (5.10) can be In index P* by Stone the such that, price substituted

InP*=Ewi'lnpi i form: in following be (5.9) P, the rewritten and estimated can of equation expression x in( Inpj Pi Eyij + + w, = a, ip*
0

(5.11)

in i base for destination With is budget the this the the year. simplification shareof where wi'

(5.12)

Equation (5.12) specifies a model in the lincar-log forin which preventsthe direct
interpretation of its coefficients as elasticities. I lowevcr, the interpretation of the signs attached to the coefficients of the model in this form can give a preliminary indication of how tile dependentvariable reacts to changesin its determinants. The coefficients of the price variables (yij ) represent the absolute change in the expenditure share allocated to commodity i due to a 1% change in the price of good j, ceteris paribus. For i=j, the coefficient's sign of the ovaibe is to negative according to the theoretical rule or negativity which price variable expected in i*j, For i theory terminology. the sign of the coeft"Icicrit economic qualifies as a normal good is expectedto be positive if i and j are substitutes, and negative if they arc complements. The Pi in in ih i, 1% the the chang absolute expenditure a share given change coeft'icients represent being held in Gencrally, AIDS prices constant. expenditure, an model capita per real framework, a coefficient Pi >0 gives rise to an expenditure elasticity above unity, and a

in interval first In (0,1). Pi to the the <0 an expenditure case, elasticity gives rise cocff icicnt the demand for commodity i would be expenditure (income) elastic and in the second case, inelastic. In income(income) theory terminology, economic with commodities expenditure "luxuries" incomc-inclastic demands demands as and commodities qualified with are elastic
3 If prices are not collinear, the linear approximation of the AIDS model obtained through the use of the Stone price index can bias the parameters' estimates of the budget share equations. However, the bias is likely to be more important in micro rather than in aggregate data as showed by Pashardes (1993).

161

indicate "luxury"'good This "necessities". to that a qualification seems while arc qualified as income imposed its budget be on arc allocation can and restrictions given up when can always "necessary" is be to a purposes, good essential to the consumer and other redistributed always budget restrictions affect only slightly the demand for this type of good. In this sense, an income-inclastic (Pi < 0) good is likely to have few competitors while an income-clastic (Pi > 0) good is likely to have many. This qualification seems to be more adequate in a tourism demand context than the theory terminology, which would qualify a tourism destination, in the first case,as a "necessity" and in the second,as a "luxury".

The model assumes and labour supply,and excludes separabilitybetweenconsumption

budget It that to their consumers allocate also assumes groups of constraints. quantity budgeting implying is in that process, of preferences each group a multi-stage commodities independent from others. Thus, it is assumedthat the expenditure allocated by UK tourists to France, Spain and Portugal is separable from expenditure allocated to other destinations and in is in decision those to the money countries made spend several stages.First, UK tourists that

in budget to tourism their southern then to tourism their other and goods; consumption allocate finally decide between France, Spain they the of other parts world; countries and neighbouring is last AIDS Ile to Portugal. this applied stageof cxpcnditurcallocation. model and 5.3.THE STATIC AIDS MODEL OF THE UK DEMAND FOR TOURISM The UK tourism demandfor France,Spain and Portugal is estimatedusing an AIDS model
for possible non-constancy of the real expenditure cocMicient over the sample period. allowing Changes in this coefficient can occur if tourists' allocation of expenditure is afTectcd, over time, by factors that modify the political and economic relationships between origin and destinationsand among the destinations themselves.The model is as follows:
WP':: (7p +YPPPP+YPSPS+YPFPF+PpE+P'p

[SE]+8pD+OpT+

up

WS: --CS+YSPPP+YSSPS+YSFPF+PsE+P's[SE]+BSD+OsT+us WF =
aF + YFPPP + YFSPS + yFFPF + PFE + PIF
ISE4]+8FD+OFT

+UF

162

budget UK The dependent the tourism country's allocatedto shareof variablesareeach thesethree countries,denotedas Wi, where i=F (France),S (Spain) and 11(Portugal).The in independent tourism the prices of effective each country (Pi), the UK real variablesare dummy destinations UK (E), the to per capita of population a variable allocated all expenditure D and a trendvariableT. The effectiveprice of tourism in country i and the UK real per capita follows: defincd arc as cxpcnditurc
CpIi /CPlUK

Ei /UKP and E=ln p*

Pi =In

Ri

10 i, index destination is is index CPIUK CPlj the the of consumer price consumer price of where the UK, Ri is the exchange rate betweencountry i and the UK, El is the nominal UK tourism
is lnll* in i, UK is index UKP Stone dcrined the tile population and country price expenditure in equation (5.11). The data for UK tourism expenditure, disaggrcgatcd by destinations and from in Business the sterling, were obtained million pounds one common source, measured

Monitor AM6 (1970-1993), continued as Travel Trends (1995-1998). Data on the UK indexes from the International Financkil rates were and exchange obtained price population, (1984,1990and 1998)., Statistics(IMF) Yearbooks The dummy variable D accountsfor severaleventsthat appearto have influencedthe in France,Spainand Portugal,during the period 1974-198 1. The first was UK tourism demand the political turmoil,that followed the Portuguese revolution in April 1974.This event had a is UK flows Portugal, the to tourism on effect negative which bclicvcd to have substantial lasted,at least,until 1979.Second,Spainwas affectedby both the eventsin Portugaland its involving the substitutionof a dictatorial regime of forty years by a changes, own political democracy. Third, had additional events which adverseeffects on the demand parliamentary for tourism all over the world and which particularly affectedweakerand unstableeconomics like thoseof Portugaland Spainwerethe oil crisesin 1973and 1979. Ideally, the modelwould includethreedifferent dummy variablesportrayingthe events that affected demandin each of the destinationcountries. However, this would mean the reducing further the already few degreesof freedom estimationof additional parameters, basisshow that tile different periodsto Testsperformedon an equation-by-cquation available. take into consideration are very close to each other (1974-1979for Portugal, 1975-1980for

163

is for France).Hence,one single dummy covcring the pcriod 1974-1981 Spainand 1975-1981 includedin the model. The coefficientsof the dummy variable arc cxpectcdto be negativcin for PortugalandSpainandpositive in the equationfor France. the equations A structural break in the influence of the UK real pcr capita cxpenditurc,on the dependent the sampleperiod into two sub-periods,1969-79and 1980-97. variablesseparates Importanteventsthat contributedto the structuralbreak arc Spain's membership of EFTA in for EC membership for in Portugal, 1980and Spainand Portugal'snegotiations which started, October 1978 and for Spain in mid 1979, marking the turning point from isolation to in in biggest the potentialmarkets the world. partnership oneof This view can be illustrated by the statementof Josd da Silva Lopes (1996, p.136): "One of the most important consequences of thesenegotiationswas the ending of a secular Portuguese From between Spanish the the economics. and an economicpoint of separation lived had, back back Before E-C, traditionally, to to the two the countries membership view both countriesallowed moretrading concessions to any other countrythen to its neighbour.In 1973Spain tried to negotiatewith Portugal the creation of an Iberian commonmarket, but believed differences dimension in Portugal declined.ThePortuguese that the authorities and be between integration that their the two were such economics would only admissible strength "' In a vast multilateralframework where other influencesandpowers could also be present. Eventsof suchimportanceaffecting Spainand Portugalinevitably affectedFranceand, licticc, amongthe countries.Thus, a dummy variable S, assumingtile the distribution of expenditure 0 is 1980-1997 by for E in I to the the and otherwise, years multiplied variable account value This new variableis dcnotedas SE in system(5.13). the structuralbreakdetected. A proxy for possible omitted dynamics, demographicshifts and deterministic nonin form the of a trend variable.A trend variable may also accountfor added stationaritywas UK behaviour i-addictcd' 'country in (giving tourists tastes of patterns such of as rise changes to a positivetrendcoefficient)or 'cultural curiosity' (resultingin a negativetrend coefficient)by into taken account other explanatoryvariables. Tourists may adherc to the already not behaviour by visiting other countries,even when known or may display more adventurous in destinationcountriesalreadyvisited. If expenditure and prices confronted with unchanged this is the caseand a trend variable is not included, the estimatesof prices and expenditure coefficients may be spuriousandthe estimationresultsmay presentsignsof autocorrelation.

164

The individual and joint statistical significance of the variables D, SE and T where testedon an equation-by-equation and a systembasis.Thesehypotheses were testedby using the likelihood ratio (LR) and the F statisticsfor the single equations, andthe Wald statistic for The LR and Wald statisticsare asymptoticallydistributedas a Quithe systemof equations. squarewith degreesof freedomequal to the number of restrictions imposed.The results of these tests are presentedin table 5.1 (p values in brackets). In-all tests, the explanatory in included (U) the andrestricted(R) modelsarc displayed. unrestricted variables Table 5.1: Testsfor the individual andjoint significanceof variablesD, SE andT
Equations in test Hypothesis LR statistic
France Spain Insignificance of D X2(1)=4.92 (0.03)

Single equation

System of equations

F statistic
F(l, 23)-4.25 (0.05)

Wald statistic
. j2(2)-42.47
(0 00) .

X2(1)=13.09(0.00) 1(1,23)-13.12(0.00)

U: PP, PS, PF, E, D

PP, PF, E R: PS, Portugal

France Spain Insignificance of SE

(0.00) F(l, 23)-20.90(0.00) X2(1)=18.74

X2(1)=5.59 (0.02) X2(l )=6.3 0 (0.0 1) F(l, 22)-4.68 (0.04) F(l, 22)-5.34 (0.03)

U: PP,PS,PF, E, D, SE Portugal R: PP,PS,PF, E, D l ---France Insignificance of T, Spain U: PP,PS,PF, E, D, SE,T Portugal R: PP,PS,PF,E, D, SE

(0.22) F(l, 22)-1.18 (0.29) X2(1)=1.51 (0.02) F(l, 21)-4.57 (0.04) X2(1)=5.71 X 1)=6.39(0.01) F(l, 21)-5.18 (0.03) (0.30) F(IO21)-0.78 (0.39) X2(1)=1.06
2(

X2(2)-5.36 (0 07) .

X2(2)-5.36 (0 07) .

France Insignificance D, SET X2(1)=24.39 F(3,21)-9.23 (0.00) (0.00) of Spain I (1)- 17.61(0.00) F(3,21)-5.85 (0.00) X2 U: PP,PS,PF, E, D, SE,T PS, PF, E R: PP, Portugal X2(l )=21.3 1(0.00) F(3,21)-7.60 (0.00) A

X2(6)-52.01 (0 00) . I

The resultsof table 5.1 show that the variablesSE and T are highly significant ill the for Franceand Spainbut not significant in the equationfor Portugal.Consideringall equations the equationsas a system,the orthodox AIDS model, involving only prices and per capita

165

involving 'unorthodox' is the the against adopted, model rejected as regrcssors, expenditure T. SE Variables D, and explanatory additional The adding-up restriction is incorporatedin the system by suppressingone of the irrespective invariant The from the estimation results are of which estimation. equations be from deleted is the the the of recovered equationcan coefficients equation excluded,and homogeneity The by the applying adding-up property. and other equations' estimates imposing by linear the tested on the appropriate required constraints symmetryrestrictionsare is Zcllncr's homogeneityThe using symmetry-rcstricted system estimated and parameters. 4 (SUR) method. (1962)seeminglyunrelated regressions An implicit assumptionin most of the literaturc concerning timc scrics rcgrcssion Nonstationarity has data been is time that stationary. of always arc series such analysis it in studies since can give rise to spuriousrelationships regardedas a problem econometric for A levels the variables. common practice avoiding tile problem of of economic among in involving between is timc-scrics the variables most regression analysis association spurious 5 introduction of a trend variable. Although this procedureis not free of criticism undcr the it in flic timc-scries thcorctical analysis, somc cascs can work on makc scrutiny of rcccnt between dependent true the the association and the explanatory rcflcct cocfficicnts regression in As the the model. shown chapter, previous appropriateway of an econometric of variables in data is time the of nonstationarity series problem cointegrationanalysis,which approaching investigates the existence of meaningfullong-runrelationships amongeconomicvariables. The stationarityof the time-scriesvariables reprcscntingtourism sharesof the three
destinations (WF, WS, WP), tourism prices in France, Spain and Portugal (I'll, PS, PF) and the UK per capita real expenditure (E) was analysed using the Dickcy-Fullcr and augmented Dickcy-Fuller tests. We found all the variables in levels to be integrated of order one, or I(l), integrated be 1(0). levels differences Therefore, first to the their of order zero, or variables' and first differences In their this chapter, we assume that the are stationary. and are nonstationary in AIDS in involved the that the model are cointcgratcd and equations system (5.13) variables
4 In the case of the unrestricted model, the OLS and Minces estimators arc equivalent since the vector of independentvariables is identical in all equations. In the case of the restricted model, Zcllnces estimator is more efficient than the OLS estimator. The efficiency gain is directly related to the correlation between the differcrit equations'disturbanceterins and between the sets of independentvariables. 3See,for example, Gujarati (1995), pp. 240-1.

166

between long-run dependent the equilibrium, relationship variablesand meaningful represent 6 I do have determinants. Although their statisticalproof of cointcgratcdequationsin we not yet the AIDS model, the estimationresultspresentedin the next sectionalreadyindicate system data-plausible (5.13) as a theoreticallyconsistent, and statistically robust econometricmodel, the presence relationships. of a genuinesteady-state endorsing 5.4. EMPIRICAL RESULTS AND THEIR INTERPRETATION The unrestrictedsystemis estimatedcquation-by-cquation using OLS. The symmetry
by linear homogeneity tested the are imposing restrictions required constraints on to the and is The Zcllncr's (1962) SUR method system restricted estimated parameters. with appropriate , for consistency with economic theory assumptions upon which the model is based.Tile Wald 2 homogeneity homogeneity, test of and symmetry simultaneously, provides Y. symmetry and 2.885,0.744 4.080 lie below the respective 5% of, respectively, and which values well statistic distribution degrees freedom (5.99,3.84 and the two, three of X2 with one and values of critical 7.8 1). This implies the non-rejection of the hypothesesunder consideration.

5.4.1. ANALYSIS AND INTERPRETATION OF THE ESTIMATION REsums

TabIc 5.2 showsthe cocf(icients' cstimatesand asymptotict-valucs in brackctsof tile

for Portugal (PT), Spain (SP) France (FIZ) using three versions of the and equations share homogeneity under and symmetry (II+S) and under homogeneity and model: unrestricted, symmetry plus the additional restriction of null cross-price cffects between the share equations * (11+S)O. The France Portugal symbols and 0 indicates the 1% ruid 5% signiricancc and of

Table 5.3 showsthe Lagrange Multiplier (LM) versionandthe F versionof level, respectively. diagnostictests for serial correlation,functional forrn, error normality and hetcrosccdasticity The LM versionstatistic follows a Qui-squarcdistribution and tile F for the modelsestimated. distribution. F The p valuesof thesestatisticsarc given in brackets. follows a standard version

6 In chapter7, we showthe non-stationarity testsfor the time seriesinvolved and provide statisticalevidenceof cointegration of the AIDS system. amongthe variables

167

Model underHomogeneityandSymmetry) Table 5.2 UK TourismDemand(Restricted

AIDS Model
Cti yip Yis YiF 81 Oi

U. 0 092 . Unrestricted 17)o 0.091 Restricted PT (11+S) (6.01)* 0.090 Restricted (6.62)o (1I+S)O 0.412 Unrestricted (7.44)o SP Restricted (11+S) Restrictcd (II+S)O 0.387 (7.86)o 0.387 (8.24)o

-0.065 (-2.23)" -0.072 (-2.99)o -0.072 (-3.5 1)* -0.06 (-0.07) 0.072 (2.95)9 0.072 (3.51)o 0.072 (0.85) -0.0004 (-0.02) 0 (none)

0.073 (2.39)" 0.072 (2.95)o 0.072 (3.51)o -0.375 (-3.92)o -0.453 (-5.62)o -0.453 (-5.98)o 0.301 (3.39 o 0.381 (5.19)o 0.381 (5.29)o

-0.018 (-0.51) -0.0004 (-0.02) (nonc) 0.526 (4.75)o 0.381 (5.19)o 0.381 (5.29)o -0-508 (-4.92)o -0.380 (-5.22)* -0.380 (-5.29)o

-0.010 (-0.81) -0.010 (-0.99) -0.010 (-0.99) 0.111 (2.90)o 0.127 (4.26)o 0.127 (4.37)o -0.101 (-2.83) -0.117 (-4.27)o -0.117 (-4.41)o

0.005 (1.10) 0.006 (1.73) 0.006 (1.74) -0.035 (-2.57)o -0.048 (-4.45)o -0.048 (-4.46)o

-0.017 (-3.60)o -0.018 (-4.28)o -0.018 (-4.28)9 -0.053 (3.52)9 0.049 . (-3.53)o -0.049 (-3.53)o

0.001 (0.88) 0,001 (1.17) 0.001 (1.25) -0.11 (-2.28)* -0.011 (-3.85)9 -0.011 (-3.92)o 7.009 (2.14)o_ 0.009 (3.59)o-. 0.009 (3.84)o

0.496 Unrestricted (9.60)o 0.522 Restricted FR (11.6)9 (11+S) 0.522 Restricted (11.8)o (II+S)O

0.071 0,031 (5.02)o (2.38)* 0.043 0.067 (4.24)o (5.17)o 0.043 I 0.067 (4.24)o (5.17)o

AIDS models Table 5.3 Diagnostictestsfor the equations of the unrestricted

I
SerialCorrelation LM version PT SP 1 FR 0.32 (0.57) 0.06 (0.81) 0.00
(0.98)

DIAGNOSTIC TESTS
FunctionalForm LM version 2.40 (0.12) 4.95 (0.03) 3.47
(0.06)

Equation

Norniality LM vcrsion 1.333 (0.51) r vcrsion Ila na na

I leteroscedasticity LM vcrsion 0.45 (0.50) 0.38 (0.54) 0.47

(0.49)

F version 0.23 (0.64) 0.04 (0.85) 0.00

(0.99)

F version 1.80 (0.20)

F version 0.42 (0.52) 0.36 (0.55) 0.44

(0.51)

. 1 4.12 0.63 (0.06) (0.73) 1 2.72 0.78

(0.12) (0.68)

The additionalrestriction of null-crossprice effects betweenthe cquationsfor France by is Portugal and suggested the estimationresults of model (11+S)and, accordingly,tested. j2 statisticvalue is 4.08 indicatingthat the hypothesis Thecorrespondent cannotbe reject cvcn

168

is included in level. When 1% the model, even though the this the restriction signiflcance at individual significance of some of the codficients increases,the changes in the magnitudes of the estimates are minimal. This indicates that the estimates of model (11+S) and those of this denote by in (11+S)o the cross-price effects null, additional restriction of same model under Table 5.2, do not differ significantly. Therefore, the interpretation of tile results focuses on the homogeneity (11+S) from the under and model symmetry since this obtained estimates its demand the theory of consumer and estimates prc-requirements specification complies with do not differ from those obtained with the model (11+S)O.

5.4.1.1.Interpretation of the coefficients' estimatesof systent

The explanatory variables are all significant at the 5% signiricancc level or less, except for the price of Portugal in the equation for France and the price of France, trend and real 7 for in Portugal. In diagnostic tests the the equation general, and tile variables expenditure basis fit cquation-by. on an cquation with the unrestricted statistics obtained goodness of All AIDS the regressions arc the a statistically robust specification. system as model, show high for indicating the independent a relatively explanatory power significant overall, is licterosccdasticity there serial evidence of correlation, or non-nornial no and variables, distributed errors. For the whole system, the null hypothesis of all cocfficiqnts being zero (equivalent to the overall significance F test on an equation-by-equation basis) was tested. Tile Wald statistic value for this hypothesis is 245.6, indicating a strong rejection of the null.

The rcgrcssions'cocfficicnts canbc intcrpretcdas follows: celerisparibus, yomcasurcs th in i the expenditure share following a 1% change in Pj and the the absolute change P'j) in (Pi i's budget Pi 1% + the measure absolute change and country sharc, per coefficients in in in 1969-1979 1980-1997, UK the the capita expenditure period and period real per change insignificant, in both UK For Portugal, these that the values are so changes real respectively. do Portugucsc However, the these not significantly affect share. capita per expenditure in for France Spain. the significant equations are and estimates coctTicients'
7 The equationfor Portugalshowsevidenceof correlationbetweenthe trend variable and the real expenditure variable(correlationcoefficientis 0.97). Whenthis equationis estimated without the trend,the coefflcientof the expenditure variableis positive and significant,but when the trend variableIs includedthe estimated cocCricient insignificant. signandbecomes of theexpenditure variablechanges

169

If Pi > 0, the shareWi increases with E and if Pi < 0, the shareWi decreases when E increases. In demandtheory terins, this would indicate tourism in Spain to be a 'luxury' and Spain tourism in Franceto be a 'necessity'.An alternativeterminology might be to categorise destination France 'secondary' (first (second 6primary' preference) and as a choice or as a would prefer to direct additional expenditure meaningthat consumers choice or preference), towards Spain rather than France.Generally,a coefficicrit Pj >0 gives rise to expenditure Di in <0 to tile coefficient elasticities unity, and a gives rise expenditure above elasticities interval (0,1). This meansthat for each 1% increasein the UK expenditureallocated to the increase, 1% (less) (France) Spain than the would respond with a more which region, shareof for increases. Spain budget UK their tourists' when preference confirms The valuesof the coefficient for the dummy variable D demonstrate that the Spanish in 1974-1981 had Portuguese the took that the oil crises and changes place period political and Portugal France. in Spain Ilcncc, this and a net positive and cffect on on effect a negative in favour France, UK Portugal Spain. The tourists' to moved preferences of relative and period, be interpreted in the trend the tile can variable as of annual average change coefficients in in take the the other place would change shares which absence of any expenditure increases Spain's by decreases France's share while share approximately variables. explanatory in increase The the Portuguese the sameamount. shareis insignificant. A more detailedanalysisof the results requiresthe relevantelasticitiesvalues.These following formulae: the using werecalculated Elasticities: Expenditure
d In q, dw, pi I= + +1 =I F, _ Inx iv, d In x

UncompensatedOwn-Price Elasticities: yij Cii= =I _I= Wi dInpj wi dInpj dInqj dwi ' w, Wi

Elasticities: Uncompensated Cross-Price dInqj CO == d In pj dw, Yjj Wil i _L d In = =p, wj wi wj

170

CompensatedOwn-Price Elasticities:
C ii = Cii + wi 90ci 7ii wi +wI '90

Elasticities: Cross-Price Compensated

+ w'i'Ci 7ii = =- +w wi 90

" destination i (i=l,..., Wi the sample'saverage shareof n) and w, represents where represents j 0=1,..., n) in the year base.For France,Spainand Portugalthe values the shareof destination following: base the these are shares andyear average of
Overallperiod average share (1969-1997) w 0.0775 0.5635 0.3590 First period average share (1969-1979)
7V69-79

Year-base share (1990)

w 90

Second period share average (1980-1997)

WBO-97

Portugal Spain France

0.0923 0.4607 0.4470

0.0574 0.6267 0.3159

0.0898 0.5249 0.3853

The ensuingdiscussionfocuseson the uncompensated elasticitiesas they tend to be

8 The for important purposcs. estimates of the expenditure and uncompensated policy more (11+S) the elasticities within restricted model and rcspcctivc t-valucs are cross-pricc and ownin 5.4.9 table presented

' The valuesof the elasticitiesquantify the sensitivity of tourism demandto changes in the expenditure budget (expenditureelasticities),prices in the destinationunder consideration(own-pricc elasticities) and prices in Compensated (cross-price In destinations elasticities). price elasticities allow for the effectsof changes alternative price changes while uncompensated price elasticities only considerprice changes. realincomewhich accompany
9 Given the elasticities' definition their variances (var), basedon which the t-valucs are calculated, are: , var(ei)-(l /w )2var(pi);

)2 PO; )=(I/W, )2var(yii)+(Wi]3/W, var(pi)- 2[wj13/(wj)71covar(yjj, Var(e,, )2 I/W, )2 var(yU)+(Wjll/W, Pj). var(pi) - 2[wjB/(wj)2]covar(yU, var((q)=(

171

Table 5.4: Expenditure price elasticities(restrictedmodel) anduncompensated

I

Expendit ure elasticities

First Period Second Period 0.947 (10.10)

Own-price elasticities
First Period -2.237 (-5.46) Second Period -1.797 (-6.87)

C ss-pricc elasticiti es PP
First Period x 0.097 (2.56) 0.033 (0,41) Second Period x 0.124 (2.74) 0.017 (0.25)

i's
First Pcriod 1.344 (2.81) x 1.376 (5.46)

IT
Second First Period Period 0.830 (2.77) x 1.077 (5.30) 0.073 (0.16) 0.517 (4.69) x Second Period 0.019 (0.06) 0.658 (4.96) x

WP

0.820 (4.51)

WS

1.150 -1.817 -1.933 1.203 (25.23) (26.97) (-12.94) (-11.74) 0.630 (7.25) 0.808 (24.52) -2.039 -1.901 (-9.25) (-10.44)

WF

5.4.1.2.Expenditure elasticities between The expenditure elasticitiesof the UK demandfor tourism differ considerably the ncighbouringcountries;thosefor Portugaland Francearc below unity while that for Spain for is France less is aboveunity. The expenditure to variationsin UK tourism share responsive for Spain Portugal. The demand those than the of or expenditureelasticity of expenditure Franceis of relatively small magnitude,so that Francebencrits (loses) less from increases in UK total tourism expcnditurcthan Spainor Portugal.The expenditureelasticity (decreases) for Portugal,although inelastic, is close to unity. Spain is, however, an expenditureelastic destinationfor UK tourists. The values of the expenditureelasticities indicate Franceas a 4secondary'
(0"'CF"I) destination

while Spaincan be viewedas a 'primary' destination(cs>1).

An interestingaspectis the differencein the magnitudes of the expenditureelasticities for Spainand Francebetweenthe first and the secondperiods.The responsiveness of the UK in real expenditure for Spainto changes demand to decrease seems while that of Franceseems to increasefrom the first to the secondperiod. This implies that, for the most rccent two (the secondperiod), Franceand Spain have been moving in opposite directions in decades 'primary' 'secondary' destinations. This view can be further to their as and roles relation by the differencesin the uncompensated own- and cross-priceelasticitiesobserved supported for in the first andsecond periods thesetwo countries.

172

5.4.1.3.Own-price elasticities
The uncompensatedown-price elasticities are all negative, as expected from normal increases for demand in prices. In tile cases of to negatively responds which commodities Portugal and France, the value of the own-price elasticity decreasesfrom the first period to the increase is Spain in the an of observed. The absolute magnitude of case while second period the decreasefor Portugal is greater (-0.44) than that for France (-0.14) or tile i ncrcasc for Spain (0.12). IIcncc, the impact of the changes in prices on the UK tourism demand for these three destinations varies not only across countries but also between tile two periods considered. In the period 1969-1979,the UK demand for tourism is more responsive to changes in Portuguese in France induces demand (a Portugal decrease 1% French prices of change or prices a or and in in Spain. I lowcvcr, 19802%, to than the than price changes ceteris paribus) period more 1997, the UK demand for tourism is more responsive to price variations in Spain than in France or Portugal. The second period values of the own-price elasticities also indicate that in Spain have in France impact UK demand for and prices a effective greater on changes in from Portugal. For all three countries in both than changes tourism equivalent would result be from increases in UK budget that the tourism the could gained returns positive periods, increased the prices. would not compensate adversecffects of Comparison of the magnitudes of the expenditure elasticities in the first and second for France increasing Portugal decreasing for Spain. On values and and a shows value periods the other hand, the own-price elasticities indicate a decreasing responsivenessof UK demand towards changes in prices of France and Portugal and an increasing sensitivity of UK demand to price variations in Spain. Hence, the estimates of expenditure and own-pricc elasticities lose for Spain France to to tendency ground and Portugal. suggesta

S.4.1.4.Cross-priceelasticities Substitutabilityand complementary among destinationsare indicated by positive and elasticities,respectively.Clear conclusionsabout the complcmcntarilyor negativecross-price destinations are not usually obtainedin studiesusing the AIDS model substitutabilityamong which haveproducedfew well-defined cross-priceeffects. However,the results in this study

173

I Icncc, they arc taken as an with a priori expectations. seemconsistentand also coincide directions in demand. indicationof the relativemagnitude the of and changes In both the first and secondperiods,all cross-priccelasticitiesare positive, indicating in However, France (Portugal) destinations. the the since price of substitutability among for in Portugal for Portugal (France) the the are coefficients expenditure equation and equation between Portugal France from different the cross-pricc elasticities and zero, not significantly is insignificance be Indeed, these the to of cross-priccelasticities statistically null. areexpected in France Table 5.4.1 lencc, Portugal Spain by t-values the and and and respective confirmed does demand for UK Portugal (France) the Spainare substitutedestinations, not reactto white (Portugal). France in the cfTective price of changes The cross-price elasticitiesfor the equationsfor Franceand Spainshow that the share
in is in Spain Spain is to than that France to changes of price changes more sensitive price of France. However, this sensitivity alters from the first to the second period, showing a decrease in the rsponsivenessof UK dcmand for Francc to price variations in Spain and an increase in the responsivenessof the UK demand for Spain to price variations in France. If Spanish prices increaseby 1%, the shareof France increasesby 1.38% in the first period and by only 1.08% in in increase by demand in if 1% France, for Spain will tourism the the second, while prices increase by 0.52% in the first period and by 0.66% in the second. This is consistent with Spain's loss of ground relative to France, mentioned above. The results also indicate Spain and Portugal as substitutes, although price changes in

Portugalhavea minor effect in the UK demandfor Spaincomparedwith tile cfTccton tile UK demandfor Portugal causedby price changesin Spain. This is not surprising given the differencein the sizesof the tourism marketsof both countries.Yet, the situation again altcrs by 1%, from the first to the secondperiod. The resultsposit that if prices in Portugalincrease by 0.10% in the first period and by 0.12% in the the demandfor tourism in Spainwill increase low demand UK for Spain These the to price small effects show sensitivity of very second. if in increase by However, Spain UK demand for 1%, in Portugal Portugal. the prices changes increases by 1.35%in the first period and 0.83% in the second.I Icncc, althoughtile sensitivity in for demand Portugal Spainremainsgreaterthan that for Spainto UK to the changes price of former diminishes in the two last decades.In contrast, tile in Portugal, the price changes for demand Spain to price changesin Portugalincreases. UK This provides the sensitivityof 174

further evidenceof the increased sensitivityof the UK demandfor Spainto price variationsin its neighbours.
5.4.2. COMPARABILITY OF THE RESULTS WITH Tliosr, FROM OTHERSTUDIE'S

Studiesthat haveappliedthe AIDS approach to tile analysisof tourismdemandarc few in number.Among them, White (1982) and O"Ilagan and Ilarrison (1984) modelled the US demand for tourism in Europe between 1964 and 1981, Syriopoulos and Sinclair (1993) investigated US and Western European tourism expenditure allocation to several (1999) studied the Mediterraneancountries in the period 1960-1987,and Papathcodorou for internationaltourism in the Mediterranean demand regionduring the pcriod 1957-1990. The elasticities estimatesprovided in this study arc only comparablewith those (S&S) Papathcodorou by Syriopoulos Sinclair in (PTI I) the casesof Spain and and obtained for fact Ilowcvcr, Portugal. this the that thcsc studics cxamine comparison must allow and different sampleperiods, different data sourcesand different sets of independentvariables. table 5.4 shows the expenditureelasticities and the uncompensatcd own- and cross-price in the studies. obtained elasticities Table 5.4. Comparison of elasticitiescstimatcsof UK tourismdcniandin Portugaland Spain
I
UncompensatedOwn-price Elasticities Expenditure Elasticities I

Uncompcnsatcd Cross-price Elasticities S&S PP _PS

3.75 1 0.88

S&S PTII -

Model (5 13) .

S&S

PTII

Model (5 13) .

PTII PP i's
0.98

Model (5.13) Pl, I's

1.34*,0.83 0.10-,0.12

m WP 1.58 0.04 0.82 ; 0.95 -2.81 T WS 0.90 1.15 1.20; 1.15 -1.11 -mmmomm - -

2.24 -2.85 . -1.80 1.30 1.82 -1.93 0.65

A noticeabledifference betweentile results concernstile expenditureelasticity for Portugal. While our resultsposit the expenditure elasticity of the UK demandfor Portugalto be inelasticbut close to unity, Syriopoulosand Sinclair's (1993) results indicate Portugalas an

175

it (1999) being Papatheodorou destination statistically shows as not and expenditure elastic be is inelastic in Syriopoulos for found Spain to different from zero.The expenditure elasticity it in Papatheodorou this Sinclair's the study qualify as elasticwith and results study,while and PTI I for by Portugal S&S While the -arc and own-price elasticitiesestimates similar values. in both flrst The lower the and second same periods. estimate show a results our similar, for Spain. The for be the uncompensated cross-pricc elasticity own-price can made comment demand different. UK The in the response of clasticities' estimates the threestudiesare rather in Spainis highly elastic(3.75) by S&S, while PTII finds that the for Portugalto price changes demand indicate UK for Portugal Our is inelastic (0.88). the to the response of results rcsponsc inelastic first in The in be in Spain the the to second. period, and response elastic price changes inelastic in is Portugal S&S (0.65) for demand Spain to UK to according price changes of the in but inelastic The this (0.88) PTII study also show obtained an much results studies. and 0.12). in both (0.10 Icr and periods smal response between The differences the elasticitiesestimates obtainedby thesethreestudiescan be for derivation in by the the the model elasticities further explained unrestricted use of, study and this one use the niodcl Syriopoulos and Sinclair's study, while Papatheodorou's inclusion in break for The the homogeneity same purpose. of a structural and symmetry under in (5.13) diffcrenccs. the also may explain some of model variable the real expenditure 5.5. CONCLUSION in other This chapterapplies the static AIDS model in two contexts not addressed first, demand: have tourism the that models of of equations model system estimated studies from features destinations, transition two to of which experienced a was applied neighbouring beginning higher developing tile the towards of sample countriesat period, characteristicof the model specificationallows levels of incomeand welfareby the end of the 1990's.Second, in demand behaviour destination in tourism for comparison each over time, in terms of changes The cross-price elasticities. ownand estimationresultsshow tile the valuesof expenditure, of indicated diagnostic by both data-consistent be tile tests to statisticallyand robust, as model in Moreover, have fit to the contrast many studieswhich of statistics. estimated goodness and homogeneity the the results are consistent with properties of models, of equation and system

176

This AIDS increases the the the accords with microfoundations of approach symmetry. and crcdibility of the elasticityvalues. The elasticities estimates obtained from the model provide interesting insights in demand behaviour lower income the two tourism the of countries,both relative concerning to each other and to their richer neighbour.Although the expenditureelasticity for Spain is in in first, lower for than Franceis higher, tile the the that second period while marginally demandfor Spain is expenditureelastic and that for Franceis expenditureinelastic in both developed "highly infrastructure" Therefore, diversified "offer tourism the a and of a pcriods. different (Papathcodorou, 1999) do be tourists" to to of satisfy groups tile seem not product inelastic demand for justification in to an elastic or response changes real expenditure, main behind hardly be lagging in Spain France Rather these tile proximity seen can as respects. since inducing be UK, in different tourism France to to the products preferably consumed a way, of for differences in be the to acceptable a more reason expenditure elasticitiesobservedin seems UK The France (which has been 'increased'with tile to the. two countries. proximity of these Channel Tunnel opening in 1994) seemsto induce a shorter averagelength of stay of UK Spain. The touristswho visit Spainappearto be willing to spend to that than assigned tourists in A length'of in in UK the Spain this time tourists glance average country. stay of and more France (see chapter 2) confirm France as a shorter-term destination than Spain. These differences, rather than a highly developedtourism infrastructure,may better explain the demand for Spain (more inelastic tourism (more tile volatile) compared with stable) elastic for France. demand tourism
in both periods, the expenditure responseof the UK demand for Spain is elastic, while

is inelastic for is Portugal for France that and close to unity. Theseresults indicate that that Spain has benefited from increasesin the expenditurebudget relative to France. Hence, in the tourism expenditurebudgetcan assistpoor countriesto 'catch-up' with their increases Portugal did However, benefit from Spain increases not as significantly counterparts. as richer different patterns in the UK tourismbudget,so that the two Iberiancountrieshaveexperienced findings These behaviour time. are clearly relevant to policy-makers who are over of demand in tourism For example, the to of with responsiveness changes expenditure. concerned in the expenditure budget,along with Spain's Portugal'srelatively stablegainsfrom increases

177

by declining share, corroborated -the negativesign of the trend coefficient, merit apparently further attentionby the relevanttourismauthorities. The estimatesof the uncompensated own-price elasticities for the first period show Spain to be a lessprice elasticdestinationthan Francebut in the secondperiod they arc more demand indicates This tourism that the of sensitivity price can vary over time. similar. in Francethan in Spain,they arc Although UK touristsremainmore sensitiveto price changes in Spain, implying that Spain may experience becoming increasinglyawareof price changes if France The UK demand instability this trend tile to continues. sensitivity of of relative more in but its to demand for Spain is increasingin response not only changes own-pricc also to its France Portugal, by in the valuesof the the and as of competitors, shown prices changes in first In UK the the tile periods. second contrast, of and sensitivity cross-priceelasticities in Spaindemonstrates to diminish demandfor FranceandPortugalto price changes a tendency betweenthe first andsecond periods. Portugalis a small country, in terms of tourism, comparedwith'its ncighbours.This
fact, combined with the political and historical events that contributed to its lagging behind its in development late Portugal's terms, European may explain awarenessof countries other based World Tourism data However, Organisation (1999) calculations on potential. tourism in Portugal has increased faster (except for France 1997, 1985 between tourism that and show in terms of arrivals and for the Netherlands in terms of receipts) than any other western or fact, in European In increased by Portugal 60% tourists country. arrivals continental southern between 1987 and 1996, only surpassedby France with an increase of 69%, and Portugal's between increased 275% 1985 1997, by Netherlands the and only surpassed receipts tourism in international 288% Furthermore, Portugal's the same period. of share of with an increase 1990, 1.5% has since classifying the country as one of the twenty exceeded tourist arrivals in 1996). in (WTO, Nevertheless, the terms world countries of receipts, the picture visited most is less favourable as Portugal drops to a position below 24 among the most important tourism destinations. In terms of tourism arrivals in Portugal, Spain is tile most important origin, followed by the UK, while in terms of receipts, the UK is the most important origin (Instituto Nacional de Estatistica, 1975-1997). Spanish tourists in Portugal are mainly short-term visitors but in huge little. UK display tourists the the opposite numbers spend very country enter who

178

behaviour,enteringthe country in much smaller numbersbut spendingmore money during their longcr-term. visits. Although Portugalundertooka major effort of modcrnisationin an attemptto adjust to its EC partners' developmentlevels, this effort, in tourism terms, has only been noticeable but during the 1990s.For most of the sample period the data rcflcct an underdeveloped between Spain, but the sea and enormous with squeezed wonderful sites country, picturesque but beaches lovely no properroads,tasty, varied and cheap clean no properaccommodations, food but no proper service,with museumsthat few visit, art that few see,history that few know. There arc, of course,places in Portugal with the cosmopolitanenvironmentseen in be but for destination to tend too they tourists countries, crowded, noisy chaotic and similar looking for a minimum of quality to accompany rising serviceprices. Thesefeaturesarc not be in for the the statistical to may an econometric model and one of capture reasons easy for Portugal, difficult the the to rendering equation elasticities of estimates more performance interpret. However,there is sufficient evidenceto indicate the changesthat Portugal has been instance, in For UK terms. tourism the the although expenditure elasticity of cxpcricncing from 0.82 in the demandfor Portugalis close to unity in both periods,its estimateincreases in 0.95 the second. If additional observationscould provide significant to f-Irst period increase in UK this tendency, that this the tourism would of growing mean an confirmation by a larger share for Portugal in the budget allocatedto the region would be accompanied indicating UK for Portugal a ceteris paribus, growing preference and assisting period second its 'catch-up' The in both to with richer neighbours. own-pricc elasticity periods the country for diminishing UK demand Portugal the sensitivity of a with respectto ownportrays clearly from first fact, In drops the to the this to a second period, elasticity estimate price changes. in France, Spain increase its below that while of shows an second period own-pricc value even clasticity estimate. The cross-priceelasticities provide useful information about the interrelationships The cross-price destinations. Portugaland Franceindicatethe lack between elasticitiesbetween for demand UK Portugal (France) in France (Portugal) to the price of changes sensitivity of both different for from The these elasticities are not significantly periods, zero. caseis since betweenSpainand Portugal.For both pcriods different for the cross-price elasticitiesestimates

179

for in demand UK in both tourism the one country responds significantly to countries, and UK demand for is Spain in However, the the expected, as only marginally other. changes price demand is UK for in its Portugal by the while more neighbour, changes smaller price affected in first demand for in Since UK Spain, the the to particularly period. changes responsive price Portugal respondsless intensively to price changesin Spain in the second period relative to the first, we can infer that the UK demand for Portugal may have achieved more stability along in in last Spain decades. independence from the two changes price with greater Political changesare likely to have an adverseeffect on tourism demand as indicated by

for dummy D For Portugal Spain. these two the variable cocfficient of and the negative in their averagesharesin the period destinations,the dummy coefficientsindicatea decrease in for France this the the 1974-1981, of variable equation showsan coefficient positive while The in the increasein its average estimated same period. coefficicnt for the trendvariable share UK for increasing France, Spain's tourism expenditure of average share while an shows also loss further decrease. Spain This the to tends already stated confirms of ground of share indicates in UK France that tourists' to and preferences, countries cannot rely on relative d'traditional' based from tourism tastes. stability on gains continuing This applicationof the AIDS model has provided new information about the long-run behaviour of the UK tourism demandfor France Spain and Portugal. For example, it has increases income benefit from lower in the that countries can although expenditure shown they will not do so automatically.Portugaland Spain, for budget relative to their neighbours, increases from in UK failed budget. Exploration to the tourism similar gains achieve example, income lower disproportionate from countries some why experience the gains rises reasons of is a topic for investigation,as is the reasonwhy countriescan fail to in tourism expenditure has The for their that advantageous position. study also shown competition maintain international tourism demand,rather than complementarity,charactcrisedthe ncighbouring Investigation degree the to which competitiveness of consideration. under riscs or countries is a further topic for research. falls asothercountriesattainmoresimilar levelsof development Nevertheless, the AIDS modelling approachadoptedin this chapterdoesnot take into features demand behaviour dynamic tourism the of and, consequently, cannot consideration for which underliesthe adjustmentprocessof the account the short-runcorrectionmechanism Although, in levels instances, focus interest their to path. the equilibrium many of of variables' 180

is long-run between dependent to the structural uncover relationships research and cconomic independent variables,the short-rundynamicsunderlying the steady-state equilibrium model in importance, have the cases where short-run particularly effects relevant also of may arc incorporation dynamics Therefore, the the of short-run within the AIDS system magnitudes. in is logical investigation demand this tourism the comprehensive next step of approach behaviour.The next chapterconcernsthe specificationand estimation of a dynamic AIDS in Spain Portugal. for demand France, UK tourism the and model

181

CHAPTER 6
DYNAMIC SYSTEM OF EQUATIONS IN TOURISM DEMAND MODELLING

6.1. INTRODUCTION One of the interesting features of economic research on consumcr dcmand behaviour is the close relationship between theoretical specifications and appropriate Given that techniques. the some notion estimation Of and consumer'soptimising rnodclling to underlie the data generatingprocess,effective evaluationof the behaviour is assumed from both judgement a theoretical and an empirical model's specification requires developed Many in the standard of now econometric procedures available were standpoint. in interpreting demand has to theory problems practical obscrvqtions utility and response for formulation data Although terminology a model and and a structure analysis. provided both theoreticallyandempirically satisfactoryis never of modelswhich are the construction by is the theory tool played economic role as a of applied economics straightforward, demand based Consumer between the theory and analysis on articulation essential. important has been introducing the theoretical core evidence of advances, new cmpirical for behavioural features is This the perspectives of economic agents. spccially modelling dynamics in integration demand theoretically the of consistent the caseof models. Economic theory of consumer behaviour provides sophisticatedmathematical dcscriptions of equilibrium relationships between demand and its dcten-ninants and, in is the steady-state changes with typically, concerned equilibrium causedby changesin However,most of the theory underlying demandsystemsdoes the demanddeterminants. by dynamic long-run the process which clearly specify equilibrium valuesarc attained. not Yet, actualdemandbehaviourover time must reflect the imperfectionsof the adjustments in their demand determinants. that consumers entertainwhenconfrontedwith changes

182

In chapter5, a tourism demand steadystatestructurewas estimated using the AIDS both from the were an econometric results satisfactory and andan economicpoint approach it is in budget However, that the current changes possible sharesof the UK of view. in demandfor tourism depend current changes on only not pricesand expcnditure,but also in disequilibrium the the variables' values extent of previous pcriods. Consumer on have been may unstableover the sampleperiod or the parameters of the utility preferences function may haveshifted over time. Evidencethat this might be the casewas providedby the estimationresultsof the AIDS model in chapter5. Therefore,an explicitly dynamic be a system should of equations consideredin accountingfor the specification within long-run towards the equilibrium. process adjustment The temporal aspectof consumerbehaviour has been largely ignored in studies
demand in Early tourism a system of equations context. rcscarch cfforts concerning in hypothesis to test the of utility theory. However, concentrate on static systems order be legitimate approachesunder special of consumer optimisation may these specifications Indeed, demand in aggregate terms circumstances only. models explaining solely restrictive likely factors to time-changing and expenditure prices are omit relevant current which of included, it dill"Icult be the to identify tile separate with variables making correlated may inclusion Yet, help for trend the to of a variable may correct some of the oinittcd cffects'. factors and thereby make results more theoretically consistent. In these special dynamics in demand be the specification of models can simple to formulate. circumstances, I lowcvcr, this simplicity is often the result of particularly strong assumptions, for example, intcrtemporal separability, unchanged preferences and invariant budget constraints over time. As stressedby Blundell (1988 p. 39), "the intertemporal separability assumption on is these models rest precisely that which rules out explicit dynamics". which The AIDS model discussed in chapter 5, gives a consistent temporal interpretation long-run despite the underlying assumption of intcrtcmpoml the relationships present, of Within identified this specification, static we a trend-like behaviour in the UK separability.

in pricesor expenditure, demandfor tourism,not associated with changes which suggested in consumer have occurredover the sampleperiod. To allow for that changes preferences 'absorbs'the effectsof time. this possibilitywe includeda trendvariablewhich supposedly dependent We also shifting factorsandmay be interpreted as a trend changein preferences.
' In a tourismdemand contextthesefactorsmay be identified with, for example,reductionsin the working to paid leaveand increased wcck,increased entitlement numbers of retiredor early-rctircdpcople

183

imposed for break by the on expenditure variable, some means of a structural allowed, Hence, be 'quasi-' budget in these the models can qualiricd as or constraint. variability 6sccmingly-dynamic' models. Still, under more general dynamic bchavioural features, the AIDS model needsto include (explicit) dynamic elements in its demand equations. If the intertemporal separability assumption is put aside, consumers adjust their

in its determinants, demandin response to intertemporal changes and an explicitly dynamic it is fact, behaviour for. is In to that consider realistic past allowed adjustmentprocess implies behaviour. Habit that affect current changes preferencesand, consequently, influence in by influenced functions turn, previouspurchases which, are consumerutility in habits Since the the unobservable, usually are associated changes purchases. present by lagged demand, prices and expenditure demand functions are usually represented learrit ) be little be from (op. Blundell However, there to cit. points out, may as variables. The in lagged to static an otherwise model. resulting specification variables adding simply implausible intelligible be bchavioural hypotheses. Unless under only may this case, implications the theoretical and specified appropriately of are general models empirical inference invalid integrated, fully types statistical may rule out plausible of are principles in lessusefuldirections. to proceed behaviour,inducingresearch behaviourhaverccognised in demand Many researchers the importance of including in demand have been dynamic systems and a number adjustments of approaches explicit implement Blundell (1983,1984) Anderson flexible For and a example, general adopted. dynamic approachwhich incorporatesa long-run solution within a system of errorfor UK demand to consumer severalgroupsof commodities, analyse correctionequations, dynamic (1996) Young Burton to use a structure model a systemof equations and and faced describes BSE the the process of meat consumers adjustment with problem. which demand dynamics tourism Unfortunately,specific research using systemsof equations, on is not abundant these and, to the best of our knowledge,only Lyssiotou (1999) addresses the dynamicsof tourismdemandbehaviour,using a issuesin an empiricalstudyconcerning AIDS to the similar model. systemof equations The objectivein this chapteris to contributean empirical study of tourism demand dynamicsand point out areaswhere the scrutiny of relationshipsbetweentheoreticaland likely insights in to Ilic this area of research. are produce new considerations empirical follows. is Section 6.2 the as provides several alternativeempirical of chapter structure from dynamic derived AIDS a general model. Section 6.3 implements spccifications 184

for bctwecn hypotheses theoretical the modelsand tests seeking consistency of statistical
the principles of consumer demand theory. Section 6.4, presents the estimation results of theoretically consistent models, which are analysed and compared with the results of the

in chapter5. Section6.5 concludes. AIDS modelestimated 6.2. DYNAMIC AIDS MODELLING OF THE UK DEMAND FOR TOURISM In a purely static demandsystem,consumers are assumcdto adjust perfectly and instantaneouslyto prices and income changes.However, habit persistence,unstable information imperfect leading incorrect to costs, or adjustment expectations, preferences, fully from is If to tlic casc, this equilibrium adjusting every consumers period. may prevent dynamic behaviour is This structure a of consumer specification of general required. more a in be for that testing more restrictive a way can modelled allows general specification itself, hypothesis dynamic forms the and static altemativc as such ricstedwithin the models
basic general model. Once a data-coherentstructure is defined, the restrictions implied by hypotheses be tested. can utility maximisation

In what follows, a flexible dynamicstructurefor the AIDS model is derived,based (1983,1984). Blundell Anderson by As these authors, the and pointed of out work on flexible itself dynamic the the general structure arc static model within and nested distributed lag the the such adjustment as partial and auto-rcgressivc specifications, (ARDL) models,which canbe testedagainstthe more general alternative. in equation(5.12). For simplicity and notation Considerthe AIDS model described Wi, In Pi i, j-1, the variables as wi = rename = we pj where compatibility, ln(x/P*) = E. Hence,equation(5.12)is written: Wit = (xi* + 7ijPj,+ PiE, n and

(6.1)

Consider equation (6.1) as the appropriatechoice for the steady state structure of the following generaldynamicstochastic specification:
nn

(ai + PiAEt +), ysijAPjt i

AWit

j-1

F, pL + + yuPjt-, jE t_I-Wil_, j-1

)+Uit

(6.2)

is ih Xi is the A the the first differcncc opcrator, coefficient of adjustment equation, where lagged indicates the t-I valuesof the variables,and ui is the ith disturbance term subscript by independent be to a singular, characterised and identicaldistribution ovcr time. assumed 185

be interpreted, L S The parameters can and respectively,as the short-run with superscript
in its dependent detenninants. long-run to the variable changes responsesof and

The dynamic specificationin (6.2) assumes that, in order to maintain the steady in (6.1), adjust the current values of their expenditure consumers given relationship state in in in the to changes explanatory response current variables and partly partly shares in desequilibrium in Young As Burton the to the previous observed period. and response (1996), we constrainspecification(6.2) to the use of an identical adjustmentcocfricicnt
i. Given for ki=% be the that, system's all singularity, such estimation may equations across delectation by the of the nthequation. arbitrary out carried

Although the dynamicspecification(6.2) is a simple first order laggcdstructure,the

data be for in loss too particular general any generating still process, resulting a may model

(1997) in Gujarati inference 'shaky' classirics as a or what statistical precision of estimation be find Hence, tests to the most restrictive may performed a of sequence procedure. budget dynamic specificationconsistentwith a particular set of T observations on shares, Examples below. of such capita real expenditure. spccifications per and arc provided prices Considerthe following equivalentform of the generalequation(6.2): Pnt-I +PisE,-PsE, + Plt-I +... +ysiPnt-yis,, AWit = Yilplt -7ili', n -,
+ %cci

Lplt_l 7il % +,

L+ % El-I -, %Wit-,+ ui, +, YinPnt-I XpL i

(6.3)

(A distributed lag RDL) model uto-regressive In the spirit of a "general to specific" approach we postulate the long-run between Y two X, relationship economic variables, say and suchthat: equilibrium
Yt POX,+PIXI-l +... +P, Xt-, +B, Y, +82YI-2+***+8myt-M --' -l +Ulf

is These ARDL behaviour the of order model m. models explicitly an consider of a which in dynamic However, time so are and nature. such models are too general over variable This form. the m remains unspecified. value of general may be rcduced to a since by for (see, I lendry Richard, 1983) several criteria applying one example, and parsimonious include definition The form (6.3) be the small of m as a number. modcl's general can which described, form hypothesis: ARDL to the under as an null rcduced

Pi' I lo:ka, =0 r) yi'j= yiLj r) li li

pL =

forallij i

186

2 to: If Hois not rejectedthe modelreduces %- I)P,El-l Wit -` 7jjPIt + Q-- 1)7iPt +-+ YinPnt + (%- I)yi, Pnl + PjEt + (,

%)Wi, + ui, -,
ARDL is first model. order which a
partial adjustment model

that the Considerthe flexible accelerator model of economictheory which assumes

is function Y*I, linear dependent level long of an variable, a say a of run or equilibrium, Xt, that such say variable, explanatory

Y* t =po+plxt+Ut
hypothesis nc partial adjustment postulates SY *t +(I - 8)Yt-l Substituting(i) in (ii) we have yt = 8po+ 8pixt +Q- 8)yt-l to be testedis Consideringmodel(6.3), the null hypothesis

(i) (ii)

110:

XpLfor all i,

to: is not rejcctedthe modelreduces If 110 LpIt Awi =%7il t Lplt_l X7L %PE, Pnt-, in il +. +... +,%7Pn, I in _ %Wi, -, +ui,

Lpj, %7 %ai i +%pLE, +... +%7Pn, + , +, ii in -1 -, -, _,

or

yilLpIt +... +XyiL,, Pn, % wi t =,%tXi i +XpLE, +(I-X)Wi, +, in

-1

+ ui,

in described (iii). is to the similar one equation model adjustment a partial which Static AIDS model is To testfor the staticmodelnested within (6.3) the null hypothesis

f lo:
(6.3) to If I Iois not rejected, reduces model

2 The hypothesis (%a, testedin this case,includesthe restrictionof null intercepts -0). I lowever,in a tourism important has intercept Therefore, demand to be tested the an economicmeaning. anotherhypothesis context, in 110. In this case.the ARDL for the ARDL model,would be not to includethe restrictionof null intercepts In chapter 4. of the null would be similar to thoseanalysed resultingfrom the non-rcjcction equations

187

Wi, =kai+yi, Pl, +... +7i,,Pn,+PiEt +u, is AIDS the model. state orthodox steady which 63. TESTING THE THEORETICAL CONSISTENCY OF ALTERNATIVE

DYNAMIC MODELS Dynamic generalisations of traditional static systemsarc an important feature of in The for areas. such gcncralisations motivation severalresearch recent empirical work largely derivesfrom the continuinglack of accordbetweenthe postulates of demandtheory demand functions Generally, do time using aggregate series. consumers static empirical and in demand immediately determinants. Hence, to their changes and perfectly adjust not dynamics is the short-run of consumers' shares expenditure appropriate modelling of hypotheses. is This because before testing the resultsof utility maximisation so essential dynamic is depend irso, tests specification a considered on whether may and, whether such it is the appropriateone. In fact, many studiesusing the orthodox AIDS approach,for (1980a), Anderson Blundell (1983) Syriopoulos Deaton Muellbauer and and and example, dynamic is (1993), Sinclair that the a misspecification. suggested probablecauseof and With theory the postulates. most economicmodels,and particularly with utility rejectionof the analysisis centredon the set of hypotheses demandsystems, relating to the long-run is independent from This dynamics rittcd to analysis short-run any coefficients. structural by Chambers (1993, 728), "the dynamic data. However, as pointed out p. structure may the to the long-run solution, and the incorporation itselfbe affectedby the restrictionsimposed dynamics h)pothesis in turn, the inappropriate testsconducted may, affect outcome of oil of the long-run parameters"., Therefore, adequatedynamic spccificationsare essential in before behavioural hypothesis demand be plausible systems, can appropriately aggregate testedagainst expenditure patterns. observed Followirfgthis line of reasoning the methodologicalstrategyin this sectionconsists
in dynamic define, first, flexible terms for the to the generic most possible, a structure of: UK tourism consumers; then, to test for more restrictive of adjustment process expenditure be believed to consistent with the sample observations: finally, to test tile specifications by on specifications maximisation restrictions not rejected tile data. utility

188

is dynamic basic 6.2. in which (6.3) structure a general Equation represents section -1 ARDL, the static and derive adjustment partial of the specifications to restrictive more used UK tourism the these In the with models of compatibility AIDS models. this section, further data, they the If subjected is data arc demand tested. the modelsarecompatiblewith The hypotheses. these these of tested non-rejcction under to utility theory constraintsand The theoretically indicates empirical subsequent the consistent. hypotheses models as data-compatible be both found to dynamic and focuses arc which specifications on analysis thcorctically-consistent. following dynamic the Hypothesis modelsprovided testingperformedon alternative
for the partial adjustment hypothesis against the more general Wald The test results. 2(g) below lies 14.51, the which = dynamic specification presents a statistic value of X level. Therefore, 5% the 16.92, the partial significance at of value critical corresponding further data.. However, by the with constrained when rejected was not model adjustment this homogeneity the was model statistical performance of symmetry, and of the restrictions hypothesis AIDS the The general against was rejccted static orthodox not satisfactory. hypothesis 19.03. The ARDL Wald with null the dynamic model, with statistic value of Wald dynamic the the with general model intercepts restriction, was rejected against itself dynamic Only 1.3 87.7 the with structure reveals compatible general of value statistic demand Therefore, theory. the remainder of the data consumer of assumptions with and the

focuses this model. on this section follows: be in (6.3) AIDS dynamic The can rewrittenas modelpresented equation Yis pSt_l ),YiF L yip L %Wi %pLEt_l pF % pp + Ii1 + +, t +XL AWit = %(Xi +, _, _I _I _I (6.4) FAPFt PsAE, iF + + + + ui, ip y, + y, 7issAPS, pAPPt break indicates in 5 the Empiricalevidence structural presence a of obtained chapter dummy the variable significanceof a in the coefficientof the real expenditure variable and 1974-1981. Portugal, for Spain France, the this in over period and (D) the shareequations by dynamic (6.4) in integrated the consideringthe spccirication general information was in break the the the of relevance variable and potential expenditure structural of a presence break is dynamic for that the However, D. the we structural dummyvariable model, assume have dummy D in long-run, the the may a significant effect variable whereas only relevant
is testedwithout the restrictionof null Intercepts, the model is (marginally) Whcnthe ARDL specification is 15.60. However, in to the this one used of specification similar value statistic X2(9) not rejectedwith a interest in the empiricalanalysis the 4, of this model. of most chapter whichremoves

189

both in the long- and in the short-run.Investigationof whetherthis is the caseis undertaken by defining an unrestrictedgeneraldynamic model including the variable under test in by (6.4). Then, this the testing tile unrestricted significance of we variable assess equation (U) modelagainstthe restricted(R) model,which excludes the variable. The statistical significancc of the variabics SEj.1, Dt.1, and ADt is tcstcd for the Wald The in the tests' statistic. using results of equations, arc presented table 6.1 system (p values in brackets).In all tests, the explanatoryvariables in the unrestricted(U) and displayed. (R) restricted modelsare D1.1 ADt Table 6.1: Testsfor the individual andjoint significanceof variablesSEt. and 1, I HYPOTHESISUNDER TEST Non-significanceof SEt.1 110: Wit. PFt. Et. SEt-j APPt APSt APFt AEt PSt. (U): PPt. 1 j 1 1 1 Wit., Et. APPt APSt APFt AEt (R): PPt.,PSt.,PFt. j 1 Non-significanceof ADt 110: Et-1 Wit. PFt. APSt APFt PSt., APPt AEt ADt Dt. (U): PPt. j 1 j 1 Et. Wit-I PFt. APPt APSt APFt AEt PSt. (R): PPt. 1 j 1 j Non-significanceof Dt-1 110: Et. SEt-j Wit. APPt APSt APFt AEt ADt D1.1 (U): PPI.,PSt.,PFt. j 1 1 Et. PFt. Wit. APPt PSt. APSt APFt AEt (R): PPt. j 1 1 1 j Joint Non-significanceof DI-1and SEt., 110: Wiw PSw PFt. Et. SEt. APPt APSt APFt AEt Dt., (U): PPt. i t i i Et. PSt. PFt-j wit., APPt APFt APSt AEt (R): PPt. j 1 1 WALD STATISTIC

(0.12) X2(2)=4.24 Not rcjcctcd

X2 (2)=2.84(0.24)

Not rcjcctcd

)?(2)=6.94(0.03) Rcjcctcd

X2(4)-10.90(0.03) Rcjcctcd

The testsindicatethat the dummyvariableD is only relevantin the long-run,sincethe nonbe AD, Ilic nonrejectedand the non-significance, of cannot of Dt., is rejected. significance for SEt-1, the structural break, in the cxpcnditure accounting of variable significance is However, the rejectionof this variable's individual significanceis not rejected. variable, joint Moreover, Dt. the Sl., is significance of strong. very and not convincingly not 1

190

4 for AIDS data. For by the thesc the static and comparison purposcs with rcasons rcjcctcd bc includcd. belicve 5, this variablc'should modcl of chaptcr wc The assumptionof equal velocities of adjustmentfor all share equationsin tile
dynamic system was also tested. The null hypothesis for this test restricts the adjustment i. Xi for Wald The be Xi that to so all statistic value =X across equations, equal coefficients hypothesis 1% does (p this 0.267 the rejected even significancc not at value=0.605), X2(l) = 5 lcvcl.

Given the considerations above,the ilh equationof the generaldynamicAIDS model for the UK tourismdemandin FranceSpainandPortugalis:
AM, %Wi, + + + + ai4PF, aisFEt-I + + ai6SEI-I ai7D, -, = a,, ai2PPt-I a,PS,-, -, -, (6.5) -, + ai8APPt + aigAPSt + ai, OAPFt+ail lAEt + ui,

FEt., k=l,... 11) UK S, P (i=F, the parameters, are represents real cxpcnditure and where aik SEt. in first (1969-1979) in the this the and represents period second variable 1 variable 6 (1980-1997). period

The gencraldynamicsystem(6.5) is now testedagainstfurther constrainedmodels

homogeneity, include the symmetry and null cross-price cffccts of restrictions which between the equations for France and Portugal, both on the long- and short-run coefficicnts in 6.2. its Table first these tests presents the variables. results showing appropriate of in its hypothesis testing; the the second column, under corresponding models, null column, in its for hereafter, Wald third the test values and column, statistic are used which notations the hypothesesunder consideration.
instability,captured by the structuralbreakin the staticmodel,revealsitself In 4 it is possible that the demand it is possiblethat, within the dynamic For instance, different waysnow, giventhe dynamicstructureadopted. is by in less break instability the expcnditurc the when expressed a structural relevant observed specification, in different it is demand if Indeed, it be the that than way. possible observed a expressed would variable instability could be better describedby the inclusion of additional interceptdummies.This possibility is in 7. chapter considered 5 T*hisis not surprisings' ince there is no evidentreasonto believe that different coefficientsof adjustment destinations have fairly for Indeed, UK the tourists the considered. of equations share similar exist should implying that they adjustto changes determinants in their demand informationaboutthesedestinations, with in is integrated The coefflcients adjustment across equal all subsequent equations of speed. constraint similar dynamic derived from the structure. general models restricted 6 The structural breakin the model is includedby dividing the sampleperiod into the sametwo sub-pcriods 1980-1997) 5 (1969-1979; definedin chapter andusingtwo dummyvariables,F and S, which assume a value first in for the and second periodsrespectively, and zero otherwise.Thesetwo dummies observations unit of SEt. integrated (6.5). in Model (6.5) and are then multiplied by Et., giving rise to the new variablesFEt. 1 1 form including is E,. former SF,., Ile includingFF,. SE,., has of a equivalent model an and and variables. 1 1 informationon the coefficientsof variableE in the first and second the advantage of giving straightforward latter, in information for has be (respectively, the the to the whereas second a6), a5 and period obtained periods by summing thecocfficientsof Et., and SEt. 1.

191

Table 6.2: Testsfor utility theoryrestrictionson the long- and short-runcocfficients 7 TEST UNDER HYPOTHESIS MODEL NOTATION WALD STATISTIC

Long-run homogeneity 110:

(II)L

X2(3)-0.541(0.91) Not rejected (0.71) X2(4)-2.16 Not rejected (0.77) X2(5)-2.56 Not llcjcctcd (7)-8.72(0.27) x2 Not l1qJcctcd
X2(8)_

110:Long-run homogeneityand symmetry Long-run homogeneityand symmetry and 110: long-run null cross-pri ce eff ect s betweenthe share Portugal France and equations of 11o:Long-run homogeneity,symmetry and null homogeneity short-run and effects cross-price Long-run homogeneity,symmetry and null 110: homogeneity and short-run and effects cross-price symmetry Long-run homogeneity,symmetry and null 110: homo d h tff i t geneity, n s or ru cross-pr ce e ec s an symmetry and null cross-pricc effects

(II+S)L

(I I+S)OL (11+S)OL & (11)S (I I+S)OL & (11+S)S (I I+S)OLS

10.35 (0.24) Not Rejected

xl(9)-10.59(0.31) Not Rejected

is rejectedby the data. The resultsin table 6.2 show that none of the hypotheses by an Therefore,oncethe dynamicsof adjustmentto equilibrium are fully acknowledged
homogeneity longbe short-run and structure, and symmetry cannot rejected at econometric is for level. hypothesis longThis true 5% the also of and short-run null significance the

France Portugal. between the of equations and effects cross-price Theseresultshave important implications for the modelling and prediction of knowledge in behaviour. They that the suggest of way which consumers adapt consumer in its determinants behaviour demand to changes requiresmore than a static systemof their long-run structural relationships.They also indicate that, for obtaining comprehensive it may triggering the processof adjustment, informationon the error-correction mechanism
The constraint of equal adjustment coefficients across equations is integrated In all subsequent restricted dynamic Therefore, from in derived the structure. general all the hypotheses tested, this constraint models holds previously. Consequently, the number of degrees of freedom for the X2 statistic includes this first restriction.

192

factors in formulations. introduce be AIDS trend the to usual static simply sufficient not dynamicmodelsrepresenting Neitherdoesit appear spccirictheoriesof sufficient to choose for A the partial adjustment example, model. correction such as, more general short-run dynamicstructureseems to be requiredto matchdataandtheory in a consistent way. 6.4. EMPIRICAL RESULTS AND THEIR INTERPRETATION The general dynamic model in (6.5) is estimatedwith SUR method. Table 6.3 This tile table the shows estimates of tile coefricicnts estimation results. presents (asymptotict-valuesin brackets)for the shareequationsof Portugal(PT), Spain (SP) and France (FR) obtained with the unrestricted(6.5), (11+S)OL and (11+S)OLS models. Tile labelled "unrestricted" the sole constraintof equaladjustmentcoefficients considers model % ' *and The 1%, 5% 10% tile represent, symbols respectively, equations. and across indicators fit levels. Goodness (RSS), the such of as residual sum squares significance log-likclihood (SLL) (ELL) log-likelihood and system values,are alsopresented. equation The statisticalrobustness spccificationsis shownin table 6.4 of the shareequations (LM) F Lagrange Multiplier diagnostic for the the tests version and version of gives which form, functional heterosccdasticity, error normality and performedon an serial correlation, for basis Tile for the thesestatisticsarc unrestricted model. p values cquation-by-equation in diagnostic Table 6.4 in The brackets. test statistics show that tile equations of tile given However,taking dynamicAIDS systemarc well-defined,statisticallyrobustspecifications. included the the as an whole, explanatory system power of variables the unrestricted by the F test of overall significance)and the goodness by tile (measured of fit (measured W coefficient)do not presentbrilliant results,particularly for the share equationof adjusted Portugal.This should not be surprising,given the individual non-signiricancc of many of in (six twelve) this equation.A solution to this problem,as suggested of out the cocfficients by'thc 'general-to-specific'practice, would be to exclude the insignificant regressors. Indeed,the imposition of theoreticalconstraintsand null cross-pricceffects betweentile Portugal improves France individual the generally and of significance of the equations imposition Therefore, insignificant the of additional null restrictions on estimates. be further to the statisticalmeaningof the dynamic strengthen expected would coefficients

193

8 dynamics However, the this spirit of general which underliestile contradict would system. information is implementation think this would we conscquently, omit which modeland, of impose further zerorestrictions. do Therefore, importantto incorporate. we not dynamicmodel and modelsI and 11 Table 6.3: Estimationresultsfor the unrestricted
Unrestricted (I I+S)OL (11+S)OLS

PT INT pp"1 pst-I

SP

FR

PT

SP

FR

PT

SP

FR 0.4299 (7.49)o 0 (none) 0.4074 (4.79)o

0.0995 0.2330 0.4599 0.0695 0.2618 0.4265 0.0671 0.2821 (4.50)o (2.66)o (7.41)* (5.41)o (3.25)e (7.94)9 (5.93)o (4.07)o 0.0494 0.0472 0 0.0717 -0.0472 -0.0494 -0.0238 -0.0479 (-0.82) (-0.63) (0.97) (2.43)" (2.43)* (none) (-2.52)" (2.52)" 0.0899 -0.5036 0.4137 0.0494 -0.4486 0.3992 0.0472 .0.4545 (2.69)" (-5.73)9 (5.16)9 (2.43)" (-5.72)o (5.38)o (2.52)" (-5.06)o 0 0.5450 -0.4882 -0.0568 (-1.22) (4.08)o (-3.79)o (none) 0.3992 -0.3992 0 (5.38)o (-5.39)* (none)
0.0359

0.4074 -0.4074 (4.79)o (4.79)o

0.0476 -0.044

M- i SE. DI., Wit.1 W APPI 1 APSt APF, AEt E RSSS ELL

L SLL

-0.0163

(-1.47)

0.0644 0 (2.23)

-0.0481 ' (-1.81)

-0.0038

(-0.50)

(1.63)

-0.0322

(-1.60)

-0.0034

(-0.46)

(2.60)" (-2.7 1

0.0295 0.0018 -0.0256 -0.0039 (-0.82) (2.41)" (-2.30)" (0.67) 0.0324 152 -0.0145 "0.0 -0.0172 (2.24)* (-2.37)* (-2.34)" 0.15) 0.792 (7.45)o -0.1013 0.1010 (0.77) (-2.1 Of

0.0188 -0.0207 0.0026 0.0177 -0.0204 (1-81), (.2.13)" (1-0-4) (1.96)' (-2.38)* 0.0263 0.0118 o271 0.0407 o. -0.0136 . (-0.87) (2.17)* (-2.28)" (. 1.77)' (2.92)* 0.758 0.779 (7.94)o (7.78)o 0.0188 -0.0569 0.0569 (0.16) (. 1.55) (1.55) 0 (none)

0.0003 -0.0700 0.0512 (0-00) (-1.58) (0.39) 0.0416 (0.96)

0.1152 -0.1506 0.0354 (1.82)' (-0.92) (0.23)

0.0286 -0.0702 0.0569 -0-1761 0.1192 (0.23) (-0.62) (1.55) (. 1.80)' (1.33) 0.1500 -0.1576 0 (1.76)' (-2.04)' (none) 0.1192 .0,1192 (1.33) (-1.33) 0.0636 . 0.0656 (1.34) (. 1.52) 0.014 0.012

0.2557 0.0076 0.0261 -0.2296 . (-0.61) (2.25)0 (-2.13)* (0.22) 0.004 (0.19) 0.001 99.42

0.1057 -0.1017 -0.0069 0.1201 -0.1132 0.0020 (1.89), (-2.04)' (-0.33) (2.18)" (-2.3 1)" (.0.12) 0.009 72.65
174.24

0.008 74.82

0.001

0.010

0.008

0.002

98.13 1 71.82 1 74.40 172.54

97.95 1 66.62 I_ 69.00 _ 167.04

I We did estimatethe model under the full set of long- and short-run restrictions and additional zero Fj, in AEt for the Portugal. The overall the the of variables and coefficients on equation constraints j increases, long-run the the and all coefficientsbecomemore significant and have model of quality statistical the correct signs. However,the short-run coefficients do not alter much, either in their magnitudesor individualsignificance.

194

Table 6.4: Diagnostictestsfor the equations of the dynamicAIDS model I

DIAGNOSTIC TESTS

Equation

SerialCorrelation LM version F ersion 0.02

(0.88)

Form Functional LM version 0.00

(0.97)

Normality LM version 0.92

. (0.63)

I letcrosccdasticity LM version 1.77 (0.18) 0.07 (0.79) 1 0.02 (0.88) F version 1.76 (0.20) 0.06 (0.80) 0.02 (0.88) 1

F version 0.00
(0.98)

F version na na na

I PT SP FR .

0.05
(0.83)

0.40 (0.53) 0.11 (0.75)

0.04 (0.85) 0.06 (0.81)

4.63 (0.03) 2.50 ) (0.11 .

2.98 (0.11) 1.47 (0.24)

0.58 (0.75) 1.97 (0.37)

For comparisonpurposes with the AIDS model of chapter5, the interpretationof denoted for (11+S)OL, focuses the model which constraintsarc on the estimationresults imposedonly on the long-run coefficients.Although the model under the full set of long(11+S)OL more statistically robust, seems model restrictions presents short-run and interpret be longto to the a considered reliable means and sufficient statisticalquality in demand France, Spain Portugal. UK behaviour tourism the and of short-run
INTERPRETATION OF THE ELASTICITIES' 6.4.1. ESTIMATES

The interpretationof the long-run cocfficients of the dynamic AIDS model is for in 5, However, AIDS the the that to codflicients of static chapter given modcl. similar former by has in information the the cstimates provided no correspondence the short-run latter. Hence, dynamic from thorough the the a analysis of model cocfricicnts, obtained is longby This short-run estimates of and a comparison worthwhile. analysis accompanied basis,leaving comparisonacrossequationsto be is carriedout on an equation-by-equation dealt with latter,wheninterpretingthe elasticities'estimates. is for 0.76 imposed by tile The adjustment estimate all share velocity equations,as This UK demand tourism suggests estimate a rapid adjustment restriction. of corresponding in demand determinants long-run the the changes to occur. Indeed,76% equilibrium, when in is is 24% the to tile next that current period, and only attained adjustment postponed of idea information, This the of almost perfect corroborates quickly circulating among period. UK tourists,concerningaspects which may influence their decisionto visit France,Spain Portugal. or

195

%) is Since by constructionof the model, the velocity of adjustmentparameter(. long-run intercept by to tile the parameters of all share obtain equations, and multiplied have divide lagged long-run to the tile the we cocfficients of coefficients of actualestimates long-run As X (0.76). by the the result, actual of, say, the a estimates of estimate variables dummy variable (Dt.1) coefficients,in the equationsfor Portugal, Spain and Francearc, 0.034. and respectively, -0.019,-0.016 For In general,all coefficients' signs are consistentwith theoreticalexpectations. have the instance,all cocff*icicnts the variable,when statisticallysignificant, expenditure of in first In longin both the the the and short-run and second periods. and signs cxpected indicating for Spain, these all coefficientsare positive, an elasticresponse of shareequation in budget. Convcrscly, in for demand Spain tourism UK to the shareequation changes the for France,thesecoefficientsare all negative,indicating an inelastic response of the UK in the UK tourism budget.In the caseof Portugal,none of demandfor Franceto changes 9 is Moreover, both longthe significant. when statistically significant, these coefficients and short-runown-price coefficientsare negative,as expectedwith normal commoditics, from destinations the positive, as expected which are are coefficients cross-price and indicate The that than the results also complements. political and rather competitors had UK demand for 1974-1981 Spain the tourism a effect negative of of events economic for France. Overall, longPortugal the the effect magnitudes a positive and and signs of and dynamic from the the model accord with corresponding obtained estimates estimates run in for long-run 5. Therefore from the the chapter static model similar values obtained be expected. should estimates clasticities For all equations, the short-runcoefficientsarc, generally,statistically insignificant, inducedby short-runchanges This may indicatethat the effectson the UK tourismdemand, in its determinants, are not of relevant magnitude. Supporting this hypothesisare the dynamic despite its the that the static model and of of robustness model short-run statistical insignificance,the consistency of the long-run estimates providedby both models,and the in tourismbudgetandprices. to changes high adjustment velocity of UK demand A moredetailedanalysisof the resultsrequiresthe relevantelasticitiesvalueswhich long-run information from the the the AIDS modelof of obtained confirm similarities may

9 The non-significance of a given coefficient does not necessarily Imply the non-significance of tile formulae for its the calculation may Include other coefficients as well as tile as elasticity corresponding averageand/or the baseyear shares.

196

before, As dynamic (11+S)OL. 5 the we compute the expenditureand model chapter and dynamic the the coefficients' estimates of elasticities using price relevant uncompensated in formulae fication, S. 6.5 Table the values given chapter shares' and showsthese and speci in brackets, for denoted AIDS t the values model estimates and respective elasticities in this chapter. (I I+S)OL in Chapter5, andfor the dynamicAIDS modeldenoted (I 1+S)O Table 6A Expenditure own- andcross-price anduncompensated elasticities
Expenditure elasticities
First Period Second Period

Own-price elasticities
First Period Second Period

PP

Cross-pricc clasticitics i's

Second

First Second First Period Period Period

IT

Period

First Second Period Period

Static ( "+S)o 0
R T LONG-RUN Dynamic 0L S (I I+ )
LONG-RUN

0.820 0.947 -2.237 -1.797 (4.51) (10.16) (-5.46) (-6.87) 0.913 1.027 -2.128 . 1.729 (5.22) (26.22) (4.61) (-5.93) 0.911 (3.35) -1.895 (-3.40)

1.344 0.830 0.073 0.019 (2.81) (2.77) (0.16) (0.06) 1.176 0.714 0.039 . 0.011 (2.62) (2.39) (0.49) (. 0.69) I 0.138 0.573 (1.01) (0.28) 1 X X 0.517 0.658 (4.69) (4.96) 0.807 0.983 (5.40) (5.38) 0.171
(1.02)

U G A L

Dynamic L (I I+S)0
SIJORT.RUN

Static ("+S)o S
P LONG-RUN

1.203 1.150 -1.817 -1.933 0.097 0.124 (25.23) (26.97) (-12.94) (41.74) (2.56) (2.74) 1.076 1.047 -1.980 -2.150 0.097 0.120 (21.39) (36.27) (40.82) (. 10.12) (2.30) (2.41)
I

Dynamic 0L ) I+S (I
LONG-RUN

Dynamic (I I+S)OL

1.213
(12.38)

[
I F

SHORT-RUN

(4.56)

-1.048

0.071
(0.32)

Static ("+S)o
LONG-RUN

0.630 0.808 -2.039 -1.901 0.033 0.017 1.376 1.077 (7.25) (24.52) (-9.25) (40.44) (0.41) (0.25) (5.46) (5.30) 0.865 0.929 -2.608 -2.336 0.012 0.007 1.730 1.400 (9.65) (25.34) (-8.76) (-9.36) (1.50) (1.93) (4.98) (5.05)
I

R I A
N C F,

Dynamic )OL (11+S

LONG-RUN

Dynamic (11+S)OL
SIIORT-RUN ,M

0.685 (5.02)

-1.298 (-5.45)

0.082 (0.25)

0.050 .
I

The long-runelasticitiesobtainedfrom the dynamicmodel arc similar in magnitude from AIDS the S. Therefore,tile discussion to those obtained model of signs chapter and about theseelasticitiesestimatesprovided then apply, in generaltcrms, to and comments

197

the long-run elasticities estimates obtained from the dynamic model. Indeed, the latter from former but behave to those the obtained similar only present values also not estimates in similar ways (increasing or decreasing), in the first and second periods. This should be is 5 AIDS "trcndcd", "apparently-dynamic" the chapter a of model as expected for factors the to temporal through correction seems allow of omittcd which specification the addition of a trend variable and the consideration of a non-constant expenditure if However, for AIDS were made using comparison an orthodox static model cociTicicnt. the same data sample, the results would show this specification not to be compatible with the data or consistent with demand theory restrictions. This result was suggested by tile form AIDS "unorthodox" tested the the when orthodox model against of rejection of further by 5, the the supported rejection of same model against (lie general and chapter dynamic model adopted in this chapter. Nevertheless, the examination of the short-run elasticities estimates and a long-run demand behaviour of shortand analysis arc pertinent within tile comparative framework. focus The dynamic will analysis on the second period of the sample general (the last two decades), given that it relates to more recent behaviour of tile UK tourism demand. While the esiimates of the expenditure elasticities in the equation ror Portugal are in long-run, in both for the the the to tile shortand corresponding unity estimates close France differences in Spain long-run their significant and present of shortand shares long-run for France is The but in to the shortexpenditure response close unity magnitudes. for is inelastic. demand France In UK for long-run Spain, the the tile clearly equation run, but is is to the unity, close short-run also response clearly elastic. response The estimate for Spain has the lowest value of all short-nin own-pricc elasticities.

'Mis indicatesthat UK tourists seemto be less sensitiveto short-term price changesin Spain than in Franceor Portugal.This information, supplemented with the fact that Spain for highest Spain the the short-run. estimate expenditure elasticity, suggests as a presents in for UK destination I lcncc, Spain tourists havc the short-run. may a comparative primary its in for to competitors and a wider scope relation manocuvreconcerning advantage involving However, decisions in price this area should not changes. short-term policies UK demand increasing long-term in Spanish the to of sensitivity changes overlook pricesas decreasing its identical in Franccand Portugal. towards sensitivity with changes compared

198

The UK tourism dcmandfor Portugalsccmsto bc morc sensitivcto short-runownin France Spain. Indeed, destination in the than this or own-pricc elasticity changes price in for Ilowcvcr, Portugal. largest the the their value equation short-run present estimates long-run elasticity estimatesfor the equationsof all destinationsarc close to -2 which indicatesthat, in the long-run, UK tourists are highly sensitiveto price changes and will for incrcasing-pricc dcstinations in the an policy. penalisc, a similar way, any of The inferencesconcerningthe cross-pricceffects drawn from the AIDS model of Indeed, long-run dynamic in the 5 the to terms, model. estimates of chapter apply, general in for France (Portugal) demand UK to tourism lack the price changes of the of sensitivity in France for demand UK the tourism in Portugal(France),the decreasing of and response for increasing demand Spain in Spain to the the to of response and Portugal price changes features in France Portugal the provided common results arc of estimation or changes pricc 0 dynamic However, the the both model permits analysisof short-runcross-pricc by models. S. AIDS the approach of chapter with elasticitiesnot possible In particular, it is worth noting an intcrcsting fcaturc of the short-run cross-pricc
in long-run. None tile tile their values of corresponding with compared when clasticities long-run levcl. 5% In is their the significant at contrast, short-run elasticities statistically from in different tile zero except, of course, case orFrancc significantly all are counterparts indicate that UK tourism demand for one destination, does These Portugal. results vcrsus in long-run in UK to tile price changes another, short-run while not respond significantly destinations be to to comparc prices across and adapt their prcrcrcliccs able tourists seem in UK demand to respond Put the the way, short-run, ability of another accordingly. in destinations is immaterial, whereas in the to changes competing cross-price significantly long-run UK tourists seem fully aware of price changes across destinations and adapt their demand with significant effects for the destinations considered. Hence, destinations are if increases in long-run likely tourism they to their to are receipts able avoid retain more from to any adverse changes price competitors within shorttheir own-prices and maintain run pcriods.

6.5. CONCLUSION The goal of this chapteris to point out areaswherethe analysisof tourism dcmand is likely AIDS to contribute new aspectsof thcorcticnl and systemapproach within an

199

is dynamic development traditional The systems of of gencralisations cmpirical relevance. The in demand behaviour important analysis. element of recent research work an from lack largely derived has for the of accordance such generalisations motivation typically betweenutility theory constraints and static demandmodels.Static specifications demand involving dynamics features of short-run rulc out theoretically plausible has between Hence, theory the evidence and empirical observed conflict adjustment. Misdynamic been to within approaches. misspecifications static attributed generally invalid to statistical specified econometricmodels give rise unreliable estimationresults, inference and inaccurateforecastingprocedures.Flawed estimatesproduce misleading inappropriate induce policy measures. cconomicanalysiswhich may Theseproblems,however,have beenlargely ignored in tourism demandresearch. feature in demand are a rare cnipirical studiesof modelling Dynamic genemlisations of has been based in demand this context on quantitative generally analysis tourism, and do the theoretical prc-requisitcs necessary of statistical and comply with not which models how knowing tourism the Nevertheless, the cxpenditurc of allocation of possibility quality. demand behaviour in how to their tourists time, order achieve a adjust and over cvolvcs for interest is tourism analysis and policy making, of considerable equilibrium state steady The unorthodoxAIDS systemestimatedin chapter 5, has specification features incorporated be in its As dynamic-like for to share equations. elements a which allow 'sccrningly-dynamic' in the this system, orthodox static with and contrast consequence, data itself the the and with utility maximisationassumptions with consistent reveals model information However, the the theory. short-run underlying about mechanism consumer of be between this within approach accessed and a clear separation cannot adjustmentprocess be dynamic Hence, long-run an explicit spccificationwas cannot made. effects short. and longboth to the found to be the appropriate and short. of means obtain reliable estimates in its determinants. demand UK to tourism changes of run responses dynamic flexible form AIDS In the this a general system. of chapterwe estimated 1'hc estimationresults show this model to be data coherentand theoretically consistent for this tourism the of robustness of methodology conducting evidence empirical providing demandanalysisin a temporal context. Moreover, the dynamic model offers dependable long'secmingly-dynamic' AIDS both the to the of reliable capacity provide evidenceon inadequacy AIDS information, the the of orthodox on and other restrictedspecific and run dynamicmodels,to reconcileconsistentlydataandtheory within their formulations. 200

The results of this extensivemodelling exercisefor the UK tourism demandarc future instance, for For indicate directions that tile research. show results and encouraging does dynamic matter when modelling systemsof equations, specification an appropriate impact have tests the the results validity of of concerning on considerable can and flexible form dynamic AIDS hypotheses. Moreover, the the general of model theoretical demand formulations. if Indeed, its for to testing prove specific models against more allows they are subjectto further testingfor evaluationof their consistency be data-consistent, with dynamic AIDS The theory. model passesall tile tests general the postulatesof utility providing estimation results which are statistically robust, empirically plausible and theoreticallyconsistent. in dynamicspccifications, the utility theory constraintsare generallytestedfor the
long-run coefficients. The motivation for testing theory restrictions, preferably in the in idea fully have the the that, consumers short-run, may not adjusted steady-state,rests on homogeneity hence, be in and symnictry and, may not observed circumstances to changing dynamic structure of the AIDS model is not Given behaviour. that the general short-run is that tourists adjust their behaviour to implication inherent data, by the the rcjcctcd lag. That is, the adjustment processtakes more dcterrninants demand in their with a changes fully. be if Hence, do to tourists the accomplished period not adjust current than instantaneously, then homogeneity and symmetry are not expected to hold in the short-run. When tested, however, this hypothesis could not be rejected, suggesting that the rationality is in behaviour for demand UK the tourists' observed of of utility maximisation postulates France, Spain and Portugal, both in the long- and in the short-run. At this point, we think it would be interesting to call upon the findings Anderson (1984) (although Blundell confronted a when with similar who, situation still rcjccting and following in homogeneity "ivith hypothesis the the make comment: short-run) and the long-run honiogcnelly imposed the coefflicients, on short-run pro(hiced a symmetry ... 15.42 implies Tile test the statistic of since only a rejection. marginal result surprising behaviour fruitful this andfurther restrictions on short-run scent would a consideration of areaforfuture research". In our case, the 'surprising' result of homogeneity and symmetry holding in the dynamic The is be to general model seems not unexpected. a sufricicritly robust short-run UK demand behaviour the to track accurately over the sample period. Thus, specification information by be fairly the the statistical this model. reliability of of sure provided can we

201

Its estimates suggest that UK tourists adjust very fast to changes in their demand determinants. As a consequence, we should expect non-signiricancc and/or irrelevant Indeed, is indication for that the the of the short-run coefficients. gcncral magnitudes insignificant AIDS by With dynamic the model. small short-run or cstimates provided by imposed Icads the the to which constraints are on statistical process model coefficients, (a hypothesis T'hercforc, 4casy' the tested. the under-rejection) possible of non-rejection an faster consumers adjust their demand behaviour, the less signiricant short-run ciTects arc, imposcd is likelier theory the the the of utility on short-run. non-re*cction postulates and This hypothesis requires, of course, further empirical support which can only be delivcred in the context of future research. The empirical results proyided in this chapter show how the estimation of a

dynamic AIDS systemcan provide new information about the behaviourof UK tourism demand. This modelling approach allows for intertcmporal rationality of consumer behaviour by explicitly consideringthe mechanismunderlying the short-run adjustment for budget Estimates tourism and price elasticities were obtained, permitting a process. hand, AIDS the the tile relative ability analysis of, on one of systemto provide comparative behaviour long-run UK information hand, the tourists the tile of on and, on other accurate longUK the and statistical relevance of and short-run sensitivity magnitudes of relative its in determinants. demand to changes tourism Nevertheless, there are theoretical and empirical issues which still have to be AIDS One the to as models quality specifications. endorse of the most important addressed include AIDS the the problem; regression since spurious systems nonstationary regards be their can results spuriousunlessthe variablesarc cointcgratcd. estimation time series, others, are linked with the a priori division of cndogenous/cxogcnous variablesassumed by the AIDS approach.Finally, an important matter in quality evaluationor econometric forecasting Statistical is be describe long. to their ability. models can good means models but if forecasters, lose they they are not equally relationships good 1111ich of run economic for policy analysispurposes. We address theseissuesin the next chapters: their relevance involves 7 framework analysis within cointegration a vector auto-regrcssivc of tile chapter definedwithin an AIDS system;chapter8 investigates long-runrelationships tile ability of the different econometricmodelspreviously estimated to predict the UK tourism budget Portugal. France, Spain and of shares

202

CHAPTER 7
VECTOR AUTOREGRESSIVE MODELLING OF THE

UK DEMAND FOR TOURISM

INTRODUCTION Static, steady-state equilibrium modelsare often termed long-run spccirications.In linked discquilibrium dynamic the arc often with concept models short-run of as a contrast, between distinction inherent long-run Thus, to the short. and models adjustment. processof is the notion of equilibrium. The long-run is a state of equilibrium where changeis not likely to occur since economicforces arc in balance.Ilowcvcr, mspointed out by Harris (19950p.5), "there is no necessity actually to achieveequilibrium at anypoint In time, even forces is is All that the towards that the economic move required system I-o act as long-run by Thus, defined is Idea the the relationship posited. what matters equilibrium ... idea is between " T'llis flinc. which variables are cvolving over relationship ofa steady-state long. theoretical the and analysis empirical economic and cstimation or of aspect a central in been has concern a main appliedeconometric research. run relationships Frequently,the estimation of long-run relationshipsinvolving time series brings from in the the that of presence non-stationary variabIcs arise modcl. problems about ignoring them and proceedingwith the estimationof a regression containingunit-roots in The is by lead data to results. spurious spurious rcgrcssion problem reflected can the long-run the existence of statistically significant suggesting relationships results estimation between variables when, in fact, all that is shown is evidence or contemporancous
but in due Simple trends to the unrelated, common, series. nonstationary correlations differentiation to remove unit roots from the data is not the answer since, while avoiding it long-run information is also removes problem, the spurious regression which crucial to levels. the the causal meaningful relations of among existence variables' establish

203

leads following it "is The spuriousregression to the question: possibleto problem infer a genuinelong-run relationshipbetweennon-stationaryvariables"?71c affirmative is in (1987) Engle Granger The through the concept and cointegration. given of answer if interpretation that two or more time seriesvariablesarc states of cointcgration cconomic linked to form a long-run stationaryequilibrium relationship then, even if the variables they will nevertheless themselves are non-stationary, presenta meaningful co-movcmcnt forces. forces These by an to time, owing underlying equilibrating arc represented over which compelsthe economicsystemto convcrgetowards its cffor-coffcction mechanism, long-run equilibrium. Thus, the conceptof cointegrationportraysthe existenceof a shortlevels dynamic the towards their which pushes equilibrating mechanism variables' run In other words, if two long-run equilibrium path, evenif thesevariablesarc nonstationary. be then there cointegratcd, variables are must an error correction mcchanisin or more underlying their co-movcmcnts and, conversely, the existence of an crror-corrcction time cointegratcd series. generates mechanism In chapter4, we addressed the importanceof distinguishingbetweenstationaryand failure do lead to to the problem of spurious so variables, since could non-stationary for in We data tested the the tiinc presence of unit roots scrics entering the regression. indications When dealing their obtained of presence. and with nonstationary model is the concept cointcgration of synonymouswith the conccpt or long-run variables, Failure to establishcointegrationoften meansthe noti-cxistenceof a steady cquilibrium. Therefore, rind ira long-run the the to variables. among step relationship tlcxt was out state linking dynamic the the variables existed, of single equation ARDL model relationship information both From long-run this model, on shortand relationships could be specified. for We long-run tested the the cointegration using residuals of specificationand retrieved. indications The linkage between that the the errorvariables were cointcgrated. obtained long-run be the and mechanism equilibrium relationship could also established. correction in the use of a single equationapproach. However,thereare severaldisadvantages In the singleequationframeworkusedin chapter4, we startedby modelling threescparntc UK demand for France, Spain the tourism which specified and Portugal as a cquations function of tourism prices and the UK real per capita income. I lowcvcr, in this specific it is possiblethat a shockaffecting tile UK demandfor tourism in tourismdemand context, F=cc also affects tourism in Spain, and a revolution in Portugal influence tourists' decisionto visit this or ncighbouringcountries.Yet, the single equationapproach doesnot

204

interrelationships between destinations for the of existence and possible account cxplicitly be testedwithin this approach. cannot cross-cquationrestrictions Furthermore, despite the statistical evidence of cointegration in tile single equations specified in chapter 4, several problems still remain within this econometric approach. When there are n>2 variables in an equation, there can be up to n-1 lincarly independent distinct long-run Only to a each corresponding equilibrium. when relationship cointcgratcd infer found is it is is If to that the this cointegrating vector possible unique. not tile n=2 is incilicicnt linear since we a single approach can only obtain one equation case, adopting information the vectors valid and not existing cointegrating about all tile combination of ' in if is Moreover, long-run the there model. relationships existing even only one possible is inefficient, the single equation approach still relationship, unless all the cointcgrating If there arc cndogcnous variables as variables are weakly exogcnous. side right-hand inefficiency. be lost However, inforl-nation information is this causing will not rcgrcssors, is instead, if, system of equations estimated allowing each cndogenous variable to a wasted Therefore, dependent from the variable. results estimated a as a singic equation appear involving if is time there series arc reliable nonstationary only a unique approach (weakly) the variables and all vector among regrcssors are cxogcnous. cointegrating In chapter 5 we estimated the UK demand for tourism in France, Spain and Portugal

dynamic A AIDS in this of equations. system version static of a model was estimated using framework 6. However, AIDS does the the methodological of approach not chapter it's include the analysis and assumptions cointegration underlying specification contemplate division be Thc that cndogenous-exogenous of variables may questionable. assumed an AIDS approachof chapters5 and 6 models the UK demandfor tourism as a systcrnof destinations' functions the expenditure shares expressing as of cxogcnous equations, include UK the per capitareal tourism expenditure and own and conipcting variableswhich is Although there some theoretical basis justifying the assumptionof tourism prices. is for fccdback the there the capita expenditure, real per possibility of effects cxogcncity for demand Yet, tourism. the AIDS approachdoesnot considerthis between the pricesand included in In tourism the the AIDS models were price variables addition, possibility. in found testedand nonstationary chapter4. Hence,the estimation resultsobtained from

I As shown in chapter4, we can obtain valid information about these issueswith the Pesaran and Shin However,ifthe testsindicatethe presence of morethan one cointegraming approach. vector,the ARDL single is inefficient. cquation

205

thesemodelscanbe deemed andthe statisticalinferenceinvalid, if no cointegrated spurious found linking the variablesof thesespecifications. are rclationship(s) if data seriesarc nonstationary, Consequently, there seemsto be a risk involved in the estimation of econometricmodels which regresscndogenousvariables on several assumed exogenous variables, without sanctioning their statistical validity with Given is that the number analysis. of potential cointegrating vectors cointcgration determination the given of simultaneous and possibility of the variables, unknown, further and establish a methodology to specify cmpirical analysis must go one step be models which can efficiently estimatedand validly testedwithin a system econometric in Summarising, the presenceof non-stationarytime series and of cquationsapproach. for estimatinglong-run relationship(s), feedbackcffects,an efficient econometric approach be which allows all potentially cndogcnous variablesto appear must a systemof equations for dependent cointcgrationanalysis. variables and appropriate as features An econometric these methodologywith was proposedby Sims (1980) and (1988). Sims' vector autorcgrcssivc(VAR) approachestablishes developedby Johansen innovative specificationmethodsthat arc consideredto be valid alternativesto both the The Johansen the traditional structural and multi-cquation equation approaches. single (1988) procedureadds cointcgrationanalysis and an cfficicnt estimation method of the long-run to the needed establish number parameters of relationships and to supply structural informationon both the short-andlong-runresponses. In this chapterwe use the VAR methodologyand the full information likelihood by developed Johansen (1988), UK demand for to the tourism approach, analysc system France,Spainand Portugal.The chapteris structuredas follows. Section7.2, addresscs the Section features Sims' 7.3, integration the methodology. of establishes order of main of the time series and the appropriate lag-length for the variables included, spccifics the VAR model for the UK demandfor tourism and presents (unrestricted) the cstimationand forecasting results obtained with this specification. Section 7.4, applies tile Johansen for determining in the tile specification number of cointegrating vectors existing procedure and presentsthe estimationresults of a cointegratedstructural VAR under exactly- and Section 7.5, presentsthe forecast results obtained with tile ovcr-idcntifying restrictions. VAR forecasts them the compares and with cointcgrated of the unrestrictedspecifications. Scction7.6 concludes.

206

7.2. MAIN FEATURES OF THE VECTOR AUTOREGRESSIVE METHODOLOGY is on estimatingthe averagevalue of the In a single equationmodel, the emphasis fixed (say, Y) the dependent valuesof the explanatoryvariables conditionalupon variable from is in to therefore The X's). run (say, assumed relationship such models cause-cffect X's in Hence, dependent the these to the nrc variable. models, the explanatoryvariables is be Y be to to assumed endogcnous. and variables cxogenous assumed Although in some economiccontextsthis assumptioncan be justificd, it is often is In cause-effect relationship many situations such unidirectional controversial. feedback "independent" in the are effects present and where particularly cases meaningless, by is dctcnilincd X's This Y terms. tile the when occurs error variablesarc correlatedwith feedback is by Y. In determined X's this opcrating, the case, a mechanism are and someof is determined. bi. In jointly there such cases, a or simultaneously making thesevariables between Y X's the the and sonic directional cause-cffcct of which renderstile relationship distinction between "dependent" and "independcnf' variables dubious. Moreover, tile important ror has X's Y the the consequences of entire and some of simultaneity observed be It that the tile variablcs oil cndogcnous shown rightcan analysis. of empirical process hand side of a single equationsare correlatedwith the error terms violating one of tile least X the that the estimation method, squares namcly, variablesare of crucial assumptions independently distributed disturbance if the terni. stochastic, are of or non-stochastic either biased is Icast-squarcs but the these not only If neither of estimatorsarc conditions met, inference based is invalid. inconsistent the on such statistical estimators and also This flaw of the single-equation approach, resultingfrom the correlationor the error is known the variables, explanatory as simultaneous of equation more or terms with one in econometrics during tile 1950S bias. This problemwas tackled by theoreticaladvances development the 1960s with of appropriatestructural which were mainly concerned and into that take tile could consideration simultaneous methods estimation and specifications These the theoreticalcfforts gave determination same equation. variables within of set of a literature in is known the the traditional structural multi-equationsystern as to what rise in is described for Johnston (1984), This textbooks, many example, approach approach. Grilliths,ct al. (1993)andGujarati(1995)andonly the main features will be outlined here. In contrast with the single equation approach, tile structural multi-equation being the that set of variables simultaneously arc viewed as methodology separates

207

determined, from the remaining set of regressors. Such models specify one equation for forming dependent jointly the variable, cndogenous a multi-equation system of or of each Hence, these models can accommodate tile endogcnous-cxogcnous relationships. structural dichotomy of the variables and further distinguish the set of endogcnous from tile set of includes lagged-cndogcnous exogcnous and which variables. predetermined variables, Once a system of equations is formulatcd, the traditional structural approach

bias. This the can which remove procedures, problem or simultaneous several suggests because in is this specificationexpresses problem present a structuralsystemof equations function of other endogcnous variables,along with variableas a separate cach cndogenous lagged (exogenous cndogenous) variablesand a stochasticcrror term. and predetermined Since the endogenousexplanatory variables are correlated with the error tcrnI, the (structural be the system structural cocfficicnts), associated with cannot cocfFicicnts by least However, be if this squares methods. estimated problem can removed consistently for the existing cndogcnousvariables, making them are solved the structural equations disturbances. dependent variablesandthe stochastic solely on the predetermined is A structuralsystemsolved for the endogcnous variables known as the reduced. form of the model, and the coefficientsassociated with this fonii are callcd mduccd-rorm Since disttirbanccs ill variables only predetermined and tile stochastic appear coefficicrits. the of reduccd-fonnequations,and since the predetermined side variablesare right-hand be OLS the terms, the to error correlated with estimationmethodappliedto the not assumed 0171ciclit I lcncc, consistent generates equations and asymptotically cstimatcs. rcduccd-forrn is bias from the reduced-fornisystem or equations. problem removed the simultaneous 11owcvcr,the estimatesobtained with this procedure arc those of the rctluced-form, andnot of the structuralcoefficientswhich are ultimately of interest.Sincethe coctTicicrits former, the the possibility exists that tile structuralcoefficients lattcr arc combinations of from Whether is be brings the this tile reduced-form. coefTicicnts. retrieved case, can about idcritification. the the the of structural problems system major approach: of problemof one The identification problem arisesbecausedifYcrcntsets of structural coefficients data; is, be the that same set of a given reduccd-formcquationmay be compatible with may different it difficult to tell which particular structural equations, with making compatible illustration, As is being investigated. an considerthe following exampleof a simple model dcmand-supply model:

208

Demandfunction: Supplyfunction:

Qd = ao + aP, + ult t Qt = PO + PIPt+ u2,

(7.1) (7.2)

Equilibrium condition: Qd = Q; =Q , is is P Q' demanded, is Qd supplied, quantity price of tile commodityand (X quantity where in jointly Q determined P is It that 0 to uI and change are as a and areparameters. easy see in both P Q demand to turn, to change. the causes and which, shift curve will cause Similarly a changein u2 will shift the supply curve which, in turn, will affect 11and Q. in bias Q 11 the Hence,there is simultaneous system, as and are correlatedwith equation both uI andu2. However, model (7.1-7.2) might be perfectly sensiblefrom a theoreticalpoint of data Q 11 intuition that two time Nevertheless, with only series on and suggest would view. four difficult be it information, to estimate unknown parameters. might and no additional by, for example,working with cantcredvariables,how can Even if we ignorethe intercepts from 01) ((xi two and a regressionwith only one structural parameters we estimate know how Q P If tile we whether can we are estimating on we regress variable? Cxplanatory demandequationor the supply equation?The identification problem consistsin seeking it is that these the to these apparent answers must precede and problem questions answers the how to structuralparameters. estimate of it is impossible is to to that Anotherway of seeingthe identification. problem realize from In this the the cocfficicnts of a system. rcduccd-fonn parameters structural retrieve be Using deniand-supply is to the under-idcntified. exampleabove,we casethe system said illustrate this point. can By the equilibrium conditionwe obtain: PIPt PO P, + u2, + + + ult oto a, = Solving for P we get the equilibrium price P, = 1710 + vt Pt given by (7.4) in (7.1) we obtain the equilibrium quantity Substituting (7.4) (7.3)

Qt=nl+Wt
where 00-aO 110 = P1 loci and alpo -(iopl III = P, CE, -

(7.5) (7.6)-

209

Equations(7.4) and (7.5) are reduccd-formequationsand rJOand I'll are reducedform coefflcicnts.It is clear from (7.6) that thereis no way of recovcringthe four unknown from the two rcduced-form coefticicnts. structuralcoefficients However,equations(7.1) and (7.2) becomeidentified if additional prcdctermined if For to the to the example, we add an variable cxogcnous system. are added variables for lagged interest (income the demand and variable rate, example) cndogcnous or equation P,., to the supplyequation, we canretrievethe structuralcoefficicnts from the reduccd-forni In and six reduccd-forni cocfficients. this case,thereare six unknownstructuralpararnetcrs for In determine the this to estimates unique structural paranicters. cocfficicnts,enablingus be form is to the cxactly-identiricd. said case, structural Adding more variablesto one and/orthe other equationwould causethe systemto
be ovcr-idcntificd and the structural parametersestimates not to be unique. In this case we inrormation, have to the too not needed supplementary exactly-identify much would for identification, The "secret" be then, to to achieving seems related parameters. structural do in For to these other which not appear an equation others. variables the adding of "missing" Intuitively, it the to the variables are zero. coefficients attached cquations, if is likely identification the the that of matrix structural parameterscontains more appears in In tcrminology, and columns. structural specific rows econometrics elements zero several in is defincd this matrix element a zero as a zero rcstriction. Without these of the presence be identificd. the systems would of structural not restrictions, most It seems to be the case, then, that much is Icf1 for the structural ccononictrician to

'irst, in decideprior to undertakingthe actualestimationproccdurcof a structuralsystem.J. is it to the to necessary system, establishwhich variablesarc cndogcnous specify and order in idcntirication, Second, to order achieve a number of zero which are exogenous. be imposed. The have to main criticism of multi-equationstructuralmodelling restrictions hasbeencentrcdon thesea priori decisions:the role of the zero restrictionassumption and into fact, In the variables exogcnous and cndogcnous. of some models the categorisation deleted from be formulated to to variables with added some equations and others seem identification, justification. On to tile other and without much economic achieve merely hand,plausibleeconomicmodelslack empirical supportbecause they are not idcntiricd. These aspects of structural modelling led Sims (1980) to consider tile zero "haphazard" differcrit "incredible" for to and and propose as a radically strategy restrictions the specificationand estimation of multi-cquation systems.Sims (1980, p. 14-5) states:

210

"Because existing large models contain too many incredible restrictions, empirical In theories too testing macroeconomic a oj? cn procceds competing research aimed at For framework it few this reason alone, appears worthwhile to equations single- or investigate the possibility of building large models In a style which does not tend to feasible be large-scale It haphazardly. to should estimate accumulate restrictions so ... treating all variables as en(h)genous". rMuced-forms, unrestricted as macro-models Sims' (1980) new methodology of vector autorcgrcssive (VAR) modelling difflers from the traditional structural approach in three main aspects: all variables are treated as imposed; is based VAR the model on any specific not arc cndogcnous; no zero restrictions by Charcmza Deadman (1997), As the third theory. and aspect out pointed cconomic derives from the other two since, "if there is no variable excludedfrom any equation of the It is that evcr)-thing causes ever)-thing else and there means exogenous, nothing and 1nodel is no room for assuming much more than very general economic priticiples as a starting is Indeed, Sims' " 157). the (p. this to of major criticisms one attached approach since, point long-run 'purcly-statistical' (1997) the Pcsaran this of much analysis within underlines, as "is the I)pe an explicit providing account of equilibrium without of conducted approach have "empirical I(' that these and applications of methodology theory that may underlie focused on the statistical properties of the underlying economic time series, ofien at the insights (p. 178). Hence, Sims' theoretical and economic reasoning" of expense its labelled by is long-run to critics, as an a-thcorctical often approach methodology cquilibrium analysis. Thc features which make the VAR approach so flexible, easy to spcciry and simple to Unlike the the traditional within which area weaknesses mark remain. also estimate is VAR because it does model unrestricted a-theoretical an systems, not use any a structural by formulation The information (unrestricted) theory. economic supplied or a general priori VAR is based on a system of equations, regressing each variable of the model on its own lagged A VAR2 little the values. rcduced-form variables' other more than requires and all include in decrrnination to the the variables model and the choice of appropriate or a little input, lag-length. With little theoretical than so one should not more expect suitable fact, In in be the the unless output. underlying model content structural can economic
2 The right-handside of all equationsin a generalVAR containsonly prcdctcrmincd(endogenous lagged) is be form This to the considered model similar to the rcduccd-fonnof a structural multi. of variables. Thcrefore, is also variables. exogenous a gcncral(unrcstrictcd)VAR specilication, system with no equation VAR. knownasreduccd-form,

211

identified from the reduced-form VAR, the economic interpretation of the estimates is difficult, if not impossible. On the other hand, the VAR is a truly simultaneous system in that all variables are regarded as endogenous and no a priori cndogenous-cxogenous division is assumed. However, the biggest practical challenge in a VAR specification is to choose the

longer lag-Icngth, for its The degrees faster lag-length the the variables. of appropriate freedomare eroded.If the lag-Icngthis p and therearc n variables,eachof tile 11 equations intercept. In limited the the plus view of number of contains n.p unknown parameters in introduction lags tile empirical analysis, most of available several generally observations for eachvariablecanbe a problem.Nevcrthel ess,appropriatelag-Icngthselectionis cnicial. is present if autocorrclation Sincea VAR containslaggeddependent variablesas rcgrcssors, in the error tcrrns,the predetermined variableson the right-handside of tile equationscan be correlatedwith the error terms,leadingto inconsistent estimators.So, purging tile error in is VAR lagging from modelling essential and autocorrclation all variablesa certain terms if is do be Yet, if is too tile this. times small, p can model can of mis-specificd; p number degrees freedom be large, the and of over-parametcrised wasted.In addition, modelcan too in Johansen the number of cointegrating establishes vectors procedure, which existing a the VAR. can be quite sensitiveto the lag-Icngthchosenfor the VAR structure.Therefore,a in defining be deal paid should p. of attention great lag-lengthis imposed,the error termsof eachequationcan be Whenthe appropriate Since VAR be to the current values of cach uncorrclatcd. a serially expresses assumed 3 function (lagged solely of predetermined variableas a variables, cndogcnous) cndogcnous theseare not correlatedwith the serially uncoffclatcderror terms, and eachequationin tile by OLS. In be OLS this the case, can estimated method provides consistentand system 4 asymptoticallycilicient estimatcS. There is an important difference betweenusing an unrestrictedVAR model for forecastingand using it for economic analysis. If economic analysis is the objective, is important.However,an unrestricted VAR modcl knowledgeof the structuralparameters is imposing (without be in there no and way undcr-identified restrictions), will which the be from VAR. On tile other hand, a the can recovered parameters reduccd-form structural
3An additionalsetof deterministic components, dummy suchas an intercept,deterministic trendand seasonal VAR is to the of a model. equations often added variables " If the appropriate lag-lengthis not the samefor all variables, IncludedifTcrcntrcgrcssors, or if the equations SURis moreefficient thanOLS for estimatinga VAIL

212

VAR modelcanbe viewedasa reduced-form systemwith no cxogcnous variablesspecially fact, if is forecasting, In for forecasting the the underlying objective purposes. adapted be its does form to the need estimated, as not parameter of model estimates arc structural In this case,evenbeing under-identiricd, irrelevantin the process an of obtainingforecasts. VAR forecasts have (unrestricted) supplies which are unbiased and appropriatelyspecified forecasting In the standard econometric unlike procedures minimum variance. addition, knowledge be forecasts upon of tile cxogcnousvariablesvalues,a must conditioned where VAR usernot only can overlook any economictheory underlying the model, but also does There know the the to variables. values cxogenous arc no cxogcllousvariables of not need in a VAR model. Sims' methodologyconsistsof regressing eachcurrent (non-laggcd)variable oil all
Consequently, lagged is in VAR times. the an unrestricted p model model overvariables lagged in be the that the the of variables sense some entering model could paramcterised insignificance. basis However, tile the advocates or this of statistical on excluded properly including himself, Sims advise against this procedure, arguing that tile methodology, imposition of zero restrictions may suppressimportant inronnation and that the rcgrcssors in a VAR specification are likely to be highly collinear, so that the t-tests on individual for down-sizing the model. reliable guides are not coefficients In practice, it is impossible to avoid the consideration of prior restrictions on n VAR

be limit Sample that there to the will constraints mean a size numberof variables system. included and the numberof lags imposed.Even if the samplesize is not a problem, tile VAR it lysis for to some structure a model and using , of giving ccononlic anti possibility
imposition forecasting Well-founded the purposes, requires of restrictions. alongside help to transform an unrestricted VAR specification into a may constraints theoretical "consistent highly detailed VAR (Charcrnza. theories" with model even econontic restricted 157). The identi 1997, for lows fication Dcadman, consideration of p. such restrictions al and interpretation in the the and economic of structural model parameters a way not possible of infortnation Furthermore, the a priori reduced-form. concerning tile parametersallows with for testing and including restrictions that can improve the precision of cstimatcs and reduce if IIcnce, interest is forecasting, tile down-sizing forecast the even the main error variance. help VAR improve to the results. can of an over-parameteriscd The specification of a reduced-form VAR can be illustrated using a simple example Consider (xi), (yt) and (wt) and assume that the time three variables. variables series with

213

by is (yt-, two the time currcnt other and past the realizations of xt affected wt) of path lag-length following Establishing the of one, we can system: specify a maximum variables.
bl3Wt bl2Yt blo + + + + 712Yt-I 713Wt-I CXt YIIXt-l x, = b2o b22X,

y, =
w,

-b23Wt

+ 721XI-l + 731Xt-I

+ Y22yt-I + 732Yt-I

+ 723WI-I + 733Wt-I

+ Ut + CWt

(7.7)

=b30-b32Xt-b33Yt

be disturbances to white-noise wherecxt , cyt andcwt areassumed first-order VAR(l). System(7.7) represent vector autorcgrcssivc model a structural

hasa in (7.7)arenotreduced-form Theequations equations sinceeach cndogcnous variable it is However, two. to transform the system the possible other effecton contemporaneous form.Usingmatrixalgebra into a moretractable wecanwrite (7.7)as:
I
b2l

bI2 I
b33

b13 b23 I

xt

blo
b20 +

711
721 731

712 722 732

713 723 ^133.

XI-I YI-I +

cxt rlyl

yt -Wt.

_b3l

_b30.

_Wt-I.

-OWJ.

or

Bz,

ro + rz, I
b12

-,

+ e,
b13b23 I-; Z, -

(7.8)
xt.

blor, =
b30.

711
Y31

712

Y13 Cl

cxl cyt
Cwt.

whcre

B=

b2l _b3l

I
b33

b20 ; ro = = Y, ;
wt. .

Y21 Y22 Y23

732 Y33.

Prc-multiplying(7.8) by B" allows us to obtainthe rcduccd-fonnVAR model: Ao + A, z, + et -, B-le, A, B-1r, B-lro Ao ; et= ; = = wherc (7.9)

If we defineajoasthe elementi of vectorA0, adasthe elementof row i andcolumn

i of vectoret we canwrite (7.9)in thecquivalcrit forni: in matrixAl andcit astheclement

x,

= alo+a, Ixt,
=a20+a2lXt-l+

+ al2Y,

y, wt

+ a, -l 3W, -, a22Yt-I + a23W, -,

+ elt +

e2, c3,

(7,10)

=a30+a3lXt-t+a32Yt-l+a33Wt-l+

System(7.7) is called the structural VAR model and system (7.10) is called the Note VAR in that the right-hand model. side of all equations (7.10) includes mduced-form As the error terms are assumed variables. predetermined not to bc scrially corrclatcd, only be in the systemcan efficicntly estimated usingOLS. eachequation

214

7.3. SPECIFICATION OF A VAR MODEL FOR THE UK TOURISM DEMAND IN

FRANCE, SPAIN AND PORTUGAL

lead in The arguments the to the conclusionthat -when previous section presented long-run the of relationships amongmore than two nonempirical analysisabout existence be doubts there time and conducted, are about the cxogencity series must stationary be the to variables and/or should of whether zero some restrictions with respect assumption imposed, the appropriate econometric approach consists of treating all variables as framework. Then, VAR using one or more of the cridogenouswithin a rcduced-form. in doubt be for the tests the of set of exogcncity variables should available, methods division is Once Johansen (1988) the the established, cndogenous-exogenous undertaken. for Tile be to test the existence of cointcgrating relationship(s). number used can procedure found long-run the establishes number of meaningful vectors relationships of cointcgrated Imposing the to the variables., restrictions cxactly-idclitiry underlying among existing long-run VAR Once the the the structural of estimates cocfiicicnts. can assess structural VAR is idcritified, over-identifyingrestrictionsmaking the VAR compatiblewith specific be tested. theories, can also economic In this section,we specify a basic generalVAR model for the UK tourism demand is Portugal, Spain r-rancc, in which used for both forecastingand economicanalysis and basic includes in The the the within model analysis variables vector purposes. z, = [w, PS,,PF,,Et 1. All variables in vector zj, arc the same as those WSt,Wt, PP,,

in chapterS. includedin the AIDS systems, alreadycharactcriscd The specification. of a VAR model startswith establishingthe order or integration its in lag-length Hence, this the variables. of section,we useunit root tests appropriate and in determine time the series vector zt arc stationary,and diagnostic tests to to whether VAR lag-length from the the the crror ternis. which eliminates serial of correlation cstablish
7.3.1. ORDER OF INTEGRATION OF THE VARIABLES INCLUDED IN THE, VAR

The issueof whetherthe variablesin a VAR needto be stationaryhasbeena matter As nonstationaryvariablescan lead to spuriousregressions for debateamongresearchers. inference, invalid (see, for the usual assumption statistical cxanipic, Grifliths el al., and 1993, p-693 and Judge et al., 1988, p.754) is that all variables in a VAR should be

215

VAR data be if Thus, tile time a contain unit roots, must series entering stationary. differentiation, for (by transformed of means example). appropriately However, differentiating nonstationaryvariables in order to achieve stationarity information important long-run concerningthe co-movements of the variables.In excludes dealing with the VAR approach,for exampleSims (1980) and Johansen fact, researchers (1988), recommendagainst differentiation even if, as stressedby Sims, the variables is determine VAR In the this to tile view, main objective author's of a roots. unit contain Hence, difTcrcntiation does long-run the variables. not add among relationships cxistcnceof instead, inrormation "throws important the to this and, model away" purpose of main much data help identify the to the can which of possible cointcgMting co-movements about in VAR, long-run If are present a vcctor(s) meaningful cointegrating rclationship(s) vectors. is the the correction mechanism and an error variables nctivc within short-run among Cxist dynamics.This informationis crucial for economicanalysisandpolicy purposes andcan be VAR. Furthermore, from Johansen (1988) for the the reduced-form. procedure rctricvcd the numberof cointegrating deten-nining to be more rcliabic vectorsin a VAR is considered involved I(I). level the are variables when all Although the usual approach adopted by VAR "aricionados" is to work with if levels, in "it is important these to recognisethe effect are non-stationary even variables distribution (Harvey, 1990, Thus, 83). the estimators" of on roots p. one should unit of VAR the caution rcgading specification special of unrestricted models which exercise include non-stationarytime series because,unless the variables arc cointcgrated, tile lead VAR to spuriousresultsand with model non-stationary a variables of may estimation invalid inference. Table 7.1 shows the statistic values and respective critical values at tile 5% DF for level, ADF levels first differences the tests the and unit of root and or significance the variablesWF, WS, WP, PP, PS, PF and E. It also showsthe AIC and SBC criteria for Ilic values in bold indicatetile the lag-lengthselectionof the Dickey-Fuller test equations. for both the test to with critical comparison value, statistic according selectioncriteria. valid The testsclearly indicate that all variablesin levels arc non-stationary and, cxccpt for APP,all variablesin first differences are stationary.This meansthat accordingto the DF PP, be be I(I) ADF tests, to except can variables, all considered variables. and

216

Table 7.1: Unit root DF andADF testsfor the variablesWF, WS, WP, PF, PS,PPand E
Variable Test Statistic AIC criterion SBC criterion Critical value

WF AWF WS AWS WP AWP PF APF PS APS pp APP E AE

ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(I) ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(l) ADF(O)
ADF(l)

-2.100 -2.008 -5.163 -3.284 -1.965 -1.756 -5.187 -3.510 -1.738 -1.458 -5.429 -4.827 -1.869 -2.311 -3.435 -3.417 -2.083
-2.428

54.21 * 50.81 49.71* 46.36 52.09* 48.71 48.08* 44.78 86.03* 81.64 81.49* 78.25 32.98* 32.24 30.59* 28.69 33.50*
31.74

52.87* 48.87 48.42* 44.47 50.76* 46.77 46.72* 42.89 84.70* 79.69 80.20* 76.36 31.64* 30.30 29.23* 26.80 32.17*
29.80

-2.971 -2.975 -2.975 -2.980 -2.971 -2.975 -2.975 -2.980 -2.971 -2.975 -2.975 -2.980 -2.971 -2.975 -2.975 -2.980 -2.971
-2.975

ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(l) ADF(O) ADF(l)

-3.788 -2.968 -1.418 -2.605 -2.480 -2.245 -2.362 -1.624 -3.696 -2.675

29.78* 27.24 31.74 34.17* 31.81 * 29.16 20.51* 18.48 18.07* 16.94

28.48* 25.35 30.41 32.23* 30.51* 27.27 19.18* 16.53 16.77* 15.06

-2.975 -2.980 -2.971 -2.975 -2.975 -2.980 -2.971 -2.975 -2.975 -2.980

217

The caseof the variablefor the price of Portugal(PP) is a specialone.As previously be for ADF tests in DF 4, testingstationarity the may not adequate and mentioned chapter 5 in a time seriesthat may presentchangingmeansover the sampleperiod. Owing to tile in in Portugal, 1970s the the meanof the the possibility of a change political upheavalof in for is likely. Hence, PP tile mean test adequate unit root allowing a shift an very variable The different Phillips-Peron be (1988) test to time seems approach. this a requires series of in PP. The Pllillips-Pcron for the test testing case of variable stationarity way an appropriate is based suchthat on a simpleDF regression, Appt = Po+ plppt-l The OLS estimationresultsfor equation(7.11) are
APPt = 0.0079- 0.14139 PPt-, (0.536) (-1.4177)

(7.11)

The t ratio of PI cqcfficient is simply the ADF(O)statistic for variable PP in Table 7.1. The by Phillips is icd by Peron to this statistic proposed and correction caff out non-parametric dcriving the White and Newey-West adjusted,covariance matrix and computing the for (7.11) The the regression results estimation using adjusted variances. adjusted follow: as are matrix covariance PP, 0.0079- 0.14139 APP,= (0.5112) (-2.9039) -, The t ratio for the PI coefficient is now the adjustedt ratio and also the valid statistic to 1-2.90391 1-2.9711 hypothesis As < the the value. we cannot critical reject of with Comparc This confirms the result obtainedwith the ADF test displaycd in I)P being non-stationary. Table 7.1 for the samevariable.However,the problemresidesin the fact that the ADF test indicatesthe first differenceof PP (APP) to be non-stationary as well. If this is the case, I(I) be Therefore, Phillips-Peron PP the test an variable. we considered perform cannot then for APP,runningthe following regression:
AAPP, = P'o +P'l APPt-I
A

(7,12)

for (7.12) are The OLS estimation results regression AAPP,=-0.001692-0.51046APP,, (-0.1211) (-2.4803) Ile estimationresultsfor (7.12)using the adjusted covariance matrix are
A

s In chapter 4 we showed thatthe PPvariableis accepted asan I(l) variablewith MacKinnoncritical valuesat Charcmza level 10% andDcadman with and significance critical valuesat the 5% level. the

218

AAPPt

0.00 1692- 0.51046 APP, -, (-0.1576) (-3.9587)

The t ratio for P, is the adjustedt ratio which is the valid statistic for comparison with the 1-3.95871>1-2.9751 being hypothesis APP As the we cannot reject of critical value. levels is in PP l(l). Phillips-Peron Hence, test, the the to variable according stationary. Tbcrcfore, we concludethat thereis sufficient statisticalevidenceto considerall variables in levels includedin the VAR as integrated of orderoneor l(l) variables.
7.3.2 DETERMINATION OF THE ORDEROF THE VAR

The maximumlag-length(p) imposedon the structureof a VAR model is gcncrally by denoted VAR(p). The VAR VAR the the order and of a plays a crucial order of termed determine its dimension in the can whether estimationmethod since analysis empirical role infcrcncc is the thus, the estimation results whether and statistical applied consistentand, in be Therefore, to the taken of choosing p, special process care ensure should valid. arc be disturbances for for high is to it the not serially correlated, and small enough enough that be VAR to over-parameterised. not tllc The observations available for this VAR modelling exercisedo not allow for an
fact, given the number of variabics involvcd and the cxistirig In lag-structure. Cxtcnsivc

Taking limitations lag-length into two. these the cannot exceed consideration, samplesize, SBC Likelihood (for AIC the the criteria selection and adjusted small and samples) we used test for selectingthe order of the VAR with tile maximum lag-Icngthpcmiittcd. Ratio (1,11) the LR test statisticandthe two selectioncritcria for choosingp. Table 7.2. presents the order of the VAR Table 7.2: AIC andSBC criteria andadjustedLR test for ScIccting Order(p) 2 0 AIC 296.91 296.59 185.90 SBC 246.37 269.37 182.01 AdjustedLit test ....... X2(36)= 37.67(0.393) X2(72) - 366.02 .

The LR test rejects order zero but cannot reject a first order VAR. 111cSBC VAR indicates be The indicates AIC the the to order of one. clearly criterion p to criterion be two, but by a very small margin. It is common for tile SBC criterion to selecta lower AIC in VAR the this case,we can be quite conridcnt with criterion so, as compared order

219

is it Yet, be the to the to always of prudent examine the residuals selected. order about in orderto checkfor statisticalevidenceof no serial autocorrelation. individual equations
VAR SPECIFICATIONor THE UK DEMAND FORTOURISM 7.3.3. Tim UNRESTRICTED

VAR for the UK tourism demandin The reducedform of the first orderunrestricted 6 be by VAR 1) (denoted France,SpainandPortugal can written as:
WF, = a,, + auWF, + al, WSt-, + al3pp, + a,, PS. + a,, Pr-, + nl6E., + Cl, _, -, _, _, -, WS, a23ppt-, a2, Pr,, + a, WSt = a2o+ a,, WF, + + + + a, a,, + c2, -, 4PS, 6E,., -, -, _, PS, llr-, + + a,, a3, a32WS, (7.12) PP, = a3, + a3,WF, + + + + a, c3, a36E, _, 3pPt-, -, -, -, -, E,., PSt a40 + a4,WF, + a42WS, -, + a, 3pp, -, + a, 4pS, -, + a,, Pl. + a4,, + c4, -', -, -, PF, a, + a,, WF, + a, 2WSt-, + a, 3"P, + a,4PS, + a,, Pr-, + as6E.,., +cS, () -, -, _, -, Et a6, + a6,WF, + a62WS, -, + a63PP, -, + a6,PS, + a6,Pl'-, + a66E, + c6, -, -, -, -,

form: in the equivalent or *t = Ao + A, z *, + -,

Ei

(7.13)

Etl PSt PFt PPt WSt [WFt wherez*t = , , , , , Sinceall observations of the sharevariables(WF, WS and Wl') sum to unity, cxact if includcd between in these three the variables' equations all exists are rnulticollinearity VAR. Hence,one of the shareequationsmust be omitted. We omit the cquation for the
(WPt). However, in Portugal 5, tile of as noted share chapter cstimation results expenditure is by invariant the share equation omitted and, whichever adding-up property, all are for be from the the estimates of the omitted can equation retrieved coctTicient estimates

first two the equations. of coefficient by estimating(7.13) and The statisticalquality of VAR I model can be accessed diagnostic Table 7.3 (t the the statistics. shows relevant estimation results ratios computing in brackets), the AIC and SBC criteria and a set of tcst statistics- adjustedR2, F statistic X2 statisticfor diagnostic functional form, tests of serial correlation, crror normality and and . includedin the basicunrestrictcd (p valuesin brackets)- for all equations heteroscedasticity VAR I structure.
'ibis basicspecification VAR (denoted by VAR 1) Is later subjected to modificationsdue of the unrestricted andthe exogeneity to the introduction assumption of dummyvariables concerning someof the regressors.

220

basic VAR I the Table 7.3.Estimationresultsandstatisticalperformance model of

EQUATIONS REGRESSORS WFj WFt., WSJ., PPI. 1 PSI-, pFt. ) Et. 1 Intercept 0.9518 (2.10) 0.6820 (1.44) -0.0712 (-1.14) 0.4620 (4.85) -0.3195 (-2.38) -0.0071 (-0.79) -0.2630 (-0.60) wSt -0.5617 (-1.14) -0.2932 (-0.57) 0.1062 (1.56) -0.4622 (-4.45) 0.3157 (2.16) 0.0000 (0.00) 0.8449 (1.77) PPt -2.0208 (-1.57) -1.9954 (. 1.48) 1.2525 (7.05) -0.1341 (-0.49) -0.4922 (-1.29) -0.0312 (-1.23) 1.9533 (1. )1

PSI
-1.0503 (-0.75) -0.7034 (-0.48) 0.2108 (1.10) 0.7453 (2.54) -0.2276 (-0.55) -0.0027 (-0.10) 0.7627 (0.57) 1

PF, -1.3501 (-1.32) -0.9772 (-0.78) 0.5033 (3.06) 0.0014 . (. 0.01) 0.0598 (0.17) -0.0318 (-1.35) 1.1671 (1.01)

E, 0.3106 . (-0.15) 1.3497 . (. 0.61) -0.4728 (. 1.62) 0.0602 (0.14) 0.6837 (1.09) 0.9315 (22.40) 1.1990 (0.59)

SLECTIONCRITERIA AND DIAGNOSTIC STATISTICS AIC SBC Adjusted

R2

61.24 56.58 0.789 17.88

58.82 54.15 0.824 22.12

31.99 27.33 0.771 16.11

29.77 25.11 0.562 6.79

34.10 29.44 0.612 8.09

18.12 13.45 0.991 511.34

F statistic

SerialCorrelation 0.78(0.38) 2.48(0.12) 0.73(0.39) 0.19(0.67) 0.37(0.54) 0.58(0.45) FunctionalForm 3.81(0.05) 8.66(0.00) 1.85(0.17) 7.39(0.01) 0.00(0.98) 0.18(0.67) Normality 1.02(0.60) 1.56(0.46) 0.56(0.75) 4.82(0.09) 11.65(0.00) 1.66(0.

0.39(0.53) 1.19(0.28) 1.33(0.25) 0.28(0.60) 0.03(0.87) 0.21(0.65) I leteroscedasticity

Thereseemsto be no statisticalevidenceof serial correlation in the error terms of However, the tests indicate VAR I. Hence,the lag-lengthselectedseemsto be adequate. in form functional in the the terms the and nonnality of of somecquations. error problems in the expenditure We areparticularlyinterested shareequations and for thesecquationsthe form These be functional are severe. problems may relatedto the assumption of problems for dummy to the omission and of variables, relevant all variables. of cndogencity

221

In the AIDS systemof equations specifiedin chapter5, the UK tourismdemandfor by the tourism shares (WS) Spain France (WF), France,Spainand Portugal,represented of dependent In be (WP), the the Portugal to only endogenous variables. assumed are and AIDS model, changesin these variablesare explained by a set of exogenousvariables PP) PS UK (PF, tourism include the tourism and and capita real per prices which in demand 3, In (E). tourism a why, context,prices may we explained chapter expenditure be consideredexogenous and, in chapter5, we explainedhow the multi-stagebudgeting (E) 'real AIDS the the sets variable per capita cxpcilditure" as approach underlying process demand determinant the shares. of a Within a VAR model framework, we are willing to question the prc-assumcd be basis However, there to theoretical the seems no obvious price variables. cxogcneityof for consideringthe UK real per capita expenditure on tourism, E, an endogcnous variable in On defined (7.13). be the to there the contrary, appears a of equations system within it (wcak-ly) If to theoretical considering empirical reasons and as cxogcnous. all numberof is bi-dircctional in VAR there endogcnous, a arc cause-cffect relationship model variables a if it is in the However, between to considerthat changes them. (fccdback) reasonable even UK real per capita expenditure affect the tourism shares of important UK holiday destinationssuch as France,Spain or Portugal, it does not seemto make much scnseto in in UK the these that their affect shares way which variations consumers allocate consider budget.The multi-stagebudgetallocationunderlyingthd mtionality of an expenditureshare bi-directional features. fact, In UK does such causc-effect consumers comprise not system (tourists)may allocatetheir incometo expenditure on tourism and other commoditiesprior item's knowledge in in (or of each current or of) any past share values their list of spite to acquisitions. Empirical evidencealso seemsto supportthis line of reasoning.It can be inferred from the estimationresultsof the equationfor Et depictedin table 7.3, that the 99% of this by lic lagged its No the model exclusively explained on value. own variable's variations is individually in In fact, F that the orjointly statistically significant. equation variable other of all explanatoryvariables,excluding1. test for thejoint significance a value of J,presents 0.86 implying that the hypothesisof these variables' coctTicientsbeitig zero cannot be VAR for WFt WSt Furthermore, the the estimation results of equations and rejected. indicatethat the laggedvalue of Et doesnot affect significantly the currenttourisin shares Spain. France or of

222

In a model, the existenceof a causal link betweenthe destinationsexpenditurc be investigated by (Et) UK (Wit) the expenditure can analysing real per capita and shares in disturbance between the term the tile the of model and stochastic error the relationships (d. ) Et Following Harris (1995, 4) data the of variable. g. p. p. generating process assumed be i1h to the sharevariableequation we consider Wit = ao + ajEt +
a2 Wit-, + U't

wherei=F, S, P

(7.14)

letting the underlyingd.g.p. be given by Assumingthat Et is a stochastic variable,and Et = PlEt-I + ct ; Pl<l and ct -* N(O,a2)

(7.15)

EVstands thatEV(uit,cs)=0 for all t, s (where if uit andctarenot correlated we canstate

for expected value,not to be confusedwith the time seriesvariable E). Then, it is possible is is, independent Et if it fixed, EV(Et, )-0. Et that that treat of uit were such uit as to I Icncc, we can treat Et as exogenous in termsof (17.14),with the current value of the UK being Granger-causc (Et) (7.14) is Wit. ]Equation to said capita per a expenditure real is E,, Wit Et being detennincd by the marginal on conditional with since model conditional is if E, PIE, (7.15) P2Wi,., EV(Ej, Note (7.15). )-0 that + as + refomiulated c,, uit = model -, Et, this variablecan only be is still valid. However,sincepastvaluesof Wit now deten-ninc in (7.14). The 1the conditional model exogenous weakly current value still considered in Granger but Wit Wit determine 1-34. the sense, since not past values of now causes As Hendry (1995, p.164) states,"a variable cannot be exogenous per so". A be interest. I'llereforc, if to with respect a set paranictcrs exogenous of can only of variable to be exogenous 1 is deemed with respectto the parameters ctj 0-0,1 s 2) in (7.14), the be is the (7.14) neglected and can conditional model model complete and marginal inferencc. Consequently, knowledge to valid sustain of the marginal model will sufficient forecasting improve the statistical and performanceof the conditional not significantly in be Only this can a set sense of variables considered cxogenous. model. As an indication of the possibility that Et can bc trcated as an cxogcnousvariabic in VAR I the to the of expenditure parameters share respect equations model, we run urith for (7.14) France, Spain Portugal the expenditure shares of and and regression regression (7.15) as a representation of the d.g.p. of 1. We retrieve tile residual seriesof thesefour for for, (7.14) the upt standing ust uFt namely residual series of respectively, regressions, , , France, Spain and Portugal, and ct standing for the residuals of (7.15). We then run individual regressions of the currentand laggedvalues(up to the fifth lag) of the residuals

223

fifth lag) (up The lagged S, (i F, P) to the the cstimation values of cl. and current on uit = indicate linear linking 90 the currentor significant no relationship all regressions resultsof lagged residuals of the conditional models to the current and lagged residualsof the indication knowledge We this that evidence statistical as an of the accept marginalmodel. forecasting improve does the statistical or performance of (he not marginal model for WSt, WFi, PPt, PSt in PFt (on Et. VAR Et treat the and equations so and as conditional 1) in VAR by (7.13). (weakly) the structure given variable cxogenous a The claim that in the relationshipsbetween the variable 'UK real per capita fccdback destination i=F, S) (Wil, (Et) the expenditure shares effects and cxpcnditure' be further VAR investigated in be this spccif"ication, can using the particular absent might (1969). The Grangcr-causality by Granger is known test test proposed a well and causality feedback between the to existence of significant establish effects used procedure widely for is A in VAR lags test to tile causality specifications. related whether of a variables in In the equation significant of are statistically another variable. other words, and variable in Griffiths (1992, definition 695), "a IT be the to et al. p. given 1arlable yll said using by y2t if current and past information on ),2t help Improlv forecasts In Granger-caused A" Therefore,if the laggedvaluesof y2t do not licip to iniprovc the forecastsof yl I, that is, if the laggedvaluesof y2t are statisticallyinsigniricantin the reducedform equationfor 7 does Granger-cause theny2t yIt. not y 11, The multivariategeneralisation of the Granger-causality conceptis known as "block

it be for Granger-causality" can used and establishingif one or more variablesin a VAR integrate Ilic the testing of the null not group of cndogenous should variables. or should hypothesisthat the coefficientsof a subsetof variablesin a VAR are zero using the log. likelihood ratio (LR) statistic, is known as block Granger non-causalitytest. This test lagged the to of extent measure which a statistical valuesof a set of variables(say provides in important 13t), predictinganothersetof variables(say Wit), oncelaggedvaluesof the are latter (say Wit.,) are includedin the model.The LR test for block Grangernon-causality of hypothesis Et tests the that the cocfficients,of E1.1 null the variable which arc null in the for WFt, WSt,PPt,PStandPFtpresents block equations the statisticvalueof)? (5) - 10.578. As the critical valuesat the 5% level is 11.071,the null hypothesis cannotbe rejected.

7Note that Granger's but ratherIs based of causalitydoesnot imply a causc-effect concept relationship oil the Charemza As Deadman (1997,p. 165)states, "predictability". and "in econometrics, of concept causalityhas moreon thelines of 'topredict'rather than 'to produce'." a meaning

224

Accordingly, we now considerEt not to integratethe set of cndogenous variablesin the set of pre-detcrmined the VAR model.Rather,it integrates and/orcxogcnous variables. Hence, we reformulatethe basic model to include this assumptionand estimatea new AIC VAR 11). The VAR (denoted the andSBC selectioncriteria andthe samesct versionof the statisticalquality of the previousmodel arc now usedto of test statisticsfor accessing basic VAR in Table 7.4. the this the of version of quality second and presented access Table 7.4. AIC andSBC selectioncriteria anddiagnostictestsfor tile VAR 11 model Selection criteria ic d diagnost an testes AIC SBC Adjusted
R2

EQUATIONS WFt 61.39 56.73 0.792 18.11 wSt 58.83 54.17 0.825 22.15 Pill 33.58 28.92 0.795 18.47 I)S, 30.02 25.36 0.570 6.97 1)1'1 34.62 29.96 0.656 8.53

F statistic

Correlation 0.83(0.38) 2.16(0.14) 0.27(0.50) 0.46(0.67) 0.47(0.49) Serial Functional Form 3.28(0.07) 8.49(0.00) 0.70(0.40) 5.56(0.02) 0.02(0.90) Normality 0.86(0.65) 1.55(0.46) 0.05(0.97) 3.04(0.22) 6.49(0.0.1)

lietcrosccdasticity0.42(0.52) 1.23(0.27) 0.84(0.36) 0.21(0.65) 0.00(0.95) I L. I I ---The selectioncriteria and the diagnostictestsindicatethis ncw spccificationof the VAR to be, overall, statisticallymore adequate than the previousone. I lowevcr, problems form functional for the the sharevariablesstill remain. the equations of with As explained in previous chapters,there is reasonto believe that political and in (change 1970s in Portugal Spain the events of political regimes and tile oil and cconomic have in VAR. In the time the the affected path of may variables included addition, crises) in break in 5, detected, the variable 'real per capitaexpenditure'(Et) a structural chapter we (linked to Spainand Portugal'sintegrationprocessin the EU) and modcllcd it with a slope dummy separating this variable's coefficient into two. The 1970seventsare takcn into dummy by DI, variable a which takesthe valueof unity in the period 1974adding account 1981 and zero otherwise.Since the modelling of slope-cocfficicnts has not been fully either theoretically or empirically within the VAR approach,to mimic the addressed detected break dummy previously, we use variablesD2 (taking the value of unity structural

225

in the period 1982-1997 andzero otherwise)and D3 (taking the value of unity in the period integration 1989-1997 the to separate otherwise) processof Spainand Portugalin and zero the EU into two periods:the integrationperiod (1982-1988)and the post-integration period leading to tile modelling of a (1989-1997).Thesedummiesare assumed to be exogenous, This basic VAR VAR is third the variables. version exogenous of system with partial denoted VAR Ill. Table 7.5 shows the AIC and SBC criteria and a set of diagnostic the statisticalquality of the VAR Ill model. statisticsto access diagnostic Table 7.5.AIC andSBC_selection testsfor the VAR III model criteria and Selection criteria testes anddiagnostic AIC -SBC R2 Adjusted F statistic EQUATIONS WFt 69.79 63.13 0.892 25.87 wSt 65.69 59.03 0.899 27.64 ppt 32.27 25.61 0.788 12.16 I)St 27.28 20.62 0.508 4.10

PF,
33,69 27.02 0.623 5.96

Correlation 0.35(0.55)0.02(0.90)0.04(0.85)0.21(0.64) 0.10(0.66) Serial Form Functional Normality 0.04(0.84)0.21(0.64)0.57(0.45)7.24(0.01) 0.14(0.71)

1.28(0.53)0.63(0.73)0.08(0.96)3.50(0.17)11.64(0.03) 1 Ileteroscedasticity0.10(0.75)0.01(0.92)0.68(0.41)0.26(0.61) 0.03(0.86) A MMMMMM ____j

The resultsof Table7.5 indicatethat the statisticalquality of VAR III ovcr-pcrfomis II, I VAR VAR particularly, with respectto the cxpcnditu and those of rc sharecquations interested in. Moreover, LR block for the Granger tile test we arc relationships arc which Et VAR III (which includes dummy the tile three performed of variable on non-causality VAR including the (not the I test of results similar confirms the performed on variables), dummy variables).In fact, the LR test presentsnow the statistic value of')? (5) - 8.376, is*well below 5% (11.071), further hypothesis the critical tile value supporting null which Et., in the statistically of are that null the block equations coefficients of the VAR. For a clear view of this and other statisticaltests performed,we presentTable 7.6 from hypothesis involving the variablesincluded testing the obtained results shows which in the VAR models. The tests concern the significance of the intercept, the joint dummy DI, D2 variables and D3 and the Grangerblock non-causalityof significanceof

226

for 7.6 hypothesis first Table Et-1. The the presents various of null cachtest column variable VAR "unrcstrictcd" hypotlicsis (U), the the the model under entering variables shows and hypothesis (R) (for "restricted" the null simplicity, the timc subscriptsarc and under is from In the the variables set of endogenous entering model separated omitted). eachcase, the set of deterministic componentsand cxogenousvariables by the symbol W. Ilic likelihood function for (ML) the the the value maximum of presents column second In log-likchhood the R) (1, third the column alternatives. and restricted ratio unrestricted is LR distributed is Under the the statistic asymptotically null, as X2with statistic computed. is hypothesis The freedom (i) degrees to the number of restrictions. null equal rcjcctcd if of the LR statisticis largerthan the relevantcritical valueof the X2distribution. Table 7.6. LR testsfor restrictivehypothesis on the interceptandvariablesof the VAR
Critical

Model
Non-significanceof intercept (INT) 110: V: WF WS PPPSPF E& INT R: W WSPPPSPFE& of dummy variables DI D2 D3 I lo: Non-slignificance U: WF WS PPPSPF E& DI D2 D3 INT R: WF WS PPPSPF E& INT Ifflock non-causalityof E with no dummy variables 110:

NIL

Lit -+ 72

(i)

value (5%)

Result

351.68 342.05

? (6)-18.05 )

1259 .

ltcjcctcd

1 385.40 351.68,

2 X (18)-67.43 28.87

Rqjccted

...... tj: WF WSPPPSPFE& INT

R: WF WS PP PS PF &E INT

1346.39 351.68

Not
X (5)-10.58
2

11.07

rcj cctcd

Block non-causalityof E with dummy variables 110: U: WF WS PPPSPF E& DI D2 D3 INT R: W WS PPPSPF &E DI D2 D3 INT 385.40 381.21 )? (5)-8.37

Not
11.07 rcj cctcd

Although VAR III is statistically more robust than VAR 1, the lattcr is a more in being fon-ncr is the restricted any way, while not model, a spcciricmodel, being a general Tliereforc, on exogenous variables. conditioned system we believeit is interestingto partial

227

in features UK for both tourism the the accessing main purposes, of comparison models use demand for France, Spain and Portugal, Within the reduccd-fonn VAR spccification.

A reduced-fonnVAR specificationis not the ideal meansof conductinga rcliablc long-run The the the relationships variables. existing among structural of economicanalysis is interpretable first, to unlike model produce estimation an results a-theoretical are reasons interpretation limits the theory; the second, economic of the structural of economic vAthin from if be identificd is the tile underlying structural model can only possible parameters included in is likely lag-structure VAR finally, VAR; the the to reduccd-form rcduced-form imprecision in the coefflicients' estimates and lead to over-parameterisation causing long-run if Yet, the the the main purposeof an parameters. of obscuring economicmeaning forecasting, is VAR has the reduccd-fonn exercise modelling specification econometric irrelevant instance, is in For the the the of structural estimation model advantages. forecastingprocesssince, althoughunder-idcntiricd,a VAR model can provide unbiased justify forecasts. These VAR I VAR the the use of reasons reduccd-form and icicnt and cff for forecasting in specifications ratherthaneconomicanalysispurposes.
I ANDVAR I [I moncis ABILITY OFTHEREDUCED-FORNIVAR 7.3.4. TjIE FORECASTING Except for the choice of the variables, VAR I mod, el does not bear any otlicr a lWort

Thus, VAR. However, "pure" this can we qualify model or restriction. as a assumption theoreticaland empirical assumptions VAR III model incorporates suchas conditioningtile dummy be Et three variables, which can considered constraintsto the more model on and interesting be forecasts It VAR. for to France, tourism the would obtain of shares general Spain and Portugalfrom thesetwo models,and comparetheir forecastingability. Fr this for in-sample including 1969-1993, tile the estimated tile are now models period purpose,
.0

in the forecastingperiod. Tables 7.7,7.8. and 7.9 report tile actual last four observations forecast France, Spain Portugal the respectively, expenditure of, shares values of and and VAR III I It forecast VAR tile and models. also reports with errors,and n number obtained (mean MAE; MSE; of absolute errors mean of statistics squared errors root of summary for RMSE) forecasting These the evaluating errors models' squared accuracy. mean of defined in 8, forecast are chapter criteria section8.3.2. evaluation

228

UK France for Table 7.7: Forecasting the expenditure share of results Actual values VAR 1 Forecast error Forecast VAR Forecast error Forecast 1994 0.39700 0.34422 0.052780 0.34274 0.054261 1995 0.38540 0.36840 0.017006 0.44818 1996 0.38967 0.39676 -0.007086 0.41524 1997 0.40481 0.42782 -0.023007 0.40121 0.003595

-0.062778 -0.025567 SUMMARY STATISTICSFOR RESIDUAL AND FOItECAST ERRORS Estimationperiod:(1970-1993) 0.015321 0.000315 0.017734 0.008510 I 1 0.000119 0.010917

Forccastperiod: (1994-1997) 0.024970 0.000914 0.030226 0.036550 0.001888 0.043451

MAE VAR 1 MSE RMSE MAE VAR III MSE RMSE

Table 7.8: Forecasting resultsfor the UK expenditure shareof Spain 1994 Actual values VAR 1 Forecast error Forecast VAR Forecast error Forecast 0.51857 0.57000 -0.051431 0.56593 1995 0.52625 0.54972 -0.023469 0.45798 0.068264 1996 0.52292 0.52281 0.000108 0.48770 0.035218 1997 0.50691 0.49058 0.016328 0.50172 0.005191

-0.047361 SUMMARY STATISTICSFOR RESIDUAL AND FORECASTERRORS MAE Estimationperiod:(1970-1993) 0.016475 0.000301 0.019780 0.010340 0.000192 0.013847

Forecast period:(1994-1997) 0.022834 0.000866 0.029422 0.039009 0.002043 0.045195

VAR I

MSE RMSE MAE

VAR III

MSE RMSE

229

Portugal UK for the Table 7.9: Forecasting share expenditure of results 1994 Actual values VAR 11 Forecast error Forecast VAR Forecast error wmsw MAE VAR MSE RMSE MAE VAR MSE RMSE Forecast 0.08443 0.08578 -0.001349 0.09133 1995 0.08835 0.08189 0.006462 0.09384 1996 0.08741 0.08043 0.006977 0.09706 1997 0.08828 0.08160 0.006679 0.09797

-0.008786 -0.005488 -0.009651 -0.006900 SUMMARY STATISTICSFOR RESIDUAL AND FORECASTERRORS Estimationperiod:(1970-1993) 0.008054 0.000088 0.009365 0.006276 0.000060 0.007770 Forccastpcriod: (1994-1997) 0.005367 0.000034 0.005850 0.007706 0.000062 0.007875

For the estimationperiod (1969-1993) and for all threecxpcnditureshareequations, than VAR I model.Nevertheless, VAR III model is shownto be statisticallymore accurate indicate be VAR better I to the forecast critcria quality model n errors and all the differences between Yet, 111. VAR device the the two setsof summary than forecasting as have fairly 0.02), both (never to we can consider models exceeding statistics are small demand for UK in forecasting the tourism. accuracy similar The fact that the variablesincluded in VAR I and VAR III modelsare I(I) implies forecasting based tests and procedures on them arc strictly valid that estimation,statistical if, and only if, cointegratingrelationship(s)link the variables.This meansthat we must long-run tests can whether which establish steadystate rclationship(s) conduct statistical (1995, Harris 76) Indeed, implication VARs. "the that the out, as p. points within exist lead least to spurious regressions unless can one cointegrating at nonstalionaryvariables is is that someform oftestingfor cointegration almostmandalorjP. means vector present However, the 'mandatory' testing for cointegrationmust bc perfornicd within a basis. Certainly, framework and not on an equation-by-cquation given systemof equations if for the single equation approach more than one cointcgratingvcctor the consequences framework the cointegration analysis within of use single equation exists, earlier

230

(performed on the ARDL models of chapter 4) must give way to tile dcten-nination of system of equations approach. Therefore, the in becomes inevitable Johansen the test step procedure next of cointegration rank demand for UK Spain France, tourism the the of and empirical analysis proceeding cointegration rank within a VAR Portugal, within a vector autoregressiveframework. Besides establishing the procedure for cointegration rank testing, tile Johansen for idcntirication, hypothesis framework testing estimation and a general provides approach in cointegrated systems (see, for example, Johansen, 1988 and Johansen and Jusclius, 1990). Hence, given a cointegrated VAR system, this approach provides a method for identifying the long-run structural relationships underlying the unrestricted reduccd-forni of hypothesis involving its kinds testing VAR, of structural parameters. all and permits tile This is of paramount importance for validly using a VAR specification for both forecasting and economic analysis purposes. Howcver, the "empirical Wntification" process suggested by Johansen to exactlyidentify the structural long-run coefficients can prove inadequate, particularly in cconornic

In testable tile these theory scrisible and cases, restrictions. strong, provides where contexts be to exact- and ovcr-idcntifyillg restrictions obtained subject must vectors cointegrating from economic theory and other relevant a priori information, rather than being subject to does which not consider the theoretical and cillpirical some normalisalion process 8 framework within which the phenomenon under analysis cvolves. Using the UK labour (199 Mizon 1) I larris (1995) Clemcnts and an example, as adopts tile same market model of line of reasoning stating that "whal is becoming Increasingly obvious Is the need to ensure by forms basis fi)r information Imposing the economic motivated arguments that prior (p. is 117). This important in the other way around' particularly not and a restrictions, demand context involving n tourism shares,n tourism prices and the origin's per capita real by In the this number of steady case, state equilibrium relationships predicted cxpcnditure. incorporated (n-1) in is, the share equations an orthodox AIDS model. theory precisely, Therefore, if theory is right, the cointegration tests involving a VAR of tile UK (with or without dummy tourism demand with variables WFt, WSt, PPt, PSt, PFt and E-1 long-run in indicate the that the model only relevant relationships existing variables) must by Furthermore, the two the cxact-idcritification those share equations. established are
' For a comparisonof Johansen's'empirical identification' processwith the AIDS thcorctical exactin a VAR modelof budgetshares, identifyingrestrictions seePesaran and Shin(1997).

231

is form that the their that of confirm steady state must the coefficients structural of process VAR III in I VAR Consequently, tile in AIDS models, either or model. equations a static identify find to the two vectors and structuralparameters to cointegrating exactly we expect identifying for the the share those used non-nalisation process that of match restrictions with long-run is If the the then AIDS this case, we can subject model. equationsof a static homogeneity hypothesis further testing, to symmetry, and such as relationships equilibrium bchaviour for basis the the consumcr validation of principalsof and contributean empirical VAR This framework, AIDS would complete specification. using a system theory within an tile in theoretical, evidence of statistical empirical and this provide and chapter our goal for basis in the AIDS analysis the a sound of and reliable offering system of robustness destinations' between tourism shares,their tourism long-run equilibrium relationships budget. tourism pricesanda specificorigin's In the next section, we use the Johansenprocedureto establish the number of
for VAR III I in VAR tile models and modcls specify rcstrictcd and vectors cointegrating interest. We for long-run then the the of use models restricted coefficients estimation of in forecasting the those obtaincd compare, with results purposes and Cconomic analysis and

this section.
7.4. THE JOHANSEN PROCEDURE Conditioninga model on the (weak) exogencityof a sub-sctof variablesimplies a dimension dimension (nj< from to the n). an n-variables a ni-variables system of reduction infon-nation is loss holds, this that (weak) such reduction without of ensures if exogeneity in interest. Therefore, the terminology the of gcncral-to-spcciric, the of parameters on is VAR III is I VAR the the more general spccific system and a of case general approach, imposing dummy including 'structure' to the In and/or variables a general addition, model. implies VAR, restrictions, other specifications and over-identifying exactly via unrestricted Whichever 'more VAR I the the general' cases of special model. constitute also which inrcrcncc based if the are present, statistical oil variables spccificationsare, nonstationary by least-onc This be to of at cointcgratingvector. evidence sanctioned evidence them needs Johansen's by be the reducedrank regression method. can provided likelihood is The Johansen approach applied to both the gcneral VAR I maximum besides finding In how III VAR this way, models. many cointegratingvectors and specific

232

imposed their in the test on relevance of several constraints also exist eachmodel, we can legitimacy the variables. the models,on a setof exogenous of conditioning and parameters, demand VAR UK to the tile The applicationof the Johansen procedure modelsof for tourismincludesthe following aspects: 1. 2. 3. 9 included in VAR. the Testingthe orderof integrationof eachvariable lag-lengthof the VAR in order to assure Testingthe appropriate white noiseerrors. Determining whether the equationsin the system should include deterministic intercept time trend. a and/or an as such components in the systemshouldincludedummyvariables. Determiningwhetherthe equations how find Testingfor reduced to out many cointegratingvectorsexist. rank ill Testingfor exactly- andover-identifyingrestrictionson the equations'parameters VAR model. the cointegrated determining VAR the the All thesestepsensure and possibility of a correctspecificationof in for It the testing cointcgrating vectors. cmbcddcd also allows the structuralrelationships involving hypotheses single and cross-equationparameters, and permits the theoretical ill Therefore, follow long-run interpretation these the coeff"icicnts. we steps of economic the the structural componentsof cointcgrating to and estimate existence establish order framework for in UK demand for the tourism r nee, vectorswithin a vectorautoregressivc SpainandPortugal. Let zt be a vector of n potentially endogcnous variables.Then, it is possible to VAR involving 1995, (see I larris, 77): to that up p-lags such p. modelzt asan unrestricted IN(0,7, ) (7.16) Apz, Alz, + + +... u, u, z, = - -p -, is Ai (n is (n 1) the and each of vector a x n) matrix of paramctcrs. x wherezt a The VAR model in (7.16) can be reformulatedas a vector crror-correction(VEC) modelsuchthat: Azt = rjAzt-1 +... + rP_jAzt_P'j + FIzt-P+ ut i=p-I A), A, ri where = -(I -... and 1-1 = -(I - A, Ap). (7.17)

4. 5. 6.

information both long-run the This specification on shortcoeffliciclits,via tile and provides rT, it is for r basis implementation the the and and of the estimatesof, respectively,
9 Although it is possibleto find cointegrated relationship(s) when there Is a mix of 1(0),1(1) variables,the is designed handle I(I) is to Johansen variables approach and more reliable under these standard circumstances.

233

Johansenprocedure. Matrix 11 can also be described as 11 = ctp', wlicrc ot represents the is long-run P 0'74. the that, matrix of such and coefficients adjustment matrix speed of p implicit in (7.17) represents up to (n-1) cointcgrating vectors in the multivariate VAR. These cointegrating vectors define the relationships which ensure that tile variables in 74 converge to their long-run solutions. If the variables in zt are nonstationary I(1) variables then, Flzt.p in (7.17), containing

for be long-run stationary ut - l(O) to be a white noisc. Illis occurs must the relationships, is, rl (n-1) in 0 has 5 there that when are vectors r: rank, cointcgrating = ctp' reduced when (meaningthat r columns in P form r linearly independent stationarycombinationsof tile (n-r) for in llcnce, ) testing nonstationary vectors. cointcgration zt, alongside variables independent linearly r] 11). in determining (tile to the of r number columns rank of amounts By the same token, testing for cointcgration amounts to determining if the last (n-r) fact, (n-1) In found, infcr 5 r: cointcgrating once vectors a are zero. of arc we can columns (n-r) last last (n-r) columnsof 0 arc tile the are zero and, consequently, of (x columns that do (7.17). In This that they the this case,we not enter specification means nonstationary. dimension (n P to x r). a and can reduce As noted by Chambers(1993), in many cases,and in particular with demand h)potheses interest is long-run "thefocus to the on a set of of relating structure of systems, independent fitted is to the empirical of shorl-run quite any namics which (1), the model, in demand UK Likewise, VAR interest 727). the the tourism case of system, our moder'(p. is focusedon the structuralequations upon which parameter restrictionsarc to be illiposcd, hypotheses theoretical aboutthe long-runequilibrium relationships. testing In sections7.3.1. and 7.3.2. we established that all variablesinvolved in the VAR
lag-length I(I) We the that appropriate and was were one. considcrcd that an system intercept, representing the average value of the left-hand side variables when all tile right. hand side variables are set to zero, was a relevant deterministic component and should be long-run (exclusively) form integrate the to of the model. In addition, we assumed added be included in the specification, as no significant trend cotrends time should that no in detected Finally, the time series. we found sufficient empirical movements were for the 1 treating possibility of not rejecting variable as an cxogc1lousvariable, evidence dummy DI, for integrating D2 D3 variables and as statistically relevant rcgressors. and Hence, we end up with VAR spccifications such as tile VAR I and VAR III models.

234

Theory predicts that there should be two cointcgrating relationships among the hypothesis III. This by Johansen in VAR I included VAR tested the was using and variables Johansen The for determining the test number of cointegrating vectors. reduccd-rank for the matrixes Ot and using reduced rank estimates consists of obtaining procedure i, X^ %^ that their > > eigenvalues and corresponding n provides regression approach 2.... J. V (l cigcnvectors -- 1-1 Those r elements in V, which determine the linear

Thercrore, the testing the cointcgratcd arc vectors. relationships, of stationary combinations hypothesis that there are at most r cointegrating vectors (and, liencc, n-r nonstationary Xi 11o: implies the testing =0; null relationships) i= r+l,..., n.

known be This hypothesis tested statistic using a as the cigctivaluctracc statistic can n-I) againstthe alternativethat r :q+ Another test for the numberof cointcgratingvectorsis known as the maximuni cigcnvaltic which teststhe null that rq ) (% the that there tests null arc r=q cointcgrating vectors againstthe which statistic .... for Asymptotic in these that critical values exist. r=q+I statistics arc providcd alternative for time seriesanalysis. used packages computer most The resultsobtainedfrom applying Johansen's reducedrank test to VAR I and III
first in 7.10. Table The the column shows eigenvalucs associated with are given models from highest lowest, 1(l) X"'. to the to ordered tile variables, necessary calculate of each hypotheses The be %t,. the to tested, column shows second various statistics. starting and e in (r VAR 1, in the the case =0 or n-r =6 case of and -5 r-0 or n-r with no cointegration by increasing followed 'I'lic following III) VAR numbers of cointcgrating vectors. and of Xtru. 5% 10% the the estimated and statistics with present rcspectivc and columns last SBC The the column presents model selection criterion. critical values.

(q = 0,1,,

235

Table 7.10:Testsof the cointegration rank for VAR I andVAR III models
g

Eigen values

110 r r statistic

Xn.,, critical

?, Uwe

4act

critical al Slic 0 1O(Y. 0/0 0o

1 5% 10% VAR I 40.53 37.65 34.40 28.27 22.04 15.87 9.16 31.73 25.80 19.86 13.81 7.53

statistic

5%

X1=0.9201 r=0 X2=0.7489 r=I 'I m, -0.5889 r=2 14=0.3273 r=3 8 r=4 X5=0.183 X6=0.0"8 r=5 Xi=0.8798 r=0 X2=0.7734 r=I X3-0.5833 r=2 A4=0.2767 r=3 %j=0.2489 mr=4
;k3

61 70.76 n-r = 38.69 =5 n-r n-r =4 n-r =3 n-r =2 n-r =1 n-r =5 n-r =4 n-r =3 24.89 11.10 5.69 2.82 59.31 41.57 24.51

153.96 102.56 97.87 274.71 83.20 44.50 19.61 8.51 2.82 75.98 53.48 34.87 20.18 9.16 71.81 290.09 49.95 292.78 31.93 291.89 17.88 287.45 7.53 283.63

VAR III 46.77 43.80 40.91 34.51 27.82 20.63 38.03 31.73 25.27 18.24

142.48 119.77 114.38 265.82 83.16 41.59 17.08 8.01 90.60 63.10 39.94 20.63 85.34 272.15 59.23 272.94 36.84 268.54 18.24

9.07 n-r =2 1 1_ 8.01 = n-r

%trw,: For VAR I model,at the 5% significancelevel, both the X .....and, statisticsgivc highcr two to the vectors corresponding cointegrating of cigcnvalucsattachedto evidence Spain. fact, In both for France and statisticsassociated with the null or the shareequations r-0 and r=I hypotheses (statistic these value > critical value) but cannotreject r reject
2 (statistic value < critical value). The SBC criterion further supports these findings by two the contains cointcgratcd relationships. which model selecting

For VAR III model, at the 5% significance level, the

statistic suggeststhe

does hypothesis the the two vectors while cointegrating statistic of not reject existence in VAR III This disagrcenictit in vector the tests model. cointegrating one apparent of only is for cointegrating rank not uncommon,particularly in cascsof small sarnpicsand added dummy variables.So, a choicemust be doneat this stage.As Harris (1995, p.89) suggcsts I'll Is importantto useany additional information that can support the choiceqf r" and wc do haveadditionalinformation that can supportthe choice of r. We haveone tcst, tile jinx *%,
in favour the existence of rejecting only clearly one of two cointcgrating vectors; statistic, VAR have SBC III the the selecting criterion model with two cointegrating vectors. We we have theory suggesting the existence of two and not one long-run relationships which is

236

in by both test the and selection statistics criteria more general supported unmistakably Hence, VAR I _: that theory evidence against given model. prediction seems of sp'ecification by settingr=2 for both VAR I andVAR III models. be ratherweak,we proceed -to Considering that both VAR I and VAR III modelscomply with theory with respect identify longthe we must now predicted, of cointegrating vectors structural the number _toIn do follow Pcsaran to the two order vectors. so, of cointegrating we and coefficients run implicit in (1997) the theoretical the exact-identifying restrictions and use Shin AIDS the model. static of a shareequations -Spc6ificationof Considerthe following notationfor the matrix of the two cointcgratingvectorsof -VAIz I model for variablesWFt WSt PPtPStPFtEt and intercept,
Pll P12 P21 P22 P31 P32 041 P42 P51 P52 P61 P62 P71 1 P72

VAR P III for WFI, two the for the vectors of cointegrating of matrix model variables ,and. intercept PFt, Et, DI, PSt, D2, D3 PPI, and WSt, ,.,
Pil P12 P21 P22 P31 P32 P41 P42 P5, P52 P61 P62 P71 P72 , P8I P82 p9l P92,

P01 P:
02

the the exactly-identify cointegrating theoretical restrictions which vectors as share fbc , AIDS specificationin both VAR I andVAR III modelsarc given by: of an cquations
P12 HAn)s P21 0 P22 01

identified VAR I VAR III the two cointegrating vectors of estimates of and Me exactly , in in (asymptotic brackets). Table 7.11 t The third given ratios are vector, models to the shareequationfor Portugal,is retrievedfrom the estimatcs the of other corresponding AIDS the the by property of adding-up model. applying two

237

Table 7.11: Long-runcoefficientestimates of the exactly-identiricdshareequations

VAR I Variables Vector I (WF) -1 0 -0.4965 (-0.86) 0.8214 (2.09) -0.2244 (-0.48) -0.0684 (-1.12) Vector 2 (WS) 0 -1 0.1781 (0.60) -0.5937 (-2.97) 0.3119 (1.20) 0.0159 (0.51) 0.3184 (0.95) -0.2277 (-0.97) -0.0875 (-0.33) 0.0525 (1.45) 1 Vector 3 (WP) Vector I (WF) .1 0 -0.0027 (-0.05) 0.2256 (4.68) -0.3394 (-3.59) 0.0153 (2.33) 0.0380 (5.27) -0.0565 (-3.65) 0.0574 (5.17) 0.8309 (2.20) 0.3818 (1.97) 1 -0.2127 0.3687 (16.26) VAR III Vector 2 (WS) 0 -1 0.1090 (1.69) -0.3075 (-5.22) 0.3044 (2.57) -0.0183 (2.29) -0.0154 (-1.74) 0.0528 (2.78) -0.0590 (4.34) 0.5443 (19.50) -0.1062 (-3.28) 0.0820 (2.90) 0.0350 (0.56) 0.0030 (0.78) -0.0226 (-5.18) 0.0037 (0.41) 0.0016 (0.24) l 0.0870 (6.33) Vector 3 (WP)

WF WS pp PS IT E DI D2 D3 INT

There is a sharp difference between the estimatesof the long-run cocfricicnts in III both VAR I VAR with models, and magnitude with and statistical obtained ignificancc and in the signs which should accord with theoretical expectations.For ., for in Portugal, VAR indicate I the the share equation estimates of model nil instance, insignificant be level, in 5% for to the the at statistically and share equations co,efficicrits intercept Spain be Spain the the to of only price and significant at the same and r.rance in for intercept Portugal, In the the the share own-price equation and addition, level. have and 'wrong' signs. Both theseestimatesand thq interceptestimatein the Cstirriatcs implausiblemagnitudes for Francepresent cquation However,in VAR I model, the price of Portugalin the equationfor Franccand the for in Portugal ditTercntfrom zero, as France the equation are not signiricantly of price in our conjecture about the competitive behaviour of these ncighbouring predicted is feature This supportedby the corresponding same destinations. estimates obtainedwith far is Yet, III this as as similarity goesbetweenthe estimationresultsof the VAR model.

238

III VAR The model arc overall statistically significant, estimationresultsof two models'.
infon-nation the and give accurate and and magnitudes expected signs plausible prcsent by dummy for demand UK the the the the variables afTcct events represented ways about

indicate DI in destinations. Indeed, the three that the political the coefficients of tourism in Portugal Spain and the oil crises,affected significantly and negativelythe and changes favouring Franceinsteadduring the period 1974LJK tourism flows of thesedestinations, 198 1. The coefficientsof D2 indicatethat the integrationprocessof Spainand Portugalin from divert France Iberian flows UK to EU the to tourism peninsula,although caused the favouring Spain more than Portugal. The coefficients of D3, representingthe post. integration period, indicate a recoveryof the sharefor Franceat tile expenseof Spain's 10 declines. for In Portugal the the same period, share steadily shows which an share increase, althoughnot statisticallysignificant. Theory suggests to the featuresof homogeneityand that restrictionscorresponding should hold within a systemof demandequations,reflecting the rationality of _symmetry behaviour. In addition, according to our own assumption of competitive consumer destinations, in between France tourism to price urism changes neighbouring bchaviour for in UK demand Portugal (France), tourism the while price should not affect (portugal) changes by imposing hypotheses These tested the appropriate Portugal. were in restrictionson both VAR I and VAR III models and using the LR statistic to evaluatetile result. If the LR in Spainshouldaffect significantly the UK demandfor tourism both in Franccand

is be than the the appropriate critical restricted value, cannot smaller model statistic do not hold. Table 7.12 showsthe testsresults for the Otherwise, the restrictions rejected. between for the effects of null cross-price share equations Franceand Portugal, hypothesis both hypotheses these symmetry and sets and of simultaneously, homogeneity The testsindicatethat the set of hypotheses cannotbe rejectedin either VAR I or both Therefore, III VAR models. models comply with the theoretical restrictions of between the symmetry and with our and assumption of cross-price null cffects homogeneity for France Portugal. and share equations

Tunnelin 1994may alsohavecontributed 10 lie openingof the Channel to this result.

239

in VAR I and VARIII models Table 7.12:Testsfor over-identifyinghypothesis

Hypothesis Null cross-price.effects Homogeneity & symmetry Homogeneity & symmetry & [null cross-price effects Model VAR I VAR 11 VAR I VAR 11 VAR I AR If II LR-+X2(i) x2(2)=1.53 x2(2)=0.36 X2(3)-3.31 x2(3)=7.28 X2(4)=4.01 2(4)=7.53 x 9.49 I 7.81 5% Critical value 5.99

for the two cointegrating The estimates vectorsof VAR I and VAR III modelsunder in identifying Table 7.13 considered above are given over restrictions and the exactbefore, in brackets). As the third vector correspondingto the share t (asymptotic ratios is from by for Portugal, the tile two estimates of other vectors retrieved applying quation c, is improvement There in AIDS the tile a general model. of property adding-up the in VAR I that they the sense arc now statistically more of model, relevant results estimation incongruent before. However, the than model still presents some plausible more seem and in for for Portugal, the share equation as, example, a such positive own-pricc results irrelevance intercept. Moreover, the the of coefficients a negative and or coefficicnt in for France and Spain, (UK Et the capita expenditure) equations real per variable ill tile equation this the statistical significance cocfflcicnt of variable's with compounded found Portugal empirical questionable, given seems evidence previouslywith tile AIDS for 6. Hence, believe 5 is VAR I ideal that tile we and, chapters model of not models for explaining the UK tourism demandbehaviourfor thesedestinations.in specification VAR Ill are overall statistically robust, theoretically the of results estimation contrast, indicating for this plausible, empirically model as an adequate vehicle and consistent in FranceSpainand Portugal. UK tourismdemand the analysing

240

Table 7.13: Long-run estimates of the shareequationsunder exact- and over-identifying for homogeneity, France Portugal null cross-pricc and symmetry effects and of rcstrictions
VAR I Variables Vector I
(WF)

VAR III Vector3 (WP) Vector I (WF) Vector 2 (WS) 0 0 0.1308 (1.76) -0.1308 (-1.76) 0 0.0361 (3.53) 0 0.2891 (4.79) -0.2891 (-4.79) 0.0091 (1.16) 0.0354 (3.83) -0.0373 (. 2.36) 0.0429 (4.56) 1 0.0895 (5.80) -0.3785 (-5.97) 0.2891 (4.79) -0.0121 (-1.38) -0.0115 (-1.13) 0.0360 (2.02) -0.4184 (-3.95) 0.5230 (16.96) -0.0895 (-5.80) 0.0895 (5.80) 0 0.0030 (0.95) -0.0239 (-7.05) 0.0013 (0.20) -0.0011 (-0.27) 0.092 (11.69) Vector3 (WI))

Vector2 (WS) 0 -1 -0.1308 (-1.76) -0.4729 (4.97) 0.6037 (6.51) -0.0001 (-0.01)

WF WS pp

-1 0 0 0.6037 (6.51) -0.6037 (-6.51) -0.0362 (-3.40)

PS

PF
E D1 D2

d D3 NT T INT I 0.6152 (10.77) 0.4916 (7.12) -0.1068 (-1.76)

0.3849 (13.17)

A detailed investigation of these results requires the analysis of the relevant We estimate the expenditure, own- and cross-pricc uncompensated values. clasticity basedon the long-run coefficicnts estimates VAR III the tourism of of shares, clasticities homogcncity,symnictry and the over-identifying restrictions of exactand under model, between for France Portugal. The the share coefficients equations and cross-price rjull formulae in 5. Table 7.14 prcsclitsthese the given chapter computed using are clasticities for comparison long-run elasticities the corresponding purposes, estimates and, clasticities AIDS 5 6 for (the the the models chapter of and with only period second obtained cstimatcs behaviour dcmand ror decades), UK these the estimates represent more two as recent of last

tourism.

241

derived from VAR III Table 7.14.Expenditure pricc elasticities anduncompensated Models
VAR 111 WP AIDS (2 period) DynamicAIDS (2nd period) W .W. WS VAR 111 AIDS (2 period) DynamicAIDS (2nd period) ww_ WF VARIII AIDS (2 period) DynamicAIDS (2nd period)
nd nd nd

Expenditure Own-pricc elasticities elasticities

1.039 0.947 1.027 0.979 1.150 1.047 1.026 0.808 0.929 -2.158 -1.797 -1.729 -1.057 -1.933 -2.150 -1.817 -1.901 -2.336

Cross-price elasticities III,

X X X 0.161 0.124 0.120 -0.002 0.017 1 0.007

i's
1.137 0.830 0.714 X X X

pr,
-0.017 0.019 -0.011 0.523 0.658

0.983 dmmm wmm 0.793 X 1.077 1.400 X

The elasticities' estimatesof the VAR III and the AIDS model of chapter 5 are
It in for Spain. the the own-pricc elasticity equation similar with one exception: rcrnarkably have in the to the elasticities of expenditure values unity close estimates every indeed, destinations. This. accords with the theoretical postulates predicting for all and model long-run expenditure (income) elasticities. Except in the equation for Spain of the unitary VAR model, the estimates of the own-price elasticities have values close to -2. If the VAR for this elasticity in the case for Spain is to be trustcd, Spain has a comparative Cstimate its in long-run, as its own-pricc elasticity estimate is to the neighbours relative advantage half that of its neighbours. The cross-price elasticities' estimates or the VAR give roughly instance, insignificant f"i for indication AIDS the those speci cations; cross. of as tilc same France Portugal between the cross-cfl'ccts and and significant of equations cffects price between for Spain, France behaviour the on the one equations and indicating competitive before, for As Portugal, Spain the the estimates of these elasticities other. and on hand, and

the price elasticities(both own- and cross-pricc)relative to the The dynamicAIDS model over-estimates for Franceand Spain,and under-estimatcs theseelasticitiesin the caseof Portugal. VAR model in the cases VAR is III in 8, in the cases be the the bestforecaster model chapter shown and Spainwhile of France As will AIDS dynamic for Portugal, As the forecasting the modelprovidesthe bestforecasts. performance in the case is it infer selection to a criterion, also that the seems reasonable model econometric elasticities estimates an of by VAR III In trustworthy dynamic by AIDS than those the more are the tile cases model provided provided is in the caseof Portugal. Spain, the France acceptable reverse while and for

242

for is demand Portugal Francc UK long-run, in to the the or more sensitive that, suggest is to price changes in Franceor Portugal. for Spain in Spain that than pricc-changcs This similarity of results should be expectedas the VAR III specification fully AIDS I lencc, the the the theoretical results underlying approach. predictions with complies for AIDS theoretically and the empirically robust means specification as an support for Moreover, demand UK tourism. the the econometric methodology of analysis economic importance VAR the tile tile these of to and substantiates approach estimates obtain applied for in finding statisticalsupport the existenceof cointcgratcdvectors, johansen procedure long-run AIDS the the the system as only meaningful shareequationsof which sanction be for included. However, the a cautionary note should added variables among relationships identification", building: "empirical Johansen VAR the tile process or model pr;actitionersof in is the to the cointcgrated cxactly-identify rclationships means appropriate not procedure Instead, idcritification long-run the tourism the share equations. of the caseof a systemof information by tourism requires share equations a priori of provided coefficients structural destinations knowledge Market conditions,policy the theory considered. of and economic location, institutionalcharacteristics, attributes and and social and geographical regulations, VAR be features tested tile to appropriately within and rcstrictcd modelled need political empirically plausibleand statistically reliable frwncwork, to obtaintheoreticallyconsistent, long-run in a the the underlying parameters equilibrium structural relationships of estimates information by Only this were we able to establishwith reasonable accuracy, using VAR. demand for in UK France, Spain Portugal. the tourism long-run of and path equilibrium the forecasting is information the complete without pcrrornlancc on 1lowever, our analysis not is This VAR task. our next models. cointegrated the of VAR SPECIFICATIONS COINTEGRATED WITH FORECASTING 7.5. VAR I specificationdid not seemto be the bestnicansfor Although the cointcgrated be a UK tourismdemandin France,Spainand Portugalit may, nevertheless, the CXplaining its forecasting device. For forecasting it assessing comparing ability and VCry competent both VAR 111, for in-sample the the that we estimate models cointegrated period of with four integrate forecasting last leaving In to the the observations period. order to 1969-1993, the these two predictive ability of of picture versions models, we complete use a or obtain for forecasting destinations' VAR VAR the the models expenditure shares: the cointegrated

243

(denoted by 'exact-VAR'); the alone and restrictions exactly-identifying models under by fuli (denoted the over-identifying and VAR models under restrictions set of exactlyforecast 7.17 7.15., 7.16. Tables the show actual and values of the and lovcr-VAR'). Portugal, Spain France, the and respectively, with obtained of shares expenditure I III the and with cointegratcd cxact-VAR and over-VAR 'cointegrated exact-VARJ and forecast before, As III tables the the the report set of errors and models. Ovcr-VAR RMSE) (MAE, MSE previously. used and statistics surnmary France UK for Forecasting 7.15: the share of expenditure Table results Actual values
VAP T I

1994 0.39700 0.34988 0.047117 0.36319 0.033805 0.38011 0.016885 0.39505 0.001942

1995 0.38540 0.33138 0.054027 0.36478 0.020626 0.37760 0.007803 0.39260

1996 0.38967 032715 0.062526 0.36285 0.026820 0.37451 0.015162 0.39485

1997 0.40481 0.32610 0.077871 0.36129 0.043523 0.37335 0.031463 0.39475 0.010054

Forecast Forecast error Forecast Forecast error Forecast Forecast error Forecast Forecast error

xact-VAR

over-VAR I mmow-wa i.;xact-VARIII ovcr. VAR III

-0.007194 -0.005178 SUMMARY STATISTICSFOR RESIDUAL AND FORECASTERIWItS Estimationperiod: (1970-1993) MAE 0.016278 0.000358 0.018907 0.017425 0.000456 0.021364 0.013428 0.000256 0.015990 0.016342 0.000362 0.019016

Forecast period: (1994-1997) 0.060595 0.003811 0.061732 0.031194 0.001045 0.032333 0.017828 0.000391 0.019785 0.006092 0.000046 0.006772

I r, -xact-VAR

MSE RMSE MAE

0vcr-VAR I

MSE RMSE MAE MSE RMSE MAE

C. Xact-VARIII Fovcr-l VAR III

MSE RMSE

244

UK Spain for the Table 7.16: Forecasting expenditure share of results

mwww-

Actual values I E-xact-VAR Ovcr.VAR I wwwwwom III E.. xact-VAR ovcr-VAR III Forecast Forecast error Forecast Forecast error Forecast Forecast error Forecast

1994 0.51857 0.56316 -0.044588 0.54121 -0.022636 0.53975

1995 0.52625 0.58313 -0-056878 0.54052 -0.014275 0.55446 -0.028212 0.52718

1996 0.52292 0.58840 -0.065126 0.54303 -0.020117 0.55924 -0.036326 0.52568

1997 0.50691 0.58956 -0.082649 0.54514 -0.038235 0.56056 -0.053649 0.52543

-0-000932 -0.002761 -0.018518 SUMMARY STATISTICSFOR RESIDUAL AND FOIZECASTER11ORS Estimationperiod: (1970-1993) MAE 0.017295 0.000422 0.020540
0.023393 0.000723 0.026893 0.014235 0.000289 0.017006 0.015845 0.000371 0.019268

-0.021177 0.51884 r-O. 000264 Forecast error77

Forecast period: (1994-1997) 0.062310 0.004074 0.063827

0.023816 0.000646 0.025410 0.034841 0.001361 0.036886 0.005619 0.000088 0.009374

r, xact-VARI

MSE RMSE
MAE

Over.VAR I

MSE RMSE MAE

III 1.; xact-VAR

MSE RMSE

Fovc

MAE MSE RMSE

III r-VAR

245

Portugal UK for the share of expenditure Table 7.17:Forecasting results A ct ual values I Exact-V 13-xact-VAR RI Over-VA III VA :%act xact. -VAR III 'Over-VAR Forecast Forecast error Forecast Forecast error Forecast 1994 0.08443 0.08696 -0.002530 0.09560 -0.011169 0.08014 1995 0.08835 0.08550 0.002850 0.09470 -0.006352 0.06794 1996 0.08741 0.08481 0.002600 0.09411 -0.006703 0.06625 1997 0.08828 0.08434 0.003942 0.09357 0.005288 . 0.06610

0.022185 0.020408 0.021164 0.004292 Forecast error Forecast 1 0.08611 1 0.08022 1 0.07947 1 0.07982 0.008124 0.007939 0.008463 Forecast error -0.001678 Estimationperiod: (1970-1993) MAE 0.008959 0.000136 0.011647 0.009119 0.000133 0.011520 0.005628 0.000055 0.007440 0.006155 0.000063 0.007914 Forecast period: (1994-1997) 0.002980 0.000009 0.003034 0.007378 0.000060 0.007713 0.017012 0.000344 0.018540 0.006551 0.000051 0.007132

SUMMARY STATISTICSFOR RESIDUAL AND FOItECAST Elk!1016

I r,. xact-VAR Fo ' I vcr-VAR F1.; III xact-VAR

MSE RMSE MAE MSE RMSE MAE MSE RMSE MAE MSE RMSE

Fo r-VARIII vc,, -

Comparisonand the choice of the best forecasting model is facilitatcd by the

for forecasting in 7.18, Table three the the accuracy statistics of pcriod of all the inclusion, lower The in the value of the statistic, the more accurate arc this chapter. VAR models used best forccasting viodel casicr, we signal Therefore, the to the forecasts. make choice of the lowest * *. the rirst, third values of statistics with, and and rcspectivcly,. second 9, tllc:

246

different for VAR the specifications Table 7.18: Forecast statistics accuracy
VAR MODELS E:Npcnditure Statistics ExactUnrestricted shares VAR I VAR I MAE France MSE RMSE MAE SPain rp '" " I tu o r ga MSE RMSE MAE MSERMSE . 0.0250* 0.0009' 0.0302' 0.0228* 0.0009' 0.0294' 0.0054* 0.0000* 0.0059* 0.0606 0.0038 0.0617 0.0623 0.0041 0.0638 0.0030o 0.00000 0.0030o OverVAR II 0.0312 0.0010 0.0323 0.0238* 0.0006* 0.0254* 0.0074 0.0001 0.0077 Unrcstrictcd Exact. VAR III VAR III 0.0366 0.0019 0.0435 0.0390 0.0020 0.0452 0.0077 0.0001 0.0079 0.0178* 0.0004* 0.0198* 0.0348 0.0014 0.0369 0.0170 0.0003 0.0185 OverVAR III 0.0061o 0.00000 0.00689 0.00569 0.00010 0.0094o 0.0066* 0.00011, 0.007 1'

to be the best forccastingdevice for the sharesof The over-VAR III model appears Portugal. 1. 'or however, for bcst Spain the this tile third the share of share, and r-rancc and for difTcrcnt VAR between the tile the quality statistics models values of are so differences , forecaster. be I Icncc we can consider them that as an cxccllcnt can considercd any of small full best device for III VAR the theoretical the set of restrictions as under specification he t, in France, Spain Portugal. for UK demand tourism the and forccasting is qualified ns VAR I sPecification However,it shouldbe notedthat the unrestricted forecasting brings best This forecasting the the model. of about matter ability or the second its form, VAR 66 of simplicity remarkable qualities albngside of estimation model pure" a is If to the main purpose of a research project provide assumptions. minimaF and case (unrestricted) forecasts the variables, estimation of share a expenditure of general rcliable lag-length is the the preferableapproach, appropriate variables and with VAR specification justify forecasts do differences tile tile the of not undergoing complexity and quality since approach.Yet, the researcher difficulty of implementingthe Johansen must be aware that based inference VAR depends the on reduced-forni statistical of systems on validity the included if thesearc nonstationary.on the cxistcncc or the time or, series of stationarity long-run the underlying the relationships(s) equilibrium path variables. of Cointcgrating

247

7.6. CONCLUSION Economicmodelsportrayingsteadystatelong-run relationshipsIxtwccn time scrics

I lowcvcr, theoretical thcsc models and empirical arc a central aspect rcscarch. of variables include in debate, to that they the theoretical sense may subject alternative often arc hypotheses about the ways in which the relevant variables nf'fcct cach othcr. Part of the role fortnal is be. to spccifications work establish appropriate which considered call of applied long-run the equilibrium path orthe relevant cconomic variabIcs means of estimating valid hypotheses those theoretical within spccirications. competing ortcsting and

Ovcr the last few dccades,tlic validity of modclling and cstimation proccdurcs time serieswithin the single cquationand the traditional structural involving nonstationary been has Ilic long-run questioned. of equations approaches cstimation or System kad involving data invalidating to may spurious nonstationary rcsults, rclationships inference forecasting Moreover, in and procedures. prc-assuming exogcncity statistical , be jointly determined basis that zero may and restrictions with no cconotnic can variables infercnceand forccasting invalid Icad to proccdurcs. jIso ; Stepstowardsresolving tlicsc probicins were taken with the work or sims o9m, (1988,1995),Johanscn 1982), EngleandGranger(1987),Johansen andJusclius(1990) and introduced, developed and such concepts as applied cointcgratiollanalysislind who OtIlcrs, dimension (VAR) to tile spccifications, adding a new autorcgression methodological vcctor 11iis economic of modelling and estimating relationships. ncw mollodology, processes its stnictural longbe usedto model, estimate to tcst can and a system of equations, applied bills, the avoiding problems simultancous of spurious regression, equations run parameters, division of cndogcnous-cxogcnous variablesand unfounded zcn) rcstrictions. a priori Ibc main linc of masoning in cointegrationanalysis is that statistical in(cmice
includes IV time which nonstationary can a system scrics only on of equations valid based its I fence, exist among variables. appropriate more cointcgrating relationships if one or be applied in order to determine whcthcr a system Is cointcgratcd tests must cointcgration if testing techniques and with so, suitable estimation proceed which can provide valid and, information its its forecasting ability. and reliable structural parameters of about cstimatcs

in the VAR approach is that "incrcdible"' zcro rcstriction The main line of reasoning
inappropriatc division Icad invalid to cndogcnous-exogcnous a priori variablcs of can and inaccuratc forccasting I lcncc, and the cstimation or reduccd. proccdure rcsults. cstimation

248

has been proposedas a form (unrestricted)systemstreating all variablesas endogenous, feasible and reliable alternative to the traditional structural approach, particularly if forecasting is the main objectiveof the modellingexercise. be integration lines The Johansen the two these viewed as of can of methodology for determining in it test the the that rank cointegration procedure establishes rcasoning in framework for system, a and provides a general number of cointcgratingrelationships idcntiflcation, estimation and hypothesistesting of cointegratcd vectors within a VAR by is identification Johansen However, the used process not always mc)dclling approach. in theory testable thcsc provides where restrictions, contexts and economic appropriate, identify be the the to of cointcgrated relationships. parameters structural used 511ould . The AIDS models specified in chapters5 and 6 are systemsof cquationswhich data seriesand assume exogcneityfor the right-handside variables. include nonstationary for UK demand features for implications the tourism the these analysis of empirical of -ilie damage Portugal Spain the validity or the estimation France, seriously could and in forecasting if inference procedures used, cointcgrating and no statistical methods, I Icncc, the this the models. such objective of chapter of variables among exist rclationships inference (or for find the otherwise) procedures and the is to empirical evidence validating For AIDS the this spccifications of previous with chapters. obtained results cStimation VAR for the unrestricted a reduccd-fonn system variables specified relevant we purpose, lag-length, We in AIDS tests to the the several statistical establish model. included -used division of the variables that the cndogcnous/exogcnous and components (jctcrministic include the generalVAR specification.Once the appropriateform of the VAR was should Johansen determine long-run test the to the cointegrated rank used number of we in place, included. the variables among relationships cointcgrated T'heeconomictheory underlying an AIDS model of n cxpcnditureshareequations (n-1) long-run (cointcgratcd) relationshipsamong its the of exactly existence predicts AIDS systemof the UK tourism demandthereare threeqxpcnditureshare In the ,variables. be for. hence, I'lic Johansen two vectors cointegratcd should accounted and, equations to test the thcorctical evidence statistical support provided prcdictions. rank cointegrated The theoretical framework of an AIDS model of the UK tourism demand also form it predicts.This form is that the the of cointegrated structural relationships Cstablishes in describe included destinations' the the tourism system which the sharesas equations of functions of own and competingtourism prices and the UK real per capitatourism budget.

249

in VAR the the cointegrated vectors Therefore, the structuralparameters corresponding of be those the matching of process normalisation restrictions with exactly-identified should AIDS identify the the tourism of static to system. shareequations used The resultingstructuralform of the cointcgratcdVAR, identical to that of the AIDS to then ovcr-idcntifying such as additional restrictions, subjected was Spccification, between the share equations of cffects homogeneity, symmetry and null cross-price for hypotheses These tested France. the parameters were or the cointcgratcd portugal and further Consequently, be evidencewas obtainedabout the VAR and could not rejected. AIDS theoretical the to the with underlying predictions comply model of capability destinations' behaviour demand UK the tourism the and conipctitivc conduct. of rationality VAR full the The estimates the the of cointegrated coefficients structural undcr of
for the then computing uncompensated used theoretical expenditure, were restrictions of sct The be tourism the to shares. results obtaincd of proved Ovvn. and cross-price elasticities from AIDS model estimated in the obtained the to values elasticity corresponding similar , is identical interpretation in 5. Hence, these to that the of results provided economic chapter long-run the 5, the that of expenditure magnitudes elasticities estimated namely, chapter have the cxpcctcd signs and that cstimatcs elasticities all own-pricc unity; approximate for being demand Portugal UK to the more sensitive with own-price magnitudes, plausible demand for France or Spain is to own-price changes in these destinations; the than cliangcs Portugal Spain hand, the confirm and oil one estimates elasticities and the cross-price increasing hand, destinations, in Spain the competing own-priccs as as other on 1:rance and lead to increasing sharesof the other, while own-pricc changes in Portugal destination one for France do (Portugal). tourism the significantly share affect not (r. rancc)

Given the theoretical and empirical consistencyof these results, the predictive full VAR the the should not model under set of restrictions conic cointcgratcd as Of ability forecasts Indeed, the this the model's criteria measuring of quality accuracy a surprise. device forecasting it the amongall VAR specifications usedin this mostprecise indicate as for demand However, UK tourism. these the to samequality criteria suggest predict chapter VAR be best (unrestricted) forecasting 11iis to the next model. the generalreduced-form for forecasting VAR to the the compctcncc claimed of support approach givcs empirical the valuable qualities of modelling simplicity cstiniation case emphasizes and and purposes, VAR. Thus,it seems if "pure" to the that, the unrestricted reasonable conclude main goal of

250

is forecasts, to obtain reliable an appropriately of a multivariatemodelling exercise is likcly VAR to suffice. reduced-fonn spccified The main goal of this chapterwas to investigate,empirically, for validating the inference AIDS the the approach and estimation and thcorctical principles underlying in AIDS implemented We did the model previous chapters. with so using a procedures VAR model frameworkand the Johansen cointegratcdvector methodology.We believewe have provided enough empirical evidence for considering the AIDS approach as a theoretically consistent and statistically robust means to produce valid and reliable between long-run the the underlying parameters rclationships equilibrium cstimatcs of
dcstinations' tourism shares, tourism prices and an origin's tourism real per capita In VAR in its the the compctcncc of confirmed addition, we model. cither cxpcnditurc. full form for the theoretical set of or under restrictions, unrestricted general providing more forecasts destinations' tourism the shares. of accurate

A good econometricmodel offers a formal quantitative framework that call be intcrprcting for understanding economic relationships,testing economic activities, useful behaviour. I lencc, theories, and policy predicting economic evaluating economic economic
is integral its judgement. an model part of of an econometric quality power the predictive forecasting in ARDI. the the of models j-jowcvcr, performance estimated ch3ptcr 4 and tile in 6 have 5 This task is carried chapters and not yet been assessed. estimated AIDS systems

in the next chapter. out

251

CHAPTER 8
THE FORECASTING ABILITY ECONOMETRIC OF ALTERNATIVE

MODELS OF THE UK DEMAND FOR TOURISM

INTRODUCTION
A fundamental objective of economic researchis to describe, explain and prcdict the

in economicbehaviour.Economicphenomena because arc complex involved ccmplcxities Shocks dynamic interrelated components. and policy changes may take andcontain tilcy are fully influences Moreover, before transmitted throughout the their economy. are time by dynamic interrelated The other variables. are affected affect, and and variables cconomic implies that technological, variables social and political changes most economic of nature in in but levels the their currentperiod also subsequent not only periods. irlipact on The nature of economicbchaviour may be capturcdby cconometricmodels, the intricacies describes between the tile the which of relationships current of value or structure lagged the dependent currcnt and/or variableand valuesor severalexplanatoryvariables, a dependent itself. lagged The the of variable values use as or appropriate well as in the construction,estimationand evaluationof such models can provide methodologies behaviour the of explaining and means predicting reliable economic under and accurate if deficient However, methodologies arc usedin any of thesethrcc basicsteps investigation. be inference the the resulting estimates may misleading, procedures of cmpirical modelling, ability mediocre. invalid andthe forecasting In previous chapters we constructed, estimated and evaluated nltcmativc for UK demand for in France, Spain the tourism modelling specifications and cconometric involved The these of evaluation statistical models portugal. sevcralquality critcria which for qualifying empirical models as accurate, reliable or useful arc usually applied

252

found According to these they the we portray. criteria, supposedly reality rcprcsentationsof in in 4, AIDS ECM the ARDL chapter models specified modelsestimated and that the in 7 VAR 6 5 theoretically the chapter constructed were coherent, models and and chapters tile representations robust of economicphenomena statistically and plausible Crnpirically difTcrcnt di(Tcrcnt In that investigation. the showed models could give addition, we under behaviour different the of economic perspectives analysed. insights and emphasize in for UK demand France, forms be the tourism they of explaining 111cr6ore, can useful for intended. I lowever, Portugal they tile the which were: Spain and under spcciric purposes is fundamental a aspect of their quality forecasting ability of econometric models has for VAR the this the of models previous chapter, aspect not yet Cvaluation and, except importance, its Given we now undertake the analysis of tile major bccn addressed. forecasting ability of the modelsspccifledpreviously. Forecastingis a crucial input of the decision-mak-ing process. Rather titan an integral it be In this this sense, this an part of process. considered should isolated activity, is not only essentialbut also an unavoidabletask of a comprehensive economic activity diverse behavioural features Forecasting the tile as of variables as are problems analysis. it is diversity This intend that to to seems unreasonable such expect a unique address. they issues. Indeed, to tackle all prediction a considerable number of bcst method is presentlyavailablefor solving tile variety of forecastingproblems tools rncthodological inherentcomplexitiesof modem economics. The wide rangeof difTercnt from the arising justifies forecasting in approaches, a wide range of alternative practice met problems fact. is In to them there that single way no right address all. every variable itnplying its has unique features,demandingcareful analysis or its past and rorccasting problem interrelated factors influenceit. behaviour the that main might and of prcscnt The problem of how to predict the future, given inrormation on the past, takes us is Understanding the time the past a soundbasefor seriesanalysis. into subjectmatter of into future. However, historical knowledgeably the cvcn after careful study of more looking be it hand, the that case more than one appropriatesolution often will at problem the is no single right forecastingapproachto time seriesanalysis, Although there cmcrges. in be than more appropriate others specific contexts,and it might some might considered different for best be that the the a combination of methods case serves solution as a Cvcn in identification The be sight. of problem what might a reasonable particular include factors. These requires consideration approach of several what nicthodological

253

forecasting for, forecasts quantities require prediction, what what are needed purposes in is is horizon relevant,what resources arc availableand what amountof cffort worthwhile forecast increased I laving from accuracy. the what must established potentialreturn Vicw of involves be far, how the the consideration bc predicted,why and of which might next step forccasts, for best accurate givcn the resources and obtaining approach methodological the information availablc. This chapterconcemstheseand other aspectsof the processof prcdicting the UK is The follows. Section Portugal. Spain France chapter structured as and of tourism shares forecasting limitations bricfly the of some quantitativc mcthods, 8.2, addresses qualitiesand demand forecasting. Section 8.3, tourism on work of research recent and gives an overview in forecast 4 the the using econometric models specificd chaptcrs obtained results prcscnts diffcrcnt forecasting 8.4 these Section the 7. ability of cconomctricmodelsand compares to 8.5 concludcs. section 8.2. MERITS AND LIMITATIONS OF QUANTITATIVE FORECASTING

r, IF,TIIODS The forecastingmethodologicalapparatus presentlyavailable includesapproaches from based luck, (1998) Hendry Clements to point out, range purc guessing and on that, as large-scale Nc%Vbold Bos building macrocconomctric systems or, as complex of and the from formal, 'firom informal 13) to the the use of unaidedhuman the very ( 1994, p. state from the very sintlVe to the the to reliance on quantitative moilels, exclusive pidgement forecasting In this tools, the wide range of of methodological spite 1jighly complex". into be divided two main groups:quantitativeand noti-quantitativc can available methods in Newbold (1994) forecasting Bos While that most succcssrul and with agreeing methods. less fon-nal fornial involve of and quantitative methods a combination cxcrciscs fon-ner in believe that the are, gencral, more appropriate in judgemental expertise,we forecasting problems involving economictime series,and Pore useful in the particular in addressing in this chapter. forecastingexercise we are interested have bccn A number of forecastingmethods based on quantitative approaches ' in literature. implemented have been Two the these of and given particular proposed
See, for example,Fair (1986) for a partial survey, Wallis (1989) for an appraisalor sonic Important Ilarvey(1993) 1980sand 1970s forastudyormore the and of recentforccastintmethods. contributions

254

forecasting tile time the and approach econometricregression series univariate attention: in its basic involves The tile predictionof variants, approach, univariate analysis approach. future valuesof a time seriesprocessexclusively on the basis of its own past values.The isolate designed involves to the of a statistical specification model cconometric approach interpretable by describe time series processes, which are among relationships and forecasting for be theory purposes. used and might economic There is a similarity betweenthesetwo forecastingprocedures springing from the based. is Assuming they that the the objective are which on models construction methodof forecasting future forecasts time these the series, valuesof an economic of to obtain reliable based which spccifics,as accurately on a statisticalmodel-buildingapproach are procedures data I'llis the generating process. underlying assumed statistical approach as possible, forecast function is in framework the through appropriate approximated which a provides is data the that tailored to the specific so prediction method of observed analysis tile
Moreover, the statistical modcls on which time a series process. of given characteristics forecasting based for time the procedures series and econometric are allow univariatc derivation of measuresof forecast uncertainty and associated tests of forecast nccuracy, not

hoc However, for the univariatc and nictliodologics econometric procedures. ad available well worth mentioning. differ substantiallyin severalotheraspects A crucial differencebetween thesemethodologies residesprimarily in the obicctives for which they are implemented.While univariate time seriesmethodscxclusivcly servc
fulfil formal forecasting purposes, multivariatc econometric modcls much more uscful rolcS forecast-generating devices. Clcmcnts I Icndry (1998, As 16) being and p. simple point than "consolidate knowledge how theoretical models existing empirical and cconometric of out, framework function, for hell) a to provide progressive research strategy and economies failures. They In to their this scruliny, are open adversary are replicable own anil, explain basis bctwccn difTcrcnccs A to research". second group of a scientific sense, offer forecasting models can be identificd with the theoretical basis and econometric univariate

degree the their they of complexity of with rest and statisticalspccifications. which on based its Thereis a wide rangeof time seriesforecastingprocedures exclusively on implemented is This typc through tile models, of often as values. qualified naive, past own Examples statistical simple models. of very of such models are no-change, Cstimation (AR), integrated AR autoregressive purc stcpwisc and autoregrcssive prcvious-changc,

255

2 The ARIMA (ARIMA) others. among classof models,usually methods, average moving is for (1970) Box-Jenkins particularly approach, useful short-termeconomic as rcfcrred . forecastingand hasbeenextensivelyappliedfor this purposein time seriesliterature.Tllcsc when comparedwith other univariatc models are generallyviewed as quite sophisticated be in different Yet, this approach can consideredas nayvcas other way, viewed a models. Indeed, forecasting the constructionof univariatc time seriesmodels processes. Univariate knowledge, information be ignores theory expertise, or may relevant whatever completely by is based It the tile time. the on evidence provided pastvaluesof exclusively at available broader in the taking account of no context which tile time the variableunderconsideration, data-analysis difier is In these this statistical and are purely models sense, scrics generated. from the formal econometricspecificationsin that "they do not begin with a conceptual framework provided by economictheory, that specylesa relationship betweciteconomic Instead, behavioural Thus, technical the equallons... are or not considereil variables. informationfrom Its is past values variablefi)rforecasting ofa use making of on nphasis C., " (Griffliths to a sophisticated future valuesby usingwhat amounts extrapolationproce(hire. involving bcticrit in 639). Nevertheless, 1993, a cost. perspective circumstances and p. et al. be to useful can proccsscs address specific methods univariate purposes, Silort-run important information Past forecasting problems. yield observations about tile future and, thesemethodsare frequently, no additionalrelevantinformation is available.Furthermore, forecasts implement for to tile of1cn provide sufficient and of quality easy relatively hand. at purposes I lowever, ignoring relevantavailableinformation cannotalways be thoughtor as n in forecasting. Reliable forecasting have be to engaging when procedures strategy good information. Forccastsof a time scrics call oflen be based on all currently available through the incorporationof information on relatedtime series.In improved considerably into is known from inclusion theory the the model-building what of cconomic addition, is importance. likely bchavioural Such is to theory postulate of paramount process ignorcd be time series which cannot without seriouslyaff: cting relationshipsamongrelated In fact, is forecasting in thcory the the process. economic csscntial of model. accuracy the Theory can suggesta particular modcl, a group or processes. building and model-selection functional form lag Morcover, thcory a suitable or can play a variables, structure. rclevant in the the of through the plausibility of the signs models evaluation selected, quality part
description 2 For an introductory NewboldandBos(1994). of severalof thesemethods see,for example,

256

hypothesis These the the testing. and outcome estimates of coefficients' of and magnitudes for behaviour indicate in is the that, time which cases of a variable series considerations likely to dependon the behaviourof other time seriesand economictheory plays a major interrelated behaviour, it is in the their to advisable move away nature of explaining role from a-theoretic naive univariate models and embrace the thcory-bascd,multivariate forecasting for purposes. econometricspecifications At this point, however, a cautionary note should be issued. The building of theory,large-scale, behavioural,high-powered complexeconometricmodelsfor forecasting information large is The the cnornious quantities of needed, purposes not alwaysadvisable. heavy their time the construction on and spent and effort computational of amount impractical testing, their them and can estimation make appear and useless rcquiremcntsof in rnany situations.Frequently,bigger is not better and successfulforecastingmodels are involving just a small numberof unknownparainctcrs. I ikcly to be parsimonious,
This brings us to the matter of comparative forecasting pcrformancc. I low accurate forecasts, how should this accuracy be evaluated and how to comparc diITcrcntly

are

important havc becti forecasts that many empirical studies arc questions addrcssing sourced decades. last Several forccasts the these the studies assess of quality of a sct of ovcr from econometric models through the comparison of accuracy with a sct of obtained from forecasts Frcqucntly, the obtained univariate models. univariate 111odels competing ARIMA for comparison are chosen models and, more rccently, their multivariate

historically, VAR Although, has been the models. univariate approach gcncralisations, in Charcmza to to well relative perform econometric methods and, as noted and claimed Deadman (1997), simple univariate time series methodsseemto provide more accurate rorccaststhanthe largeand complexmodelsconstructed underthe traditional methodology, is While for the conclusive. evidence not results of several studies, example tile available Cooper (1972), underline the superior perfon-nancc of univariatc forccasting incthods, indicate for Zamowitz (1978), that structural econometricrnodcls out. example others, by ARIMA for forecast substantial quite models margins, particularly extended perform horizons. While in Granger(1981), the irrelevanceof the univariatc forecastingapproach is pointedout, it is also recognizedthat the sameremarkcan be made for policy purposes ba(Ily "a that econometric specified models and poorly s1wifle(I econoinctricmodel about
because dangerous it may appear to have something to say about alternative is even more be (p. 127). very may misleadine, actually andyet policies

257

In studies which address the issue of comparative forecasting performance, for it has been (1998), forecasting Hendry Clements that the acknowledged and example, is likely be due to to to econometric approaches relative models of univariate success dynamic misspccifications in the econometric models and/or to the failure of adequately because being Perhaps disappointing for data the the of used. results of early accounting been have dynamic to tile extra paying attention econometricians specifications evaluations, it their and seems not a, coincidence that recent research on the subject or models, of indicates favourable forecasting a more performance position towards tile comparative forecasting ability of well-spccified econometric models when compared with that of naivc 3 However, as Newbold and Bos (1994) difTerencc in the stress, models. given univariatc be involved, it is difficult impressed findings to these overwhelmingly effort with and scale in further improvement forecasts for the quality of econometric still remains. and scope In the specific context of recent tourism demand analysis, several empirical studies ,
forecasting based forecasting the accuracy matter of comparative on alternative address methodologies. Many of these studies only cmphasisc univariatc nicthods, othcrs focus

include few forecasting the comparative pcrronnancc only on econometric approaches and 4 The

of econometric versus univariate procedures.

results of most studies comparing the

forecasting accuracy of econometric and simple univariate methods seem systematically to favour naive and ad hoc forecasting techniques (such as the no-changc model) relative to forecasting "more to-be complex" procedures (such as tile AIUMA what are claimed method) and what is usually qualificd as the "regression" or "econometric" approach. I lowcvcr, in several cases,the "regression" models used for comparison are not what could bc viewed as structured multivariate econometric causal spccirications, but rather some simplified version of single equation models, with no account for integrated data, dynamics and/or for the theoretically establishedfull rank of explanatory variables. Some studies, although stating their intention of nddrcssing alternative tourism dcrnand forecasting methods, do not explicitly do so. For example, Fujii and Mak (1981) analyse alternative estimation techniques for forecasting purposes, rather than alternative forecasting models. In their study, the alternative estimation techniques (ridge regression,
3 Flur debate on the forecasting accuracy an enlightening of large-scale macrocconometric modelsseeKmenta anQ Ramsey (1981). For a surveyin the comparison forecasts of time seriesand econometric seeGrangerand NcWbold(1986).For a reviewof econometric forecasting (1990). methods seeGranger

' FOr forecasting demand in tourism for example, Archer (1976), Uysal Crompton (1985) see, surveys and and Witt andWitt (1995).

,)58
Ak.

GLS and OLS) are applied to a single equation model which regresscs the number of US its lagged Hawaii transportation to real airfare values, on cost and real per capita visitors JJS income. The study concludes that the ridge regression method out-perfornis the GLS Other for forecasting Martin Witt OLS (1987) purposes. examples arc methods and and and LCc, Var and Blaine (1996). The fomcr study, although under the title "Tourisin demand forecasfing models", focuses on the suitability of alternative measures of the explanatory latter, importance forecasting 'cost the tourism' the although underlining of and of variable in tourism demand analysis, focuses on specification aspects and estimation methods for in forecasting for demand than engaging rather procedures predicting regressions, tourism dependent future the variable. of values tilc An example of a comparative analysis of alternative forecasting methods is Gonzalez and Moral's (1996) study of forecasting trends of international tourism demand for Spain. This study analyses the forecasting performance of a Basic Structural Model (13SM) which is supposedto capture the trend, seasonal,cyclical and irregular components Box-Jcnkins dependent ARIMA "classic linear time the variable series, a model a and of from is derived BSM by imposing the model model" which some restrictions on rcgression its parameters. The explanatory variables of this 'econometric' regression arc a linear deterministic trend and a seasonal component. The forecasting performance of tile IISM is found be to ARIMA that the superior when compared with of rnodcl "deterministic regrcssioW'modcls. In other studies, for example, an early work by Gcurts and lbrahirn (1975) and a by Witt (1994), included ror et one al. recent econometric specifications not are more forecasting In formcr, Ilox. Jcnkins the the purposes of accuracy. and the comparison forecast incoming to the approaches tourists smoothing are applied number of cxponcntial besides latter, In forecasting Ilawaii. these the two mcthods, fivc other univariate to (no-change, Gompcrtz trend, constant-changc, and stcpwisc autorcgrcssive) are approaches in for UK flows. the tourism comparative purposes predicting outbound used Several research studies of tourism demand forecasting do not address forecasting and the

instance, For than an econometric approach. other a group or paperson tourism racthods demand forecasting, using solely an econometric approach to predict long-tcmi international tourism flows, includes Smeral et al. (1992), Smcral and Witt (1996) and Smcral and Weber's (2000) studies. In the study by Smcral el al. (1992), a system of is demand equations estimatedover the period 1975-1988and used to generate tourism

259

imports flows for industrialised 18 forecasts tourism and exports countrics, of cconometric (1996) Witt's Smeral 2000. the study modifics the previous model to allow to and year up for destination-specific demand structures in the export functions, extends the estimation forecast horizon four by 1992 2005, to the the to up and year and considers years up period income The Smcml Weber's different (2000) and scenarios. study contemplates growth two for 1975-1996, 20 the over estimated countries period model and extends the a similar forecasting horizon up to 2010 taking into consideration, for the alternative scenarios in European Union. Monetary the the countries' participation adopted, Examples of studies explicitly comparing the forecasting performance of

include Martin Witt (1989), that naive models univariatc with of models and cconometric Witt and Witt (1992), Kulcndran (1996) and Kulcndran and King (1997) Kim mid Song (1998), Songetal. (1999),Songclal. (2000)and Songand Witt (2000). While Martin and Witt (1989), and Songet al. (1999) suggest that the forecastingability of naive no-change forecasting including type ARIMA tends to techniques other of over-pcrfonn and Inodcls Witt (1992), Kulcndran Witt (1996) Kulcndruni approaches, and and and King cconometric (1997) conclude that ARIMA models can be powerful competitors with econometric (1998) Song Kim Song (2000) findings these and et al. contradict and specifications. bettcr forecasts that to crror-correction econometric models are able than produce showing Using RMSE), (MAE Song two time Witt quality criteria series models. and and univariate ECM, unrestrictedand (2000) comparethe perfortnance of severalmodels- Engic-Grangcr VAR, ARDL, (TVII) time varying parameter and no-change(NC) - ill Cointcgratcd forecastingthe demandfor tourism of two origin countries(UK and USA). Dependingoil horizon, forecasting tile ranking of tile best quality criteria and the origin considered, tile For instance,while the onc-year-allead forecastsfor forecastingmodel variesconsiderably. is UK that the no-change model ranked first using MAE and tile TVII model is show the for USA first both is first TVP the thc, using using'RMSE, criteria, and tile no. ranked forecasting is In both two-years-ahead third. model and using critcria, the Johansen change is for best UK VAR is for the best the and tile USA. 'I'lle change no model cointegrated is forecasting TVP first for for USA the the tile UK, it of performance ranked overall while, depend to on the quality criteria used. seems Given theseratherconfusingresults,research on comparativeforecastingissueshas Nevertheless, long to go. early evidenceproducedon tile poor forecasting way still a performanceof econometricmodels may be attributed more to misspecirications of the

260

defect intrinsic Furthermore, than to the of econometric used, rather some approach. rnodels it in the specific research that more than simple all are concerned we with, matter appears hoc no-change models or any other univariate extrapolation procedure, however forecasts is UK for France, to the tourism accurate needed of obtain sophisticated, ticinand Spain and Portugal. Indeed, based on the estimation results obtained from tile ccononictric models in behaviour that the we showed previous chapters, of the dependentvariable spccif*icd is linked for UK demand behaviour the tourism I to tile structurally of scvcM, representing in implying have that these changes variables, rcgrcssors a scparatc and cxplanatory impact dependent Moreover, for the significant on variable. all the econometric additional, specifications used previously, we provided empirical evidence of thcir theoretical implying that these and statistical robustness, specificationsarc valid nicansor consistency describing, interpreting and forecastingthe UK tourism demand for France,Spain and 11ortugal.Hence, we believe that a univariatc model, basing the explanation of tile dependent variable's behaviour solely on its laggcd values (with or without lagged disturbances),is a mis-spccificdrepresentation of the underlying data gcncratingprocess. in i fi Ille consequences cation statisticalmodelsare well known: the cstinintion of mis-spec inference invalid. As procedures are statistical a result, we consider that univariatc and behaviour the models are not appropriate means of predicting of the UK tourism time-scrics betweenthe forecastingability demand in France,Spainor Portugal.Hence,comparisons data-colicrcnt, that the thcorctically these naive models with of consistcrit, wcll-spccirlcd ()f be to robust econometric statistically models of previous chaptcrs sccin either a trivial and -con flinn-thc-obvious"task,or a pointlessexerciseof self-indtilgence. In what follows we concentrate on the derivation,quality evaluationandcomparison demand in UK France, forecasts Spain Portugal from tourism the ARDL of and obtained of 4, AIDS 5 6, models chapter of chapters of and and VAR modelsof chapter7. models TRIC MODELS OF 8.3. THE FORECASTING PERFORMANCE OF ECONONIE, THE UKDEMAND FORTOURISM IN FRANCE.,SPAINAND PORTUGAL In order to explain how the forecastsof the UK tourism demandin 1. 'rancc, Spain derived, Portugal it is important evaluated and are compared, to identify .vhich timc and is be forecast to to the terminology usedto describethe basic concepts and clarify series

261

include These in-samPlc forecasting derinition the the procedures. of and methodsused'in forecasting forecasting horizon, the the period, and out-of-sample period estimation forecasting equationsvalid in each case,the type of forecast(ititcrval/point, cx.post/ex. for the static/dynamic), and statistical critcria used conditional/unconditional, ante, forecasts. the the quality of assessing
8.3.1. FORECASTINGPROCEDURES:BASIC CONCEPTS

Econometric forecasts are quantitative estimates of a variable's future values, included in information from the spccification of all econometric past and current obtained beyond its derive forecasts for By this the estimation model period, projecting we model. future values of the variable under analysis. Two types of forecasts can be obtained by this forecast interval single value, which predict a and process: point-forecasts -forecasts which define a confidence interval used to provide a margin of error around tile point-forccast. The time series variable for which we are interested in obtaining econometric forecasts is the level of the UK demand for tourism in France, Spain and Portugal. The derivation of these forecasts is basedon the ARDL models specificd in chapter 4, tile AIDS in VAR 5 6 7. For the and and models this chapters spcciricd chapter specified in models forecasting exercise, the existing sample of 29 observations over the period 1969-1997 is divided into two sub-periods: the in-sample estimation period comprising 25 observations from 1969 to 1993, and the out-of-sample forecasting period involving the last four both these sub-periods, the actual values of all variables For available. observations irivolvcd in the estimation and forecasting equations are known. I lowevcr, if tile forecasting had to be extended beyond 1997, no information about the actual values of these period 5 This fact be distinguish between to available. allows would us cx-post and ex. variables forecasting and conditional procedures, unconditional ante and Both ex-post and ex-ante procedures predict values of the dependent variable However, the the of period model. estimation while in the former the actual values beyond dependent and the explanatory variables are known with certainty, in the latter both the of be known Therefore, the may or may not variables of with values certainty. an cx-ante the forecasting procedure permits the checking of the forecasts against existing data and, lictice, immediate forecasts. the of evaluating means the an accuracy of provides In thecase s -rbis is truewith oneexception. 5, thefuturevalucs of theAIDS models of chapters of thetrcnd known. all equations the of side are on right-hand previously variable

262

the valuesof all the explanatory In an unconditionalforecasting variables procedure,

forecast is Hence, known. forecasting any ex-post also an unconditional in the equation are be for if, forecast However, may also unconditional example, the forecast. an ex-ante forecasting lagged In include trends values. a conditional or explanatory variables be known. So they must procedure, values of one or more explanatory variables are not 6 introduces forecast Illis dependent in be forecast the to variable cquation. used previously for dependent forecasts into this particular the the variable which, of further uncertainty forecasting exercise, we want to avoid. Therefore, to obtain point-forccasts of tile UK levels from demand the econometric models previously speciricd, we use cx-post tourism 1994-1997. forecasting the out-of-sample period over procedures unconditional
FORECASTS CRITERIA EVALUATION OF 8.3.2.

In any forecastingevaluationprocess a fundamental measure of accuracyon which difference between is based, the the the time actual value or are criteria evaluation most is difTercrice known rorccast This forecasted its as error value. and corresponding and series forecast is "the defined forecast the best yields error with ininlinum one which as the 204). Therefore, 1998, Rubinfeld, (Pindyck once econometric p. an modcl and variance" dependent important for the been to variable, statistic one and used predict have estimated forecast its is forecast the the error varianceand associated conridcticc precision evaluating

interval 100(1-a)%.
Assumethat the time seriesvariableto be predictedis Y1,wheret=l,..., T, and that
its known for further determinants Assume t. this all and all arc variable the actual values of has been its detcnninants Yt to using used a sample of model regress on econometric an that N out of the T available observations (N<I), leaving the remaining n observations for for forecast We define h-step T). this variable as (N the +n= ahead forecasting purposes 7 in I Icncc, h=l,..., this the the of variable same period asyN+h. value YfN+h, actual n, and is. to forecast this ahead period error corresponding the

(1998). in (Newbold Rubinfeld Yet, Bos, Pindyck follow 6 T'hese textbooks, those and other and of concepts 1994, for example),conditional and unconditionalforecastsare defined in the oppositeway; that Is, In in known rorccasting forecasting the the the values all variables of and unconditional of are values conditional forecasting have be forecast. for to the period variables explanatory 7 if a regression variableon its right-handside. the forecastof this variable, containsthe laggeddependent is is generallyknown as a dynamicforecast.If no lags of the by to create used what period, period updated the forecasting amongthe regressors, process generates dependent variablearepresent static forecasts.

263

eN+h*-- YN+h- Y ,f-N+h -

interval is 00(l I the confidence associated and -cc)%

[YfN+ht, /2.sd(eN+h) ; YfN+h+

t,, n.sd(eN+h)]

(8.2)

ftccdom degrees for is the t the and the of appropriate statistic of value critical where t,,,. a forecast If deviation is level, tile the error. standard of estimated and sd(eN+h) significance for interval, forecast lies forecasted this this the the within variable of the actual value levcl. be is the to at a% significance accurate actual value considered forecast In addition or alternatively,three other measures quality are of average
frequently used in the literature: the mean absolute'prediction errors (MAE), the mean These (RMS). (MSE) the squared crrors prediction root mean and squared prediction errors forecast If follows. defined the errors ct are available, mean absolute n as are measures

is predictionerrors
MAE = 2: n ,.,
n2

letl n

(8.3)

is predictionerrors the meansquared C 2: MSE = n ,., is the errors prediction root meansquared and RMSE n
2 t

(8.4)

(8.5)

These measuresare all useful statistics for assessingthe average quality of a set of

The set or forecastsand,providedthat competingsetsare available,they can be compared. be better. is be j for to the statistic forecasts which criterion smallerwill udgcd In additionto thesegenerallyusedmeasures, the predictivefailurc test can also be a forecasting For the accuracy. of a method's assessment complementing valuable meansof between test the this time similarity observations pertainingto evaluates series, the relevant forecasting in-sample those to the period, and pertaining out-of-samplc estimation the in hypothesis The that the the used states out-of sample observations null period. have do outlier characteristicswhen comparedwith the set of forecasting period not in-sample estimation period and, statistically, these two sets of the of observations be homogeneous is If to to the the pcrtain same considered sample. null observationscan be have been both from As to said can the sets withdrawn population. same rejected, not (1985), failure in the be Pesaran test al. predictive et can also used as a general shown 264

Fhas this test Under the exact an test. assumptions, statistic classical error specification lies below F if the the Hence, the statistic critical distribution. relevant observed value of 8 be value, the null cannot rejected.

When comparing different sets of forecasts, it is important to assurethat the is information. it Hence, homogeneous is based to essential sets of on exercise comparative is forecast being the time the that and same prediction variable series same guarantee by different forecasts is being to methods. generated horizon considered compare
8.3.3. TIIE FORECASTING PERFORMANCE OF ALTERNATIVE ECONONIETRIC MODELS 9

8.3.3.1. ARDL models

In chapter4 we specifiedsingleequationdynamicARDL modelsof the UK demand for tourism in France, Spain and Portugal. The estimation results obtained with these in Tables 4.4 4.2 (France), (Spain) the are presented available observations models using all demand for France,Spainand Portugalarc UK level The tourism (Portugal). 4.6 the of and forecasting in interested these and variablesare represented time are we variables series the irl the ARDL modelsby the log of the UK real per capita expenditureallocatedto cacli by for ARDL tile these models are obtained estimating destination. Forecasts variables of four forecasting last for leaving 1969-1993, the in-sample purposes. observations period tile forecasted dependent (real 8.3 8.1 tile the to variable values Tables of per actualand report for destination) in and a numberof summarystatistics evaluating Capitaexpenditure each 10 for forecasting France, Table 8.1 forecasts the results accuracy. presents the accuracyof for 8.3 Portugal. for Spain 8.2 table and table

I This test is easily performedon an equationby equation basis when the cstimation method Is OLS. is is considered equations and another estimation method rcquired,the value or of a system when 11owever, its is distribution F-distribution. Thcrcrorc, is to and statistical not an compute exact easy test not statistic this basis. by OLS is the models are estimated test with an equation only when equation on performed this 9 The forecastsand summarystatistics for all the modelsanalysedin this section were computcdusing (1997)Microfit 4.0. If laggedvaluesof the dependent variablearc explicitly includedas pesaranandPesaran ARDL, dynamic VAR in AIDS Microfit the the of cases and as models, automaticallycomputes rcgressors, it forecasts. forecasts, generates static otherwise namic are single equationmodelswhich can be estimatedusing OLS estimation As the ARDL specifications is for failure test thesemodels. reported the method, predictive

265

log in France UK for the the Table 8.1: Forecasting of per real capita expenditure results 1994 Actual values Forecast Forecast error castel 1.4848 1.4813 0.00357 1995 1.5080 1.4966 0.01138 1996 1.4785 1.4960 1 -0.01746 1997 1.5215 1.5450 -0.02357

SUMMARY STATISTICS FOR RESIDUAL AND FORECAST ERRORS

MAE SE Ms M SEE

Estimationperiod:(1970-1993) 0.015081 0.0003083 0.017559

Forecast period:(1994-1997) 0.013995 0.0002506 0.015831

Predictive Failure Test: F(4,15) = 0.28717(0.882)

log for UK in the the Table 8.2: Forecasting of real per capita expenditure Spain results 1994 Actual values Forecasts Forecast error 1.5515 1.5092 0.042350 1995 1.5817 1.5583 0.023414 1996 1.5383 1.5478 1997 1.5477 1.5597

-0.0095275 -0.011944 FORRESIDUAL AND FORECAST ERRORS SUMMARYSTATISTICS Estimationperiod:(1970-1993)_ 0.024182 0.0008754 0.029588 1 Forecast period:(1994 -1997) 0.021809 0.0006438 0.025373

MAE PFRIMSE MSE

Failure Predictive Test:F(4,16)= 0.403(0.804)

a-

for Forecasting 8.3: the log of the UK real per capita expenditure in Portugal Table results

1994

1995

1996

1997

0.71954 0.76522 0.74808 0.72156 Actual values, 0.73205 0.73337 0.68552 0.71755 Forecasts 012517 0.031846 0.062560 0.004001 Forecast error FORRESIDUAL AND FORECAST SUMMARYSTATISTICS ERRORS MAE MSE RMSE Estimationperiod:(1970-1993) 0.033324 0.0017025 0.041262 Forecast period:(1994 -1997) 0.027732 0.0012752 0.035709

Failure Predictive Test:F(4,17) 0.30196(0.873)

266

8.3.3.2. Static AIDS models In chapter5 we specifieda staticAIDS model for the UK tourism demandusing as dependentvariable the expenditureshareof each destination. We estimatedthis model by [denoted (I I+S)] homogeneity and under restrictions and symmetry unrestricted,under [denoted between France Portugal and effects the additional restriction of null cross-price in 5.1 table The (11+S)O]. reported and show that this additional are by results estimation does have improves the the of estimates and a significant not precision restriction only forecasts Hence, tile obtained the with coefficients. the estimated of magnitudes on cffect from forecasts diverge tile the to (11+S) with obtained much expected are not model the " is forecasting In tile the ability of unrestrictedmodel expectedto (I I+S)O contrast, model. diverge significantly ftom that of the restricted(11+S)model. Therefore,for comparison I+S) (I interesting the tile to the unrestricted model and model. are use models purposes, The forecastability of these models is comparedby estimating the unrestrictcd by **) for (denoted 1969-1993, by *) the the vcrsions period restricted (denoted and Tables forecasting. for 8.4,8.5 8.6 four last the and report the actual 1caving observations Spain Portugal, for France, the forecasts the and respectively, and shares of values and 12 for forccasts. the those evaluating accuracy of used previously statistics same summary

11 Both thesemodels were used to forecastthe relevant variable and the results obtainedshow that the forecasts is minimal, neverexceeding 0.0003. between two the of sets difference AIDS canbe estimated 12The unrestricted usingeitherSUREor OLS on an equationby equationbasis,as the in for However, including this equivalent the case. are methods restricted version crossestimation two is SURE Therefore, failure the method applicable. the the we predictive report constraints, results of equations for the unrestricted version. only test

267

for UK the Table 8.4: Forecasting expenditure shareof France results 1994 Actual uI values Fc orecasts* Forecasterror** Forecasts* Forecast error* 0.39700 0.37775 0.019251 0.40027 1995 0.38540 0.38044 0.004959 0.39706 1996 0.38967 0.40188 -0.012202 0.42707 1997 0.40481 0.38832 0.016484 0.46015

-0.011659 -0.003272 -0.037393 -0.055338 FORRESIDUAL AND FORECAST ERRORS SUMMARYSTATISTICS Estimationperiod:(1970-1993) 0.018839 0.000548 0.023407 0.016122 0.000450 0.021212 PredictiveFailureTest*: F(4,17) Forecast pcriod: (1994-1997) 0.013224 0.000204 0.014281 0.026916 0.001152 0.033938 0.423(0.79)

MAE** MSE** RMSE** MAE* M SE* MS MS RMSE*

UK for the Table 8.5: Forecasting expenditure shareof Spain results 1994 Actual values Forecasts* Forecasterror** Forecasts* Forecasterror* 0.51857 0.52613 -0.007558 0.50068 0.017892 1995 0.52625 0.52235 0.003893 0.50406 0.022187 1996 0.52292 0.49786 0.025054 0.46910 0.053812 997 0.50691 0.51260 -0.005696 0.42879 0.078121

FORRESIDUAL SUMMARYSTATISTICS AND FORECAST EIIRORS MAE** MSE** RMSE** MAE* MSE* RMSE* Estimationperiod:(1970-1993) 0.021050 0.000637 0.025236 0.016681 0.000500 L 0.022362 Forecast period:(1994 -1997) 0.010550 0.000183 0.013532 0.043003 0.002453 0.049525

ictive Failurc Test*: F(4, 17) -=O593(0.673) Pred .

268

UK for the Table 8.6: Forecasting expenditure shareof Portugal results 1994 Actual values * Forecast* Forecasterror** Forecasts* Forecasterror* 0.084433 0.096126 -0.011693 0.099053 1995 0.088348 0.097201 -0.008853 0.098877 1996 0.087411 0.10026 -0.012852 0.10383 1997 0.088283 0.099071 -0.010788 0.11107

-0.010529 -0.016420 -0.014620 -0.022785 FORRESIDUAL ERRORS AND FOIZECAST SUMMARYSTATISTICS Estimationperiod:(1970-1993) 0.006626 0.000056 0.007508 0.006259 0.000052 0.007277 PredictiveFailureTest*: F(4,17) Forecast period:(1994-1997) 0.011047 0.000124 0.011143 0.016088 0.000278 0.016684 0.469(0.758)

MAE** MSE** RMSE** MAE* MSE* RMSE* .00-OMM

S.3.3.3. Dynamic AIDS models

In chapter6 we estimateda dynamic AIDS model for the UK tourism demandin , Portugal homogeneity Spain the France, unrestricted, under and restrictions of and bet imposed France Portugal cross-price effects and null ween only on the and symmetry ' and underthis sameset of restrictionsimposedoil both the long-run [denotedas (H+S)OL], We presented the estimation long- and the short-runcoefficients [denotedas (11+S)OLS]. 6.3. For in dynanlic the table the comparison of elasticities estimates obtained with results AIDS with those obtained with the AIDS model specified in chaptcr 5, wc uscd the I lowcvcr, tile dynamic model with restrictionsimposedonly on the long-run coefflicients. for dynamic AIDS imposed both the theoretical the which of restrictions are on tile version long- and short-run,may be a betterrepresentation of the behaviouralcharacteristics of UK for imposed these long-run. than the tile which version restrictions are on only tourists former is believe the that version 1-Icnce, we able to forecastbetter than the latter. I [cncc, imposed dynamic AIDS with restrictions model only on tile long-run and the version the imposing the samerestrictionsboth on the long- and short-runare estimatedfor the period

269

for forecastingpurposesand a comparison 1969-1993,leaving the four last observations

between the forecasting performances of these two versions will complcment tile analysis of the predictive ability of this class of models. However, if comparison of forecasting performances is to be carried out across all the models specified in previous chapters, we must assure that the time series variables

The being intcrestcd in time the same. series are variabIcs predicted we are arc which France, Spain levels UK Yet, in Portugal. tourism the the shares of of and are predicting J dynamic AIDS model specified in chapter 6 the dependentvariable, representingthe in first differences instead levels. is defined A of sirnple shares, algebraic -destinations' , first difference dependent (AWj=Wjr-Wjt. in transforms the variable original manipulation j) its level forni (Wit), which is the variable of interest for forcasting purposes.This from Wit., 1cft-hand the transferring the tile variable side of n consists of rnanipulatio for its Except Wi,.,, the to the coefficient of side. variable right-hand explanatory cquation V X, all the is (I+%) instead the
regression coefficicnts remain exactly is innocuous in This transformation the terms of the model's that operated shows the same. which now of original

for forccasts Therefore, this transformed model obtaining we use of the cstimation results. levelsbasedon the dynamicAIDS model. UK tourism demand Tables 8.7 to 8.9 report the actual and forecastedvalues of each destination level dynamic (Wit) AIDS the the with obtained share model, and usual expenditure forecast The full the the results obtained with statistics. model under set of summary long-run by imposed denoted forecast the the on solely coefficients are 9; restrictions full imposed both longthe the the under model on with set obtained of restrictions results denoted by Table forecasting 8.7 the the coefficients are short-run *4,. presents results and for France,table'8.8for Spainandtable 8.9 for Portugal.

270

for UK France Table 8.7: Forecasting the expenditure share of results 1994 Actual values Forecasts* Forecasteffore Forecasts*o Forecastefforoo 0.39700 0.34472 0.052281 0.34490 0.036838 1995 0.38540 0.30937 0.076035 0.32012 0.038495 1996 0.38967 0.32640 0.063271 0.32887 0.042336 1997 0.40481 0.29533 0.10948 0.30352 0.10129

FORRESIDUAL AND FORECAST EfUlORS SUMMARYSTATISTICS MAEo MSEo RMSEo MAEoo MSEoo RMSE9* Estimationperiod:(1970-1993) 0.010997 0.0001751 0.013233 0.011914 0.0001986 0.014094 Forecast period:(1994-1997) 0.075267 0.0061260 0.078269 0.069866 0.0052329 0.072339

UK for Forecasting 8.8: the expenditure share of Spain 117able results

1994 Actual values Forecastse D Forecast crrore Forecastsee 0.51857 0.56625 -0.047682 0.56555

1995 0.52625 0.60544 -0.079192 0.59186

1996 0.52292 0,58724 -0.064320 0.58387

1997 0.50691 0.61457 -0.10766 0.60660

Forecasterror** -0.046982 -0.065614 -0.060958 -0.099696 SUMMARYSTATISTICS FORRESIDUAL AND FOIIECAST ElZRORS MAE* MSE* RMSEe MAEe9 MSE** RMSE** Estimationperiod: (1970 -1993) 0.010966 0.0002060 0.014353 0.012295 0.0002572 0.016037 Forecast period: (1994-1997) 0.074713 0.0060682 0.077898 0.068312 0.0050419 0.071006

271

Portugal UK for the Forecasting 8.9: share expenditure of Table results

wwmm

1994 Actual values Forecasts* Forecast crrore Forecastsee 0.084433 0.089032 -0.0045990 0.089544

1995 0.088348 0.085192 0.0031559 0.088015

1996 0.087411 0.086362 0.0010490 0.087253

1997 0.088283 0.090104 0.0018211 . 0.089875

0.0003331 0.0001578 Forecasterror** -0.0015922 -0.0051110 EltRORS AND FORECAST FORRESIDUAL SUMMARYSTATISTICS Forecast Estimationperiod:(1970-1993) period:(1994-1997) MAE* MSE* RMSE9 MAE99 MSE** RMSE&* 0.0064103 0.0000597 0.0077283 0.0063505 0.0000602 0.0077599 0.002656 0.0000088 0.0029802 0.001799 0.0000072 0.0026830

8.3.3.4. VAR models in chapter7 we modelledthe UK demandfor tourism using the VAR methodology forecasted in AIDS We diffcrcnt included the the model. and cstiniated samevariables and forecasts VAR to that the the good were model and came conclusion of specifications . forecasts best VAR, ('pure') the the and were unrestrictedmduced-form obtained with VAR the cointcgrated cxogcnous variables conditioned on under structural with obtained forecasting for The full the results and over-identifying restrictions. set of exactlytiie fornicr by (denoted Spain Portugal France, the obtained with and tc)urism sbaresof forecast The for 7.9. in 7.7,7.8 1), the tables results and tjnrestricted-VAR were presented by latter (denoted 111) destinations' the obtained over-VAR were with shares same 7.17. in 7.15,7.16 tables and presented We now report this sameinformation for the cxpcndituresharesof France,Spain 8.12, is denote The in 8.10,8.11 0 the Portugal to tables and symbol used respectively. and forecastingresultsobtainedwith the unrestricted-VARI model. The symbol DO is usedto denotethe forecasting resultsobtainedwith the over-VAR 111.

272

France UK for the Table 8.10: Forecasting expenditure share of results

wmwmm

1994 Actual values Forecastsc] Forecast errW Forecasts" Forecast erroi3o 0.39700 0.34422 0.052780 0.39505 0.001942

1995 0.38540 0.36840 0.017006 0.39260

1996 0.38967 0,39676 -0.007086 0.39485

1997 0.40481 0.42782 -0.023007 0.39475 0.010054

-0.007194 -0.005178 AND FORECAST ERRORS FORRESIDUAL SUMMARYSTATISTICS Estimationperiod:(1970-1993) 0.015321 0.000315 0.017734 0.016342 0.000362 0.019016

MAEP MSE" RMSE' MAE13E3 NlSt] RMSE13E3

Forecast period: (1994-1997) 0.024970 0.000914 0.030226 0.006092 0.000046 0.006772

UK Spain for the expenditure Table 8.11:Forecasting share of results 1994 Actual values Forecastso Forccasterroro Forecasts"o 0.51857 0.57000 -0.051431 0.51884 1995 0.52625 0.54972 -0.023469 0.52718 1996 0.52292 0.52281 0.000108 0.52568 1997 W 0.50691 0.49058 0.016328 0.52543

Forccasterrd3o -0.000264 -0.000932 -0.018518 -0.002761 EIIRORS FORRESIDUAL AND FORECAST SUMMARYSTATISTICS MAE MSEP F RMSEO E MAE'3'3 "' MsIf" RM SE00 Estimationperiod:(1970 -1993) 0.016475 0.000301 0.019780 0.015845 0.000371 0.019268 Forccastperiod:(1994-1997) 0.022834 0.000866 0.029422 0,005619 0.000088 0.009374

273

Table 8.12: Forecasting results for the UK expenditure share of Portugal

1994 Actual values Forecasts3 Forecast errW Forecastsoo 0.08443 0.08578 -0.001349 0.08611

1995 0.08835 0.08189 0.006462 0.08022

1996 0.08741 0.08043 0.006977 0.07947

1997 0.08828 0.08160 0.006679 0.07982

0.008124 0.007939 0.008463 Forecast erroi3o -0.001678 FORRESIDUAL AND FOItECAST EftRORS SUMMARYSTATISTICS MAEO MSE(3 RMSff3 MAEo
MSff313

Estimationperiod:(1970-1993) 0.008054 0.000088 0.009365 0.006155 0.000063 0.007914

Forccast period: (1994-1997) 0.005367 0.000034 0.005850 0.006551 0.000051 0.007132

RMSEE30

8.4. FORECASTING PERFORMANCE OF ECONOMETRIC MODELS '011 Tilt', JJK TOURISM DEMAND: COMPARATIVE EVALUATION For eacheconometric model previously specified,we now comparctile forecastsof its dependent forecasting the with actual values variable over period 1994the relevant 1997. The objective is to see how closely the forecaststrack its corrcspondingactual , different for forecast in The 8.1 8.12 the tables to statistics crrors reported nicasure values. dependent to the econometric specification of each predict variable valuesovcr tile ability forecasts The forecasting the the the the statistics, closer smaller value period. of arc to the The the variable. quality statistics associatedwith each model, of values observed the itidicate that the forecastsof the.UK demandfor tourism in France,Spain and Portugal data has that the the track and suggest actual each of models good prcdictivc ability. closely , the predictionquality of one particular Yet, howeverusefulthesestatisticsmay be to assess information better forecasts be to about whether obtain might achievedwith specification, have to forecasting devices their have compare we values models, across all we qltcmative ,

14

available. 274

We need to go further in our analysis in order to compare the predictive power of different econometric models and to come to a conclusion about which one provides the best forecasts of the UK tourism sharesfor France, Spain and Portugal. Indeed, we have to forecasts the and respective quality statistics obtained with the sets of competing contrast ARDL models of chapter 4, the AIDS models of chapters 5 and 6 and the VAR models of be judgement for Yet, 7. to this valid, we need to assurethat the competing sets of chapter forecasts are homogeneous, in the sense that they represent predicted values of the same horizon. for time the same time seriesvariable projected The forecasts obtained with each model refcr to the same time, liorizon, which

include the four last observations of the out-of-sampleforecasting period (1994-1997). I-lowcver, the variable we are interested in forecasting (UK demand for tourism) is different in the to alternative econometric spccification according ways measured in For AIDS 5 the the example, while prediction process. equations of chapter underlying demand UK for France, Spainand Portugal, dependent tourism the variable measuring the is the share of the UK tourism expenditurein each destination,in tile ARIJI, models of is log dependent UK 4 the the the the same entity, of measuring variable real per chapter Since forecasts in destination. AIDS the tile obtained with each expenditure models capita 6 VAR 7 for level 5 the those the of chapter models are valuesof tile of chapter and and IJK expenditure share of eachdestination", the only model providing forecasts percentage for a different time seriesis the ARDL model of chapter4. Consequently,for forecast comparison purposes,the dependentvariable in the the level of the UK ARDL modelshas to be transformedinto a time seriesrepresenting destination. budget The transformation process,although time share each of tourism is Taking forecasting ARDL the the original straightforward. models and consuming, dependentvariable for the relevant forecastinghorizon, we obtain the forecastsgiven in We forecasts 8.3.3.1. then the these compute antilog of and multiply each antilog section deflatorthat was originally usedin the ARDL model to forecastedvalueby the appropriate into In level this the real expenditure. expenditure way we nominal nominal obtain change destination. UK Summing tourism to the thesevalues allocated expenditure each of values
13Although the original dependent variablesin the AIDS model of chapter6 are the first differencesof the destinations'shares andnot their level valuesas required,we showedin section8.3.3.3.how to transformthe into desired level first difference the variables variables.In the same section we computedthe original sharesin levels and not in first differences.Hence,theseforecasts forecastsfor the destinations' are readily AIDS the those static modelof chapter5 and VAR modelsof chapter7. of with comparable

275

forecasting for destinations the period, we obtain the total nominal per eachyear of across capita tourism expenditure allocated to the whole set of destinations. Dividing tile by in level forecast this total, tile we obtain each country value of the cxpenditure for in destination forecasting Note the that tile each year of share each period. expenditure invariant level is total per capita whether values are or expenditure considered. shares' By this processwe are able to transformthe original "real per capita expenditure" forecasts of the ARDL model into the "expenditure shares" forecasts,which tire the for forecasts as they are tile time seriesprcdictcdwith all tile purposes comparison rclevant for ARDL Hence, 4, 8,13,8.14 the tables models of chapter models. econometric and c)ther 8.15 presentthe actual and 'transformed' forecast values of the expenditureSharesof France, Spain and Portugal,respectively,and the summarystatisticspreviously used for forecasts. the the accuracy of cvaluating
Table 8.13: Forecasting results of the ARDL model for the expenditure share of Francc

EFo

1994

1995

1996

1997

0.3854 0.3970 0.3897 0.4048 Actual values ' 0.3926 0.4140 0.3951 0.4170 recasts Forecast 122 c recast error -0.0072 -0.0170 -0.0054 -0.0 FORRESIDUAL SUMMARYSTATISTICS AND FORECAST ERRORS MAE SE ms M RMSE 1 Forecast period:(1994-1997) 0.01045 0.0001297 0.011389

Table 8.14: Forecasting resultsof the ARDL model for the expenditure sharc, of Spain I 1994 1995 1996 1997 Actual alues a v values Forecasts rccasts t error aS 0.5186 0.4946 0.0240 0.5263 0.5215 0.0048 0.5229 0.5206 0.0023 0.5069 0.5085

-0.0016 FORRESIDUAL SUMMARYSTATISTICS AND FORECAST ERRORS Forecast period:(1994 -1997) MAE 0.008175 MSE M 0.0001517 SE RNM 0.012318

276

for Portugal ARDL the Forecasting the 8.15: Table model cxpenditurcshareof resultsof 1994 Actual values Forecasts Forecast error 0.0844 0.0914 1995 0.0883 0.0859 0.0024 1996 0.0874 0.0843 0.0031 1997 0.0883 0.0746 0.0137

-0.0070 AND FOIZECAST EM1016 FORRESIDUAL SUMMARYSTATISTICS Forecast period: (1994 -1997) 0.006550 0.0000630 0.007938

MAE MSE RMSE

We can now gathertogetherall previousinformation,and comparethe ability ofthe VAR forecast UK AIDS dynamic AIDS, to tourism the models and ARDL, static Portugal, 1994 Comparison Spain France, the over and period of -1997. expenditureshares facilitated including is by in 8.16,8.17 forecasting best tabics the model the choiceof and Spain Portugal, France, for forecasting the 8.18, and respectively, of the shares results and AIDS the unrestricted static specifications namcly: relevanteconometric obtained with all homogeneity, AIDS by (denoted under symmetryand null crossstatic-AIDS);static model France by Portugal [denoted between the of and equations static. price effects dynamic AIDS model under homogeneity,symmetry and the same null AIDS(11+S)OI; [denoted dynamiclong-run imposed the only on cociTicients cross-price effects imposed both AIDS dynamic the set of restrictions model undcr same on AIDS(11+S)OL]; [denoted dynamic-AIDS(11+S)OLS]; long-run coefficients unrestricted short-run and the ARDL (denotedby ARDL); unrestrictedreduccd-form('purc') VAR model (denotedby VAR under the full set of exactly- and over-idcntifying restrictions VAR I); cointegrated (denotedby ovcr-VAR III).

277

UK France for the Table 8.16: Forecast expenditure share of results FRANCE Actual values Static-AIDS O Static-AIDS(H+S) O L Dynamic-AIDS(11+S) DynamicAIDS(II+S)OLS ARDL VAR I Error 1994 1995 1996 1997 M A E RM SE

0.3970 0.3854 0.3897 0.4048 Forecast 0.4003 0.3971 0.4271 0.4602 -0.0033 -0.0117 -0.0374 Forecast 0.3778 0.3804 0.4019 0.3883 0.0193 0.0050 -0.0122 0.0165 Forecast 0.3447 0.3094 0.3264 0.2953 Error Error 0.0523 0.0760 0.0633 0.1095 0.0368 0.0385 0.0423 0.1013 Error

0.0269 0.0339 0.0132 0.0143 0.0753 0.0783

Forecast 0.3449 0.3201 0.3289 0.3035

0.0699 0.0723

Over-VAR III

Forecast 0.4140 0.3926 0.3951 0.4170 1 0.0105 0.0114 Error -0.0170 -0.0072 -0.0054 -0.0122 Forecast 0.3442 0.3684 0.3968 0.4278 0.0250 0.0302 -Error 0.0528 0.0170 -0.0071 -0.0230 Forecast 0.3951 0.3926 0.3949 0.3948 0.0068 v. vuu Error 0.0019 -0.0072 -0.0052 0.0101

278

Table 8.17: Forecast shareof Spain resultsfor the UK expenditure SPAIN Actual values Stati c-AIDS ic-AIDS(fl+S )o Stat OL i ( 11+S ) AIDS Dynam cO i ( H+S ) LS AIDS Dynam c ARDL VAR 1 Error Error 1994 1995 1996 1997 MAE RMSE

0.5186 0.5263 0.5229 0.5069 Forecast 0.5007 0.5041 0.4691 0.4288 0.0179 0.0222 0.0538 Forecast 0.5261 0.5224 0.4979 0.5126 0.0039 0.0251 0.0057 -0.0076 1 Forecast 0.5663 0.6054 0.5872 0.6146 0.1077 -0.0477 -0.0792 -0.0643 . Forecast 0.5656 0.5919 0.5839 0.6066 -0.0470 -0.0656 -0.0610 Forecast 0.4946 0.5215 0.5206 0.5085 0.0240 0.0048 0.0023 -0.0016 Forecast 0.5700 0.5497 0.5228 0.4906 0.0001 0.0163 -0.0514 -0.0235 Forecast 0.5188 0.5272 0.5257 0.5254 Error 0.0 85 -0.0003 -0.0009 -0.0028 . -0.0185 Error Error Error Error

0.0430 0.0495 0.0106 0.0135 0.0747 0.0779

0.0683 0.0710 0.0082 0.0123 0.0228 0.0294

VAR 111 Over-

0.0056 0.0094]

279

Portugal UK for the Table 8.18: Forecast expenditure share of results

PORTUGAL Actual values Static-AIDS F o ) Static-AIDS(11+S) c 0 ic Dynam -AIDS(II+S) L EDynamic-AIDS(11+S) O LS 1994 0.0844 Forecast 0.0991 Error 1995 0.0883 0.0989 1996 0.0874 0.1038 1997 MAE 0.0883 0.1111 RMSE

-0.0146 -0.0105 Forecast 0.0961 0.0972 -0.0117 -0.0089 Forecast 0.0890 0.0852 Error 0.0032 -0.0046 0.0880 Error

0161 0 0 0167 - . . -0.0164 -0.02281 0.1003 0.0991 0.0110 0.0111 -0.0129 -0.0108 1 0.0864 0.0901 0 0027 0 0030 . 0.0010 -0.00181 . 0.0873 0.0899 0.0002 -0.0016 0.0843 0.0746 0 0066 0 9 7 00 . 0.0031 0.0137 . 0.0804 0.0070 0.0795 0.0079 0.0816 0.0067 0.0798 0.0085 0.0018 0.0027

Forecast 0.0895 Error

0.0003 -0.0051 Forecast 0.0914 0.0859 0.0024 -0.0070 Forecast 0.0858 0.0819 Error Error 0.0065 -0.0013 0.0802 0.0081 -0.0017

ARDL

VARI

0.0054 0.0059

Forecast 0.0861 Over-VAR 111

0.0066 0.0071

The first thing to bc said aboutthe forccastingperforniariccof the modelsis that it be if fact, forecasts had In fairly be to to only obtained using one of overall. accurate seems for forecasts companison purposes, any the modelswithout availability of competingsetsof forecasting device. However, be as confronted them a good qualified with would ()f forecasts, from decide information tile we can which other sets of of supplementary budget UK for forecasting is tourism the sharesof econometric models more reliable France,SpainandPortugal. Analysis is facilitated by interpretationof both the MSE and RMSH statistics in forecasted. is T'his the the variable variable measurement cach of of original units tcrins of destination's shareof the UK tourism budget and is measuredin percentage points (pp). Ilence, the statistics' values representthe averagedeviations of the forecastsfrom the in instancc, For (0.0066) MSU the points. percentage measured and values of actual values, F_MSE(0.0071) for the over-VAR III model forecastingthe share of Portugal may be by the interpreted as which the forecasts of this variable dcvintc from the amount

280

from forecasts deviate In this then, the the the case, true series. of values corresponding less (0.7 by (absolute) than of erfor one percentage pp) on point variable's actualvalue an average. Another complementarymeasureof accuracy can be derived by comparing the This be forecast the the values can actual series. of measure average errors with average forecasting interpretation for the the correct and comparison of models" crucial less (absolute) interpretation Indeed, than the computed average a of of error performance. ipp for a seriesof observedshares averaging8%, is quite different from that of the same in interval because is (0.000; This 50%. for the value so a statistic a seriesaveraging crror 0.005), which meansa maximum average(absolute)error of half of a percentage point, leads to different (relative) error intervals dependingon the averagesize of tile share's forecasting 1994-1997, For the period a statistic value considering example, actual values. 5.7% (relative) (0.005) the to half corresponds a error of average percentage point of a of Yet, 0.005/0.0871=0.057. less for (0.0871), Portugal the same statistic value means as share for Spain (0.5187), 0.005/0.5187-0.0096. We 1% to the share as average error relative than justifies (absolute) the transformation importance these the of average that aspects of think into intervals (relative) by intervals the average statistics, error Which quality given crror Since, into the to the the errors relative actual series. of overall, weight consideration take both the MSE and RMSE quality criteria give similar information,we concentrate attention hereafter. RMSE the statistic on
For the purpose of comparing absolute errors with relative errors, we present table 8.19 which shows, for the destinations' average share values in the period 1994-1997 (denoted byW94-97). the absolute error range of the RMSE statistic, the corresponding for forecast lay the the models which errors within these ranges. error range and relative

281

for destination forecast the Absolute 8.19: shares error range Table andrelative
E RMSE t absolute r ro rof c rrorrange Destination
W94-97

RMSE. Relative error range (0%-1.5%) (0%-1%) (0%-6%) (1.5% - 2.5%) (1%-2%) (6%-12%) (2.5% - 4%) (2%-3%) (12%-18%) (4%-6%) (3%-4%) (18%-23%)

Models and respective RMSE absolute and relative errors

France (0.000; 0.005) Spain Portugal France (o. 005; 0.010) Spain Portugal France 6-01 0.015) n2 nI (0.010; (a , Spain Portugal France (0.015; 0.020) Spain Portugal

0.3942 0.5185 0.0871 0.3942 0.5185 0.0871 0.3942 0.5185 _ 0.0871 0.3942 0.5185 0.0871

DrainicAIDS(I I+S) L: 0.0030(3.4%) AI DS(II+SVI.S: 0.0027(3.1%) Ovcr-VAR 111: 0.0068(1.7%) 0.0094(1.8%) Ovcr-VAR 111: VAR 1:0.0069(7.9%) Ovcr-VAR 111: 0.0071(8.2%) ARDL: 0.0079(9.1%) ARDL: 0.0114(2.9%) (3.6%) Static DS(I I 0.0 143 +S)O: , -Al ARDL: 0.0123(2.4%) 0.0135(2.6%) Static-AIDS(11+Sf: Static-AIDS(I[+S)O: 0.0111(12.7%)

Static-AIDS: 0.0167(19.2%)

differ from Table 8.19showsthat mostmodelsproduceforecasts which, on average, indicates This low 1.5 by less than these points. average error percentage series actual the below 1.5 devices. However, forecasting while an absolute error percentage models asgood 4% France Spain, below to the tile relative average shares or same of error an means points 18% Portugal. to to the of error up an relative average share of absoluteerror can mean suchas ,I*his indicatesthat modelswithin the samequality rangefor all destinationshares, 1.5 from I to with model an average absolutc errors ranging tile static-AIDS(11+S)o 282

from destination, producesaveragerelative crrors ranging points whateverthe percentage 2% to 3% of thc'sharcfor Spainand 12%to 18%of tile 2.5% to 4% of the sharefor France, has between hand, ARDL I On for the Portugal. to the an absolute error range other share has France for Portugal, Spain the the 1.5 percentage shareof and and, of shares points-for less for former This 1 0.5 the lower than to gives of a relative error model pp. the rageof 4%, while for the latter the relative error is 9.1%. IIcncc, the forecastingperformanceof in differ the they relative crrors produce, even when they are considerably can models interpreting As in the tile the when accuracy a result, of range. error same absolute qualified both be the of absolute and average errors aware relative must a researcher statistics, quality for a more reliablerankingof the models. We now turn to the analysis of the results shown in table 8.18.'An immediate is indication given by the resultsof the forecastquality statisticsfor all destinations'shares, full than taking those worse perform account or partially,restrictedmodels that unrestricted betwccn behaviour. The difTerence bounds the theoretical performance of consumers of the VAR fully AIDS is the of static specifications and restricted models and unrestricted between difference fully the The the partially restricted and performance remarkable. is is bccausc, AIDS This dynamic the model not as significant. so restricted versionsof does AIDS VAR the the'static the and models not considcr version of unrestricted while by dynamic theory, the the economic suggested partially constraints rcstrictcd any of in in does long-run. Thus, the them the the so short-run, considering not model, although imposition indicate theorctical that the to tile appropriate restrictions of oil seem results forecasting bcttcr leads in to performance. models econometric of coefficients in does 'what is forbid that where researchers often as physics say nature not econometrics, forbiddcn, is is in 'what (statistically) (modelling) that theory', say we not compulsory do Therefore, theoretical models which not reject appropriate constraints compulsory'. describe forecast better hypotheses, to be thcm, these tile and allowing must modelledunder underlying economicreality. Another straightforwardindication given by the forecastingresultsreportedis that is forecasting best The best the the system of equations model. cointcgratcd worst and of VAR model underthe full set of exactly- and over-identifyingrestrictions,which produces for below 1 destination percentage point all errors shares,and average absolute average for France 2% below Spain 8% the the average of shares and and around of errors relative The is for Portugal. AIDS the worst, unrestricted share model, which static average

283

for destination 1.5 5 to percentage all ranging points produces averageabsoluteerrors 8% for 9% France, the of average exceeding share of the errors relative shares,and average for Portugal. I Icncc, for 19% Spain the the average share remainder of and of share average AIDS forecasting the this we exclude unrestricted static model exercise, comparison of from the comparative analysis. The secondbest forecastingsystemseemsto be the rcstrictcd static AIDS model. This model suppliesforecasts absoluteerror not exceeding1.5 pcrccntagc with an average below 4% destination for the of avcragcshare relative errors average and shares, all points for France,3% of the averagesharefor Spain and 13% of the averageshareof Portugal. basis, the ability of the 11owevcr,if we analysethe systemson an equation-by-equation Portugal, Spain France is forecast by AIDS the to of shares and out-perfornicd model static by in in Portugal, ARDL the the cascsof the the and case of model models other that of all Franceand Spain.This singleequationmodel,produces averageabsolute errorsbelow I pp for the shareof Portugalandbelow 1.5pp for the shares of Franceand Spain.Furthermore, I forecasts 2.9% this of model's errors relative arc average of the average the corresponding for Spain 9.1% for for 2.6% France, the the share and of avcrage average share of share Portugal. Nevertheless, the analysisof the predictive ability of a systemshouldbe carried out basis, basis, than that an cquation-by-equation rather as of a single cquation system a on basis be Indeed, there on equation carried out a separate and not as a systcm. should rnodel. by in be considering all can only answered equations a systemas a whoic arc querieswhich best forecasting from fact instance, For the that the the and worst aside modelscall Cntity. to be no easyway be readily pointedout from the information reportedearlier,thereseems between in I Icnce. distinguishing the their there other models relative predictive power. of for in be tile to some originality method of ranking space conceiving a models seems forecasts. their to the of accuracy according by is The forecastingperformance the of sevenmodels ranked awarding, for cach destination share,a maximum of 7 marks to the spccification with the lowest average lowest is 6 in This done to the next and so on. marks error, such a way that the absolute highest I Then, the average absolute error gets only mark. summingthe marks with raodel destination be total to shares, we obtain across a which used model can comparethe each of forecasting accuracyof the models. For example, while the dynaniic VAR system or full the (denoted by set of exactly and over-identifying under restrictions ovcrequations

284

VAR III) gets 7 marksin the shareequationfor France,7 marks in the shareequationfor Spain and4 in the shareequationfor Portugal,summinga total of 18,the static unrestricted AIDS system(denotedby static-AIDS)obtainsonly 3+3+1=7marks. Figure 8.1 showsthe total sum of forecasting models accordingto their specification featuresand associated in brackets. marks Figure 8.1: Models' specificationfeatures and rank.

STATIC

UNRESTRICTED=: > AIDS (7) RESTRICTED=: > Al DS(II+ S) (12) IUNRESTRICTED=:>VAR 1(13)

SYSTEM

N40DEL

PARTIALLY RESTRICTED=* Al DS(II+ S)L (8) DYNAMIC, AIDS(II+S)LS(ll) RESTRICTED=: Ovcr- VAR 111 (18)
I IDYNAMIC STATIC SINGLE (UNRE STRICTED =* ARDL (15)

With this unorthodoxmethodof ranking the predictionquality of diffiercrit models, better forecasters that than systems of establish equations are single equation easily can we , fully dynamic out-perfonn models static models, and restricted specirications rnodels, forecasts Wc than unrestricted partially restricted models. or can also more accurate provide forecaster best III 18 the the over-VAR model with spot marks- and the worst readily 7. the forecaster static-AIDS unrestricted model with This schematic way of ranking the modelsalso permitsthe identification of aspects inconsistencies. instance, For if bcttcr as appear systems may of cquations are that than single forecasters modelswhy doesthe ARDL model out-perforin the AIDS -equation The in fact UK dynamic the explanation that the the resides of equations? aspects or system basis,than the demandfor tourism seemto be more relevant,on an equation-by-cquation

285

destinations basis. interactions are existing among on a system of cquations potential Consequently, an econometric specification such as the ARDL model that takes into dynamics the of the adjustment process, although not contemplating short-run account forecast is interactions, better does to than expected a model which not take equations important dynamics. These behaviour UK tourists' are characteristics of account of explicit included it a model's structure, can within make once a powerful and reliable which, forecasting device. However, the single equation form of the ARDL model does not allow for testing the theories of consumer demand behaviour and this limitation call affect its qualities. On the other hand, the ARDL specification differs substantially from all the other

UK demand it behaviour in destinations tourism the the towards tile way portrays models UK budget This destination tourism the to models specification allocated each considered. income. does It destination function function not model of and prices sharesas a or as a Ilencc, "transform" tourism tile original ARDL although we could expenditure. and prices in poundssterling) into tourism shares (measured in forecastsof tourism budgets(measured itsclrand did directly "transform" the model we cannot not points),we compare pcrcentage from different dependent Apart formation" "trans this variables. with which made models forecasts ARDL different fron) the model of possible, remains a approach all comparison Unless the same variables critcring tile AIDS and VAR models. the other econometric demand be UK for framework ARDL to the tourism used can model the within an systems between former latter limited. Tile ARDI, the the and are rnodclling with tile comparisons included in is AIDS is VAR but it the for variables and models possible, a matter same future research. Nevertheless, an ambiguity needingclarification still remainsin the model ranking dynamic displayed If forecast better to than above. specifications are expected scheme does AIDS dynamic the the restricted why static ones, out-perfonn model restricted static AIDS? Given the inherentdynamicnatureof tourism demand,one would cxpcct a general dynamic AIDS specification to forecast better than a static one, irrespective of the destination. Yet, a comparisonof the rcstrictcd forms of the static and dynamic AIDS favour dynamic in the France to than the static the rather seems cases of and Spain, models Portugal in drawn. the be the Two setsof reasons of opposite case conclusion can can while to explainthesefindings: one can be qualifled as statistical;the other embeds be suggested theoreticalandempiricalarguments.

286

The staticAIDS model specifiedin chapter5 appears to be statisticallymore robust 6. both Although dynamic AIDS the of chapter modelsrespondpositively to all model than the diagnostictestsperformedfor evaluatingtheir statisticalquality, the static AIDS model dynamic instance, For do than the to model. severalof the convincingly so more seems in dynamic insignificant. be the to the system appear statistically equations coefficients of yet, a solution to improve this model is not the simply deletion of the insignificant dynamic implcmcnt jeopardise the to this general structure we wanted would variables as in the first place,and disturb the invarianceof the cocfflcicnts' cstiniatcsconfcffcd by thc hand, On SUR tile the other retaining tile estimation method. within adding-up property insignificant variablesmay leadto imprecisionof the estimates which may take away sonic indicate These to that tile the seem considerations accuracy. more statistically model's of for it be forecasting is, to the appears morereliable purposes. robust a model However,if we comparethe estimationresultsof the cquationsfor Portugalin the
in dynamic AIDS indicates bcttcr the the static models, spccification and statistically static dynamic dynamic AIDS the bcttcr than the and yet, still results model provides estimation forecasts of the UK demand for Portugal than the static model. It seems, then, that the just in than the Tile tile statistical areas robustness models. reside more of must explanation j in found dynamic be two the ma or aspects of short-run may adjustment explanation for each destination: the factors that arc mainly demand behaviour tourism underlying responsible'for being used to dcpict this adjustment mechanism. nicsc will now be tools econometric considered. One key aspectof the particular static AIDS model spcciried in chaptcr 5 is that it is dragging the dependent variable towards its equilibrium path, and tile

The Deaton Muellbaucr (1981) AIDS AIDS model. and orthodox orthodox an model, riot Tile destinations' "unorthodox" forill on and per capita expenditure. shares prices regresses in dependent 5, AIDS the the same variable chapter rcgresscs adopted on prices, model of for break trend time tile variable, and a structural a models cocfficicnt of and xpcnditure c, Recall AIDS income that the the orthodox version of variable. model was convincingly the data, by form 5, tested the the when against either unorthodox of chapter or against rejected in form 6. dynamic Therefore, chapter specified we concludethat on the one the general hand, a trendvariablemust be includedin an AIDS specificationwhich doesnot explicitly dynamicsof tourism demandand, on the other hand, inherent for the short-run account

287

is trend the the explicitly modelled, variable mechanism adjustment equilibrium when becomes superfluous and the orthodox AIDS model is rejected.

In the AIDS model of chapter 5, the trend is a highly significant explanatory in but Spain for In Portugal. for France in tile the the not so and equation equations variable be key factor in Spain, for France trend to the variable seems a mimicking and cquations demand behaviour, in for dynamic inherent tourism the while equation of nature the to fail this objective.Apparently,the dynamic processby which tourists' portugal it seems behaviour adjusts to their long-run equilibrium in the share equation for Portugal, is different from thosefor Franceand Spain.This may be a reasonwhy tile AIDS model or in dynamic AIDS France Spain but better 5 forecasts the the than cases of and not chapter Portugal. in the case of so long-run cquilibrium niodel and its A static model is, by definition, a steady-state
coefficients information give long-run impacts the that changes in the explanatory about if for Therefore, dependent destination have the the variable. any on particular variables long-run effects happen to be more important than the short-run ones, a steady state parsimonious is be likely to AIDS trend more adequate than the complex model with static

dynamic in AIDS. If, the the trcild general of addition, structure over-parameteriscd and features of the demand behaviour, in inserted the model can capture steady-state variable like habit persistence, which are not accommodated by other rcgressors, we have a main for believing that for destinations for which these features arc particularly relevant, a reason demand is describe better behaviour its to the trend than able underlying a with model static

be dynamic We believe for Francc to the this counterpart. case 4 and over-parameterised Spain. For Portugal, however, the relevant dynamic factors seem to be linked with other features habit do be to UK the and persistence not appear a characteristic of reasons, in Hence, lag demand for this this country. case, an explicit structure of the tourism better dynamic demand behaviour. be This be to the to capture able sccms might regressors in insignificance by for in Portugal the trend the the tile of variable equation confirmed by forecasting AIDS dynamic the the superior and performance model, of general static form of this model for this particular destination. However, as noted previously, the comparative analysis of tile predictive ability or be conducted considering each system as a whole, rather than should econometric systems in an equation-by-equation perspective. From this point of view, the forecasting ability of the unrestricted VAR model out-performs that of all AIDS specifications, being second best

288

its VAR. formal Given fully to the simplicity and ease of restricted cointegrated only (pure) has VAR desirable to the the all make unrestricted qualities reduced-form. estimation, its choice an easyone for forecastingpurposes.Yet, if the most accurateforecasts of tile destinations'-sharesare required, then the predictive ability of tile fully restrictedVAR is fully VAR This the that not surprising since restricted other models. of all overshadows incorporatesthe threemain features of a quality forecastingmodel: it takesexplicit account hypotheses for includes dynamics, theoretical the rational accounting of short-run behaviour of consumersand uses a formal specification matching theory and data in a is both framework empirically plausibleand statisticallyrobust. which quantitative 8.5. CONCLUSION Kierkegaardsays that life must be lived- forwards, but can only be understood
backwards. This applies also to economics, whcre the reality of a particular phenomenon its behaviour, by in become it the past understanding and ways apparent which only may if behaviour be future. Yet, foreseen in be 'lived' in the economic could can only Cvolves

its knowledgeable future it to approach conscqucnccs. a more the present, would permit important. I'licy are fundanicritalinputs in This is the reasonwhy accurateforecasts are so decision-making policy strategy. and economic processes An appropriatelyspecifiedeconometricmodel can be projectedinto the future to forecast the behaviour of the underlying economicreality it portrays. Yet, econometric forecasting devices. formal These than simple models arc more quantitative are rnodels data, linking theory observed and which allow us to understandeconomic relationships, behaviour, interpreteconomicfacts, test economictheoriesand evaluateeconomicpolicy. Forecasting,however,hasa crucial role in economicanalysisand is an integral part of the Therefore, the this to econometric chapter of modelling. goal of process was evaluation forecasts from key the to the econometric previously address models specified and obtain by forecasting their their statistical quality comparing performance. of evaluating problem Many methodscan be usedto obtain time seriesforecasts. I lowevcr, irrcspcctivcof have be into in there to taken aspects are which selected, method account order to the in in forecasts the the method employed and reasonable reliability accuracy ensure First, the time the has path of to crnbcdcertain phenomenon to we want predict produced. in the future. Then,theseregularitiesandthe features 'regularities' which self-reproduce of 289

by depicted be the selected adequately captured and must process their self-reproducing forecastingmethod.The former is a characteristic of any economictime series.The lattcr future. into forecasting the to the depends the adopted project series model specific on in previouschaptersindicatesthat the quantitative The empirical evidencegathered
frameworks chosen for modelling the UK demand for tourism were theoretically coherent, the underlying of robust representations regularities statistically and plausible empirically in 4, investigation. The ARDL chapter model specified the economic phenomenon under in 6 VAR 5 7 in the AIDS used ch3ptcr and and approach chapters systems estimated the level in tile be, values their own right, appropriate means of anticipating were considered to Portugal. Under Spain for France, demand this UK we used and assumption, tourism the of forecasts in-sample 1969-1993, to the period obtain point of those models, estimated over 1994-1997. in destinations' the period out-of-sample shares the The statistics used to measure the accuracy of the forecasts indicate that most

fully VAR ARDL the the model, unrestricted and restricted specifications, namely AIDS forecasts the system, produce with average static restricted specifications and be below 1.5 can viewed which as reasonably accurate. points, percentage errors absolute Beyond this range, for one or more destinationshares,are the unrcstrictcdstatic atid tile indicate dynamic The AIDS that also dynamic results modelspcrronn better specifications. forecasts fully tilan tllcir models generate more precise restricted than static ones, VAR the counterparts, cointcgratcd and spccification restricted unrestrictedor partially best forecasting is full the the model. setof restrictions under Therefore,if the goal is to obtain the most accurateforecastspossibleof the UK
for France, Spain Portugal, budget the appropriate econometric and shares tourism dynamics and the imposition of appropriate theoretical into has take to account niodelling framework Furthermore, the the variables within which share quantitative arc restrictions. involve the variables arc a system of equations where not a preferably should rnodelled priori assumed as exogenously given, and an appropriate account of qualitative events dependent All the the the of variables sample path over period. affect which significantly is VAR these the characteristics specification, of cointcgmted with model econometric is form to those the exactly-identifled with used restrictions matching structural which identify the shareequations of a steady-stateAIDS system. However, if a cost-benefit perspective is important and less accurate forecasts are ('pure')VAR be has the the to reduced-form unrestricted model considered a problem, not

290

best predictor.The model's remarkable qualities of simplicity of form, caseof estimation fairly its forecasting it accurate assumptions alongside ability, give prime and minimal if Yet, interpretation to the models. a sensible other econometric economic relative of place the long-run relationshipsis also required,the estimationof this modci's structural form does not help. Rather,its coefficientsarerevealedas theoreticallyinconsistent, empirically implausible and statistically insignificant estimates of the corresponding long-run fully In VAR the the of contrast, estimation restricted cointegratcd gives parameters. long-run linking the of estimates structural plausible and significant rclationships coherent, theseestimates arc similar to thoseobtained the variableswithin the model. Furthermore, for AIDS be this to the additional support providing system model restricted with behaviour destination the tourism the a reliable of as means explaining of shares considered 1969-1997. the period over

291

CHAPTER 9
CONCLUSION

has The key role that tourismcanplay in assistingeconomicgrowth in destinations Accordingly,research beenbroadly acknowledged. on the economics of tourismhasbeena demand forecasting interest, have been tourism modelling and studies and growing areaof literature. in frequently Ilowcvcr, the majority of thesestudies economics reported more fail to apply theoretical and methodological tools which are fundamental to tile for models reliable explaining and predicting tourism construction of accurate and limitations has The demonstrated these recognition of models' the recently phenomena. identify hand, for improvement that, the the that clearly one areas studies on require need hand, the which can amelioratethe reliability of the results. other provide methods and, on to tourism demandanalysisconsistsof both extendingtile thcorctical This new approach fornial quantitative models content of previous and of presenting empirical alternative and describe demand bchaviour. that tourism These and can predict validly specifications be based theoretical sound on grounds and subjectedto extensive must specifications or otherwise,their empirical findings. testing to substantiate, in In this thesis,we examinethe methodologies applied the existing literature on demand limitations. Then, we discuss alternative their and analysis point tourism out both to crilarging the theoretical basis or approaches which contribute methodological implementing in and recent advances econometric modelling, currently used models, hypothesis forecasting Throughout testing the study we and evaluation, procedures. quality based how theoretical on sound models grounds and constructed tinder the show rules can withstand rigorous statistical scrutiny and provide a appropriate'econometric description demand its tourism and reliable of and accurate predictions consistent of behaviour. This study makes several important contributions. The thcsis analysesthe UK demand for tourism in its geographically proximate neighbours, France, Spain and The choice of tile countricsinvolved took portugal, using data for the period 1969-1997.

292

into account the fact that the UK is a major tourism origin which is of particular importanceto France,Spain and Portugalas destinations.Moreover, Spain and Portugal for interesting they were experiencing considerableeconomic as analysis cases are developmentduring the sampleperiod.At the beginningof the period, they demonstrated flinal but by the the they sample, countries year of of underdeveloped characteristics many developed European France In included be the economics. contrast, more alongside could for between developed the sample period, allowing whole comparison country over was a it and its poorer neighbours.Furthermore,Portugal and Spain underwentmijor political in 1980s (the 1974, Portuguese in 1970s the the and revolution changes economic and in democracy 1970s integration in from dictatorial tile tile to tile mid and regimes change in the models.In addition, F-U in 1986).It was importantto take accountof thesechanges development different by Spain the stages attained and and neighbourhood of the concept interesting time of portugal over aspectsof the changesIn tile permittedtheanalysis intcr-dcpcndcncics, destinations by dynamic behaviour their and means or of competitive frameworks. modelling The majority of early empirical studiesof tourism demandhave rclicd on static

based hoc demand These thcory arc models not on ad and, models. equation single dynamics, data, fail to take of nonstationarity and potential of account generally, inference in invalid forccasting bias in statistical and poor resulting ability. simultaneity deficiencies, flaws, inadequacies 3, theoretical technical the and specification chapter demand tourism tile traditional modelling were cxamincd and with rcquircd associated following inadequacies In these the chapters, models' wcre solutionswerepointedout. by the of alternative construction and estimation econometric spccifications overcome be both 11lcsC to theoretically consistent and cnipirically were shown plausible. which incorporate from dynamics; derived theory, the consider economic are specifications interrelationships within system and test utility theoryliypotlicscs amongdestinations of frameworks; question pre-assumed cridogenous/cxogcnous constraintsand equations bias The using a vector autorcgressivc modclsestimated approach. simultaneity overcome to rigorousquality scrutinyusingboth tile traditionallyquality in this studyweresubject testing, constancy methodologies suchas structural criteria andmorerecenteconometric The testing and exogencity cointegration. and analysis. resultsshowed statistical causality long-run between UK the of the a tourism existence of equilibrium relationship evidence

293

forecasters better than the demandand its determinants some models of arc and, although fairly goodpredictivequality. others,all present introduced in each The theoretical,methodologicaland econometricapproaches from to the the structure complex, providing an expanding more simple proceed chapter by towards tools incorporates, of new concepts and analytical a a number step, step which destination demand for its UK tourism the ncighbouring examinationof comprehensive issues detailed however Yet, the theoretical coverage of practical and wide and countries. is in this study, there are important matters that, for one reasonor another,had to be inclusive direction The this structure ignored or addressed of study's only superficially. had to be decidedat the beginning,and the choiceof one direction automaticallyexcludes interesting fruitful Moreover, be the tile as and wide one explored. as may which others for demand suitable approximating consumer range of quantitative specifications behaviourare suchthat a completelycomprehensive analysisof them would require more in lack by The this the thesis of consideration of such study aspects subject. on no than one but for inclusion in importance diminishes the their their enhances rather need means future research. The objective of this concluding chapter is, therefore, to identify the main in the the theoretical crilarging scope of and econometric chapters of previous contributions demand in tourism analysis,supplyinga rigorousand original applied currently approaches demand for its based UK tourism tile the southern ncighbours on of economics of analysis issues directions in tile which of sonic rcqu*lrc statistical resultsobtained,and pointing out further research. Chapter2 explains,by meansof basicstatistics,graphsand tables,the evolution of its for demand UK tourism southernneighbouringcountriesover the sampleperiod the The analysisfocuseson the significanccof UK tourists' preferences 1969-1997. relative to importance Spain Portugal, for France, these the of countries and on and origins other for for 17his destinations UK tourists. tourism to guidance analysis provided relative other in adopted subsequent chapters since the appropriate the quantitative approach definition linking the the and variables of quantitative of relationships characterisation features facts trends, on a good understandingof and them, are generally grounded behaviour A the time tile thorough the of series under study. examination of affecting is, information importance in Yet, therefore, of paramount quantitative analysis. available in tourism researchcontexts, detailed and reliable information in suitable quantities is

294

difficult to gather.The main reasonis that the relevant time seriesof tourism data are levels basis, for at very aggregate and relatively short generally available on a yearly demand describing implies This the that time. quantitative approaches can periods of "which Is do tourism than treat commodity as* a generally undifferentiated other scldom by destination is by than purchased consumers who are unspecified other and which other 1997, 215). Yet, Stabler, by (Sinclair tile considerationof tile p. and than nationaliV' behaviour features and the application or more powerful of consumers' microcconomic demand depend data the tourism to analysis on availability of niethodological approaches link demand level, household disaggregated types can specific of with specific the which at issues is importance for Future these tourists. of paramount concerning research types of in differcntinte by tourists identifying the preferences market nichcs which structure such by its income level, the tourism or gender, and age group, product social as criteria Nevertheless,aggrcgatedata pen-nit different componentsand mixes of characteristics. importance for levels are also of grcat economic and policy analysis, which of other findings by in is demonstrated the this thesis. empirical provided as cxamination Chapter3 beginswith an overview of early empirical research on tourism analysis,
have becil the that econometric specirications used to explaining and critically evaluating functional forms demand. The the the type usually adopted, examines chapter of estimate from differcrit included the results obtained static single equation estimation and variables different The that slightly spccirications can produce considerably results show models. different inconsistent thereby providing results upon which no reliable estimates, from lack The disparities be based. the to seem emerge of a sound conclusions can framework theoretical and/or a prevalent within Which plausible methodology empirical hypotheses be fully integrated behaviour Static tested. can single equation and consumer interdependencies demand destinations, tourism among competing neglect models of basis lack dynamics theoretical a sound within which consumers' preferences and overlook by Empirical be faults these specifications constrained modelled. methodological are can bound to produce biased and inconsistent estimation results. Perhaps, as Lcamer (1987) is if 'elimination the of static single altogether' equation unavoidable model suggests, indications for is to reliable provide more economic analysis and policy purposes. research The critics of the static single equation modelling argue that the nlost significant is its failure integrate this to general approach the dynamics of tourism of aptly weakness demand. An appropriate approach within the single equation framework would have to

295

form, involving functional determinants dynamic flexible the all of relevant a consider for long-run demand the separate estimation of shortand allowing tourism and dynamic implemented in in 4, The a such model was construction of chapter coefficients. in Thus, 4, "general to to the methodology. specific" chapter a more reliable accordance by derivation ARDL the considered, was of approach cffor-corrcction equation single dynamic dimension incorporates demand behaviour the tourism of absent which models from the staticversion. 1 Prior to the estimationof the parsimoniousARDL models, we testedthe order of integration of all time seriesinvolved and obtainedindicationsthat all were nonstationary. Nonstationarytime seriescan give rise to spuriousregressions, unlessit can be shown that ARDL From 4, derived the the long. models of chapter cointegrated. we are the variables for indicated The tested the tests stationarity of and residuals. run equilibrium regressions hence, for demand the existenceof cointegrated residuals and, equation's each stationarity for in demand UK destination its between determinants. tourism the cach and relationships The statisticalvalidity of the estimationresults obtained from both the long-run and the by battery diagnostic further tests confirmed a of was which specifications short-run ARDL the to modelsas robust,structurally stableand classify evidence provided sufficient ARDL Therefore, 4 the be to models of chapter proved specifications. reliable %vell-defined instruments for the investigation of the UK tourism demand and reasonablyaccurate findings These its behaviour. future forecasters show that, evenwithin a single equation of into dynamics functional forms incorporation huge the the can of make a structure, difference in the reliability of the estimation results, demonstratingthat appropriate fundamental to the accuracyof econometric testing are models. modelling and have been dynamic ARDL be The to single equation models claimed more reliable long-run equilibrium relationshipsthan the two-stagcsmethod of means of estimating Yet, demand former (1987). Granger tourism the ICIngle studies applying are practically and literature. interesting be in future in An to the aspect analysed research would non-existent forecasting the statistical and robustness performance of the ARDL consist of comparing in 4, derived derived those chapter with of single equation as speciflcations with the raodels Engle-Granger method. two-stages inhibits Nevertheless, the equation-by-equation these the fortnal nature of models interdependencies destinations, for testing among of and precludes restrictions rnodelling involving cross-equation coefficients.Hence,the models' structuremustchangeto cnibracc

296

features be included. In 5 demand these can chapter we which within equations a systemof based for demand UK tourism the within a system equations of approach on tile examined AIDS model of Deatonand Muellbauer(1980). The orthodox versionof tile AIDS system UK budget function destination's tourism the as a solely of tourism shareof specifieseach however, by This version, clearly was rejected the prices and real per capita expenditure. data, againstthe"unorthodox' AIDS model adoptedin chapter 5. This model includes a for the expenditure the of a non-constant coefficient consideration trend variable and for in UK The tourists' tastes trend accounts changes variable not variable. explanatory in The breakinto by the coefficient the variables. structural otherexplanatory taken account for in the the changes allocation of tourists' expenditure accounts variable expenditure of factors due to that time, modificd the political and economic over occurred which destinations destinations between the the themselves. and among and origin relationships Unlike earlier studiesusingthe orthodoxstatic AIDS approach which rejectedutility theory hypotheses and,more of1cnthan not, revealedscrial correlationand other mis-spccification bias, the "unorthodox"AIDS modelof chapter5 was shownto be statisticallywell-defined, Tile AIDS theory the data-coherent utility postulates. systemofclllptcr with and consistent 5 has specificationfeatureswhich allow for the incorporationof dynamic-likc cicilicnts in its equationsand, in contrast with the orthodox system used in previous research,this Iscemingly-dynamic'model was consistent with the data and with the utility maximisation Hence, theory. this model was consideredto provide reliable consumer assumptionsof information aboutthe long-runbehaviourof the UK tourism demandfor France,Spainand Portugal. However,the AIDS approachadoptedin chapter5 doesnot explicitly considerthe
dynamic features, of tourism demand behaviour and cannot account for the short-nin demand Illus, the which underlies process. adjustment a clear correction mechanism be long-run in between Although, and cannot assessed. shorteffects many separation instances, the focus of interest of economic research is to uncover the long-run structural independent dependent between and variables, the short-run dynamics relationships importance, in the equilibrium model steady-state are also of particularly cases underlying have the relevant magnitudes. Consequently, an explicit dynamic effects short-run where be found information to the appropriate was means of obtaining reliable about specification both the long- and short-run responses of the UK tourism demand to changes in its

297

determinants,and the incorporationof short-run dynamics within the AIDS systemwas investigation. be logical the to the of next step considered In chapter6 we estimated a flexible generaldynamic forril of the AIDS system.Tile be data to this coherent and theoretically consistent, model cstimation results showed AIDS for the the robustness of of empirical evidence methodology undertaking providing tourism demandanalysisin a temporalcontext. In addition, the dynamic model provided both 'sccmingly-dynamic' AIDS of the the capacity of on robust evidence statistically information, long-run inadequacy 5 the to and on of the orthodox supply reliable chapter AIDS and other restrictedmodelsncstcd in the more general specification,to conciliatc forinulations, data Indeed,the long-run their theory quantitative within and consistently dynamic from AIDS the general obtained model were similar to thoseobtained estimates in 5, from the steady-state to supportthe latter as an chapter wliicil seems estimated version long-run behaviour for describing UK demand. the On the tile tourism of model adequate AIDS hand, in the the more generalform and spccific orthodox model models ncstcd other dynamic AIDS the were unmistakably rejected against the more general form. In of AIDS dynamic the system conformed with all quality criteria and general addition, hypotheses tests,providing estimationresults which were statistically robust, crnpirically These theoretically resultsare encouraging consistent. and and indicatedirections plausible For instance, dynamic specificationdoes for future research. they show that an appropriate have of equations, systems when'modclling and can considerableinipact oil tile matter hypothcscs. tests the theoretical validity of of concerning results The utility theory constraintsin dynamic specificationsarc generallytestedfor the long-run coefficients.The motivation for testingtheory restrictions,preferablyin the steady is based in the that, the short-run, consumersmay not liave fully assumption on state, hence, homogeneity to be circumstances changing and, and symmetry may adjusted not in behaviour. The from dynamic AIDS the short-run results obtained general observed behaviour lag. homogeneity Hence, tourists that their adjust suggest with a model and hold in to the short-run. However, when tested, these are not expected symmetry hypothesescould not be rejected, meaning that the utility inaximisation postulatesarc in in both longthe short-run.This result is not completelyunexpected. the and observed Given that the estimates of the generaldynamic model suggested that UK tourists adjust in demand fast determinants, their to changes very non-significant and/or irrclcvant for magnitudes the short-runcoefficients should be expected.Indeed,that was the gcncral
IMR

by insignificant With the indication of the estimates model. short-runestimates, provided imposed leads by the are on coefficientsgenerally the statisticalprocess which constraints to an 'easier' non-rejection(a possibleunder-rejection)of the null. Therefore,the faster less behaviour, demand the their significant short-runcoefficientsarc, and adjust consumers imposedon them.This statement the likelier is the non-rejectionof utility theory postulates be in delivered further the support, which can only context of empirical course, of requires, future research. The similarity of the long-run estimatesobtainedfrom the generaldynamic AIDS
in from 5 the those steady-state version estimated chapter and tile obtained model with both to seemed models, support the existence of a structural of robustness statistical UK demand its between dacmunants. Yet, tourism the there and equilibrium relationship issues had be AIDS that to to the theoretical addressed endorse and empirical still were important issue One regards the spurious, regressioil model as a quality spccification. includes its The AIDS time that nonstationary series, so system estimation results problem. issues Other be the cointcgratcd. are variables arc related with tile a spurious unless can by AIDS Since division the variables assumed approach. of cndogcnous/exogcnous priori its levels the AIDS that current of explanatory variables are the system assumes is bias determined, the this case, unless mis-spccificatioll can occur," exogenously invalidating its estimation results and statistical inference. Therefore, when data series are nonstationary, the estimation of econometric models

which regressendogcnousvariables on several assumedcxogcnousvariables, without be I lcncc, their validity with statistical cointegration may questionable. analysis sanctioning data fccdback and potential cffccts, an efficient ccononictricapproach given non-stationary for estimatinglong-run relationship(s),is a systemof equationspermitting all variablesto I'lic be treatedas endogenous and appropriate analysis. vector nutoregressivc cointcgration is solutionto overcometheseproblems. approach an appropriate The vector autorcgrcssive approachestablishes a model specification which'is a both dynamic the to single equationand the traditional structural multivalid alternative Johansen The (1988) approaches. procedurepermits cointcgration analysisand cquation identifying tile number of cointegrating vectors and method of establishesan cfficient Tllis is the parameters. structural methodology applied to systemof equations estimating and avoids the problems of spurious regression,simultaneousequation bias, a priori division of endogenous-exogenous variablesand unfoundedimpositionof zero restrictions.

299

The AIDS model specifiedin chapter5 is a systemof equationswhich incorporates for data Tile their cxogeneity assumes right-hand side variables. series, and nonstationary implications of thesefeaturescould underminethe statistical validity of this model if no its found linking fence, I tile objectiveof chapter variables. were relationships cointegrating 7 was to find empirical evidenceof cointcgration in order to support, or otherwise, the For this inferenceprocedures and estimationresultsobtainedwith the AIDS spcciFication. including VAR the same unrestricted system a reduced-form specified purpose, we lag-Icngth, AIDS to tile the pertinent used procedures and establish model, variables as division or tile. Once tile deterministic components and the endogenous/exogcnous in Johansen VAR form tile the place, cointcgratcdrank test was usedto was of appropriate determinethe numberof cointegrating vectors.The test provided statisticalstipport for the AIDS the model, which predicts tile existenceof exactly cconomic principles underlying in (n. 1) long-run (cointegrated) relationships a systemof n expenditureshareequations.In in VAR the the the of cointcgratcd parameters vectors structural model were addition, those of tile normalisation processused to matching with restrictions cxactly-identified identify the tourism shareequationsof an AIDS system. Finally, the resulting structural VAR, identicalto that of the AIDS spccirication,was testedunder form of the cointegrated homogeneity, such as symmetryand null cross-pricc restrictions additional over-identifying In hypotheses these not were rejected. and addition, the elasticities estimates cffects VAR full the the tile of cocfricients cointegrated with structural under set of obtained theoreticalrestrictionsproved to be fairly similar to tile corresponding elasticitiesobtained from the AIDS modelsestimatedpreviously. Consequently, strong evidencewas obtained to support the AIDS model's ability to portray accurately the theoretical predictions UK the the tourism demand behaviour and the destinations' of rationale underlying competitiveconduct. VAR under the ful I set The theoreticalandempirical consistency of the cointegratcd implied be This the that model's would predictive ability good. was, indeed, of restrictions indicatedit the accuracyof this model's forecasts the caseas the quality criteria measuring forecasting device, the most precise amongall VAR spccificationsusedin chapter7. The as (unrestricted) VAR best the that tile general reduccd-form suggested was next criteria forecastingmodel.This finding gaveempirical supportto the claimedcompetence or VAR for forecasting purposes,and endorsedthe qualities of modelling simplicity and models "pure" VAR. the Given theseresults,we concludethat if the unrestricted of ease estimation

300

is forecasts, fairly to exercise only provide modelling accurate main goal of a multivariate rather than to estimatethe structuralrelationshipsamong the variables,the rcduccd-ronn VAR modelmay constitutethe bestchoice, Given the results obtainedin chapter 7, we believe we have contributed enough hand, for, AIDS the tile considering on one approachas a theoretically evidence empirical consistentand statistically robust meansof producing valid and reliable estimatesof the long-run equilibrium parameters underlyingthe relationshipsbetweendestinations'tourism for hand, its determinants the other conrin-ningthe competence or tile and, on sharesand VAR model, either in its more generalunrestrictedform or under the full set of theoretical destinations' forecasts the tourism to of shares. accurate provide restrictions, The forecastingability of econometricmodels is a fundamentalaspect or their
8 4 Therefore, 7, by tile to evaluates chapter models of chapters evaluation. quality forecasting 1994-1997. The the their over ability out-of-sampic relative period comparing

fully VAR the that the of spccirications, most namely showed restricted results 'unorthodox' ARDL AIDS tile the and restricted modcl systcm, produce specification, below I be forecasts percentage errors points, which can viewed as with average absolute 5
fully VAR forecasts destinations, For the restricted produces with average all accurate. 0.6 I between ARDL between the and percentage points; errors ranging models, absolute 0.7 and 1.3 percentage points, and the restricted 'unorthodox' AIDS model, between 1.1 Only VAR 1.5 AIDS the the reduced-form unrestricted points. and percentage models and

for destinations. Tile indicate this that tile range, one or more results produceerrorsoutside fully restrictedmodels generate more preciseforecaststhan their unrestrictedor partially dynamic bettcr than static ones, and the counterparts, models pcrform restricted full VAR is the the out-performsall the othcr niodcls. under set of restrictions cointegrated The conclusion drawn from these results is, that appropriately spccificd econometric forecasters. Moreover, are excellent claims that the pcrformancc of univariate models if is hold, to to thesemodcls nre tcstcd econometric models superior arc unlikely models econometric spccifications. againstwell-defined,robustandconsistent A good econometric model providesa formal quantitativeframeworkwhich can be interpret to understand activities, economic relationships, test validly used cconomic behaviour. The theories, economic evaluate policy and predict economic main cconomic drawn from be the extensivemodelling exercisecarried out in this that can conclusion is be derived that from appropriatcly spccificd consistent and reliable results can study

301

features, integrate knowledge based theory and of specific postulates on modelswhich are facts and eventsthat affectedthe economicrelationshipsbetweenorigin and destinations, from The different destinations the themselves. results obtained specifications and among in hoc 7 4 to the tile to contrast singic equation across models, (W are consistent of chapters dcrivcs by former from 3. The the the of results provided consistency approachof chapter their foundationof pertinent economicprinciples and sound econometricmethodologies. Thus, these models constitute the necessary and reliable basis for quantirying, behaviour demand. long-run UK tourism the the of andpredicting understanding Each econometricspecificationprovides information about particular aspectsof For instance, the ARDL single demandbehaviourbut, inevitably, leavesothersunattended. incorporation dynamics, the of cointegration analysis and give equation models pen-nit direct informationaboutown-priccandincomeelasticities.With this specirication, we were impacts UK induced by demand long-run tourism the to on changes able assess short- and in such variables as destinations' prices, UK tourism expenditure in ncighbouring ion structure does not destinationsand UK per capita income. Yet, its cquation-by-cquat imposed incorporation for theoretical constraints the on cross-cquation parameters. of allow The AIDS systemof equationsallows for the incorporationof theseconstraintsas dynamic The infornlation it for the structure. a of empirical provides modelling well as instead income it the elasticities and, of elasticity, and cross-price supplics concerns ownform The in the tourism elasticities. specific of models expenditure adopted of estimates break in includes 6 5 the expenditurevariable cocwlcicnt, which a structural chapters and UK for the time, the the tourism of of changing sensitivity, over examination allowed demandelasticities.Nevertheless, the assumptions underlying the AIDS model include a division hold, and variablesthat of exogenous/endogenous variables, which may not priori in its The tile that tile model are variables validation of requires results are nonstationary. in AIDS forming the system; a number of vectors as predicted cointegrating cointegrated, that is, a system of n equationsmust have n-I cointegmung vectors. In addition, tile identification of their structuremustmatchthat underlyingthe AIDS equations. Theseissues usingthe VAR approach, can be examined which treatsall variablesas for identification hence, for the the of and allows cointegrating vectors and, cridogenous long-run structural coefficients which are the main intcrcst of the analysis. The VAR the AIDS the validity of sanction can model and provides the same type of approach information. However, the modelling, testing and inferenceproceduresinvolved in the

302

it VAR approachare more complex than those used in the AIDS models. Nevertheless, be AIDS be the that system can only sanctionedthrough cointcgration noted should VAR to test the undertaken estimate and a model. analysiswhich requires procedures The main conclusionis, therefore,that a comprehensive analysisof tourism demand di(Tcrcnt include specified models which can provide a wide rangeof appropriately must insights about the main featuresof interestand supply information about the consistency from different If the tile models comply the approaches. obtained of results and validity be their are consistent,we can conrldclit of their results with all quality critcria and for This for tile the and analysis policy purposes. case was economic accuracyand validity Hence, draw in findings the to this can we use results conclusionsabout study. cmpirical behaviourfor France,Spainand Portugal. the UK tourism demand The resultsshow that the long-run expenditureelasticities,althoughclose to unity for all destinationcountries,tend to increasefor Franceand Portugaland to dccrcasefor Spain, in the morerecentyearsof the sample.In contrast,Spaindisplaysthe smallestshort. its indicate This those of ricighbours. with that compared may run expenditureelasticity Spain has a comparative advantage,relative to its ncighbours, in terms of short-run in the UK tourismbudgetbut seems to be losing groundto Franceand Portugalin increases the long-runpreferences of UK tourists. In the more recent yearsof the sample,the estimatesor the sensitivity or tic UK demandto own-pricechanges providedby the static AIDS modeland by the fully restricted indicating long-run for France Portugal, VAR, and a value close to are similar cointegrated indicates for former in Spain, However, the the equation model an own-price elasticity -2. latter indicates Given VAR the to that to the while value a close clon: model -2. -1, forecaster, being best the quality criteria as with all well as we assumeits complies for demand UK The information the more reliable. as given reactionto estimationresults be indicates for Spain less sensitivethan that demand to the short-runown-price changes for its ncighbours.Hence, Spain seemsto have a comparativeadvantagerelative to its increases, in in in both longthe the the short-run. context of price and neighbours in Portugaland Franceshows a However,as the UK demandsensitivity to pricc-changcs in Spainseems to diminish over time, while that to price-changes to be increasing. tendency the comparative of Spainmay be short-lived. advantage The information given by the AIDS and VAR modelsconcerningthe competitive is destinations the remarkablyconsistent.The long-run estimatessupport our conductof

303

likelier is rind destinations the to the evidenceof are, closer geographically conjecturethat in indicate between Indeed, behaviour them. the all cases, results significant competitive destinations in UK Spain Portugal tourists' to France relative are competitive that and between I lowcvcr, found France Portugal. link but tile and was significant no preferences, demand UK in tourism tile no significant reaction of cases, suggest, all short-run estimates in its ricighbours.This may indicatethat in the shortfor one destinationto price-changes its likely level UK destination in decreasing to are not affect significantly of prices one run, to be a meansof producingnet gains from UK tourists.Tile demandand, therefore,seems France is for demand UK Portugal long-run, in the the or more that, results also show is demand for in Spain in Spain, than the to to price changcs sensitive price changes Portugal or France. This indicates that the UK demand for Spain may be considered 'cut-throat' from its long-term insensitive to pricc-dccrcasing policies relatively indicate, for Yet the the the more recent years of sample, a results again, ncighbours. diminishing sensitivity of the UK demandfor France and Portugal to price changesin Spain, and an increasing sensitivity of the demand for Spain to price changesin its for UK Spain tourists' Therefore, the to secular prcfcrcncc sccills again, once ricighbours. towardsfavouring its neighbouring be changing,slowly but progressively, competitors. Although the context of the analysiscarried out in this study concernstile tourism demand of one origin for three destination countries, any of the models spccificd ill larger destinations be 4 7 to to number of and origins readily extended a can chapters Furthermore, VAR AIDS loss the systemsof equationscall and of generality. without any tourism productat the be appliedin contextsother than the demandfor an undifferentiated for demand level. Indeed, these to can we use models explain andpredict particular national individual items, specific country, resorts within one region and even within one regions local tourism type area. within one or accommodation attractions, as of such As long as tourists' spending on diversified products can be divided into its determinants dependent between the the variable and expenditureshares, relationships framework. The VAR be AIDS the of share within a system equations as modelled can do but they the general, as not necessitate more use share equations are even of models form demand. The VAR tourism measuring of other with any only probicni accept is data-intensive AIDS in that they tourism models arc and, and models specifications data arerelatively scarce. contexts,

304

The ARDL models, as single equation specifications arc less data-intensivc than the

freedom to the sameextent as tile other modcls,when former and do not erodedegrees of included. Hence, be for incorporating they can useful are other regressors additional interest, such as transport costs or advertising expcnditurc, when relevant, variables of in interesting issucs All losing these their extensions estimates. constitute precision without for future research.

305

REFERENCES
Achen, C. H. (1982) Interpreting and using regression, Sage Publications, 13evcrly I fills, California.

Anderson, G. and Blundell, R. (1983)Testingrestrictionsin a flexible dynamicsystem:an in Canada, ReviewofE,conomicStudies, 50,397-410. application to consumers expenditure Anderson, G. and Blundell, R. (1984) Consumernon-durabicsin the UK: a dynamic demandsystem,TheEconomicJournal, 94,3 3-44.
Archer, B. H. (1976) Demand forecasting in tourism, Occavional Papers in Econonlics, 9, Bangor, University of Wales Press.

Artus, J. R. (1972) An economctricanalysisof intcmationaltravcl, InlernationalAfonetar' Fund StaffPapers,19,579-614.

Attfield, C. L. F. (1997) Estimating a cointegrating demand system. European E*collonlic Review, 41,61-73.

Bancrjee,A., Dolado,J. J., Galbraith,J. W. and Hendry, D. F. (1993) Cointegralion,errorcorrection and the econometricanalysisof non-stationarydata, Oxford Unimsity l1rcss, Oxford. Banks, J., Blundell, R. and LcwbcI A. (1997) Quadratic Engel curves and consurner demand.ReviewofEconomicsand Statistics,79,527-39. Barnett, W. A., Lee, Y. W. and Wolfe, M. D. (1985) I'lic thrcc-dinicnsional global properties of the Minflex Laurent, gencraliscdLconticf and translog flcxible functional forms. Journal ofEconomelfics,30,3-3 1. Barnett, W. A., Lee, Y. W. and Wolfe, M. D. (1987) I'lie global propcrtics of the two Minflex Laurentflexible functionalforms.Journal ofEconomarics, 36,281-298. Bamett, W. A. and Yue, P. (1998) Scmi-paramctric estimationof the asymptoticallyideal model: The AIM demandsystem,in Nonparainetric and Robust hiference. &II-ances In Econometrics7, G. Rhodes ), JAI Press, Greenwich,229-52. andT. Formby (cds. Barten,A. P. ( 1968)Estimatingdemand Economelrica, 36,213-5 1. equations, Barten, A. P. (1969) Maximum likelihood estimation of a complete systcm of dcriland EuropeanEconomicReview,1,7-73. equations,
Barten, A. P. (1977) The systems of consumer demand functions approach: in a review, Michael D. Intriligator (ed.) Frontiers of Quantitative Economics, North-11olland Publishing Company, Amsterdam.

306

in British An (1971) K. tourist O'Hagan, K. expenditure Barry, econometricstudy of and 3,143-161 Ireland,Economicand SocialReview, Barton, M. and Young, T. (1996), The impact of BSE on the demandfor bccf and othcr 28,687-93 Economics, Applied Britain, in Great meats Blundell, R. (1988) Consumerbehaviour: theory and cmpirical evidence -a EconomicJournal, 98,16-65. survey,

Box, G. E. P. and Jenkins,G. M. (1976), Time series analysis,forecasting and control, Holden-Day,SanFrancisco. Brown, R. L., Durbin, J. and Evans,J. M. (1975) Tccliniqucs for testing tile constancyof Ro), StailsticalSociely R, 37,149-92. Journal time, ofthe al relationsover regression impact for BSE beef The (1996) T. Young, M. the Burton, on of and otlicr ticniand and Economics, 28,687-93. Applied in Great Britain, meats Traveland Tourism,Government Statistical Monitor MA6 (1970-1993)Overseas Business Service,CentralStatisticalOffice, IIMSO, London. L. D. (1980) Global propertiesof flexible functional fornis. Caves, D. and Christensen, AmericanEconomicReview,70,422-32. Chalfant, J. A. (1987) A globally flexible Almost Ideal System, Journal of Business EconomicStatistics,5,233-42. Chalfant,J. A. and GallantA. R. (1985)Estimatingsubstitutionelasticitieswitli (lie Fouricr Journal 28,205-22. Carlo Monte function: results, ofEconometrics, some cost
Chambers, M. J. (1993), Consumers' demand in the long-run: some cvidencc froill UK data, Applied Economics, 25,727-733 Chambers, M. J. and Nowman, B. K. (1994) Forecasting with the almost ideal dcmand University i. Discussion 426, Papers Series, Department or -'sscx. ofEconomics system. Charcmza, W. W. and Dcadman D. F. (1997), New directions /it econometric practice, 2"d Cheltenham. Publishing Limited, Elgar Edward edition, Christ, C.- F. (1966) Econometric models and methods, John Wiley & Sons, Inc., Ncw 'York. Christ, C. F. (1975) Judging the performance of econometric models or tiic us economy. international Economic Review, 16,54-74. Christensen, L. D., Jorgenson, D. W. and Lau, L. J. (1975) Transcendental logarithmic Economic Review, American 65,367-383. functions. utility

307

Clements,M. P. and Hendry D. F. (1998) Forecasting economictime series, Cambridge Cambridge. University Press, time series: Clements,M. P. and Mizon G. E. (1991)Empirical analysisof macroeconomic VAR andstructuralmodels,EuropeanEconomicReview,35,887-932. Cooper,R. L. (1972),The predictiveperformance of quarterly econometricmodelsor tile ), EconometricModels of Cyclical Behaviour, 36, United States,in Hickman, B. J. (ed. Studiesin Incomeand Wcalth, 813-947,Columbia National Bureauof EconomicResearch New York. University Press, demand (1996) A R. K. McLaren, R. J. Cooper, equationssatisfying system of and Statistics, 78,359-364. Review ofEconoinics and cffcctively global regularityconditions. Darnell, A. C. and Johnson, P. S. (2001) Repcat visits to attractions: a preliminary 22,119-26. Management, Tourism economicanalysis, Davidson,J. H., 11endry,D. H., Srba,F. and Yco, S. (1978) Econometricmodelling or tile income 'in between the time-series consumers' and cxpcnditurc relationship aggregate Journal, 88,661-92. United Kingdom, TheEconomic De Mello, M., Sinclair M. T. and Pack A. (2001) A system of cquationsmodel of UK forthcoming. Economics, 41)plied in demand countries,, neighbouring tourism ) in Griliches M. D. Intriligator Z. (cds. Ilandbook Demand (1986) A. and Deaton analysis, Elsevier, 30, Amsterdam. Chapter III, ofEconometrics,vol. Deaton,A. and Muellbauer,J. (1980a)Economicsand ConsumerBehaviour, Cambridge Cambridge. University Press, Deaton,A. and Muellbauer,J. (1980b) An Almost Ideal DcmandSystem,The 'Itnericall EconomicReview,70,312-26.
Dickey, D. A. and W. A. Fuller (1979) Distributions of the estimators for autorcgrcssive 74, Journal American Statistical Association, 427-31 the pp root, of a unit time serieswith Dickey, D. A. and W. A. Fuller (1981) Likelihood ratio statistics for autorcgrcssive time 49, 1057-72 Econometrica, pp root, a unit series with Dickey, D. A., Jansen,D. W. and Thornton, D. 1. (1991) A primer on cointcgration with all income, Economic Federal Bank St. Louis, Review, Reserve to of and money application March-April, 57-71. Engel, R. F. (1984) Wald, likelihood ratio and Lagrange multiplier tests in econometrics, in Z. Griliches and M. D. Intriligator (cds.), Handbook of Econometrics, Vol. 2, NorthHolland, Amsterdam. Engel, R. F. and Granger, W. J. (1987) Cointegration and error correction: rcprcscntation, Econometrica, 55,251-76. testing. estimation and

308

) (1991), Long-run economicrelations: reatfingsin Engle, R. F. and C. W. J. Granger(eds. Oxford Press, Oxford University cointegration, Journal Engel, R. F. and Yoo, B. S. (1987)Forecasting and testing in cointegratcd systcins. 35,143-59. ofEconometrics, 51,277Engel, R. F., Ilendry, D. F. and Richard,J. -F. (1983) r. -xogeneity,Econometrica, 304. Eurostat (1987) Tourism: Annual Statistics,Theme 7 (Servicesand transports),series C (Accounts,surveysandstatistics),Eurostat, Luxembourg. Fair, R. C. (1986), Evaluating the predictive accuracyof models, , in Griliclics, Z. find ), Handbook of Economctrics,vol. 3,1981-95, North-Holland, Interligator, M. D. (eds. Amsterdam. Fisher,D., Fleissig,A. and Serletis,A. (2001) An empirical comparison of flexible demand Econometrics, 16,59-80. forms. Journal functional ofApplied system tourism demand;somc methodologicalissues, Fujii, E. T. and Mak, J. (1981) Forecasting 15,72-82. 4nnals ofRegional Science, .
D. and Ibrahim, I. B. (1975) Comparing the Box-Jcnkins approach with tile Hawaii Journal forecasting to tourists, application model: of smoothed exponentially Marketing Research, 12,182-8 8. Geurts, A

Gonzalez,P. and Moral P. (1996),Analysis of tourism trendsin Spain,"Innals of Tourism 23,739-54. Research, Granger,C. W. J. (1969) Investigatingcausalrelationsby ccononictricmodelsand cross. 6. Econometrica, 37,24-3 spectralmethods,
Granger, C. W. J. (1981) The comparison of time series and econometric forecasting in large-scale Conlriblitions theory macro-econometric 1() models: practice, and strategies Economic Analysis, 141,123-28. North-flolland, Amsterdam.

in the study of cointcgratcdccononlic variablcs, Granger,C. W. J. (1986) Developments Oxford Bulletin ofEconomicsand Statistics,48,213-28. in the conccptof causality.Journal of Granger,C. W. J. (1988) Somerecentdevelopments 39,199-211. Econometrics,
Granger, C. W. J. (1995), Modelling non-lincar relationships between extcndcd-nicniory 63,265-79. Econometrica, variables. Granger, C. W. J. (1997) On modelling the long-run in applied economics. The Economic Journal, 107,169-77.

309

In Modelling (1990) ) J. (ed. W. C. econometric series: readings economic Granger, Oxford. Press, University Oxford methodology, in Journal Spurious (1974), Newbold W. P. C. regrcssion cconomctrics, Granger, of and 35,143-59 Econometrics, Granger, C. W. and P. Newbold (1986), Forecasting economictime series, 2"d cdition, New York. Academic Press, Canada, by Unitcd Statcs international demand the travcl (1966) The and H. P. Gray of International EconomicReview,7,83-92. Griffiths, W. E., Hill R. C. and JudgeG. G. (1993) Learning andpracticing econometrics, John Wiley & Sons,Inc., New York. Guilkey, D. and Lovell C. (1980) On the flexibility of the translog approximation. international EconomicReview,21,137-47. for Dickcy-17-ullcr Extended (1989) P. tests, tabulations Schmidt, K. D. Guilkey, and EconomicsLetters,31,355-57. Gujarati, D. N. (1995), Basic Econometrics,McGraw-Hill International Editions, New York. in Tourism Singaporct Demand (1986) tourism C. Bocy, 11. of Gunadhi, and elasticities 7,23 9-253. Alanagement, Harris, R. I. D. (1995) Using cointegration analysis in econometricmodelling, Prentice Hall, HemelHempstead. Harvey, A. C. (1990) The econometricanalysis of time series, 2"d edition, Philip Allan, Hemel Hempstead. Harvey, A. C. (1993) Time series models, 2nd edition, Harvester Whcatshcaf, Ilemel Hempstead. Economica, 47,387-406. Hendry, D. F. (1980)Econometrics science?. or alchemy Hendry, D. F. (1983) Econometricmodelling: the consumption function in retrospect, 30,193-220. ScottishJournal ofPolitical Economy, Hendry, D. F. (1987) Econometricmethodology:a personalperspective,in .4drances In ), CambridgeUniversity Press, TrumanF. Bewley (ed. Fifth World Congress, Econometrics Cambridge. Oxford University Press,Oxford. Hendry,D. F. (1995)DynamicEconometrics,

310

Hendry, D. F. and Mizon, G. (1978) Serialcorrelationas a convenientsimplification, not a for England, demand by Bank the the of money nuisance: a comment on a study of EconomicJournal, 88,549-63. Hendry, D. F. and Morgan, M. S. (1995) The foundations of econometric analpis, Cambridge. CambridgeUniversity Press, in dynamic formulation On (1982) F. J. Richard, the F. D. Hendry, of empirical models and 20,3-33 Journal ofEconometrics, econometrics, Hendry, D.F. and Richard,J.F. (1983) The econometricanalysisof economictime scrics, International StatisticalReview,51,3-33. ) (1984) Econometricsand quantitative economics, Hendry, D. F. and Wallis, K. F. (eds. Basil Blackwell, Oxford. Statistics Yearbook, International Monetary Fund (1980-1998) International 171nancial DC. internationalMonetaryFund,Washington, Estatisticasdo Turismo,Instituto Nacional dc Instituto Nacional de Estatistica(1975-1997) Estatistica,Lisboa. Johansen,S. (1988) A statisticalanalysisof cointcgration vectors.Journal of Economic Dynamicsand Control, 12,231-54. Johansen,S (1996) Likelihood-basedinference in cointcgration vector outorcgrcssive Oxford. Press, nd University Oxford 2 models, edition, Johansen,S. and Juselius,K. (1990) Maximum likelihood estimation and inference on demand for Economics Oxford Bulletin the to and money. of cointegrationwith application Statistics,52,169-210. Johnson,P. and Ashworth, J. (1990), modelling tourism demand :a summary review, 9,145-160.. LeisureStudies, Jorgenson, D. W., Lau, L. and Stoker, T. (1980) Welfare comparison and exact Economic 70,268-72. Review, 4merican aggregation. . 11.(1974) Internationaldemandfor Latin Americantourism, Growth Jud, G. D. andJoseph, 25-30 Change, January, and judge, G. G., Hill, R. C., Griffiths, W. E., Lutkepohl,11.and Lee, T. C. (1988) Intro(hiction 2ndedition, John Willey & Sons, Inc., New to the theory and practice of econometrics, York. Kim, S. and Song, H. (1998) An empirical analysis of demand for Korean tourism: a Tourism 3,25-4 A 1. approach. correction error and nalysis, cointegration

311

Transportation Canadian tourism, A (1981) L. overseas Kliman, M. quantitativeanalysisof Research,15,487-97 nd New York. Macmillan, 2 Elements Kmenta,J. (1986) edition, ofEconometrics, ) (1981) Large-scale models:thcory macro-cconometric Kmenta, J. and J. B. Ramsey(eds. North-11olland, Amsterdam. Analysis, Economic in Contributions to and practice, Kulendran,N. (1996) Modelling quarterly tourism flows to Australia using cointcgration 2,203-22 Economics, Tourism analysis. flows international Forecasting tourism (1997) M. L. King, N. quarterly Kulendran, and Forecasting, Journal 13, International of using error correction and time seriesmodels. 319-27 Annals Kulendran,N. and Witt, S. F. (2001) Cointegrationversusleastsquarcsregression. 28. Research, Tourism of York. Wiley & Sons, Inc., New John Specification E. (1978) E. Leamer, searches, American EconomicReviciv, Learner,E. E. (1983a)Let's take the con out of econometrics. 23,31-43. Learner,E. E. (1983b) Model choice and spccification analysis, in Z. Griliclics and M. 1, I lolland Company, North Publishing Econometrics, (eds. ), Handbook vol. Intriligator, of Amsterdam. for causal inference?,in K. Brunner and A. Learner,E. E. (1985) Vector autoregression ConferenceSeries on Public Policy, vol. 22, North), Carnegie-Rochester Meltzer (eds. Holland, Amesterdam. fi'Jfth Worl(l Learner,E. E. (1987), Econometricmetaphors,in AdvancesIn Econometrics Cambridge. TrumanF. Bewley (ed. ), CambridgeUniversity 11'ress, Congress, Lee, C. K., Var, T. and Blaine, T. W. (1996) Deten-ninants of inbound tottrisni Tourism 23,52742 Annals Research, of expenditures. Journal ofEconometrics,36,311-337. Lewbel, A. (1987)Fractionaldemandsystems. Lewbell, A. (1990) Full rank demandsystems.International EconomicReview,31, May. 289-300. Lewbcll, A. (1991) Utility functions and global regularity of fractional demandsystems. International EconomicReview,36, November,943-61. Lcwbell, A. (1995) The rank of demandsystems:theory and non-paramctricestimation. 59,711-30. Econometrica,

312

Lim, C. (1997) An econometricclassificationand review of internationaltourism demand Economics, 3,69-81 Tourism models, ), Tourismin Asia: the Lin, T. and Y. Sung(1983) flong-Kong, in T. Lin and E. Pyc (cds. Press, Singapore. University Singapore impact, economic Little, J. S. (1980) International travel in the US balance of payments,New England EconomicReview,May/June,42-55 Loeb, P. D. (1982) Internationaltravel to the United States:an econometricevaluation, 9,7-20. Annals of TourismResearch, desde1960,Gradiva,Lisboa. Lopes,J. S. (1996)A EconomiaPortuguesa Lovell, M. C. (1983)Datamining. ReviewofEconomicsand Statistics,65,1-12. Lyssioutou,P. (1999) Dynamic-analysisof British demandfor tourism abroad,Empirical Economics,15,421-36 MacKinnon,J. G. (1991) Critical valuesfor cointcgrationtests,in R. F. Engeland C. W. J. ), Long-run economic relationships: readings In cointegration, Oxford Granger (eds. Oxford. University Press, Mak, J., Moncur, J and Yonamine,D. (1977) Determinantsof visitors cxpendituresand US lengths lawai L I Journal of Travel to analysis of visitors cross-section of stay: a visitors 15,5-8. Research, Martin, C. A. and Witt, S. F. (1987) Tourism demand forecasting models: choice or tourists' cost of living, TourismManagement, 8,23346. variablesto represent appropriate Martin, C. A. and Witt, S. F. (1988) Substituteprices in modelsof tourismdemand, Annals 15,255-68. of TourismResearch, Martin, C. A. and Witt, S. F. (1989a) Forecastingtourism demand:a comparisonof tile InternationalJournal ofForecasting,5,7-19. accuracyof severalquantitativemethods, Martin, C. A. and Witt S. F. (I 989b) Accuracyof econometricforecasts of tourism,Annals Research, 16,407-28. Tourism of in in Hendry, Mizon, G. E. (1984) The encompassing D. F. nnd approach econometrics, ), Econometricand QuantitativeEconomics, Wallis, K. F. (eds. BasiIB lackwd 1,Oxford. for autocorrelationcorrectors:Don't. Journal of Mizon, G. E. (1995) A simple message 69,267-88. Econometrics, Mizon, G. E. and Richard,J. -F. (1986) The encompassing principal and its application to Econometrica, hypothesis 54,657-78. tests, non-nested

313

Muellbauer, J. (1976) Community prefercnccs and the rcprcscntativc consurna, 44,979-99. Econometrica, P. (1988)Testsof dynamicspccificationand homogencityin J. and Pashardes, Muellbauer, for Fiscal Series, Institute 88/6, Studies. Papers IFS Working demand system. a
Newbold, P. and Bos T. (1994) Introductory business and economic forecaving, Cincinnati, Ohio. Co., Publishing South-Westem edition, 2"d

Newbold, P. and Davis, N. (1978) Error miss-spcciricationand spurious rcgrcssions, Intemational EconomicReview,19,513-19. in 133-urope: US tourist O'Hagan,J. andHarrison,M. (1984)Market shares of expenditure till 16,919-3 Economics 1. Applied analysis, econometric S. and Witt, S. F. (1985),A marketinganalysisof foreign tourism in Grcccc, Papadopoulos, S. Shaw,L. Sparksand H. Knynak in Proceedingof the SecondWorldAfarketingCongress, ), University of Stirling, 682-93. (cds. A. (1999) The demand for international tourism in the Mediterranean Papatheodorou, 0. Applied 31,619-63 Economics, region, P. (1993) Bias in estimationof the Almost Ideal DemandSystemwith tile Stolle Pashardes, Journal, 103,908-916. Economic index approximation, Perron, P. (1989) The great crash, the oil price shock and the unit root hypothcsis, Econometrica, 57,1361-401. Perron,P. (1990) Testing for a unit root in a time serieswith a changingmcan,Journal of Business and EconomicStatistics,57,1361-40 1.
Pesaran, M. 11. (1998) The role of economic theory in modelling tile long-run. Econornic Journal, 107,178-19 1. Pesaran, M. 11.and Pesaran,B. (1997) Working wilh microfit 4.0: interactive econometric Oxford University Press,New York. analysis,

distributedlag niodclling approach Pcsaran,M. 11.and Shin, Y. (1995)An autorcgrcssivc Seriesno. 9514, Departnicnt to cointegration analysis.DAE WorkingPapers.4malgainaled University of Cambridge. of Applied Economics, Pesaran,M. H. and Shin, Y. (1997) Long-run structuralmodelling, revisedversion of the DAE Working PapersAmalgamatedSeries no.9419, Departmentof Applied Economics, University of Cambridge. Pesaran, M. 11. and Shin, Y. (1996) Cointegration and the speed of adjustment to 71,117-43. Journal ofEconometrics, equilibriurfl.

314

M. H., Shin, Y. and Smith, R. J. (1996) Testing for the existenceof long-run Pesaran, Series, 9622, Department Applied Amalgamated Working Papers DAE of no. relationships. University of Cambridge. Economics, Pesaran,M. H., Smith, R. P. and Yeo, S. (1985) Testing for structural stability and School, 53,280-95. Manchester failure: a review, predictive in econometrics, Journal of Phillips, P. C. B. (1986) Understanding spuriousregressions 33,311-40. Econometrics, Phillips, P. C. B. (1987a)Time seriesregressions with unit roots, Econometrica,55,277302.
Phillips, P. C. B. (1987b) Towards a unificd asymptotic theory of autorcgrcssion, Biometrika, 74,535-47

Phillips, P. C. B. and Perron,P. (1988) Testing for a unit root in tinic scrics rc8rcssion, Biometrika,75,335-46. P (1991)Econometric forecasting R. S. Rubinfeld Pindyck, modelsand economic cd., and Ramajo,J. (1994) Curvaturerestrictionson flexible functional fornis: an applicationor the Minflex Almost Ideal Demand System to the pattern or Spanish demand,Journal of 6. 4,431-3 Statistics, Economic Business and
Richard, J. -F. (1980) Models with several regimes and changes in cxogcncity, Reviciv of Economic studies, 47,1-20. Reimers, H. E. (1992) Comparisons of tests for multivariate cointcgration, Statistical Papers, 33,335-59. Romilly, P., Liu, X. and Song, 11.(1998) Economic and social dctcri-ninantsof intcrnational tourism spending: a panel data analysis. Tourism Analysis, 3,3-15.

Sheldon,P. J. (1990) A review of tourism expenditureresearch,in C. 11.Coopcr (cd.) Progressin tourism,recreationand hospitality management, vol. 2, Belhavcii,London 48,1-48 Sims,C. (1980)Macroeconomics andrcality. Econonietrica, Sims, C. (1987) Making economicscredible, in Advancesin EconomelricsI"Jflh If"orld University Press, Cambridge. TrumanF. Bewley (ed.), Cambridge Congress, Sinclair,'M. T. and Stabler,M. J. (1997) The Economicsof Tourism, Routlcdge,London New York. and
Smeral, E. and Webar A. (2000) Forccasting intcmational tourism trcnds to 201O.Annals of Tourism Research,27,982-1006.

315

Annals demand 2005. forecasts Econometric to tourism S. F. (1996) Witt E. of Smeral, and 23,891-907. Research, Tourism of tourism trendsto 2000. Smcral,E., Witt S. F. and Witt C. A. (1992)Econometricforecasts: 19,450-66. Annals of TourismResearch, demand An X. (2000) tourism P. Liu, H., Romilly, outbound Song, empirical study of and 32,611-624. in the UK. Applied Economics, Song, H. and Witt, S. F. (2000) Tourism demand modelling and forecasting: modern Amesterdam. Pergamon, approaches, econometric Stone,J. (1954), Linear expendituresystemsand demandanalysis: an application to tile 64,511-27 Journal, Economic demand, British of pattern. Syriopoulos, T. (1995) A dynamic model of demand for Mediterranean tourism, 9,318-336. International ReviewofApplied Economics, Syriopoulos,T. and Sinclair, M.T. (1993) An econometricstudy of tourism demand:tile tourism in Mediterranean AIDS model of US and European countries,Applied Economics 25,1541-52. Tansel, A. (1991) Economic impact and demandfor tourism: a survey, Mau StudiesIn 1. 18,3,255-8 Development, vol. 33,67-87. to demandanalysis,Econometrica, Theil, 11.(1965),The informationapproach Tourism Policy and International Tourism in OECD (1995-1997) Tourism In OECD OECD, Paris. countries, member Travel Trends: A report on the 1994 (5/6/7/8/9) international passengersurvey, StatisticalService,CentralStatisticalOfflice,I IMSO, London. Government Tremblay, P. (1989) Polling internationaltourism in WesternEurope.Annals of Tourism 16,677-91. Research, in Taiwan: US international tourism Tsai, P-L. and Wang,K-L. (1998) Competitiveness of 1. Applied 30,631-64 Economics, Japanese visitors, versus Uysal, M. andCrompton,J. (1984) Determinants of demandfor internationaltourism flows December1984,288-97 to Turkey, TourismManagement, Uysal, M. and J. L. Crompton (1985), An overview of approaclicsto forccast tourism 23,7-15. Journal of TravelResearch, demand, for demand Thailand's cxPorts the Vogt, M. and Wittayakorn,C. (1998), Determinants of 30,711-15. Economics, Applied tourism, of forecasting: Wallis, K. F. (1989)Macroeconomic a survey.EconomicJournal, 99,28-6 1.

316

White, K. J. (1985) An international travel demand model: US travel to Wcstcrn Europe. Annals of Tourism Research, 12,529-45 Witt, C. A., Witt S. F. and Wilson N. (1994), Forecasting international tourism flows, Annals of Tourism Research,21,612-28. Witt, S. F. (1980a) An econometric comparison of UK and German foreign holiday behaviour, Managerial and Decision Economics, 1,123 -3 1. Witt, S. F. (1980b) An abstract mode-abstract (destination) nodc model of foreign holiday demand,Applied Economics, 12,163-80. Witt, S. F. and Martin, C. A. (1985) Forecasting future trends in European tourism demand, Tourist Review, 4,12-19.

Witt, S. F. and Martin, C. A. (1987a)Internationaltourism demandrnodcls: inclusion of 8,3340 Management, Tourism marketingvariables, Witt, S. F. and Martin, C. A. (1987b) Econometricmodels for forecastinginternational 25,23-30 Journal of TravelResearch, tourism demand, Witt, S. F. and Witt, C. A. (1992)Modelling andforecastingdemandIn lourism. Academic Press,London.
Witt, S. F. and Witt, C. A. (1995) Forecasting tourism demand: a review of crnpirical 11,447-75. International Journal ofForecasting, research.

(1996)Dalos Esenciales, WTO, Madrid World TourismOrganisation (1997)Europe,WTO, Madrid World TourismOrganisation World Tourism Organisation(1999) Destination country totals according to a single WTO, Madrid. indicator -Arrivals of touristsfrom abroad 1987-1996, World Tourism Organisation(1999) Destination country totals according to a single WTO, Madrid. indicator - International tourismreceipts1985-1997, World Tourism Organisation(1999) Tourism Highlights, World Tourism Organisation, Madrid. World Tourism Organisation(1977-1980)Trendsof Tourisin Afovenients and Iapnents, Madrid. World Tourism Organisation, World Tourism Organisation(1980-1999)Yearhookof tourisin statistics, World Tourism Madrid. Organisation, Yavas, B. F. and Bilguin Z. (1996) Estimation of tourism dcmand in Turkcy: a poolcd Tourism Analysis, 1,19-27 time analysis. series and cross-section

317

Zarnowitz, V. (1978) On the accuracyand propertiesof recentmacro-cconomic forccasts, Papersand Proceedings ofthe AmericanEconomicAssociation,63,313-19. Zellner, A. (1962) An efficient methodof estimatingseeminglyunrelatedregressions and bias,AmericanStatisticalAssociationJournal, 57,348-68. test for aggregation

TINCJ

.A. SITY

318