Desenvolvimento Económico Português No Espaço Europeu - 2006

Conferência do Banco de Portugal
10 de Fevereiro de 2006
Desenvolvimento Económico
Português no Espaço Europeu
Portuguese Economic
Development in the European Context
comunicações
proceedings
Banco de Portugal
Av. Almirante Reis, 71
1150-012 Lisboa
Economic Research Department
Printed by
Tipografia Peres, SA
Number of copies printed

800 issues
Legal Deposit no. 236983/06
ISBN 972-9479-90-9
Nota de Apresentação
Foreword
Portuguese Economic Development in the European Context V
NOTA DE APRESENTAÇÃO
Com a realização desta 3ª Conferência sobre o «Desenvolvimento Económico

Português no Espaço Europeu», o Banco de Portugal reafirma o objectivo de contribuir
para estimular a investigação económica aplicada aos problemas do crescimento real da
economia portuguesa. Este é um problema de médio e longo prazo que se julgaria
relativamente distante da esfera de actuação de um Banco Central. No entanto, a
participação bem sucedida do país na união monetária europeia também depende da
capacidade de ajustamento da economia real ao novo enquadramento em que tem de
funcionar. Nesta perspectiva, atravessamos uma fase difícil da nossa vida económica após
vários anos de fraco crescimento. Muitos factores contribuíram para esse desempenho
económico, mas o que aparece como mais preocupante é a possibilidade de estarmos
perante problemas estruturais de falta de competitividade. Esta é uma questão associada à
desaceleração continuada da nossa taxa de crescimento potencial e tem que ser analisada
numa perspectiva de médio prazo.
Em termos da teoria tradicional esperar-se-ia que a economia portuguesa
continuasse numa trajectória de convergência com os nossos parceiros europeus mais
desenvolvidos. A participação na união monetária deveria não só criar condições favoráveis
ao crescimento – que na verdade criou – como deveria também aumentar a pressão para
que se produzissem as transformações da estrutura produtiva e o aumento de produtividade
que a concorrência acrescida veio tornar mais prementes. No entanto, a redução da restrição
financeira que acompanhou a passagem a um regime de baixa inflação e baixo custo do
capital, contribuiu para atenuar aquela pressão. Erros cometidos na política orçamental e no
comportamento dos custos unitários do trabalho contribuíram também para o período mais
recente de crescimento insuficiente. Isto não invalida, porém, que a participação no euro
tenha criado circunstâncias favoráveis ao nosso desenvolvimento desde a diminuição do
custo do capital, à garantia de estabilidade macroeconómica e ao funcionamento mais
eficiente de mercados abertos e concorrenciais. Para aumentar o potencial de crescimento
haveria, porém, que juntar a estas condições o progresso da qualificação dos recursos
humanos, o aumento e melhoria do stock de capital e o aperfeiçoamento da gestão dos
factores e processos produtivos. No contexto do espaço económico da união monetária a
VI Vítor Constâncio
economia portuguesa pode atrair projectos de investimento que integrem estas

características e que, desse modo, consigam atingir bons níveis internacionais de
produtividade, como acontece em tantos casos de investimento estrangeiro existentes no
país. O ponto é que no quadro da actual concorrência mundial existem muitas localizações
alternativas no espaço europeu com factores competitivos assinaláveis. Para responder a
isso temos, pois, que melhorar nos aspectos cruciais de que depende o salto de
produtividade de que necessitamos para resolvermos o problema do nosso desenvolvimento
económico. A produtividade é uma variável complexa que é uma resultante final de um
vasto conjunto de condições que envolvem todos os subsistemas sociais com influência na
economia. Contribuem aspectos tão variados como a qualificação dos recursos humanos, a
qualidade das instituições judiciais e políticas, a excelência da investigação e da inovação
tecnológica, o aproveitamento de economias de escala, a concretização de efeitos de
composição na estrutura produtiva ou o investimento em tecnologias de informação e
comunicação.
Os trabalhos apresentados nesta Conferência abordam temas relevantes para
avaliar algumas das condições de que depende o crescimento a prazo da economia
portuguesa. Por exemplo, a consideração da estrutura produtiva sectorial revela a
importância crucial dos serviços e a necessidade de expandir a respectiva produtividade
através de maior investimento em tecnologias de informação e comunicação. Os trabalhos
dedicados à educação e qualificação dos recursos humanos salientam a importância dos
factores familiares nos resultados do processo educativo e a necessidade de investir mais
nos anos iniciais do percurso formativo. Por outro lado, a quantificação do elevado retorno
da formação profissional proporcionado pelas empresas reforça o caso do seu interesse para
a melhoria da produtividade. A influência do sistema fiscal na oferta de trabalho ou a
influência das instituições democráticas na disciplina orçamental da importante
componente do sector público constituído pelas autarquias, são outros aspectos com
interesse para a análise de aspectos relevantes, ainda que necessariamente parcelares, que
condicionam o nosso desenvolvimento. A verdade é que não existem fórmulas simples de
sucesso para os problemas da economia portuguesa. O caminho do ajustamento real para
vencer os desafios da nova concorrência internacional será longo e difícil. O contributo dos
economistas tem sido importante para esclarecer os problemas e as grandes orientações das
Portuguese Economic Development in the European Context VII
políticas necessárias à sua resolução. Os trabalhos desta Conferência constituem uma prova
adicional desse contributo
Os estudos que estão na base da Conferência provieram quer de um concurso
público dirigido aos Departamentos de Economia das Universidades portuguesas, quer de
convites directos a alguns economistas. Seguiu-se um processo de selecção dos estudos,
tarefa que foi da responsabilidade do Comité Científico, composto por Mário Centeno
(Banco de Portugal e Instituto Superior de Economia e Gestão), Isabel Horta Correia
(Banco de Portugal e Universidade Católica Portuguesa), José Ferreira Machado
(Universidade Nova de Lisboa e consultor externo do Departamento de Estudos
Económicos do Banco de Portugal), Carlos Robalo Marques (Banco de Portugal e
Universidade Nova de Lisboa), Maximiano Pinheiro, (Banco de Portugal e Instituto
Superior de Economia e Gestão) e Pedro Portugal (Banco de Portugal e Universidade Nova
de Lisboa).
Agradecimentos são devidos a todos pelo trabalho realizado, tal como ao
Departamento de Estudos do Banco de Portugal é devido o nosso reconhecimento pela
tarefa de organização da iniciativa.
Da qualidade e conteúdo dos trabalhos seleccionados dá conta esta publicação,
onde são divulgados os estudos apresentados na Conferência, que se realizou a 10 de
Fevereiro de 2006, na Fundação Calouste Gulbenkian, em Lisboa.
Vítor Constâncio
Governador do Banco de Portugal
Portuguese Economic Development in the European Context IX
FOREWORD
With this third Conference on “Portuguese Economic Development in the

European Area”, Banco de Portugal strengthens the objective of stimulating economic
research applied to the problems of the real growth of the Portuguese economy. This is a
medium and long-term problem that might be considered relatively far from the scope of
action of a central bank. However, the successful participation of the country in the
European Monetary Union also depends on the adjustment capacity of the real economy to
its new operational framework. In this perspective, Portugal is undergoing a difficult phase
of its economic life after several years of weak growth. This economic performance was
due to a number of factors, but the most worrisome seems to be the possibility of facing
structural problems of lack of competitiveness. This is associated with the continued
deceleration of the Portuguese potential growth rate and should be analysed from a
medium-term perspective.
In terms of the traditional theory, the Portuguese economy would be expected to
continue on a convergence path with Portugal’s most developed European partners. The
participation in a monetary union should not only create favourable conditions to growth –
which indeed it did – but also increase pressure towards changes in the productive structure
and a rise in productivity, which the increased competition made more pressing. However,
the reduction of the financial restraint, which accompanied the shift towards a low inflation
and low capital cost regime, led to the easing of such pressure. Mistakes in the fiscal policy
and in the behaviour of unit labour costs also led to the most recent period of insufficient
growth. This notwithstanding, the participation in the euro area created favourable
conditions for development, from a reduction in the cost of capital to the guarantee of
macroeconomic stability and to the more efficient functioning of open and competitive
markets. However, to increase the growth potential, these conditions should be
supplemented by progress in the qualification of human resources, the increase and upgrade
of the capital stock and the improved management of factors and productive processes. In
the context of the economic area of a monetary union, the Portuguese economy may attract
investment projects with these characteristics that are able to reach good international
productivity levels, as it is so often the case with foreign investment in the country.
X Vítor Constâncio
However, under the current global competition environment, there are many alternative
locations in the European area with remarkable competitive factors. In response, Portugal
has to improve the key aspects that may give the required boost to productivity to solve
Portugal’s economic development problem. Productivity is a complex variable, which is
the final result of a large set of conditions that involve all social subsystems with a bearing
on the economy. Such aspects include the qualification of human resources, the quality of
political and judicial institutions, the excellence of IT research and innovation, the
exploitation of economies of scale, the materialisation of composition effects in the
productive structure or investment in information and communication technology (ICT).
The research papers presented in this Conference are focused on relevant issues to
assess some of the conditions upon which depends the long-term growth of the Portuguese
economy. For instance, the taking into consideration of the sectoral productive structure
reveals the crucial importance of the services and the need to expand the respective
productivity through greater investment in ICT. The research papers on the education and
qualification of human resources emphasise the importance of family factors in the results
of the learning process and the need to invest more in the first years of schooling. On the
other hand, the quantification of the high return on vocational training provided by
companies, strengthens their interest in improving productivity. The influence of the tax
system on job offers or the influence of democratic institutions on the fiscal discipline of
local authorities, which are an important component of the public sector, are other factors
of interest for the analysis of the relevant aspects, albeit partial, which condition
development in Portugal. The truth is that there is no easy success formula for the
problems of the Portuguese economy. The path of real adjustment to address the challenges
of the new international competition is hard and difficult. Economists have given an
important contribution to the clarification of problems and of the broad policy guidelines
required to overcome them. The research papers presented in this Conference are an
additional proof of such contribution.
The research papers presented in this Conference came from both a public call for
research papers addressed to the Departments of Economy of the Portuguese universities
and from the direct invitation of some economists. The scientific committee in charge of
the selection process was comprised of Mário Centeno (Banco de Portugal and Instituto
Portuguese Economic Development in the European Context XI
Superior de Economia e Gestão), Isabel Horta Correia (Banco de Portugal and

Universidade Católica Portuguesa), José Ferreira Machado (Universidade Nova de Lisboa
and external advisor to the Economic Research Department of Banco de Portugal), Carlos
Robalo Marques (Banco de Portugal and Universidade Nova de Lisboa), Maximiano
Pinheiro, (Banco de Portugal and Instituto Superior de Economia e Gestão) and Pedro
Portugal (Banco de Portugal and Universidade Nova de Lisboa).
Our sincere thanks are due to all the contributors for the work they have
undergone. Similarly, we place on record our recognition to the Bank's Economic Research
Department for organising this Conference.
Attesting the quality of the research produced, this publication puts together the
research research papers to be presented in the third Conference “Portuguese Economic
Development in the European Context”, that will take place on 10 February 2006, at the
Calouste Gulbenkian Foundation in Lisbon.
Vítor Constâncio
Governador do Banco de Portugal
Índice
Contents
Contents XV
Nota de Apresentação / Foreword
Nota de Apresentação......................................................................................................... V
Foreword ............................................................................................................................ IX
Vítor Constâncio
Conferência do Banco de Portugal sobre “Desenvolvimento Económico Português no

Espaço Europeu” / Conference held by Banco de Portugal on “Portuguese Economic
Development in the European Context”
Third Conference on “Portuguese Economic Developments in the European Context:

A Synthesis......................................................................................................................... 3
José A. Ferreira Machado (Faculdade de Economia, Universidade Nova de Lisboa)
Sessão 1
Transformação Estrutural e Produtividade Agregada em Portugal .................................... 9
Margarida Duarte (Federal Reserve Bank of Richmond)
Diego Restuccia (Federal Reserve Bank of Richmond e University of Toronto)
Impostos e Decisão de Trabalho: Portugal, Europa e Estados Unidos ............................... 41

André C. Silva (Faculdade de Economia, Universidade Nova de Lisboa)
Budget Setting Autonomy and Political Accountability..................................................... 65

Susana Peralta (Faculdade de Economia, Universidade Nova de Lisboa e CORE-
UCL)
Sessão 2
Equality of Opportunity and Educational Achievement in Portugal .................................. 97
Pedro Carneiro (University College of London, Institute for Fiscal Studies e Center
for Microdata Methods and Practice)
The Internal Rate of return to On-the-Job Training............................................................125

Rita Almeida (World Bank)
Pedro Carneiro (University College of London, Institute for Fiscal Studies e Center
for Microdata Methods and Practice)
Will the East Follow Portugal?...........................................................................................165

Cátia Batista (Department of Economics - University of Oxford)
XVI Contents
Sessão 3
Pequenas e Médias Empresas em Portugal: Factos, Teoria e Política Microeconómica ....223
Luís Cabral (New York University e CEPR)
Asymmetric Information in the Stock Market: Economic News and Co-movement

between US and Portugal ...................................................................................................257
Rui Albuquerque (Boston University)
Clara Veja (University of Rochester)
Terceira Conferência "Desenvolvimento Económico
Português no Espaço Europeu" - Uma Síntese
Third Conference on "Portuguese Economic Development in

the European Context"
– A Synthesis
THIRD CONFERENCE ON “PORTUGUESE ECONOMIC DEVELOPMENT IN
THE EUROPEAN CONTEXT”: A SYNTHESIS
José A. Ferreira Machado
A nation’s standard of living is determined by the productivity of its workers that
is, by the amount of goods and services produced for each hour of a worker’s time. For a
given state of technological knowledge, this productivity depends essentially on the
quantity of factors of production per worker and on the efficiency of their utilization.
Generally speaking, the papers presented to the Conference address some basic forces
behind productivity namely, human capital accumulation, market competition, the role of
the services sector and the efficiency costs of taxes
In the second-half of the twentieth century the Portuguese GDP per capita grew at
a remarkable pace in excess of 4% per year. Taking the U.S. as a benchmark, this translated
into a significant convergence of standards of living. The paper by Margarida Duarte and
Diego Restuccia, shows that this convergence reflects mainly the increase of the
productivity in manufacturing since productivity growth in agriculture and services lagged
behind the US. Since services tend to gain weight in GPD as economies progress, further
convergence of GDP per capita will be in jeopardy unless productivity in services starts
growing faster.
The previous analysis is probably too aggregate to draw specific policy
recommendations on how to increase productivity. Nevertheless, the focus on the service

4 José A. Ferreira Machado
sector is quite useful and important. Indeed, since services industries are usually more
regulated and shielded from market forces than manufacturing and are also more human
capital intensive, faster productivity growth will demand increased competition and faster
human capital accumulation. These are the topics of five of the communications to the
conference.
A significant part of the differences observed in workers’ productivity (measured
by their wages) is explained by diverse levels of education in the workforce (Pedro
Carneiro’s contribution). Educational policy is thus critical to increase productivity.
Unfortunately, the paper reveals that the most important determinant of students’
performance is the socio-economic family background and, consequently, traditional
policies based upon increasing the amount of resources in schools have a very limited
effect. The family, Carneiro concludes, must be seriously taken into account when
designing new policies. But, not only formal education increases the stock of human
capital: as Rita Almeida and P. Carneiro note, more than 50% of the human capital
accumulated during a typical life time results from post schooling investments in particular
from on-the-job training. The rate of return on this type of investment was found to exhibit
a huge dispersion: there are firms for which investing in training is a bad idea while others
have suboptimal levels of investment. The sources of this inefficiency are left unexplained
but, one may safely infer that incentive schemes to formation and training that do not
accommodate the existing heterogeneity in returns should be avoided. Besides these direct
effects, education and training have also indirect effects on productivity. As the paper by
Cátia Batista remarks the aggregate technology in Portugal is characterized by the
complementarity between physical and human capital; this means that the more qualified
Third Conference on “Portuguese Economic Development in the European Context”: A Synthesis 5
the work force is, the higher the return on investment and, consequently, the more attractive
will Portugal be to new capital.
Improving the productivity of workers and firms is certainly a way of improving
aggregate productivity. The paper by Luis Cabral draws attention to an alternative: the
selection process. Within each sector, firms with low productivity coexist with those with
relatively high productivity: 10 to 20% of firms have productivity levels that more than
double the sector average. So, one way of improving aggregate productivity is to ensure
that competition works on a leveled playing field so that the most efficient firms survive
and prosper. Properly functioning markets are thus key to productive efficiency. The capital
market is preeminent in this respect; distortions in this market imply that the best projects
may lack funding and less efficient firms survive. The paper by Rui Albuquerque and Clara
Vega studies how the Portuguese capital market reacts to local and U.S. news. The authors
conclude that the co-movements between indices of the two markets are consistent with
efficiency in the sense of adequate reaction to fundamentals.
Governments’ interventions have also major effects on the incentives to an
efficient use of resources and on productivity. Susana Peralta deals with the issue of
government decentralization. Specifically, she analyses the case for decentralizing taxes in
a context where local governments already decide on the provision of local public goods.
From an accountability perspective, the basic trade-off is between ability to identify bad
governments – bigger under tax autonomy – and discipline – bigger under a centralized tax
regime. No system dominates but, autonomy tends to be preferred the higher the proportion
of good politicians. André C. Silva analyses the impact of taxation on income consumption
on the labor supply decisions. The increase in these taxes is shown to explain the reduction
6 José A. Ferreira Machado
in hours worked in Portugal over the last decade and a half. Taxes displace people from the
market into less productive activities and, thus, lower productivity.
As a very personal summation, four major lessons may, in my view, be drawn
from these communications to the Bank of Portugal 3rd Conference:
▪ Services are critical to continued convergence of productivity and standards of

living.
▪ Income and consumption taxes destroy incentives to work and this effect in
empirically relevant.
▪ It is not enough to invest additional resources in education in order to improve

results; policies should address the issue of the student’s family background.
▪ Sound competition policies may provide a significant boost to aggregate

productivity.
Sessão 1
Session 1
THE STRUCTURAL TRANSFORMATION AND AGGREGATE PRODUCTIVITY
IN PORTUGAL∗
Margarida Duarte
Federal Reserve Bank of Richmond
Diego Restuccia
Federal Reserve Bank of Richmond and
University of Toronto
September 2005
Abstract
In this paper we document the structural transformation - the reallocation

of labor between agriculture, manufacturing, and services - undergone by Portugal
between 1956 and 1995 and the U.S. between 1870 and 1995. We document that
together with the process of labor reallocation across sectors, Portugal has reduced
the aggregate productivity gap with the U.S., from a relative productivity of 0.26 in
1956 to 0.55 in 1995. We calibrate a simple general equilibrium model of the
structural transformation to data for the U.S. We use this model to gain insight into
the factors accounting for the structural transformation in Portugal. We show that
during this period Portugal featured low and roughly constant relative productivity in
agriculture and services (around 22 percent), low and growing relative productivity
in manufacturing (from 44 to 112 percent), and a time varying wedge on non-market
activities. We find that the aggregate consequences of Portugal closing its relative
productivity gap in services to levels consistent with European data would be
dramatic. In addition, while productivity growth in manufacturing accounts for the
reduction of the aggregate productivity gap with the U.S., unless services start to
grow faster, the model implies no further closing on this productivity gap.
Keywords: productivity, structural transformation, agriculture, manufacturing, services.
JEL Classification: O1,O4.
∗
The views expressed in this article are those of the authors and not necessarily those of the Federal Reserve
Bank of Richmond or the Federal Reserve System. Contact Information: Research Department, Federal
Reserve Bank of Richmond, Richmond, VA 23220, USA. E-mail:margarida.duarte@rich.frb.org; and
diego.restuccia@utoronto.ca.
10 Margarida Duarte - Diego Restuccia
1. Introduction
We address the long-run economic performance of Portugal and its process of

structural transformation, whereby the agricultural sector is replaced in importance by the
manufacturing sector and later by the service sector. Portugal has undergone a substantial
process of structural transformation over the last 50 years, similar to the past experience of
other developed economies. Together with the process of labor reallocation across sectors,
Portugal has reduced the aggregate productivity gap with the U.S. over this period, from a
relative productivity of 0.26 in 1956 to 0.55 in 1995.
In this paper, we investigate the role of each sector in the process of structural
transformation in Portugal and their implications for aggregate productivity. We calibrate a
simple general equilibrium model of the structural transformation to data for the U.S. Our
analysis produces sharp conclusions about the factors accounting for the structural
transformation in Portugal during the 1956-95 period and about the factors accounting for
aggregate productivity growth relative to the U.S. Agriculture and services observe much
lower relative productivity than manufacturing (about half) in 1956 and no systematic
growth during the period, while relative productivity in the manufacturing sector increased
substantially. We find that the aggregate consequences of Portugal closing its productivity
gap in services relative to the U.S. to levels consistent with European data would be
dramatic. As a result, while productivity growth in manufacturing accounts for the
reduction of the aggregate productivity gap with the U.S., unless services start to grow
faster, the model implies no further closing on this productivity gap.
We build a three-sector model of the structural transformation. Following
Rogerson (2005), in our model, labor reallocation across sectors is driven by two channels:
income effects due to non-homothetic preferences as in Kongsamut, Rebelo, and Xie
(2001) and substitution effects due to differential productivity growth as in Ngai and
Pissarides (2004). We calibrate our benchmark economy to the structural transformation of
the U.S. between 1956 and 1995 and use this benchmark economy to determine the factors
that underly the structural transformation in Portugal. We use this model to measure
productivity differences between Portugal and the U.S. across sectors in 1956 and to gain
insight into the factors accounting for the structural transformation in Portugal. We show
The Structural Transformation and Aggregate Productivity in Portugal 11
that during this period Portugal featured low and roughly constant relative productivity in
agriculture and services (around 22 percent), low and growing relative productivity in
manufacturing (from 44 to 112 percent), and a time varying wedge on non-market
activities. These features are essential in accounting for the structural transformation in
Portugal. Our model of the structural transformation in Portugal allow us to study the
aggregate productivity implications of the factors leading to the structural transformation as
well as other counterfactual situations.
The paper is organized as follows. In the next section we document the long-run
performance of the Portuguese economy relative to the U.S. and the process of structural
transformation in both countries. In section 3 we build a general equilibrium model of the
structural transformation. We describe the calibration of the model in section 4 and present
the results in section 5. In section 6 we discuss our findings and we conclude in section 7.
2. Transformation and Long-run Performance
In this section we document the economic performance and the process of

structural transformation in Portugal relative to the U.S. from 1950 to 2000. We focus on
labor productivity (GDP per worker) as our measure of economic performance and
document the behavior of GDP per worker in Portugal relative to its behavior in the U.S.
2.1. The Behavior of Aggregate Labor Productivity

We find a substantial process of convergence in aggregate labor productivity from
1950 until the mid 1970's in Portugal relative to more developed economies. However, this
process of convergence in productivity slowed down considerably in the mid 1970's. Figure
1 plots the trend GDP per worker in Portugal relative to the same measure in the U.S. and
Ireland from 1950 to 2000.1 While the two measures differ substantially, both depict a
marked slowdown in the process of Portuguese convergence occurring in the mid 1970's.
Aggregate labor productivity in Portugal relative to the U.S. (represented by the solid line)
increased steadily from 1950 to 1975, from about 0.22 to 0.45. The rate at which GDP per
worker in Portugal converged to that in the U.S. slowed-down markedly around 1975: In
1
Trends are obtained from the data by using the Hodrick-Prescott filter.
the latter 25 years of the sample period, GDP per worker in Portugal grew from 0.45 to
0.55 relative to the U.S.2 The dashed line documents GDP per worker in Portugal relative
to Ireland. In 1974, after a period of strong convergence, Portugal and Ireland had
comparable levels of GDP per worker. Since then, however, GDP per worker in Portugal
has declined steadily relative to that in Ireland, reverting the convergence accomplished in
the first half of the sample period. This process accentuated markedly during the 1990's, a
period of substantial growth in Ireland.
To gain insight about the driving forces behind movements in output per worker
we consider an aggregate production function that is common to all countries. Let output
Y in a given country be characterized by a Cobb-Douglas production function that depends
on the total capital stock K , total hours worked Lh , and total factor productivity (TFP)
A:
Y = AK α ( Lh )
1−α
. (1)
In this expression, L represents the number of workers employed, h represents average

hours worked per employed person, and α represents the share of payments to capital in
total income Y (when factor markets are competitive). Notice that TFP is not directly
observable in the data. Given observations on the capital stock, employment, hours worked,
and output and given an estimate for the share of payments to capital in total income, we
can obtain a measure of TFP as the residual in equation (1).
Equation (1) can be re-written in terms of output per worker Y / L as:
α
1 ⎛ K ⎞
1−α
Y
= A1−α ⎜ ⎟ h. (2)
L ⎝Y ⎠
This equation shows that movements in GDP per worker Y / L can be decomposed into
movements in TFP, movements in the capital-output ratio K / Y , and movements in the
average number of hours worked h .
2
Later on in the paper we compare the long-run performance of Portugal relative to the U.S. from 1956 to 1995
due to data restrictions on sectoral employment. For this period, relative labor productivity increased from 0.26
in 1956 to 0.55 in 1995.
Empirical evidence suggests that capital to output ratios are remarkably stable
over time for many countries, including the U.S. and Portugal.3 Therefore, we abstract from
movements in capital-output ratios as a driving force of relative movements in GDP per
worker across countries and focus on the contributions of relative movements in TFP and
hours worked across countries in the observed behavior of GDP per worker in Portugal
relative to the U.S.
In the U.S., average hours worked per year fell from 2, 008 hours (about 39 hours
per week) in 1950 to 1,878 (about 36 hours per week) in 2000. Thus, in the U.S.,
movements in hours worked contributed negatively towards growth in GDP per worker in
this period. In Portugal, average hours worked fell more than in the U.S. over the period
1950-2000. Average hours worked were 2,344 (about 45 hours per week) in 1950 and they
had fallen to 1, 715 by 2000 (about 33 hours per week). Thus, the pattern of hours worked
in Portugal relative to the U.S. suggests that hours worked contributed negatively towards
the observed convergence of GDP per worker in Portugal relative to the U.S. during this
period. The process of Portuguese convergence relative to the U.S. is thus consistent with
higher TFP growth in Portugal relative to the U.S., which more than offsets the fall in hours
worked in Portugal relative to the U.S.
Given the behavior of hours worked and GDP per worker in Portugal relative to
the U.S. over the period, we conclude that movements in aggregate productivity (as
measured by TFP) were the main driving force behind the behavior of GDP per worker in
Portugal relative to the U.S.
2.2. The Process of Structural Transformation

Underlying the pattern for relative labor productivity in Portugal depicted in
Figure 1 is a substantial process of reallocation of resources across sectors and different
patterns of labor productivity by sector. This reallocation process from the agricultural
sector into the manufacturing sector and from this sector into the service sector is typically
3
See for instance Kaldor (1961), Cooley and Precott (1995), Kongsamut, Rebelo, and Xie (2001) for the U.S.
and Cavalcanti (2004) for Portugal.
referred to as the structural transformation of the economy in the development literature.4

The process of structural transformation has been extensively documented.5 This
process is typically characterized by a substantial fall in the share of employment in
agriculture to less than 10 percent, by a steady increase in the share of employment in
services, and by an hump-shaped pattern for the share of employment in manufacturing.
That is, the typical process of sectoral reallocation involves an increase in the share of
employment in manufacturing in the early stages of the reallocation process, followed by a
decrease in the share of employment in manufacturing in the later stages.
Different economies have started the process of structural transformation at
different points in time. In Figure 2 we report the shares of employment in agriculture,
manufacturing, and services in the U.S. from 1869 to 1970, which are broadly consistent
with the general characterization described above. By the middle of the 20th century, a
substantial degree of sectoral labor reallocation had already taken place in the U.S.: While
in 1869 employment shares in agriculture, manufacturing, and services were 0.48, 0.24, and
0.28, by 1948 these shares were 0.10, 0.34, and 0.56. In the second half of the century, the
process of labor reallocation from agriculture and manufacturing into services continued, as
Figure 3 documents. From 1956 to 1995, the share of employment in agriculture in the U.S.
fell from about 10 percent to 2 percent, the share of employment in the manufacturing
sector fell from about 38 percent to 24 percent, while the share of employment in services
increased from 52 percent to 74 percent.6
4
In this paper we refer to manufacturing and industry interchangeably. In the appendix we describe in detail our
definition of sectors in the data.
5
See, for instance, Kuznets (1966), Maddison (1980), among others.
6
Figures 2 and 3 use different data sources and the shares of employment for 1956 do not match. Nevertheless,
the two figures are consistent regarding the pattern of structural transformation in the U.S.
Portugal has experienced a process of structural transformation that is also broadly

consistent with the experience of other economies. Figure 4 documents the share of
employment for the agricultural, manufacturing, and service sectors from 1956 to 1995. As
this figure shows, Portugal has undergone a substantial process of sectoral labor
reallocation in the last 50 years. The share of employment in agriculture has fallen from
roughly 48 percent in 1956 to 10 percent in 1995. The share of employment in the service
sector has increased steadily throughout this period, from 33 percent in 1956 to 65 percent
in 1995. Similar to other countries, the share of employment in manufacturing has an
hump-shaped pattern in this period: It increased from 1956 to 1980 (from roughly 20
percent to 26 percent) and it decreased thereafter (to 22 percent in 1995). This pattern of
reallocation of employment across sectors over time is similar to the characterization of the
structural transformation for the U.S. economy and other economies. It is interesting to note
that the structural transformation in Portugal from 1956 to 1995 resembles closely the
structural transformation in the U.S. economy from the period 1870 to 1956. Although
Portugal is lagging behind the process of structural transformation in the U.S., it has
accomplished about the same reallocation of labor across sectors than the U.S. in less than
half the time (39 years as opposed to 89 years in the U.S.).
These distinct patterns of employment across sectors observed both in Portugal
and the U.S. over the period 1956-1995 are associated with distinct patterns of labor
productivity across sectors. In the U.S., labor productivity increased in all sectors, specially
so in the agricultural sector. Figure 5 documents the GDP per worker in each sector from
1956 to 1995. The annualized growth rates of labor productivity over this period were 3.4
percent in agriculture, 2.4 percent in manufacturing, and 1.4 percent in the service sector. In
1956 labor productivity in manufacturing and services were 2.2 and 1.7 times that of
agriculture. By 1995, labor productivity in manufacturing and services had fallen to 1.5 and
0.8 times that of agriculture.
The pattern of labor productivity across sectors in Portugal is distinct from that in
the U.S. Figure 6 plots the GDP per worker in each sector from 1956 to 1995 in Portugal.
The annualized growth rates of labor productivity over this period were 4.6 percent in
manufacturing, 4.0 percent in agriculture, and 1.7 percent in the service sector. The pattern
for labor productivity and share of employment across sectors is determinant for aggregate
labor productivity. In the following sections we build and calibrate a general equilibrium
model of the structural transformation. We use this model to asses the role of each sector in
the process of structural transformation in Portugal and their implications for relative
aggregate productivity.
3. Basic Model
We develop a simple model of the structural transformation of an economy where

at each date three goods are produced: agriculture, manufacturing, and services. Our
baseline economy features linear technologies in labor with potentially differential
productivity growth across sectors and preferences for agriculture and services that feature
income elasticities different than one. We calibrate our benchmark economy to U.S. data
from 1956 to 1995. We show that this basic framework captures the salient features of the
structural transformation in the U.S. for the calibrated period.
3.1. Description of Economic Environment
Production At each date there are three goods produced: agriculture ( a ),

manufacturing ( m ), and services ( s ) according to the following constant returns to scale
production functions:
Yi = Ai Li , i ∈ {a, m, s}, (3)
where Yi is output in sector i , Li is labor allocated to production in sector i , and Ai is a

sector-specific technology parameter. When comparing our model to data, we associate the
labor input Li with the employment level of sector i . We assume that there is a continuum
of representative firms in each sector that are competitive in output and factor markets. At
each date, given the price of good- i output pi and wages w , a representative firm in
sector i solves:
max { pi Ai Li − wLi } , (4)

Li ≥ 0
where Li is the demand of labor in sector i .
Population The economy is populated by an infinitely-lived representative household of

constant size over time. (Without loss of generality we normalize the population size to
one.) The household has preferences over consumption goods as follows:
∑β u (c , c
t =0
t
t a, t ), β ∈ (0,1),
where ca , t is the consumption of agricultural goods at date t and ct is the consumption of

a composite of manufacturing and service goods at date t . The per-period utility is given
by:
u (ct , ca , t ) = log(ct ) + V (ca , t ),
where V (ca , t ) is non-homothetic, i.e., there is a subsistence level of agriculture a below

which the household cannot survive. This feature of preferences has a long tradition in the
development literature. It has been emphasized as a quantitative important feature leading
to the movement of labor away from agriculture in the process of structural transformation.
(See for instance Echevarria (1997), Laitner (2000), Kongsamut, Rebelo, and Xie (2001),
Caselli and Coleman (2001), Gollin, Parente, and Rogerson (2002), and Ngai and
Pissarides (2004) among many others.) Following Gollin, Parente, and Rogerson (2002),
we simplify the specification of V by assuming that in addition to the subsistence feature
of preferences, V is such that households only care to consume the subsistence amount.
Formally, V (ca ) = −∞ when ca < a , and V (ca ) = min{ca , a } when ca ≥ a . This feature of
preferences makes our analysis much more tractable. We shall see in Section 5 that this
simple preference specification captures the share of employment in the agricultural sector
in the data remarkably well.
The composite consumption ct is given by:
1
ct = ⎡⎣bcmρ , t + (1 − b)(cs , t + s ) ρ ⎤⎦ ,
ρ
where s > 0 , b ∈ (0,1) , and ρ < 1 . Given s , these preferences imply that the income
elasticity of consumption of service goods is greater than one. The variable s can be
interpreted as a constant level of production of service goods at home. Kongsamut, Rebelo,
and Xie (2001) assume similar preferences but they abstract from reallocation due to
differential productivity growth between manufacturing and services. Our model allows
both channels to be operating during the structural transformation and our calibration
strategy described in the next section disentangles the contribution of each factor. We
found that for reasonable parametrization of preferences and productivity growth, the
movement away from manufacturing to services in the U.S. from 1956 to 1995 cannot be
captured by the model without an income elasticity greater than one for services. Our
approach to modeling the home sector for services is reduced form. Rogerson (2005)
considers a generalization of this feature where people can allocate time to market and non-
market production of service goods. However, we argue that our simplification is not as
restrictive as it first may appear since we abstract from labor hours in our model.
The problem of the household is also static, at each date and given prices, the
household chooses consumption of each good to maximize the per period utility subject to
the budget constraint. Formally,
{ }
1
max log ⎡⎣bcmρ + (1 − b)(cs + s ) ρ ⎤⎦ + V (ca ) ,

ρ
(5)
ci ≥ 0
subject to
pa ca + pm cm + ps cs = w.
In what follows we normalize the wage rate to one.
Market Clearing We assume that the household is endowed with one unit of productive
time that supplies inelastically to the market, so the demand of labor from firms must equal
this exogenous supply at every date:
La + Lm + Ls = 1. (6)
This specification implies that labor inputs in the model Li are associated with the shares
of employment in the data. Also at each date the market must clear for each good produced:
ca = Ya , cm = Ym , cs = Ys . (7)
3.2. Equilibrium of the Model
Equilibrium A competitive equilibrium is a set of prices { pa , pm , ps } , allocations

{ca , cm , cs } for the household, and allocations { La , Lm , Ls } for firms such that: ( i ) Given
prices, firm's allocations { La , Lm , Ls } solve the firm's problem in (firms), ( ii ) Given prices,
household's allocations {ca , cm , cs } solve the household's problem in ( hh ) , and ( iii )
markets clear: equations (6) and (7) hold.
Characterization The first order condition from the firm's problem implies that the benefit
and cost of a marginal unit of labor must be equal. This implies that prices of goods are
inversely related to productivity:
1
pi = ,
Ai
since the wage rate is normalized to one. Our specification of V (ca ) implies that ca = a 2
and, therefore, the resource constraint of agricultural goods implies that the labor allocation
in agriculture is determined solely by the subsistence constraint and labor productivity in
agriculture, i.e., labor in agriculture must satisfy:
a
La = . (8)
Aa
The first order conditions for consumption of manufacturing and service goods implies:
ρ −1
b ⎛ cm ⎞ pm
⎜ ⎟ = ,
(1 − b) ⎝ cs + s ⎠ ps
and using the market clearing conditions for output in manufacturing and services and for
labor we obtain:
(1 − La ) + s
Lm =
As
, (9)
1+ x
where
ρ
1
⎛ b ⎞ ρ −1 ⎛ Am ⎞
ρ −1
x≡⎜ ⎟ ⎜ ⎟ ,
⎝ 1 − b ⎠ ⎝ As ⎠
and La is given by (8).
Discussion Note that when s = 0 , equation (9) can be written as Ls / Lm = x . If ρ = 0

then the composite consumption good is a Cobb-Douglas aggregate of consumption of
manufacturing and service goods and differential productivity growth across these two
sectors will cause no reallocation of labor in this model. Our model is consistent with the
observed labor reallocation across the manufacturing and service sectors as labor
productivity grows in the manufacturing sector relative to services when the elasticity of
substitution between these goods is low ( ρ < 0 ). When s is strictly positive, however, the
model can imply a given amount of labor reallocation from manufacturing to services as
labor productivity in services grows for higher elasticity of substitution ρ .
4. Calibration
We calibrate our benchmark economy to U.S. data for the period from 1956 to
1995. Our calibration strategy involves selecting parameter values so that the equilibrium
of the model matches a given set of statistics in the data.
4.1. Description
We assume that a model period is one year. We need to select the following
parameters values: b , ρ , a , s , and the time series of productivity for each sector Ai , t for
t from 1956 to 1995 and i ∈ {a, m, s} . A summary of our calibrated parameters and targets
is in Table 1.
Our baseline calibration strategy is to restrict the parameters values to match the
structural transformation of the U.S. for the 1956-1995 time period. Since labor allocation
in agriculture is determined independently of the state of the other sectors in our model, our
calibration procedure can be roughly divided in two parts. First, we calibrate productivity
and subsistence in agriculture so that the equilibrium of the model matches the share of
employment in agriculture. Second, we calibrate the other parameters of the model to

match the share of employment in manufacturing (thus, the share of employment in
services is also matched by market clearing).
In particular, we proceed as follows. First, we normalize productivity levels across

sectors to one in 1956, i.e., Ai = 1 for all i ∈ {a, m, s} . Second, given the normalization for
Aa , 56 = 1 , we choose a to obtain the share of employment in agriculture in U.S. data in
1956. Third, given a we use equation (8) to choose Aa,t to match the share of
employment in agriculture in U.S. data at every date from 1957 to 1995. This calibration
implies an annual average labor productivity growth in agriculture during this period of
3.4%, which is consistent with the data for labor productivity growth in U.S. agriculture.7
In the second component of our calibration, we need to restrict s , b , ρ and the

annual average growth rates of productivity for manufacturing and services to match the
share of employment in manufacturing. Therefore, the data on employment in
manufacturing cannot uniquely restrict all these parameters. We proceed as follows. We
assume that labor productivity growth in manufacturing and services are 2.4% and 1.4%.
These estimates are consistent with our own calculations using the Groeningen Growth and
Development Centre data and the Penn World Table. Given ρ and b , s is chosen to
match the share of employment in manufacturing in the U.S. in 1956. Then b is chosen so
that, given the time path for relative productivity, the model matches the time path for the
share of employment in manufacturing. Since ρ determines how much relative
productivity growth is needed to produce a given reallocation of labor across sectors, ρ
will induce different patterns of aggregate productivity growth. We choose ρ to match
average aggregate productivity growth during the period (at 1956 prices). We calculate that
the average labor productivity growth in the U.S. between 1956 and 1995 is 1.7% in the
Penn World Table. Our calibrated value for ρ is -1.5. This value is consistent with the
assumed value in Rogerson (2005) in his analysis of the U.S. and European employment
and in the middle range of values in Ngai and Pissarides (2004).
7
Annual average growth rates over the period 1956-1995 are computed as γi = ( )
Ai ,95 1/ 39
Ai ,56
−1
4.2. Results of Benchmark Economy

Our calibration restricted preference and technology parameters of our model to
match the data for the U.S. structural transformation during the period 1956-1995. The
share of employment implied by our model are reported in Figure 7 (dotted lines), together
with data on the shares of employment in the U.S. (solid lines). The equilibrium
employment shares for manufacturing and services implied by our model match closely the
process of structural transformation of the U.S. over this period.8 The model implies a fall
in the employment share in manufacturing from about 38 percent in 1956 to 26 percent in
1995, while the employment share in services increases from about 53 percent to 72
percent. We found that, given the observed growth rates of labor productivity in the U.S.,
this process of labor reallocation between manufacturing and services could not be
accomplished in the model without an income elasticity greater than one in services. (See
Kongsamut, Rebelo, and Xie, 2001 and Ngai and Pissarides, 2004 for a detailed discussion
of relative productivity vs. income effects in the process of structural transformation.)
5. Quantitative Analysis
In this section, we use our quantitative model calibrated to the U.S. to gain
understanding into the process of structural transformation in Portugal. Our calibrated
benchmark economy puts discipline on technology and preferences, allowing us to gain
insight into the role of productivity differences in the process of structural transformation in
Portugal. In this section, we perform four experiments aimed at gaining insight into the
structural transformation in Portugal. Then, we perform counterfactual exercises that assess
the aggregate implications of different factors that drive the process of structural
transformation in Portugal. Our main findings in this section are that manufacturing
productivity growth accounts for the reduction in the aggregate productivity gap with the
U.S. during the 1956-1995 period and that the lack of relative growth in services has kept
Portugal lagging behind in aggregate productivity relative to the U.S.
8
Note that our model matches the share of employment in agriculture by construction.
5.1. Structural Transformation in Portugal

We take four steps aimed at understanding the structural transformation in
Portugal. First, we consider an economy identical in terms of preferences and technologies
to the benchmark economy except in the level of economy-wide productivity, in order to
match the observation that output per worker in Portugal was 26 percent of the U.S. level in
1956. Second, we allow for different relative productivities across sectors in 1956 that are
consistent with the observed shares of employment across sectors in Portugal in 1956.
Third, we consider an economy that in addition to the features described above has
productivity growth across sectors that is driven by observations on sectoral productivity in
Portugal. Finally, we consider a time-varying wedge between market and non-market
activities that allows us to match the process of structural transformation in Portugal.
Economy-wide Productivity Our first experiment involves using the benchmark economy
but with lower aggregate labor productivity in 1956. As documented in section 2, output
per worker in Portugal was 26 percent the output per worker in the U.S. in 1956. Our first
experiment assumes that labor productivity in each sector was 26 percent of the respective
sector in the U.S., i.e. Ai , 56 = 0.26 for i ∈ {a, m, s} instead of 1 in our calibration of the
benchmark economy. The results of this experiment in terms of the share of employment
across sectors are reported in Figure 8 where the solid lines represent the data and the
dashed lines the model. Focusing first on the share of employment implications for 1956,
the model implies too little employment in agriculture and services and too much
employment in manufacturing relative to the data. In other words, in the context of our
model, Portugal being 26 percent of the frontier productivity in all sectors implies much
less employment in agriculture and much more employment in manufacturing than in the
data, suggesting that Portugal may be less than 26 percent productive at agriculture and
more than 26 percent productive in manufacturing in 1956 relative to the U.S. This result
suggests our second step in constructing the structural transformation in Portugal.
Relative Sectoral Productivity in 1956 We set sectoral productivity in 1956 so that the
model matches the share of employment across sectors in Portugal in 1956 (in addition to
the relative aggregate productivity of 26 percent in 1956). Our calibration of this
experiment implies that agriculture, manufacturing, and services must be 21, 44, and 22
percent as productive as in the benchmark economy in 1956.9 The results of this experiment
are reported in Figure 9. While the model matches the share of employment across sectors
in 1956 exactly by construction of our calibration in this experiment, the time path of the
shares of employment are different than in the data, specially in agriculture. This result is to
be expected since we know from the sectoral productivity observations that productivity
growth in agriculture was slower in Portugal than in the U.S. from 1956 to about 1975 but
faster thereafter. This slower productivity growth should reduce the amount of labor
released from agriculture by the model over time. This leads to our next step in
constructing the structural transformation of Portugal which is to consider the sectoral
productivity growth observed in Portugal.
Productivity Growth in Portugal As noted above, Portugal is not riding along the same
technological process as the U.S. While in 1956 relative sectoral productivity in Portugal
were all below the U.S. level, Portugal experienced higher annualized rates of labor
productivity growth in all three sectors than the U.S. In this experiment, we use the growth
rates in agriculture, manufacturing, and services in Portugal implied by the smoothed path
of these variables between 1956 and 1995, together with the features of the two previous
experiments. The employment shares implied by the model are plotted in Figure 10. The
employment share in agriculture implied by the model matches very closely the data. This
result suggests that the simple characterization of preferences for agricultural goods that we
use in our model has implications for labor allocations in the agricultural sector that are
consistent with the data. The share of employment in services implied by the model grows
faster than in the data, while the share of employment in manufacturing implied by the
9
Recall from Figures 2 and 4 that Portugal underwent a structural transformation in agriculture between 1956
and 1995 that resembles closely the structural transformation that the U.S. underwent between 1870 and 1956.
Hence, an alternative calibration of a and productivity growth in agriculture to our benchmark would be to
select these parameters to match the structural transformation of the U.S. between 1870 and 1956. Normalizing
the productivity level of agriculture in 1870 to one, this alternative calibration would imply that a = 0.48
and a level of productivity in agriculture in 1956 of 4.8. This implies an annualized average rate of
productivity growth in agriculture of 1.84%. (Note that this rate of growth is less than half the observed growth
rate of productivity during the 1956 to 1995 period in the U.S.) If this level of productivity represents the
frontier in the world, then Portugal in 1956 should have observed a share of employment in agriculture of 10%
as opposed to the 48% in the data. We conclude that Portugal is not riding along the same technological
process as the U.S. There are factors (either institutional or policy driven) that lead to a large share of
employment in agriculture in Portugal in 1956.
model declines. We conclude that there may be factors, perhaps not technological, that are
preventing the “normal” movement of people from manufacturing to services. We consider
as our next step a wedge between non-market and market activities that summarizes all the
possible forces that prevent reallocation to services. Prescott (2004) and Rogerson (2005)
argue that taxes on market activities may be behind the employment problem in European
countries.
Structural Transformation in Portugal The previous discussion implies that our simple
framework does not capture the process of labor reallocation between manufacturing and
services observed in Portugal between 1956 and 1995. In this experiment we add a time-
varying barrier to market activities so that the model matches the Portuguese structural
transformation in this period, as reported in Figure 11. The resulting time varying barrier to
market activities has the feature that it grows almost monotonically from 1 in 1956 to 5 in
1995. Our benchmark economy with lower relative sectoral productivity in 1956, faster
productivity growth, and a time varying wedge in non-market activities is able to reproduce
closely the pattern of labor reallocation observed for Portugal in 1956-1995. We use this
economy as the basis of our counterfactual experiments in the next subsection.
Relative Sectoral Productivity Our model of the structural transformation implies patterns
of sectoral labor productivity relative to the U.S. that are consistent with aggregate data.
These relative sectoral productivities are plotted in Figure 12. The model implies that labor
productivity in manufacturing in Portugal converged fast relative to the U.S. during this
period, from 0.44 in 1956 to 1.12 in 1995. Labor productivity in agriculture and services in
Portugal, however, experienced very limited improvement relative to the U.S. during this
period: In 1956 relative productivity in agriculture and services were 0.21 and by 1995 they
were 0.26 and 0.25. Notice that the model imposes discipline on the relative sectoral
productivities in 1956. Their pattern thereafter is implied by data on productivity growth by
sector in each country.
5.2. Counterfactuals
Our previous analysis suggests that productivity in agriculture and services in
Portugal are behind the aggregate productivity differences of Portugal relative to the U.S.
In this section, we use our general equilibrium model to isolate the importance of each
sector in the structural transformation for aggregate productivity and sectoral labor
allocations.
Closing the Productivity Gap in Agriculture Our model implies that productivity in
agriculture in Portugal was 21 percent in 1956 and 26 percent in 1995 relative to the U.S.
In sharp contrast, Rogerson (2005) suggests that the productivity gap in agriculture
between Europe and the U.S. required to reproduce relative labor allocations in this sector
in 2000 was 0.97. Therefore, we ask what the aggregate productivity implications woud be
in the case in which Portugal closes the productivity gap in agriculture to 0.97 by 1995. To
produce this catch up in productivity, labor productivity in agriculture in Portugal would
need to grow at an annual rate of 7.5 percent instead of the 4 percent observed in the data.
The labor allocation and aggregate productivity implications of this counterfactual are
summarized in Table 2 and in Figure 13. While closing the productivity gap in agriculture
produces an important reallocation of employment from agriculture to services by 1995
(the shares of employment change from 10.6 to 2.9 percent in agriculture and from 65.4 to
72.5 percent in services) the aggregate productivity implications of this change are
relatively small: The annualized growth rate of aggregate productivity increases from 3.7
percent in the Benchmark Portugal to 3.8 percent in this counterfactual and aggregate
productivity in Portugal relative to the U.S. in 1995 increases from 0.55 to 0.57. The
intuition behind this result is that while improving productivity in agriculture produces an
important reallocation of labor, this reallocation shifts labor mostly towards services. As we
documented previously, the service sector in Portugal has a roughly similar relative
productivity to that in agriculture. In addition, the direct effect of the sharp improvement in
agricultural productivity in the aggregate falls over time, as the associated fall in the share
of employment in agriculture reduces the weight of this sector in the aggregate economy.
Closing the Productivity Gap in Services Our model implies that productivity of services
in Portugal relative to the U.S. was 21.7 percent in 1956 and 24.9 percent in 1995. As in the
case of agriculture, these numbers are in sharp contrast with the productivity gap in services
between Europe and the U.S. suggested by Rogerson (2005), 0.89, required to reproduce
relative labor allocations in 2000. In this counterfactual we ask what the labor allocation
and aggregate implications are of a change in productivity in services from 21.7 percent in
1956 to 89 percent in 1995. This remarkable change in relative productivity in services
generates almost no effect in the shares of employment across sectors. The reason is that
this improvement in the relative productivity of services generates no effect in the
allocation of labor in agriculture in our model and two opposing effects in the allocation of
labor across industry and services. On the one hand, higher productivity in services relative
to manufacturing, all else equal, reallocates labor towards services due to the low
substitutability between these two goods in preferences ( ρ < 0 ) . On the other hand, higher
productivity in the production of market services relative to non-market services (constant
s ), all else equal, reallocates labor towards manufacturing. In this counterfactual, these
two opposing effects roughly cancel each other and the effects on labor allocations are
small. However, this change in relative productivity has an important effect in aggregate
productivity because the improvement in productivity occurs in a large sector of the
economy that is growing in size due to the structural transformation. The growth rate in
aggregate productivity increases to 5.1 percent annually, leading to a relative aggregate
productivity of 0.90 in 1995 as documented in the dotted-dashed line in Figure 13.
Closing the Productivity Gap in Agriculture and Services While we found that
improving agricultural productivity by itself did not have large aggregate productivity
effects, when combined with improvements in the service sector, improving agricultural
productivity can have important aggregate effects. In this counterfactual we combine the
improvements in productivity described in the previous two counterfactuals. As
documented in Table counterfactuals, higher productivity in agriculture implies that there is
a substantial release of labor from agriculture to services (as in the first counterfactual). In
turn, higher relative productivity in services implies that this reallocation of labor has a
higher aggregate effect than in the second counterfactual. In this counterfactual, relative
aggregate productivity in 1995 is 0.97 compared to 0.9 in the case of improvement in the
service sector only.
5.3. Discussion
Our analysis produces sharp conclusions about the factors accounting for the
structural transformation in Portugal during the 1956-95 period and about the factors
accounting for aggregate productivity growth relative to the U.S. Agriculture and services
observe much lower relative productivity than manufacturing (about half) in 1956 and no
systematic growth during the period, while relative productivity in the manufacturing sector
increased substantially. While our analysis is silent about the factors accounting for this
behavior of relative productivity over time, we suspect that competition within each sector
may be responsible for this differential performance across sectors. Manufacturing is
composed of a number of tradable goods while agriculture, and specially services, are
composed of a number of non-tradable goods. Our analysis suggests that finding ways of
improving labor productivity in the service sector relative to the U.S. would have large
consequences for aggregate productivity. However, these policies cannot rely on foreign
competition. In addition, while manufacturing productivity has accounted for all the
aggregate productivity growth in Portugal relative to the U.S. during the period, its role in
determining aggregate productivity in the future is mitigated by its decreasing share in
employment. (Recall that Portugal has already started a second phase of structural
transformation whereby employment is moving from manufacturing to services.) We
conclude that only relative productivity growth in services can provide further closing of
the aggregate productivity gap with the U.S.
We documented that aggregate productivity in Portugal relative to the U.S. slowed

down since 1975. Our analysis suggests that this relative slowdown is accounted for by
relatively faster growth in manufacturing in the U.S. starting in 1975 and by the decline in
the share of employment in manufacturing in the process of structural transformation in
Portugal.
It is well known that distribution services represent a large portion of final-good
prices in developed economies. For instance, the U.S. Department of Agriculture reports
that out of every dollar spent on food in the U.S., eighty cents correspond to distribution
and marketing services, while only twenty cents correspond to the producer price that
farmers receive. Therefore, low relative productivity in services may be partly responsible
for the observed low relative productivity in agriculture. While our model does not
explicitly account for the role of distribution services, the last counterfactual in the previous
subsection suggests that if improvements in productivity in services go along with
improvements in productivity in agriculture, the aggregate productivity implications are
amplified.
Our analysis suggests that there has been an increasing wedge on non-market
activities. This wedge could represent an increasing role of taxes and other regulations in
the market economy. While this wedge is important in the model to account for some
features of the structural transformation in Portugal, as discussed in subsection 5.1, we
emphasize that this wedge is not important for aggregate productivity in Portugal relative to
the U.S. In fact, the counterfactual situation where this wedge is not present would imply
that employment would have moved faster out of manufacturing into services. Since
manufacturing in Portugal is relatively more productive than services, aggregate
productivity growth would have been lower in this counterfactual situation than with the
wedge.
6. Conclusions
From 1956 to 1995, GDP per worker in Portugal relative to the U.S. increased
from 0.26 to 0.55. This reduction of the aggregate productivity gap with the U.S. was
associated with a process of labor reallocation across sectors of production, typically
referred to as the structural transformation. In this paper we build a general equilibrium
model of the process of structural transformation. Using this model we are able to
disentangle the role of sectoral labor productivity growth in the reduction of the aggregate
productivity gap of Portugal relative to the U.S. We find that relative labor productivity in
manufacturing increased substantially and played an important role in this process. In turn,
relative labor productivity in agriculture and services lagged behind. We find that the
aggregate consequences of Portugal closing its productivity gap in services relative to the
U.S. to levels that are consistent with relative productivity levels in Europe would be
dramatic.
References
Caselli, F. and Colleman, W. J. “The U.S. Structural Transformation and Regional
Convergence: A Reinterpretation,” Journal of Political Economy, 2001, 109: pp.
584-616.
Cavalcanti, T. (2004) “Business Cycle and Level Accounting: The Case of Portugal,”
manuscript, Universidade Nova de Lisboa.
Cooley, T. and E. C. Prescott (1995). “Economic Growth and Business Cycles,” in
Frontiers of Business Cycle Research, ed. T. Cooley. New Jersey: Princeton
University Press.
Echevarria, M. “Changes in Sectoral Composition Associated with Growth,” International
Economic Review, 1997, 38: pp. 431-52.
Gollin, D., Parente, S., and Rogerson, R. “The Role of Agriculture in Development,”
American Economic Review Papers and Proceedings, 2002.
Groningen Growth and Development Centre and the Conference Board, August 2005, 
http://www.ggdc.net .
Heston, A., R. Summers, and B. Atten. (2002). Penn World Table Version 6.1, Center for
International Comparisons at the University of Pennsylvania (CICUP), 
http://pwt.econ.upenn.edu .
Kaldor, N. (1961) “Capital Accumulation and Economic Growth,” in The Theory of
Capital, ed. F.A. Kutz and D.C. Hague. New York: St. Martins.
Kongsamut, P., S. Rebelo, and D. Xie (2001), “Beyond Balanced Growth,'' Review of
Economic Studies 68, pp. 869-882.
Kuznets, S. (1966), Modern Economic Growth, Yale University Press.
Laitner, J. “Structural Change and Economic Growth,” Review of Economic Studies, 2000,
67: pp. 545-61.
Maddison, A. (1980) “Economic Growth and Structural Change in the Advanced
Countries,” in Western Economies in Transition, eds.: I. Leveson and W. Wheeler.
London: Croom Helm.
Ngai, L. R. and C. A. Pissarides. (2004) “Structural Change in a Multi-Sector Model of
Growth,” manuscript, London School of Economics.
Prescott, E. C. (2004). “Why do Americans Work So Much More than Europeans?”

manuscript, Arizona State University.
Rogerson, R. (2005). “Structural Transformation and the Deterioriation of European Labor
Market Outcomes,” manuscript, Arizona State University.
U.S. Census Bureau, Department of Commerce. Historical Statistics of the United States:
Colonial Times to 1970 (Part I). Washington, DC: U.S. Government Printing
Office, 1975.
A Data Sources and Definitions
We use the following data sources:

• Banco de Portugal, Séries Longas para a Economia Portuguesa.
• Groningen Growth and Development Centre and the Conference Board, August 2005,
http://www.ggdc.net. We use the following databases: Total Economy Database and
10-Sector Database.
• Historical Statistics of the U.S. (1975): Persons Engaged in Production, by Industry
Divisions: 1869 to 1970.
• OECD Employment Database.
• Penn World Tables Version 6.1.
We adopt the following sectoral definitions. We include in agriculture the
following sectors: agriculture, forestry, and fishing. We include in manufacturing: mining,
manufacturing, public utilities, and construction. Services include: wholesale and retail
trade; transport and communication; finance, insurance, and real estate; community, social,
and personal services; and government services.
All observations (except the historical shares of employment in the U.S.) are
smoothed using the H-P filter.
Figure 1
Relative Labor Productivity in Portugal
1
PT/US
PT/IR
0.9
0.8
Relative Labor Productivity
0.7
0.6
0.5
0.4
0.3
0.2
1950 1960 1970 1980 1990 2000
Years
Source: Penn World Tables Version 6.1.
Figure 2
Share of Employment by Sector – U.S. Historical Data
0.8
Agriculture
0.7 Industry
Services
0.6
Share of Employment
0.5
0.4
0.3
0.2
0.1
0
1879 1899 1919 1929−37 1944−48 1953−57 1960−69
Years
Source: Historical Statistics of the U.S. (1975)

Figure 3
Share of Employment by Sector – U.S.
0.8
Agriculture
0.7 Industry
Services
0.6
Share of Employment
0.5
0.4
0.3
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Source: OECD and author’s calculations.
Figure 4
Share of Employment by Sector – Portugal
0.8
Agriculture
0.7 Industry
Services
0.6
Share of Employment
0.5
0.4
0.3
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Source: Banco de Portugal and author’s calculations.

Figure 5
Labor Productivity by Sector – U.S.
7
Agriculture
Industry
Services
6
Value Added per Worker
1
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Source: Groningen Growth and Development Centre (2005) and author’s

calculations.
Figure 6
Labor Productivity by Sector – Portugal
7
Agriculture
Industry
6 Services
Value Added per Worker
1
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years

Figure 7
Share of Emplyment by Sector – Model vs. U.S. Data
0.8
0.7
0.6 Ag. Data

Ag. Model
Share of Employment
Ind. Data
0.5
Ind. Model
Svc. Data
0.4 Svc. Model
0.3
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Source: OECD and author’s calculations.
Figure 8
Share of Emplyment by Sector – Model vs. PT Data
0.8
Ag. Data
0.7 Ag. Model
Ind. Data
Ind. Model
0.6
Svc. Data
Share of Employment
Svc. Model
0.5
0.4
0.3
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years

Figure 9
Relative Sectoral Productivity in 1956 – Model vs. PT Data
0.8
0.7
0.6
Share of Employment
Ag. Data
0.5
Ag. Model
Ind. Data
0.4 Ind. Model
Svc. Data
0.3 Svc. Model
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Figure 10
Sectoral Produtivity Growth PT – Model vs. PT Data
0.8
0.7
0.6
Share of Employment
Ag. Data
0.5
Ag. Model
Ind. Data
0.4 Ind. Model
Svc. Data
0.3 Svc. Model
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years

Figure 11
Non- market Wedge – Model vs. PT Data
0.8
Ag. Data
0.7 Ag. Model
Ind. Data
Ind. Model
0.6
Svc. Data
Svc. Model
Share of Employment
0.5
0.4
0.3
0.2
0.1
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Figure 12
Sectoral Productivity (PT/US)
1.4
Agricuture
Industry
1.2
Services
1
Productivity (PT/US)
0.8
0.6
0.4
0.2
0
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
Figure 13
Counterfactuals on Sectoral Productivity
1
Data
0.9 Ag. Productivity
Svc. Productivity
Svc. & Ag. Productivity
Aggregate Productivity (PT/US)
0.8
0.7
0.6
0.5
0.4
0.3
0.2
1955 1960 1965 1970 1975 1980 1985 1990 1995
Years
These counterfactuals refer to changes in sectoral productivity from 1956 to 1995.

Ag. Productivity refers to the counterfactual situation where the relative productivity in
agriculture changes from 0.21 to 0.97 Svc. Productivity refers to the situation in which
relative productivity in services changes from 0.22 to 0.89. Ag. And Svc Productivity refers
to the counterfactual situation in which both relative sectoral productivity change as
describe above.
TAXES AND LABOR SUPPLY: PORTUGAL, EUROPE,
AND THE UNITED STATES
André C. Silva∗
Faculdade de Economia, Universidade Nova de Lisboa.
November 2005
Abstract
I relate the hours worked with the taxes on consumption and labor. I
propose a model and compare its predictions for Portugal, France, Spain, United
Kingdom and United States. Hours per worker in Portugal decreased from 35.1 in
1986 to 32.6 in 2001. With only the parameters and the taxes for Portugal, the
model predicts the hours worked in 2001 with an error of only 12 minutes from the
actual hours. Across countries, most predictions differ from the data by one hour or
less. The model is able to explain the trend in hours with only the changes in taxes.
Keywords: labor supply, consumption tax, labor income tax.
∗
Faculdade de Economia, Universidade Nova de Lisboa. Campus de Campolide, 1099-032 Lisboa, Portugal.
Email: acsilva@fe.unl.pt. Tel: 351-21-380-1600. I thank Pedro Pita Barros, Susana Peralta and an
anonymous referee for useful comments. All remaining errors are my own.
42 André C. Silva
1. Introduction
You may cook your dinner when arriving at home or you may heat prepared
food. Heating prepared food requires less time. Regardless of your decision, when the
price of prepared food increases it is more likely that you choose to cook your dinner. To
do this, you will stop working earlier to arrive earlier at home or you will decrease your
leisure time. Probably you will combine the two. Unless you decide to cut entirely from
your leisure time, you will decrease your hours of work in the market.
Taxes influence our behavior because they change relative prices, that is,
benefits and costs of different alternatives. The example above runs for a tax on
consumption. We have a similar effect of a tax on labor income. We care for the goods
that we can produce with our time. When the labor tax increases, we can buy less goods
with the same amount of labor in the market. Therefore, there are more incentives to use
our time in nonmarket activities as doing more domestic tasks, or increasing the leisure
time.
Hours worked per worker in Portugal decreased from 35.1 hours per week to
32.6 hours per week from 1986 to 2001. The data are from the Organisation for Economic
Co-operation and Development (OECD). Consumption and labor taxes increased during
the same period. If we adjust a model to the average levels of hours and taxes for Portugal
during 1986 to 1989 and let taxes vary but fix all other parameters, we can predict the
number of hours worked and compare the predictions with their actual values. This
exercise yields a prediction of 32.4 hours worked in 2001. According to these results,
with only the change in taxes and abstracting from any other change in Portugal during
these 15 years, we can predict the number of hours worked with an error of a little more
than 10 minutes, or less than 1 percent from the true value! The same model predicts a
decrease in the number of hours to 31.9 hours per week if there is a further increase of
one percentage point of the taxes on consumption and labor.
The model does not make as precise predictions for all periods. It follows the
general trend up or down of hours worked, but it misses, for example, a peak of hours
worked in 1995. Nevertheless, the model is surprisingly precise to follow the general
trend. Figure (1) shows the data and the predictions of the model in each year from 1986
Taxes and Labor Supply: Portugal, Europe, and the United States 43
to 2001. The results in this figure are for the parameters of the model calculated for
Portugal. We will also calculate the parameters taking into account other countries and
compare the new predictions with the data.
Figure 1
Portugal
36.0
35.5
Hours per week per worker
35.0
34.5
34.0
33.5
33.0
32.5
32.0
31.5
1986 1988 1990 1992 1994 1996 1998 2000 2002
Actual Hours Predicted Hours
Predicted hours obtained from the model developed in sections 2 and 3.

Source of actual hours: OECD.
This paper is about the effects of consumption and labor taxes on labor supply. I
will make the model and the definitions more precise in the following sections. I will
focus on Portugal but I will also compare the predictions and the results with other
countries. To do this we need to use data the most comparable as possible. For this
reason, all data discussed here are from the OECD publications Revenue Statistics and
National Accounts and the OECD labor market statistics. I expect that using the same
source of data we can minimize the chance of committing errors as trying to explain
something that is only an effect of different measurement standards. The appendix
contains more details about the data sources.
Notice the implications of a decrease in hours worked in the market caused by an
increase in taxes. One effect is the decrease in total production. With less products we
44 André C. Silva
also have less consumption and investment and less welfare. But it can be the case that
the total number of hours stays constant but simply flows from market to nonmarket
activities. Even in this case, we can still have welfare consequences.
When I say hours worked in the market I refer to hours used in the legal market
for production. Hours not used in the market can be used for leisure or for nonmarket
work. Leisure refers to activities that do not increase the quantity of goods and services
available. Nonmarket work refers to domestic labor, as cooking your own dinner or
teaching your children, or to work in the underground economy.
We use markets to specialize in some activities and increase our ability to
produce more goods. I can produce an apple planting an apple tree in my backyard or
buying it in a supermarket with my labor income. I prefer to use markets, that is, to use
the supermarket, because I know that I will use my time better in this way.
Therefore, even if the total quantity of time devoted to work does not change
when hours worked in the market decrease, we would probably be more efficient using
the time in market activities. Moreover, the decrease in market work can also correspond
to an increase in the underground economy. As taxes increase, the incentives for
nonmarket work increase and so the economy works less efficiently.
I use the annual hours of work from the OECD and divide by 52 to obtain the
weekly hours of work. I focus on the decision of hours worked and not worked in the
market. If the workers in a country have less vacation for the same number of hours per
working week then their annual hours of work in the market is higher. Consequently, their
average number of hours per week is also higher.
The work is inspired by Prescott (2004) “Why do Americans work so much
more than Europeans”. He argues that the main factor able to explain the difference of
income per capita in the United States and in the group of European countries Germany,
France, Italy and the United Kingdom is the number of hours worked per person. The
number of hours worked, in turn, can be almost fully explained by the higher taxes in the
group of European countries.
Here I focus on the effects of taxation on the labor supply for different years. I
investigate whether only taxes are able to explain the behavior of the number of hours
worked in each year, and whether it can explain the difference in hours between
countries. In addition to Portugal, I focus on the countries France, Spain, United Kingdom
and United States.
I find that the increase in taxes during 1986 to 2001 is very good in predicting
the behavior of labor supply in Portugal. Differences in taxes alone are not as precise in
predicting differences in labor supply across Portugal and the other countries, but they are
still able to explain a great part of the difference.
In section 2 below I discuss the effects of taxation on labor supply with an
economic model. We then have to calculate the labor and consumption taxes for Portugal
and other economies to compare the predictions of the model with the actual values of
labor supply. There are several taxes in each economy with incidence on consumption
and labor. We have to find a way to aggregate all these taxes in one number for
consumption and another one for labor. This is done in section 3. In section 4, I use the
calculated taxes to discuss the implications of the model, and discuss additional evidence
on the effect of taxes on labor obtained in other settings.
2. Taxes and the decision of consumption and labor
What to expect from an increase in taxes regarding the decision of consumption

and labor? We need to build a model with the following ingredients: a decision of how
much to consume and work, taxes on consumption and labor, a government sector, and a
production sector.
The elements of the following model are standard, it is the same model used by
Prescott (2004). Its advantage is to model explicitly the decisions of consumers and firms,
and the possibility to confront the predictions with data. Different versions of the model
have been used in several applications. They can be found, for example, in the papers in
Cooley (1995).
I treat all consumers in the economy as if they were represented as an average
consumer or a representative consumer. When I mention the consumer, I am in fact
making a reference to this representative consumer.
The only taxes that affect the decision of consumption and labor in each period
are the consumption and the labor taxes. We can have other taxes in the economy as taxes
on capital gains or taxes on investment but only consumption and labor taxes are
46 André C. Silva
important for our purposes. It is interesting to look first to the budget constraint of an
average household to understand this behavior.
The budget constraint defines the resources available and their possible uses. Let
us talk first about the resources. If the consumer works the number of hours ht at time t
for a wage wt then the labor income will be equal to wt ht . If there is a labor tax h then
the labor income after taxes, available for consumption, is equal to (1 − τ h ) wt ht . The time
t subscript stresses that the consumer must make a decision for each period. If the same
consumer has a certain quantity of capital kt and the interest rate during the period is rt
then the capital income will be equal to rt kt . If there is a tax on capital τ k then this
income will be subtracted by τ k ( rt − δ ) kt , where δ stands for the depreciation of capital
caused by its use. The government may also redistribute part of the taxes back to
consumers in the form of transfers Tt , which are also part of the resources.
The possible uses of the resources are consumption, ct , and investment, xt .

Consumption stands for consumption of goods and services. If the tax on consumption is
c then the price of a unit of consumption must be multiplied by (1 + τ c ) . Similarly, if
the tax on investment is x then the price of a unit of investment must be multiplied by
(1 + τ x ) .
Therefore, a consumer in this economy faces the budget constraint
(1 + τ c ) ct + (1 + τ x ) xt = (1 − τ h ) wt ht + rt kt − τ k ( rt − δ ) kt + Tt . (1)
The left-hand side shows the possible uses for the resources: consumption and
investment. The right-hand side shows the different resources: labor income, capital gains
and transfers. The consumer has to decide the levels of ct , ht and xt such that the
constraint above is satisfied. The important variables for our purpose are ct and ht .
Everything is in terms of the price of the consumption good. Therefore, if the

consumer works one hour, it will be possible to consume (1 − τ h ) wt units of the
consumption good.
The tax systems in the world are much more complicated than what we have in
the budget constraint above. There are tax brackets, exemptions, subsidies for some
investments and so on. But considering these complications will not change the basic
point of the analysis. In section 3, I map the tax revenues and the tax bases to the relevant
taxes in the model.
The quantity of capital and the quantity of investment are related. The quantity
of capital tomorrow is equal to the quantity of capital today minus the depreciation from
the use of capital plus any investment made today. In symbols, write
kt +1 = (1 − δ ) kt + xt . (2)
What is invested today will be available tomorrow in the form of capital and will
generate an additional form of income tomorrow. Therefore, investment is only a way to
connect consumption today with consumption tomorrow. As consumption and investment
are connected, apart from the taxes τ c and τ x , the price of a unit of investment is equal to
the price of a unit of consumption. It is only a decision of consuming today or tomorrow.
The decision of working and consuming, on the other hand, affect the consumer
in the same period. As capital will sooner or later be transformed in consumption, we can
consider only the quantity of consumption and labor as the variables of interest in each
period.
It is reasonable to admit that any additional amount of consumption makes the

consumer happier. It is also reasonable to say that the quantity of time available for the
consumer is fixed, and that any increase in the time available after working also makes
the consumer happier. The time not working can be used to a nonmarket activity, as the
production of goods to be consumed at home, or to be enjoyed as leisure. The number of
hours ht refer to the quantity of work in the market. The model, therefore, considers the
decision of consumption, labor and savings of an employed worker. It does not consider
the decision of entering or not in the labor force as a job searcher.
If the quantity of hours worked in period t increases one unit then the quantity
of leisure decreases. If it were for only this effect the consumer would be worse off. But,
from the budget constraint, the amount of resources will increase by (1 − τ h ) wt and this
will allow the consumer to consume more. If this additional resource is used for
48 André C. Silva
consumption, the consumer will be able to consume

(1 − τ h ) w units of the consumption
(1 + τ c ) t
good.
The ratio
(1 − τ h ) w is key for this analysis. It relates how many units of the
(1 + τ c ) t
consumption good we are able to consume when we give up one unit of leisure. If taxes
on labor or on consumption increase then the number of consumption units decrease.
There are two important aspects to pay attention. The first is that taxes on capital income
do not appear in this ratio. The second is that taxes on labor and on consumption have
similar effects: a unit more of work will bring less goods if the labor tax is higher or if the
consumption tax is higher. In fact, for small values of τ h and τ c , the number
(1 − τ h ) / (1 + τ c ) can be approximated by 1 − τ h − τ c .
We need to be more explicit in the way the consumer weights the benefits of
consumption and labor in order to make predictions as the ones contained in Figure (1).
Suppose that consumption and labor in each period yield a level of welfare given
by
log ct + α log (100 − ht ) . (3)
The number 100 comes from the assumption that the consumer has 100 hours available
in each week. The number of working hours have to be taken from these 100 hours. The
parameter α is a positive constant. It will be used later to calibrate the model to the
actual economies. The logarithmic function log X is increasing in X . Therefore, as
consumption or leisure hours increase, the welfare in each period increases.
Our consumer evaluates consumption and leisure for all periods. But recent
periods are valued more than later periods. That is, the period t is discounted by the
value β t , where β is greater than zero but less than one. The welfare of all periods is
given by the sum of the terms β t × ( log ct + α log (100 − ht ) ) for t = 0 , 1 , 2 etc. As the
decision of consumption and labor depends on the values within a certain period, the
parameter β will not play an important role for our purposes
To close the model we need to define how products are produced and the role of
government in this economy.
Products are produced by firms. They combine capital and labor to produce
goods. Capital corresponds to the machines, tools and instruments used to manufacture
the products. Capital and labor must be hired in the market, their prices are given
respectively by rt and wt . Let yt denote the value of production. Capital, labor and
production are related by the production function
yt = At ktθ ht1−θ . (4)
The constant θ is positive and between zero and one. So, when capital or labor
increase, production increases. θ is equal to the fraction of capital income in the total
income. Correspondingly, 1 − θ is equal to the fraction of labor income in the total
income. This parameter does not vary much from country to country and during different
time periods. The parameter At is called total factor productivity. According to this
model, if two countries use the same quantity of labor and capital but produce different
quantities of the consumption good it is because At is different between them. But At
will not affect the quantity of hours worked and will not be important for our purposes.
Profits are given by the difference between revenue and costs. Profits are
therefore given by
At ktθ ht1−θ − wt ht − rt kt . (5)
A useful result is that, in order to maximize profits, each firm will set the quantity of
capital and labor so that wages and the production function are related by
yt
wt = (1 − θ ) . (6)
ht
There is also another condition for the interest rate. As interest rates do not enter in the
decision of consumption and labor within periods, we will not use this additional
condition here.
50 André C. Silva
Total production is used for consumption and investment. The government also
uses part of the production to provide goods for the citizens. Let gt denote the quantity of
goods consumed by the government and left to the consumer. Some examples of this kind
of good are public schooling, police service and the maintenance of public parks. Total
consumption in this economy is given by ct + gt . As consumption and investment must
come from production, we have
ct + xt + gt = yt . (7)
The government collects taxes to pay for government consumption. Any

difference between the government's revenue and the government consumption is given
back to consumers in the form of transfers Tt . Therefore, the government budget
constraint is given by gt + Tt = revenues from taxes.
Solving the model
The variable that we do not know and need to find is the number of hours
worked in each period, ht . Solving the model means writing ht as a function of the tax
rates and of the parameters θ and α . We do not know ct , xt and kt either and a full
solution also involves obtaining these variables. Here we are concerned with the behavior
of hours worked and so we will concentrate on ht .
The value of ht depends on how consumers value consumption and leisure, and
on how firms value the use of labor and capital.
From the standpoint of consumers, when they increase one unit of labor they
α
have a decrease in welfare of as implied by how they value consumption and
100 − ht 1−τ h
leisure in (3). On the other hand, this unit of labor allows them to buy wt units of
1 1−τ h 1+τc
the consumption good and yields an increase in welfare of wt . Consumers will
ct 1 + τ c
increase the quantity of labor until the benefits are greater than the loss in welfare.
Therefore, they will choose consumption and labor so that
α 1 1 −τ h
= wt . (8)
100 − ht ct 1 + τ c
Firms choose capital and labor so that wt = (1 − θ ) yt / ht . With this, we have
α 1 1 −τ h y
= (1 − θ ) t ,
100 − ht ct 1 + τ c ht
or
1−θ
ht = 100 × , (9)
c 1
1−θ + α t
yt 1 − τ
1 −τ h
where 1 − τ = .
1+τc
Equation (9) is the key to understand how taxes influence labor supply. If taxes
on labor or on consumption increase, then τ increases, the denominator increases and so
the number of hours worked, ht , decreases. The ratio of consumption to production,
ct / yt , depends on the expectations about the future. For example, if taxes are higher in
the future relative to today then it is better to work more now, save the additional
proceeds, and use the savings in the future. The additional savings will decrease the ratio
of consumption to production.
Figure (2) shows data on hours per worker for Portugal and for the United States
from 1986 to 2001 and the values of τ for the two countries in the same period. We see
that hours per worker decreased more in Portugal and that taxes increased more in
Portugal. This observation agrees with the model. To be more precise, we have to
substitute the values of τ and ct / yt in the formula above and compare the predictions of
the model and the actual values of hours per worker. This is the subject of the next
sections. First, we have to calculate the tax rates to be used in equation (9).
52 André C. Silva
Figure 2
Hours Tax Rates
36.0 46%
35.5 44%
35.0 42%
34.5 40%
34.0 38%
33.5 36%
33.0 34%
32.5 32%
32.0 30%
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 1986 1988 1990 1992 1994 1996 1998 2000 2002
Portugal United States Portugal United States
Source for Hours: OECD. Tax rates: calculations by the author following section 3.
3. Taxes on labor income and consumption
We have to calculate the value of taxes over the years. Consider the tax on
consumption for example. We have different taxes with incidence over the various goods
and services traded in the economy. Usually taxes are small for basic food and high for
luxury goods. In the model, the unique good represents all kinds of goods and services.
Therefore, the tax on consumption τ c must represent the tax on all goods and services.
The same idea applies to the labor tax. The income tax systems are progressive
over income and allow for exemptions and deductions. We also have social security taxes
paid by workers. Moreover, we have to consider the difference between marginal tax
rates and average tax rates. Consumers decide over consumption or labor according to the
tax rate paid over one additional unit of labor. This is the marginal tax rate on labor.
Hence, consumers make decisions according to the marginal tax rate, and not according
to the average tax rate.
The calculations follow the same principle: dividing the revenues from the tax
by the tax base. We have data of tax revenue detailed by the origin for all OECD
countries from the OECD database. We also have data for National Accounts aggregates
as consumption and GDP from the same database. Using this data and the procedure
described in detail in the next paragraphs, I obtained the taxes in figures (3) and (4).
Figure 3
Taxes on Consumption
30%
25%
20%
15%
10%
5%
0%
1970 1975 1980 1985 1990 1995 2000
Portugal France Spain

United Kingdom United States
Source: author’s calculations based on OECD data.
Figure 4
Marginal Taxes on Labor
45%
40%
35%
30%
25%
20%
15%
10%
1970 1975 1980 1985 1990 1995 2000

Source: author’s calculations based on OECD data.

54 André C. Silva
The idea of using National Accounts aggregates to calculate effective tax rates
has its origins in Lucas (1990) and was extended by Mendoza, Razin and Tesar (1994).
Prescott (2002) and Prescott (2004) also used this method and I am following closer his
calculations. There are some differences in my calculations in order to yield longer series
and being the most comparable as possible. Calculating directly using Mendoza-Razin-
Tesar's method yields slightly different values. But the differences do not modify the
conclusions as I calculate the taxes for each country using the same procedure.
The sources of data are from the OECD publications National Accounts,
Detailed Tables: vol II and Revenue Statistics of OECD Member Countries. The codes of
the items used in the calculations are in the appendix.
For the tax on consumption τ c I divide the revenues from indirect taxation by
household final consumption expenditure minus indirect taxation. The revenues
considered for the consumption tax come from general taxes from goods and services and
from excise taxes. Final consumption expenditures takes into account indirect taxation in
the National Accounts. Therefore, we have to subtract these taxes to find the relevant tax
base.
As Mendoza-Razin-Tesar note, part of the indirect taxation comes from taxes
paid by the government. This requires summing the tax base by the difference between
government consumption and wages paid to government employees. As the denominator
is bigger with the government sector, the tax on consumption calculated in this way is
smaller. However, I compared the calculations with the two methods and the two series
are very close, moving almost in parallel. For Portugal, the OECD database would allow
us to calculate consumption taxes only for the period 1995-2001. The difference between
the calculations with or without the government sector are about five percentage points
for this period. The advantage of using the simpler formulation is that it is possible to
construct a longer series.
For the tax on labor income τ h we have to calculate the social security tax and a
measure for the marginal income tax not including the social security tax.
The social security tax is obtained by dividing the social security contributions
by the labor income. From equation (6) of the model in section 2, the share of labor
income over total production is equal to 1 − θ . Therefore, I use (1 − θ ) times GDP less
indirect taxes as an estimate of the labor income. We have to subtract indirect taxes from
GDP because the National Accounts figures include the payment of taxes. I set
1 − θ = 0.7 . This share is approximately constant over countries and time and the number
used is conventional in models of this type. Gollin (2002) obtains further evidence for this
number and for the stability of the labor share across countries. The difficulty for the
calculation of the labor shares results from the estimation of the labor income of the self-
employed. Gollin (2002) finds that the labor shares are approximately constant, in the
range of 0.65 and 0.80 , once the labor income of the self-employed is treated carefully.
The conclusions of the present analysis do not change if we consider other values for this
parameter in this range.
The average income tax is calculated by dividing the revenue from taxes of
individuals by GDP less indirect taxes less the depreciation of capital. Consumers care for
the tax payments of an additional unit of labor. This marginal tax rate is higher than the
average tax rate because tax rates are usually higher when incomes are higher. To obtain
the marginal tax rate, I follow Prescott (2002, 2004) and multiply the average tax rate by
the factor 1.6 . The labor income tax is then obtained as the sum of the social security tax
and the marginal labor income tax.
4. Data and predictions
We now turn to the predictions of the model. I repeat the key equation for the
predictions of the model for convenience
1−θ
ht = 100 × , (9)
c 1
1−θ + α t
yt 1 − τ
1 −τ h
where 1 − τ = . The taxes τ h and τ c are the labor income and consumption taxes.
1+τc
The objective of this section is to use the same model for different countries and
verify if the difference in the tax rates is able to explain most of the difference in hours
worked. In addition to Portugal the countries chosen for this study were France, Spain,
United Kingdom and United States. I first focus on four periods: 1970-1974, 1983-1986,
56 André C. Silva
1993-1996 and 2000-2002. The first period is the first that we have data available for
France, United Kingdom, and United States. The first observation available for Portugal
is in 19861, therefore I included the period 1983-1986. The period 2000-2002 is the last
period that we have data available. Prescott (2004) focuses on the periods 1970-1974 and
1993-1996.
Several other important features may affect the number of hours worked and
labor market statistics. For example, Bover, Garcia-Perea and Portugal (2000) and
Blanchard and Portugal (2001) analyze how institutional structures such as
unemployment benefits, unions and other factors can affect unemployment and
unemployment duration. Here I focus on taxes and the differences in hours worked.
I used two estimates for ct in equation (9). In the first case I considered only
private consumption net of indirect taxes. In the second I also considered government
consumption. So, ct = Ct − Indirect Taxes in the first case and ct = Ct + Gt − Indirect
Taxes in the second case. Ct and Gt are the figures of private and government
consumption in the National Accounts data. The value of yt is the same in the two cases:
yt = GDP − Indirect Taxes.
The justification for considering government consumption as part of ct is that
this value may substitute part of private consumption. However, I do not have a clear
estimate of the degree of substitution between private and government consumption.
Therefore, I calculated the predictions for the two extreme cases, with zero or one-to-one
substitution, and compared with the actual values.
The parameter α sets the weights between consumption and leisure. It should
have the same value for all countries to assess how much of the difference in hours is
explained by the difference in taxes only. This parameter is obtained so that the difference
between the actual and the predicted values of hours worked is minimized for the period
1970-1975 for France, United Kingdom and United States. It yields α = 1.39 and
α = 1.02 when we consider respectively private consumption only and private and
government consumption. The predictions of figure (1) were obtained for a value of α
set to minimize the difference during the period 1986-1989 for Portugal, with only private
1
Labor income taxes are available since 1989 for Portugal. I maintained the labor income tax rate during
1986-1988 equal to its value in 1989 to calculate τ from 1986 to 1988.
consumption. This yields α = 1.47 and is used to assess how much of the change in
hours during the period 1986-2001 may be accounted for changes in the tax rates for
Portugal.
Therefore we estimate the parameter α in two ways. In the first way we focus
on Portugal and obtain the first predictions in figure (1) as a starting point. In the second
way we do not use the data for Portugal and use other countries and another period. In
this second way the parameter α is set independently from the data for Portugal. Any
difference in the predictions of hours worked across countries and through time is
because of the difference in taxes and the difference in the ratio of consumption to GDP. I
now focus on this second estimation of α .
Table 1 compares the actual hours and the predicted hours using the model for
the countries studied. Each period has the average of the actual and the predicted values.
Each column for a prediction fixes the same value of α for all countries. Therefore, I fix
tastes on consumption and leisure and all kinds of institutional arrangements across
countries. I allow only taxes to vary and the corresponding equilibrium value ct / yt .
Table 1. Actual and Predicted Hours

Actual Pred 1 Pred 2 Actual Pred 1 Pred 2 Actual Pred 1 Pred 2
1970-74 - - - 35.9 37.2 36.4 - - -
1983-86* 35.1 36.2 37.7 31.6 31.9 29.6 36.0 35.3 36.8
1993-96 33.9 33.6 33.5 30.2 32.6 29.9 34.9 33.9 34.0
2000-02** 32.6 33.5 32.4 28.3 32.4 29.8 34.9 34.6 34.7
Actual Pred 1 Pred 2 Actual Pred 1 Pred 2
1970-74 36.6 34.8 34.8 36.3 37.0 38.1
1983-86 33.5 34.4 33.2 35.1 35.2 36.5
1993-96 33.3 33.7 33.4 35.2 33.8 35.9
2000-02 32.8 32.3 32.3 34.8 32.1 34.5
Hours per week per worker. Pred 1: predictions with private consumption. Pred 2: predictions with private and
government consumption. * 1986 only for Portugal. ** 2000-01 only for Portugal.
Source of actual hours: OECD.
Although the model abstracts from several potentially important features of the
labor market, it is able to generate very reasonable estimates based only on the tax rates.
Most of the numbers are around one hour of error from the actual value of hours worked.
58 André C. Silva
Only five predictions misses the actual numbers for more than two hours.
According to table 1, the predictions for hours worked for Portugal are above the
actual values for 1986 and are much closer for 1993-96 and 2000-01. For Spain, in only
one period the model predicts the hours with more than one hour of error. For the United
Kingdom it misses the period 1970-74 but it is good in predicting the other periods. For
France, the model predicts more hours of work than the actual values specially for the
period 2000-2002. It can be an effect of the new legislation on the reduction of the
working week in France, and it can have a greater role than the change in taxes for this
period.
Therefore, the model is good in predicting the general pattern of hours worked
even though it misses some observations. As everything is fixed across countries and only
taxes are allowed to vary, this shows that taxes have an important effect in the number of
hours worked. The good fit is more surprising for Portugal and Spain as I have not used
these countries to estimate the value of α , and we could believe a priori that they are
much different from the other countries apart from the difference in taxes. See figures (5)-
(8) for the comparison of actual and predicted hours year by year.
Figure 5
Portugal France
39 39
38 37
37
35
36
33
35
31
34
33 29
32 27
1986 1988 1990 1992 1994 1996 1998 2000 2002 1970 1975 1980 1985 1990 1995 2000
Source of actual hours: OECD. Pred 1: predictions using only private consumption net of indirect taxes. Pred 2:
predictions with private and government consumption net of indirect taxes
Figure 6
Spain United Kingdom
41 38
40
37

39
36
38
37 35
36 34
35
33
34
32
33
32 31
1975 1980 1985 1990 1995 2000 1970 1975 1980 1985 1990 1995 2000
Source of actual hours: OECD. Pred 1: predictions using only private consumption net of indirect taxes.
Pred 2: predictions with private and government consumption net of indirect taxes.
Figure 7
United States
39
38
37
36
35
34
33
32
31
1970 1975 1980 1985 1990 1995 2000
Actual Pred 1 Pred 2
Source of actual hours: OECD. Pred 1: predictions using only private

consumption net of indirect taxes. Pred 2: predictions with private and
government consumption net of indirect taxes.
I am using hours per worker as a benchmark to compare the predictions of the

model. We can also think about using hours per capita, or hours per person in the active
population. In the model, the worker decides how much to work in one job, and therefore
hours per worker seem appropriate to use as the benchmark. It states how many hours
someone will work, given the fact that the decision of entering the labor market has
already been made. If the consumer is an average worker, taking into account those that
60 André C. Silva
are in the labor force and those there are not, hours per capita may also seem appropriate.
Hours per working age person in Portugal are on average 2.4 hours lower than in
the United States during 1986 to 2001. During the same period, however, hours per
person were more or less constant as the number of people in the labor force increased.
Prescott (2004) uses hours per working age person (by convention, people aged 15 to 64)
and he also finds an important effect of taxes on the supply of labor. He notes, however,
that hours per capita increased in the United States during 1970-1990 although taxes
increased or remained approximately constant. He then notices that households switching
from one wage earner to two wage earners may explain this behavior. In my analysis, I
abstract from this possibility.
Davis and Henrekson (2004) show further evidence on the impact of
consumption and labor income taxes on labor supply. They use an empirical analysis to
study the relation between taxes and labor supply. They find that a 12.8 percentage points
increase in the tax rate implies 122 market work hours per year per adult, a decrease of
4.9 percentage points in the employment-population ratio, and an increase in the shadow
economy equal to 3.8 percent of GDP. It also decreases employment in retail trade and
repairs, and in eating, drinking and lodging. The increase in taxes affect more heavily
these industries with less-skilled labor as it is easier to substitute market work for
domestic work in these sectors.
There are also other explanations for the difference of labor supply in Europe
and in the United States. Alesina, Glaeser and Sacerdote (2005) state that policies of the
unions and labor market regulations are able to account for most of the difference in labor
supply.
5. Conclusions
Tax rates affect the decision of labor supply by changing the relative benefits
and costs of work and consumption. The effect is large. For Portugal, France, Spain,
United Kingdom and United States, taxes are able to predict the pattern of the labor
supply over the years.
I first estimate the tax rates for the countries considered during 1970 to 2001.
For all countries, taxes increased during the period and the number of hours worked
decreased. I then compare the predictions of the model on hours worked per worker with
the actual values for each country. Most of the predictions are different from the actual
values by one hour or less. The model is able to follow the general pattern of labor
supply.
Hours refer to hours working in the market. When the number of hours worked
decreases, the number of nonmarket hours increases. These hours may be leisure hours or
hours used in the production of nonmarket goods, as domestic production. Production is a
function of hours worked. With a decrease in the number of hours worked, the number of
consumption goods decreases. Therefore, when consumers decide to work less hours,
they consume less products in the market and increase their consumption of leisure or
nonmarket goods.
If we are more efficient trading goods in the market, a decrease in market hours
makes consumers less efficient. A result in Economics is that decisions are efficient when
the costs and benefits are equalized by consumers and firms. With taxes, consumers
receive a wage but can buy goods only with the after-tax wage. Therefore, firms and
consumers do not equalize the costs and benefits of the labor decision.
GDP per capita in Portugal increased from US$ 10,700 to US$ 17,100 during the
period from 1986 to 2003. These are figures in dollars of 2000 corrected for purchasing
power. Relative to the United States, GDP per capita in Portugal increased from 41 to 48
percent during this period. In order to continue growing at high growth rates we have to
look for ways to make the economy more efficient. Paying attention to the incentives that
make consumers decide on their use of time is one of them.
Appendix - Data Sources and Formulas for the Tax Rates
The sources of data for National Accounts and Revenue Statistics are from the
OECD (Organisation for Economic Co-operation and Development) publications
National Accounts, Detailed Tables: vol II and Revenue Statistics of OECD Member
Countries. I use the OECD online dataset SourceOECD. For the labor force, I use the data
available online from the OECD following the path Statistics/Data by
Topic/Labour/Labour Force Statistics - Data, “Average actual annual hours worked per
person in employment”. I divide the annual hours by 52 to write the values in weekly
62 André C. Silva
hours. For the figures of GDP per capita, I use the “Comparative tables based on
exchange rates and PPPs”.
The formulas for the tax rates on consumption and labor are
IT
τc = ,
C − IT
τ h = τ SS + 1.6τ havg .
τ c and τ h denote respectively the tax rate on consumption and the marginal tax rate on
labor income. τ SS and τ h denote respectively the social security tax and the average tax
rate on labor income. C denotes consumption expenditures and IT denotes indirect
taxes on consumption. In the National Accounts, the final price of the products are used
to calculate aggregate consumption expenditures C . Therefore C takes into account the
indirect taxation. Therefore, we have to subtract indirect taxation from consumption
expenditures to find the relevant tax base.
The social security tax and the average labor income tax are obtained by
Social Security Contributions

τ SS = ,
(1 − θ )( GDP − IT )
Direct Taxes of Individuals

τ havg = .
GDP − IT − Depreciation
The parameter θ is the fraction of capital income in the total income. Hence, 1 − θ is
equal to the fraction of labor income in the total income. In the Revenue Statistics, social
security contributions are given by the item 2000, IT is given by the sum of the items
5110 and 5121 and Direct taxes of individuals is given by the item 1100. We have to
subtract indirect taxes from GDP because the value of the GDP in the National Accounts
includes indirect taxes. See section 3 for the explanation of the formulas.
References
Alesina, Alberto, Edward Glaeser and Bruno Sacerdote (2005). “Work and Leisure in the
U.S. and Europe: Why So Different?” NBER Working Paper 11278.
Blanchard, Olivier and Pedro Portugal (2001). “What Hides Behind an Unemployment
Rate: Comparing Portuguese and U.S. Labor Markets.” American Economic
Review, 91(1): 187-207.
Bover, Olympia, Pilar Garcia-Perea and Pedro Portugal (2000). “Labour Market Outliers:
Lessons from Portugal and Spain.” Economic Policy, 31: 379-428.
Cooley, Thomas F., ed. (1995). Frontiers of Business Cycle Research. Princeton:
Princeton University Press.
Davis, Steven J. and Magnus Henrekson (2004). “Tax Effects on Work Activity, Industry
Mix and Shadow Economy Size: Evidence from Rich-Country Comparisons.”
NBER Working Paper 10509.
Gollin, Douglas (2002). “Getting Income Shares Right.” Journal of Political Economy,
110(21): 458-474.
Lucas, Robert E., Jr. (1990). “Supply-Side Economics: an Analytical Review.” Oxford
Economic Papers, 42: 293-316.
Mendoza, Enrique G., Assaf Razin, and Linda L. Tesar (1994). “Effective Tax Rates in
Macroeconomics: Cross-Country Estimates of Tax Rates on Factor Incomes and
Consumption.” Journal of Monetary Economics, 34(3): 297-323.
Prescott, Edward C. (2002). “Prosperity and Depression.” American Economic Review,
92(2): 1-15.
Prescott, Edward C. (2004). “Why Do Americans Work So Much More Than
Europeans?” NBER Working Paper 10316.
BUDGET SETTING AUTONOMY AND POLITICAL ACCOUNTABILITY∗
Susana Peralta
FEUNL and CORE-UCL ∗∗
Abstract
The autonomy of local governments in deciding their revenue level varies a
lot worldwide, and is very low in Portugal. We analyze the consequences of this
autonomy from the viewpoint of political accountability. We study a two-period game
in which elections take place at the end of the first period, in a model where local
officials may be public or self interested. We show that a greater autonomy improves
selection (i.e., voting out bad incumbents), while it decreases discipline (i.e., giving
incentives to the bad incumbent). Electoral turnover is expected to be higher with
greater autonomy. We analyze the effect of tax setting autonomy on expected voter
welfare.
∗
This paper was written for presentation at the conference “Desenvolvimento Económico Português no Espaço
Europeu”, to be organized by the Portuguese Central Bank in February 2006. I am grateful to the Comité
Científico for useful comments and to Banco de Portugal for financial support. I thank Isabel Dias at the
Comissão Nacional de Eleições for providing me with data on Portuguese local elections, Fátima Teixeira at
the Direcção Geral das Autarquias Locais for help with municipal finance data and Sofia Fernandes for
research assistance.
∗∗
Campus de Campolide P-1099-032 Lisboa. Phone: +351 21 3801642. Email: peralta@fe.unl.pt
66 Susana Peralta
1. Introduction
Fiscal decentralization is a policy objective advocated by international

organizations such as the World Bank (World Bank, 2000) and the OECD (OECD, 2001,
2002). Moreover, it has become a dominant trend in several countries (Epple and Nechyba,
2004). However, the degree to which it is implemented, both on its expenditure and tax
collection aspects, varies a lot. Different institutional arrangements on the sharing of
competencies between central (or federal) and local (or state) governments exist (see, for
instance, Ter-Minassian, 1997, and OECD, 1999).
A recent study by the OECD (OECD, 1999) has looked in detail at the tax setting
autonomy of local governments in 19 countries. It focuses both on taxes collected at the
local level and tax-sharing arrangements. These latter refer to taxes collected by the central
government, whereof part of what is collected by (or arising from activities located in) a
local government's territory accrues automatically and unconditionally to it. Taxes and tax-
sharing revenues are classified into eight categories of decreasing autonomy. The largest
autonomy occurs when local governments are free to chose both tax rates and bases, while
the lowest level refers to centrally set tax rate and base. Table 1 summarizes the results of
the study. For each level of local government in each country, the percentage of revenues
falling into each category is computed. Table 1 displays, for each country, the unweighted
average across all local government levels (i.e., for Portugal, autarquias, on the one hand
and autonomous governments from Madeira and Azores, on the other).
Table 1 allows one to conclude that local governments are far from having full
control of their fiscal revenue. It is noteworthy that transfers from the central government
and supra-national institutions like the European Union, virtually totally out of local
governments's control, are outside the scope of the OECD study. Also, in some countries
the central government decides on an allowable range for the tax rates on revenue sources
falling into categories (a) and (b). Hence, if anything, the figures in Table 1 are too
optimistic about local governments' autonomy. Table 1 also makes clear that there is
considerable variation among countries. In this study, Portugal shows up as one of the
countries where local governments enjoy less revenue setting autonomy.
Budget Setting Autonomy and Political Accountability 67
The question of revenue setting autonomy falls into the more general debate on
whether government should be decentralized or not. The literature on decentralization is
abundant, but it usually treats public good provision and financing together, i.e., either both
are left to the central government or both are decentralized. One notable exception are
papers studying transfers from the central to local governments. They build a case for local
governments' lack of revenue-collection autonomy both on efficiency (avoiding inter-
jurisdictional externalities) and equity (avoiding fiscal imbalances) grounds (see, e.g.,
Bucovetsky and Smart (2004) for a recent contribution where both objectives are shown
not to be incompatible).
This paper seeks to look at revenue autonomy from an alternative viewpoint,
namely, political accountability. One may think that giving more autonomy to local
68 Susana Peralta
governments increases the opportunities for corrupt local officials to extract rents for
private purposes. As an illustration, one may read in the web-based edition of “Inside
Indonesia” that “corruption in the provinces stems in part from local politicians’ access to
the budget.”1 The fact that local governments may be more prone to corruption is the basis
of the analysis of decentralization by Bardhan and Mookherjee (2000), who also recognize
it to be a prominent issue in the political debate about decentralization, dating back to the
foundation of the US (see Bardhan and Mookherjee (2000) and the reference to the
Founding Fathers therein). The potential danger of decentralization due to the poor quality
of local politicians is also recognized in the political science literature (see, for instance,
Bird, 2000).
On the one hand, if one accepts the idea that local governments are more prone to
corruption than the central one, revenue-collection centralization can be seen as a way to
fight corruption, in that it hampers the capacity of local officials to extract rents. On the
other hand, if the central government sets the local budget, we may reasonably suppose that
voters observe it only imperfectly. In fact, fiscal laws are often complex and confusing
(Bird, 2000), and it may be in the interest of politicians to hyde information from voters
and render the budget unclear (this is suggested by Besley (2005, page 193), in the context
of the use of debt). When the voters actually pay for the public good with local taxes, their
awareness of its actual cost increases. For instance, in a World Bank report about
intergovernmental fiscal relations in the Czech Republic (World Bank, 2003), one may read
“(...) local authorities have little autonomy over revenue. (...) predictable and transparent
preparation of local government budgets has been limited by lack of synchronization with
the central government budget, uncertainty about basic budget parameters, and insufficient
information on central government guarantees and contingent liabilities. (...) Budgets
would be more predictable if basic parameters (...) were defined in organic laws rather
than the annual budget law.” This lack of transparency is also a recurrent issue in the
Portuguese political debate. This is clear in the following quotation, taken from the
Portuguese newspaper O Publico (November 11, 2005, page 10), citing the intervention of
the Minister of Finance in a parliamentary debate “everyone who has assumed
governmental positions or has had any contact with local governments knows that, besides
1
http://www.insideindonesia.org/edit80/p9-10steele.html
the transfers predicted by the Law of Local Finances, there are numerous additional
mechanisms to transfer funds for local governments, (...) transfers that are outside of the
control of the Minister of Finance and are performed without any transparency. (...) The
proposal in this year's budget aims at imposing a maximum value and increasing the
transparency of such transfers”. This lack of transparency at the local level is likely to have
an impact on voters' capability to discipline local politicians. The idea that financing
expenditures by raising revenues locally leads to greater accountability is explicitly put
forward by the World Bank (World Bank, 2003).2
What is the impact on corruption and accountability when budgets are under the
full control of local governments? Such is the object of this paper. We build on the model
in Besley and Smart (2003). A public good is provided at the local level. The provision cost
may be high or low. There are two types of politicians, public-interested and self-interested,
unobservable to voters. If revenue is collected at the local level, voters observe both the
public good and the revenue level. If it is collected centrally, the central government knows
the true provision cost, sets the budget accordingly, and voters cannot observe it, i.e., they
only know the public good level. The corrupt politician can misbehave in two ways, as in
Besley and Smart (2003) and Hindriks and Lockwood (2004). Firstly, he may provide no
public good and extract maximal rents, thereby revealing is type (separating equilibrium).
Secondly, when the true provision cost is low, he may pretend it is high and extract the
equivalent (smaller) rent (pooling equilibrium). In the former case, voters vote him out of
office for sure, which is not true of the latter. As usual in this type of models, there is a
trade-off between selection (the bad incumbent is revealed) and discipline (the bad type
mimics the good one). While voters want to discipline a bad politician (incentive), his good
behavior hampers the voters' capability to identify him as bad, thus they are less likely to
vote him out of office (selection).
We begin by comparing the two regimes in terms of rent extraction and obtain that
none is better under all circumstances. Rents are higher in the decentralized regime under
the separating equilibrium, but lower when the bad incumbent pools. Indeed, when the
local official has full control of the budget and decides to reveal his type, he extracts
2
One may also find explicit references to this idea in the Local and Municipal Governance and Finance program
(sub-program of the Governance and Anti-Corruption Learning Program), details on http://web.worldbank.org.
70 Susana Peralta
maximal rents. This is in line with the common wisdom that increased autonomy leads to
more rent diversion. On the other hand, if the politician decides to pretend that the
provision cost is high, he takes advantage of the lack of voters' information in the
centralized regime and extracts more rents in that case. It is not a priori clear whether it is
better to have the budget set by the central or local government.
Given the above, it is clear that the politician looses less from pooling in the
centralized regime (conversely, has more to gain from separating in the decentralized one).
This explains our main result, namely, that separation occurs under a greater parameter
range in the decentralized regime than in the centralized one. Conversely, the centralized
regime disciplines bad incumbents under a greater parameter range. This means that, in
terms of expected voter welfare, no regime is clearly preferred. In fact, for the parameter
range under which discipline and selection are the same in both regimes, centralization
dominates decentralization. More interestingly, when the parameters are such that
decentralization improves selection (at the cost of discipline) when compared to
centralization, we obtain that this latter is preferable if the average quality of politicians is
low and the future is heavily discounted. Intuitively, the value of selection comes from the
possibility of finding a good politician to replace a bad one in the future.
Literature review
This paper is related to the literature on (fiscal) decentralization, whose roots date
back to the seminal contributions by Tiebout (1956) and Oates (1972). Traditional
approaches focus essentially on inter-regional spillovers and preference heterogeneity and
suppose, in general, aggregate utility maximizer governments.
As we have already pointed out, the literature on fiscal competition has provided a
rationale for decreasing the autonomy of local governments (see, for instance, the survey by
Wilson, 1999). Fiscal base mobility may be detrimental, as it increases the marginal cost of
taxation, thereby creating a downward pressure on the public sector. This constitutes a
welfare loss if governments are benevolent and is a reason to centralize tax collection.
This paper is more in line with the recent contributions that take a political
economy viewpoint: benevolent governments are abandoned. The relative merits of
centralized and decentralized systems have been analyzed, inter alia, by Bardhan and
Mocherjee (2000), Besharov (2002), Besley and Coate (2003), and Hindriks and Lockwood
(2004). The first two focus on whether one of the government levels is more prone to
capture by special interests, the first using a model of political competition and the second a
menu-auction one. In Besley and Coate (2003), centralized decision is undertaken by an
assembly of locally elected representatives. Hindriks and Lockwood (2004) use a political
agency model. Voters have incomplete information about the quality of politicians and
must choose whether or not to reelect an incumbent, using past performance to infer his
quality. A decentralized system (each region run by a politician) is compared with a
centralized one (one politician for both regions). A key feature of decentralized systems is
the possibility for yardstick competition: with correlated economic contexts, voters can
condition their reelection strategy on policy outcomes in the other region. This is studied by
Belleflamme and Hindriks (2002), Besley and Smart (2003) and Hindriks and Lockwood
(2004), among others.
The remainder of the paper is organized as follows. The next section presents a
brief description about the Portuguese system of local public finance. Section 3 introduces
the model and equilibrium under each regime. Section 4 looks in detail at the relative
merits of both regimes in terms of discipline and selection. Section 5 compares voters'
expected welfare under the two regimes. In Section 6 we present some preliminary data on
Portuguese local elections. Section 7 concludes.
2. A closer look at Portuguese local finance
Portuguese local government is divided into two levels: municipalities

(municípios) and a smaller unit called freguesia.3 These two government levels are called
autarquias. Additionally, the archipelagos of Madeira and Azores have their own
autonomous governments. As of 2004, there were 304 municípios and 4281 freguesias. The
average município has an area of 299 square km and 34 thousand inhabitants. Freguesia is
a smaller unit, with an average area of 21.6 square km and 2438 inhabitants. A closer look
at the data also allows one to conclude for the existence of considerable heterogeneity in
the sizes of municípios and freguesias (Direcção Geral das Autarquias Locais - DGAL,
2004b). Both government levels have their own elected officials, with elections taking
3
The Portuguese Constitution mentions an additional level of local government, the Regions, encompassing
several municipalities, which have not, to date, been implemented.
72 Susana Peralta
place every 5 years.

In the OECD publication mentioned in the introduction (OECD, 1997), Portugal
shows up as one of the countries conferring a smaller autonomy to local governments in the
determination of their fiscal revenue.
Taking a closer look at the Portuguese official publications (DGAL, 2004a,

2004b) one may actually conclude that autonomy is even more restricted than what is
suggested by OECD (1997). We have attempted a reclassification, taking into account the
detailed description of local governments' tax setting autonomy in DGAL (2004a, 2004b)
and Baleiras (1997). The result is Table 2. This table refers only to autarquias, i.e.,
Madeira and Azores autonomous regions are excluded.
It is noteworthy that taxes only represent 28% of total revenue (figure for 2002),
as shown in Table 3, which displays the composition of revenues of Portuguese autarquias.
This raises the question of the capacity of local governments to influence the amount of
funds they receive as transfers. As is apparent from a close inspection of the transfers' rules,
local governments have little or no power to influence the amounts. Local governments are
entitled to 30 % of the average revenues of personal income tax, corporate income tax and
value added tax collected in the country two years before. This total amount is then
distributed to local governments as follows.
• 62.1 % for the Municipalities General Fund (Fundo Geral Municipal), split
according to a formula that accounts, in decreasing order of importance, for
the number of inhabitants, hotel and camping occupation rates, younger than
15 inhabitants, geographic area, number of freguesias and amount of personal
income tax collected in the municipality.
• 16.7 % for the Municipalities Cohesion Fund (Fundo de Coesão Municipal),
which targets the poorest municipalities, i.e., the ones with a per capita fiscal
revenue below the national average and the ones performing worse in terms of
quality of life indicators (health, life expectancy, and education).
• 13.6 % for the Municipalities Base Fund (Fundo de Base Municipal), equally
split amongst the municipalities;
• 7.6 % is distributed to freguesias.
Local governments are also entitled to conditional transfers (i.e.,targeted for
specific projects) from the central government, varying from 60% to 90% of the total cost
in situations like unexpected public calamity, projects related to urban reconstruction,
environment and natural resources, transportation infra-structures, subsidized housing,
among others. Finally, municipalities may apply to European Structural Funds and
Regional Development Funds. In both types of conditional transfers, the scope of local
governments to influence the total amounts received relies almost exclusively on their
capacity to submit projects which are in line with the funding priorities defined by the
central government or the European Union.
The brief description above allows one to conclude for a “weak autonomy of local
governments vis-à-vis the design of their own revenue sources” as “tax proceedings are
virtually insensitive to local policy-makers decisions concerning tax parameters” and, as
regards transfers, “the discretionary autonomy of policy makers is very limited” (Baleiras,
1997, pages 4 and 5). Local governments in Portugal do seem to enjoy considerable
freedom in choosing the provision of local public goods of which they are in charge (a non-
exhaustive list includes parks, transportation and road system, sports and leisure, consumer
protection, housing; see Baleiras (2002) and DGAL (2004b) for a detailed description).
74 Susana Peralta
3. The model
The base model is adapted from Besley and Smart (2003). A local government
decides on the quantity of public good G . The cost θ of the public good is uncertain and it
can be high ( θ = H ) or low θ = L ), with H > L . The probability of a high cost is q .
Voters derive utility from the consumption of the public good and dislike high local
budgets. The local budget is denoted x ∈ [0, X ] , i.e., there is a no-debt constraint and a
maximum budget size. The utility function is W (G, x) = G − C ( x) , where C is a strictly
increasing and strictly convex function. The C function is meant to capture inefficiency
costs of tax collection. We define GHg and GLg as follows
Gθg = arg max G − C (θ G )
and let xθg = θ Gθg . Convexity of the C function ensures that xLg > xHg , i.e., the public sector
is optimally larger when the provision cost is low. This, together with L < H , implies that
GLg > GHg , i.e., the quantity of the public good provided decreases with the provision cost.
We also suppose that X > xLg .
There are two types of politicians, good ( g ) and bad ( b ). Good politicians
always pursue the interests of the electorate, while bad ones care about the rents r they
manage to extract. The proportion of good politicians is π . This may be interpreted as a
measure of the quality of the polity.
The timing of the model is as follows. There are two periods. In the first period,
there is an incumbent in place, who implements a given policy. At the end of the first
period, an election takes place. The incumbent is either approved by the electorate and
stays in office one further period or he is voted out, in which case nature randomly selects a
politician to be in office in the second period. The game ends at the end of the second
period, with no further elections. The future is discounted by β .
The local budget x may be set by the local government or by the central one. We
shall denote the former regime as D (mnemonic for decentralization) and the latter as C
(standing for centralization). When the budget x is set locally, local officials decide x and
G , and both are observed by the electorate. If the budget x is set by the central
government, local officials only decide G . In this situation, voters have less information
about the size of the budget. This may be because the law is unclear or too complicated to
be understood by the electorate, with lots of exceptions, or even because there is no clear
written rule as to the size of the budget. We capture this idea in its more straightforward
form, by supposing that voters do not observe x . The central government knows the true
provision cost and sets the budget accordingly, that is x = xLg if θ = L and is x = xHg if
θ = H . Hence, voters know that the budget is xHg ( xLg ) with probability q ( 1 − q ).
3.2. Preliminary results

Voters observe the policy implemented by the incumbent and use Baye's rule to
compute the posterior probability Π that he is good, given the observed record. If Π is
greater than the probability that the randomly selected official is good, π , the incumbent is
reelected. Otherwise, he is voted out of office.
What about the politicians? As usual, we solve the game by backwards induction.
We begin by looking at regime D. A good politician implements (Gθg , xθg ) in both periods.
Hence, any other policy vector perfectly signals a bad politician. A bad politician extracts
maximal rents in the second period (since he is no longer concerned by re-election), i.e.
x = X and G = 0 . We now look at the bad politician's behavior in the first period. He may
implement one of three policy vectors: (GHg , xHg ) , (GLg , xLg ) or (0, X ) 4 To simplify notation
4
Any other policy vector is a perfect signal of his type and is dominated by (0, X ) .
76 Susana Peralta
we shall refer to the policy vectors only by the G component. Also, let Π G denote the
posterior probability of the good type given an observed quantity of the public good of G .
If the provision cost is H , the bad incumbent gets a negative rent (equal to ( L − H )GLg ) by
implementing GLg , no rent by choosing GHg and the maximal rent when he implements 0 .
The two former strategies are dominated by the last one.5 If the provision cost is low, a
similar dominance argument allows one to eliminate GLg . The politician either implements
0 , extracting a rent of X , or GHg , extracting a rent of ( H − L)GHg . We let λ denote the
probability that he takes the latter action.
We may summarize the behavior of the bad politician as follows. He never
provides the correct amount of the public good. If the cost is high, he extracts maximal
rents and is voted out of office. When the cost is low, he may use one of two possible
strategies. He either separates himself from the good type by providing no public good and
being voted out. Or he pools with the good type under a high cost and keeps his chance of
re-election. In so doing, he foregoes some current rents in return for a probability of
reelection (hence, future rents).
From the above, we immediately see that GLg can only be implemented by the
good politician, hence Π G g = 1 > π and the voters reelect the incumbent. Also,
L
Π 0 = 0 < π , hence the incumbent is voted out of office. Finally, we have that
πq
ΠGg =
H π q + (1 − π )(1 − q)λ
As regards regime C, a good politician will, again, implement Gθg in both periods.
The bad politician will again extract maximal rents in the second period (i.e., implement
G = 0 and get a rent equal to the budget set by the central government). As for the first
period, when the provision cost is high, the central government sets a budget of xHg . A
dominance argument of the kind used above allows one to conclude that the politician
implements G = 0 and extracts r = xHg . If the provision cost is low, again we may
eliminate the dominated action GLg and we are left with G = 0 (with a rent of xLg ) or GHg
(with a rent of L(GLg − GHg ) ).
5
Indeed, 0 yields a payoff of X , whereas the other two actions give at most ( L − H )GLg + β X and
βX , respectively.
Voters can only observe the level of G . They do not reelect when G = 0 and
reelect with probability one if G = GLg . Conditional on observing GHg , the probability that
the incumbent is good is again given by Π G g .
H
Comparing regimes C and D, one concludes that maximal rents are higher in the
latter, which is not surprising, since the local official has full control of the budget.
However, L(GLg − GHg ) = xLg − LGLg > ( H − L)GHg = xHg − LGHg (since xLg > xHg ) and the rent
he extracts when pooling his higher in regime C. This is due to the lack of transparency,
since voters can only observe the level of G , and not the budget. We will use the following
notation
r D = ( H − L )GHg
r C = L (GLg − GHg )
We have the following preliminary result regarding the bad politician's behavior.
Proposition 1 When the bad politician signals his type to the electorate, he extracts higher
rents under the decentralized regime. When the bad politician mimics the good one, he
extracts higher rents under the centralized regime.
3.2. Equilibrium
Having outlined the strategies of the players, the equilibrium of the game is
straightforward to obtain. Let σˆ denote the probability that voters approve the incumbent
after observing GHg and λ̂ the probability that the bad politician implements GHg when the
cost is low, for regime D . For regime C , σ and λ have analogous meanings.
Equilibrium in regime D is described in the following lemma.
Lemma 1 (Besley and Smart, 2003) When the budget is set at the local level, an
equilibrium exists for all values of parameters and is generically unique.
1. A pooling equilibrium, with λˆ = σˆ = 1 , exists if and only if
1
q≥ and r D ≥ (1 − β ) X
2
2. A hybrid equilibrium, with λˆ = q /(1 − q) and σˆ = ( X − r D ) / β X , exists if and

78 Susana Peralta
only if
1
q< and r D ≥ (1 − β ) X
2
3. A separating equilibrium, with λˆ 0 and σˆ = 1 , exists if and only if
r D ≤ (1 − β ) X
Proof See Besley and Smart (2003).
The equilibrium strategy of the bad incumbent is ruled by the trade-off between
current (obtain r D and forego X ) and future rents ( β X ). When behaving according to
voters' interest, the incumbent bears a cost of X − r D , with a gain of βσ X . If the cost is
very high, he prefers to reveal his type by extracting maximal rents and be voted out of
office. If the cost is low enough, he mimics the good type to gain reelection.
While future rents are given by X in regime L , irrespective of the state of nature,
this is no longer true when the budget is set by the central government. In this case, the bad
incumbent will extract a rent of xHg when the cost is high and xLg when the cost is low. To
compute equilibrium strategies, we have to make an hypothesis about the realization of the
shock in the second period. To keep things simple, we suppose that the provision cost is the
same in both periods. Hence, the rent extracted by the bad incumbent in the second period,
if he keeps in office, is equal to xLg .
Lemma 2 When budget is set at the central level, an equilibrium exists for all values of
parameters and is generically unique.
1. A pooling equilibrium, with λ = σ = 1 , exists if and only if
1
q≥ and r C ≥ (1 − β ) xLg
2
2. A hybrid equilibrium, with λ = q /(1 − q ) and σ = ( xLg − r C ) / β xLg , exists if and

only if
1
q< and r C ≥ (1 − β ) xLg
2
3. A separating equilibrium, with λ = 0 and σ = 1 , exists if and only if
r C ≤ (1 − β ) xLg
Proof See Appendix.
The intuition underlying the equilibrium is the same as in the D regime. It

amounts to a trade-off between current and expected future rents. Equilibrium of both
regimes is summarized in Figure 1, where i = C , D and max rent i stands for X when
i = D and xLg when i = C . When rents from pooling are too low, separation occurs. If
rents are high, some pooling occurs at equilibrium. If the probability of a high cost is high
enough, voters are willing to believe that GHg was implemented by a good politician and
they reelect the incumbent with probability one. Otherwise, both voters and the politician
play mixed strategies.
4. Discipline and selection
While both regimes are equivalent in the equilibrium structure, they do no yield
the same payoffs. A closer inspection of Figure 1 reveals that r i , max rent i and σ differ
between regimes. We know from Proposition 1 that max rent D = X > max rent C = xLg and
r C > r D . It is therefore natural to expect that the separating equilibrium arises more under
regime C, which is exactly what one sees in Figure 2, where the (q, β ) space is divided
80 Susana Peralta
into five regions, according to the prevailing equilibria. As the value of the future, β ,
decreases, the bad politician prefers to extract maximal rents in the first period and be voted
out. This will happen for lower values of β under regime C.
It is now clear that, when the budget is set centrally, less separation occurs at
equilibrium. Hence, lack of autonomy in budget setting improves discipline, at the cost of
decreasing selection.6
Proposition 2 The discipline of the bad incumbent is improved when the budget is set at the
central level.
We now turn to the selection issue, i.e., bad incumbents being voted out of office.
In a separating equilibrium, selection is perfect: the bad incumbent signals his type and is
voted out. In a pooling equilibrium, there is no selection, since the incumbent is always
reelected, irrespective of his type. Hence, when β is intermediate and q > 1/ 2 we may
safely conclude that selection is improved in the D regime. When q < 1/ 2 , that is still the
case. A bad incumbent is always voted out of office in case D and only sometimes in
regime C .
6
Note that, under the hybrid equilibrium, λ̂ = λ , hence both regimes yield the same discipline.
When the equilibrium of both regimes is hybrid, we have
xLg − r c X − r D xLg r D − Xr C
σ − σˆ = − = <0
β xLg βX β xLg X
And again the probability that the bad incumbent is voted out (1 − σ ), is greater in
the decentralized regime.7
The fact that σ < σˆ is a result of the mixed strategy equilibrium. Given that in
regime D the gain from separating is higher, voters must reelect with a higher probability
to keep the incumbent indifferent between current and future rents.
Proposition 3 Elections are more effective in selecting out bad incumbents under regime
D.
We may interpret devolution of budget setting power to local governments as
making more information available to voters. Our results would then imply that improving
voters' information is bad for discipline but good for selection. This is reminiscent of the
result in Besley and Smart (2003). In their model of local budget setting, they introduce a
probability that voters become informed about the provision cost and show that increasing
this probability improves selection but decreases discipline.
Propositions 2 and 3 highlight the basic trade-off between the two budget-setting
systems. It is not clear, a priori, which system is better. In the next section we compare
expected welfare under both regimes.
One implication of this proposition is that we should observe more turnover when
budgets are set locally. We present some preliminary evidence regarding turnover in
Portuguese local elections in Section 6.
7
For the sake of precision, one must reckon that in a hybrid equilibrium there is a probability to vote out the
good incumbent, when the cost is high. This is equal to q (1 − σ ) . The probability to vote out the bad
incumbent is q + (1 − q)(λ (1 − σ ) + 1 − λ ) . Hence, the net selection effect of elections is equal to
2
1 − λσ + qσ = 1 − 1q− q σ , i.e., the greater is σ , the lower is selection.
82 Susana Peralta
5. Welfare comparison
We begin by computing expected welfare (at the beginning of the first period) for
both regimes, EWC and EWD . Recall that the utility function is given by G − C ( x) . For
the decentralized regime, this poses no problem, as voters observe both G and x . This is
no longer true of regime C, as voters only observe G . However, the central government's
budget is funded by taxes, a part of which is born by the local voters. Our objective is to
concentrate on the trade-off between selection and discipline. Hence, we prefer not to make
any explicit assumption about the regional distribution of the central government tax
collection. In order to keep the two regimes fully comparable, we shall use the same utility
function for both regimes. Voters may care about the budget size even if they do not pay
for it entirely, or they do not foresee the link with their tax bill. Alternatively, one may
interpret the utility function as reflecting the trade off between public good provision and
budget size which a benevolent planner would take into account for the purpose of
comparing the two regimes.
Before proceeding, it is useful to introduce some notation. Let Wθg = Gθg − C xθg , ( )
Wθ = −C ( xθ ) and W = −C ( X ) . Also, denote EW = qW + (1 − q)W . Now let WCo and
b g b
X
i i
H L
i
WDo denote the expected per-period utility in the absence of elections under each regime,
i.e.
WCo = π EW g + (1 − π ) EW b and WDo = π EW g + (1 − π )WXb
Expected utility in regime C is given by
EWC = (1 + β )WCo + (1 − π )(1 − q)λGHg +

(1)
( )(WLg − WLb )
β (1 − π )π qσ (WHg − WHb ) + (1 − q)(1 − σλ )
While expected utility in regime D equals
EWD = (1 + β )WDo + (1 − π )(1 − q)λˆ (WHg − WXb ) +
( )
(2)
β (1 − π )π qσˆ (WHg − WXb ) + (1 − q)(1 − σλ
ˆ ˆ )(WLg − WXb )
The two last terms in (1) and (2) give the impact of elections on voters' welfare.
The first one is the discipline effect, the increased utility obtained when the bad incumbent
refrains from extracting maximal rents, which happens with probability λ when the
incumbent is bad (probability 1 − π ) and the provision cost is low (probability 1 − q ). The
second term is the selection effect, representing the increased utility from identifying and
voting out bad incumbents. This may happen when the cost is high or low. In the former
case, a good politician is voted out with probability 1 − σ and a bad one with probability 1 .
Hence, selection improves voters' welfare with probability σ . In the latter case, a bad
politician is replaced by a good one either if he pools and is voted out ( λ (1 − σ ) ) or if does
not pool ( 1 − λ ) . The total probability equals 1 − λσ .
One should note at this stage that incentive and selection only matter to the extent
that some bad politicians exist. If π = 1 , the existence of elections has no impact on
welfare. This is a natural result in a model where there is consensus, i.e., all voters agree on
their preferred policy.
The welfare difference between the two regimes may be sub-divided into three
parts. The first one is the baseline difference, i.e., what would obtain in the absence of
elections,
(
(1 − π )(1 + β )ΔC = (1 − π )(1 + β ) C ( X ) − qC ( xHg ) − (1 − q)C ( xLg ) ) (3)
The second one pertains to the discipline effect of elections
(
(1 − π )ΔD = (1 − π )(1 − q) λGHg − λˆ (WHg + C ( X )) ) (4)
The third one is related to selection
[(
(1 − π ) βπΔS = (1 − π ) βπ q σ (WHg + C ( xHg )) − σˆ (WHg + C ( X )) + )
(5)
( )(WLg + C ( xLg )) − (1 − σλ
(1 − q) (1 − σλ ˆ ˆ )(WLg + C ( X )) ) ]
Summing up the three effects we obtain
EWc − EWd = (1 − π ) ( (1 + β )ΔC + ΔD + βπΔS )
A few observations are in order. First of all, in the absence of bad politicians
( π = 1 ), both regimes would be equivalent in terms of welfare, since good politicians have
the same behavior in both. Secondly, in the absence of elections, regime C dominates
84 Susana Peralta
regime D, since there is no discipline at all and only maximal rents matter. Finally, ΔC ,
ΔD and ΔS are independent of π . The average quality of the polity, π , then determines
the relative weight of selection on the welfare difference. Indeed, the gain from voting out a
bad incumbent is only good to the extent that there is a high probability that he be replaced
by a good one.
Table 4 summarizes the relative performance of each regime in giving proper
incentives and selecting bad incumbents.
Even if regime C is better in the absence of elections, regime D is, in general,

more effective in getting the most out of elections. Excluding the positive effect of
discipline in SH and SP (and selection in HH, which we cannot sign), regime C is always
outperformed by regime D. Note that this is not incompatible with Proposition 2. It is still
true that regime C provides more discipline, in the sense that it does so for a greater range
of parameter values: in regions SH and SP, D provides no discipline at all while C provides
some. However, when both regimes provide discipline, i.e., in SS, HH and PP, the expected
impact on voters' welfare is higher under regime D. This is because the stakes of
disciplining the incumbent, i.e., the difference between maximal rents and rents extracted
when pooling are higher in regime D. Given the difference in maximal (i.e., second period)
rents, the stakes of voting out a bad politician (the selection effect) are also higher under
regime D. This, together with the result in Proposition 3, immediately implies that the
impact of selection on welfare is higher under regime D.
We summarize our conclusions in the following two propositions.
Proposition 4 When equilibrium is separating or pooling in both regimes, expected welfare

is higher in the centralized regime.
Proof See Appendix.
When equilibrium is separating under both regimes, bad incumbents are always
voted out, and second period utility is the same as first period one. As bad incumbents
always extract maximal rents, regime C is better. As regards case PP, it is easy to see that
maximal rents also play an important role. On the one hand, bad incumbents are never
voted out and always extract maximal rents in the second period. On the other hand, the
probability of a high cost q , is high, and bad incumbents also extract maximal rents in the
first period when the provision cost is high. It is interesting to notice that the advantage of
the centralized regime stems in both cases from the welfare difference in the absence of
elections. Indeed, inspection of Table 4 reveals that the impact of elections on welfare, i.e.,
ΔI + ΔS is negative in both cases SS and PP.
We know, however, that a change in the tax setting regime may induce a change in
the equilibrium outcome of the game. In particular, it is possible that selection is improved
in the decentralized regime (cases SH, SP and HH). We now investigate what is the likely
impact of this equilibrium change in the expected welfare of the voters.
Proposition 5 When selection is improved in the decentralized regime, expected welfare is

higher under this latter if the value of the future is high and the average quality of the
polity is low.
Proof See Appendix.
The intuition behind proposition 5 is easy to grasp. In the two cases in which
ΔD > 0 and ΔS < 0 , i.e., SH and SP, we just have to recall that the weight of the selection
effect is increasing with the average quality of politicians and that selection matters when
the weight given to the future is high enough. Hence, when politicians are on average very
good (and future is not too heavily discounted), the regime that performs better in terms of
selection (D) is better than C. As regards the hybrid equilibrium, we have that, excluding
the selection effect, the centralized regime outweighs the decentralized one. Imposing one
further condition (amounting to an upper bound on the rent difference of the decentralized
regime, X − r D ), we ensure that the impact of selection on the expected utility is negative,
and again the intuition goes through.
86 Susana Peralta
6. Turnover in Portuguese local elections
Our analysis has highlighted several implications of the lack of tax setting
autonomy. One of those is related to electoral turnover, which is expected to be lower than
with a higher degree of local autonomy. In this section, we present some preliminary data
on turnover in Portuguese local elections. A richest empirical analysis would entail either
an international comparison (countries conferring different degrees of autonomy to local
governments) or a comparative analysis of Portuguese municipalities, exploring the
variation in revenue sources. Even if the general level of discretion is very low, there are
some revenue sources for which municipalities enjoy a greater autonomy. A full-fledged
empirical analysis being outside the scope of this paper, we simply show some descriptive
statistics to suggest that turnover at the local level seems indeed to be restricted in Portugal.
We present our descriptive statistics in Table 5. Local elections in Portugal took
place in 1976, 1979, 1982, 1985, 1989, 1993, 1997, 2001 and 2005. Table 5 gives the
number of different presidents that have been in office in each municipality during the last
three decades. It refers to the identity of the president, as opposed to his party.8
Although the Presidente da Câmara is not the only elected official in each
municipality, he is the chief of the executive branch of the local government, the Câmara
Municipal, and seems to enjoy a considerable latitude in decision making. Indeed, the
8
This distinction is relevant, although not very important empirically, as it has happened that an incumbent runs
for reelection representing a different political party.
following quote, adapted from DGAL (2004b, pages 45-47), suggests that the Presidente
da Câmara enjoys considerable power. “The Câmara Municipal is a permanent executive
office, in charge of the organization and functioning of the municipal services, urbanism
and public works, as well as relations with other local government bodies. It executes the
decisions of the Assembleia Municipal [legislative and consultative body], administers the
employees and the patrimony, decides the local budget, concedes licences [for construction
and economic activities], and gives support to Freguesias. (...) It is the Presidente da
Câmara who decides the division of policy areas amongst the elected members of the
Câmara Municipal (vereadores). Within certain limits, he may also decide the number of
vereadores who are actually responsible for a policy area, as opposed to merely attending
the meetings, and whether they work part or full-time. He is also competent to supervise the
administration of the employees at the service of the municipality”.
Table 5 allows one to conclude that most Portuguese municipalities have indeed
had a very low number of different presidents. Indeed, with a history of 9 elections, the
average number of different presidents is 3,76. Around 73% of the municipalities have had
4 or less presidents. This means more than two mandates per president. No municipality at
all has changed president in every election.
This piece of evidence is to be taken with care, given that it does not take into
account several important aspects. First of all, it may well be the case that turnout is low
because the incumbent is, on average, good. If this is the case, however, our welfare
analysis suggests that increasing local autonomy will have a positive impact on voters'
welfare. More importantly, our model is one of political consensus, that is, all the voters
agree on the best policy to be implemented. In reality, the Presidentes da Câmara belong to
different parties and propose different platforms to the electorate and this is, admittedly, an
important determinant of turnover.
7. Concluding remarks
This paper looks at the relationship between budget setting autonomy and political
accountability at the local level. The usual focus of the literature is on decentralization of
both expenditure and revenue collection. However, both functions are not necessarily
decentralized to the same extent. In particular, there is evidence of revenue collection being
88 Susana Peralta
only partially left to the autonomous initiative of local governments. The usual rationale for
such a lack of autonomy is the internationalization of inter-jurisdictional spillovers as, e.g.,
externalities stemming from tax competition. This paper studies the issue from the
viewpoint of political accountability.
We use a political agency model to compare two budget setting regimes: the
centralized (budget set by central government) and decentralized one (budget set by local
government). In both regimes, expenditure is set at the local level, i.e., local governments
decide the quantity of the local public good.
No regime dominates the other in terms of rent seeking. While maximal rents, that
reveal a bad incumbent, are higher with decentralization, a corrupt local official can extract
higher rents in the centralized regime without revealing his type. Hence, the relative gains
from separating vis-à-vis pooling are higher with decentralization. Therefore, the
decentralized regime outperforms the centralized one in terms of selection, while it is
outperformed in terms of discipline.
With the background of the relative merits of the two regimes in terms of
discipline and selection, we proceed to a comparison of expected welfare. When both
regimes yield exactly the same discipline and selection (i.e., equilibrium is either separating
or pooling in both), we show that centralization is preferred to decentralization. In both
types of equilibrium, maximal rents play an important role, be it because the bad incumbent
always plays them in the first period (separating equilibrium) or because he is never voted
out of office, hence they are very often played in the second period. Therefore, the regime
with the lowest maximal rents is preferred. The most interesting cases arise when a change
in regime switches the type of equilibrium, namely, discipline is improved in the
centralized regime. That is, the decentralized one has a separating equilibrium, with no
discipline at all, and at least some discipline arises under centralization, with a hybrid or
pooling equilibrium. In these cases, the decentralized regime dominates the centralized one
if the average quality of politicians is high and if the future is not heavily discounted.
Indeed, it makes more sense to invest in selection if there are a lot of good politicians in the
world and if second period utility matters a lot; conversely, if most politicians are bad, or
future is very discounted, then it pays more to give them a stronger discipline.
The analysis undertaken in this paper constitutes a first look at the issue of
different degrees of autonomy in the revenue and expenditure functions of local

governments. It gives interesting insights about the main tradeoffs driving the choice of the
degree of autonomy in budget setting. The analysis could be extended in a number of ways.
Firstly, we could analyze a game with more than two periods. Indeed, in Portugal, local
politicians are not term limited and it is not realistic to suppose that they cannot run for re-
election at the end of the second period. With a finite number of periods, the last period
effect (i.e., the bad politician extracting maximal rents) always arises. The gain from
pooling is however higher as the politician reckons that he may serve for more than one
additional period. Hence, we could expect restraint to be more often observed. This is likely
to shift the balance in favor of the decentralized regime, where rents extracted when
pooling are lower than in the centralized regime. If we consider, instead, an infinite number
of periods, then matters become more complicated, since we are likely to run into multiple
equilibria issues, and it is not clear how to compare the regimes. Another interesting
extension would be to consider endogenous entry of politicians, that is, suppose that each
individual in the society is either public or self-interested and may decide to become a
politician. The pool of politicians is likely to be better in the decentralized regime, as
public-spirited individuals know that they have a higher chance of being elected, due to the
increased turnover, and this increases their utility from becoming politicians. One may also
envisage extensions that will have an impact in the information of voters, like the
introduction of debt or the possibility of yardstick competition. These can change the
relative merits of the two regimes if there are reasons to believe that they have an
asymmetric impact in information available to voters under the two regimes.
The most interesting predictions implied by our analysis may be summarized as
follows. Firstly, there should be more turnover at the local level when budget setting power
is held by local officials. Secondly, we should observe more rent seeking at the local level
in countries where local governments enjoy less budget setting autonomy. Finally, one
should observe less budget setting autonomy of local governments in countries where the
average quality of the politicians is lower.
The relative merits of increased local autonomy seem to have made their way into
Portuguese politics. Indeed, the Portuguese Government has recently created a Commission
for the revision of the law that regulates local public finances (Lei das Finanças Locais),
90 Susana Peralta
where the idea of “giving increased autonomy to, and increasing the responsibility of, local
governments” plays an important role (Secretary of State for the Local Administration,
cited by the online newspaper Portugal Diário, October 25, 2005).
Appendix
Proof of Lemma 2
We look at each type of equilibrium in turn.
Pooling With λ = 1 , we have that Π G g ≥ π if and only if q ≥ 1/ 2 . With q < 1/ 2 , voters

H
vote against the incumbent when they observe GHg , and the best reply from the bad
incumbent is λ = 0 . With q ≥ 1/ 2 , we have to make sure that GHg dominates 0 for
the bad incumbent, i.e.
r C + β xLg ≥ xLg
Hybrid Voters must be indifferent between reelecting or not, i.e., Π G g = π , which, solving
H
for λ , yields λ = q /(1 − q) . The bad incumbent must be indifferent between pooling
or separating, i.e.
r D − xLg
r C + σβ xLg ≥ xLg ⇔ σ =
β xLg
Separating Playing GHg must be dominated 0 with σ = 1 , i.e.
r C + β xLg ≤ xLg
With   0, we have GgH  1   , hence reelecting is a best reply from

voters.□
Proof of Proposition 4 For region SS, it is straightforward to obtain
EWc − EWd = (1 − π )(1 + β (1 − π ))ΔC > 0
Under PP, straightforward manipulation yields
(
EWc − EWd = (1 − π ) q(C ( X ) − C ( xHg )) − (1 − q)(C ( xLg ) − C ( xHg ))
+(1 − q) β (C ( X ) − C ( xLg )) + (1 − π )q β (C ( X ) − C ( xHg )) )

where q (C ( X ) − C ( xHg )) − (1 − q )(C ( xLg ) − C ( xHg )) > 0 because

(C ( X ) − C ( x )) > (C ( x ) − C ( x )) and q > 1/ 2 .□
g
H
g
L
g
H
Proof of Proposition 5
We utilize the following assumption.
Assumption 1 The rents of the decentralized game are such that
⎛ ⎧ ⎫⎪ G g ⎞
⎜ max ⎨⎪ ΔC ΔC ⎟ X < X − r D < GH GL − GH
g g g
⎬+ g
H
,
⎜ qG g (1 − q)GLg G ⎟ GLg WLg − WHg
⎝ ⎩⎪ H q <1/ 2 ⎪
q >1/ 2 ⎭ L
⎠
As regards SP and SH, we have
EWc − EWd = (1 − π ) ( (1 + β )ΔC + ΔD + βπΔS ) > 0
if
(1 + β )ΔC + ΔD
π <π =
− βΔS
To complete the proof, we investigate whether, in each case, π < 1 . Let

EWc − EWd
ΔW (π ) = 1−π
. Also, denote β = GHg / GLg and β = 1 − r D / X .
For SP, we have, after simplification,
(
ΔW (π ) = (1 + β )ΔC + (1 − q)GHg − βπ ΔC + (1 − q)GLg )
ΔW is decreasing in π and we have that
ΔW (0) = (1 + β )ΔC + (1 − q)GHg > 0

ΔW (1) = ΔC + (1 − q)(GHg − β GLg )
ΔW (1) is decreasing in β , and it is positive when β = β and negative when β = β ,

under Assumption 1. This implies that there exists a β such that π < 1 when β > β .
For SH, we have
ΔW (π ) = (1 + β − βπ )ΔC + qGHg (1 − βπ ) − βπ qσ GLg − GHg ( )

GHg
ΔW is decreasing in π and, using σ = β GLg
, we have
92 Susana Peralta
ΔW (0) = (1 + β )ΔC + qGHg > 0

⎛ Gg ⎞
ΔW (1) = ΔC + qGHg ⎜ Hg − β ⎟
⎝ GL ⎠
Following a similar argument as for SP above, we obtain that there exists a β

such that π < 1 when β > β .
As regards the hybrid equilibrium, we have that, under Assumption 1, ΔS < 0

and, after simplification,
ΔW (0) = (1 + β )ΔC + q(C ( xHg ) − C ( X )) > 0

ΔW (1) = β q(C ( X ) − C ( xHg )) + (1 − q)(C ( X ) − C ( xLg )) +
⎛ Gg ⎞
q ⎜ β X (WLg − WHg ) − Hg (GLg − GHg ) ⎟
⎝ GL ⎠
Under Assumption 1, the last term in EW (1) is negative. Since it is increasing in β , as

soon as it is negative for β = β and positive for β = 1 , there exists a β such that π < 1
when β > β .□
References
Baleiras, R., 1997. Local finance in Portugal: rules and performance, mimeo, Faculdade de
Economia da Universidade Nova de Lisboa.
Baleiras, R., 2002. Finanças e fiscalidade regional e local. In J. S. Costa, ed. Compêndio de
Economia Regional. Coimbra, Portugal: Associação Portuguesa para o
Desenvolvimento Regional, 647-684.
Bardhan, P. and D. Mokherjee, 2000. Capture and governance at local national levels.
American Economic Review, 90(2), 135-139.
Belleflamme, P. and J. Hindriks, 2002. Yardstick competition and political agency
problems. CORE DP 2002/29, forthcoming in Social Choice and Welfare.
Besley, T., 2005. Principled Agents? The Political Economy of Good Government.
Available online at http://econ.lse.ac.uk/staff/tbesley/book/completebook.pdf.
Besley, T. and S. Coate, 2003. Centralized versus decentralized provision of local public
goods: a political economy approach. Journal of Public Economics, 87, 2611-
2367.
Besley, T. and M. Smart, 2003. Fiscal restraints and voter welfare, mimeo, LSE.
Besharov, G., 2002. Influence costs in the provision of local public goods, mimeo,
Department of Economics, Duke University.
Bird, R., 2000, Intergovernmental fiscal relations: universal principles, local applications.
Working paper 00-2, International Studies Program, Andrew Young School of
Public Studies, Georgia State University.
Bucovetsky, S. and M. Smart, 2004. The efficiency consequences of local revenue
equalization: tax competition and tax distortions, Journal of Public Economic
Theory, forthcoming.
Direcção Geral das Autarquias Locais (DGAL), 2004(a),Finanças Municipais 2002,
Ministério das Cidades, Administração Local, Habitação e Desenvolvimento
Regional.
Direcção Geral das Autarquias Locais (DGAL), 2004(b), Estrutura e Funcionamento da
Democracia Local e Regional, Ministério das Cidades, Administração Local,
Habitação e Desenvolvimento Regional.
Epple, D. and T. Nechyba, 2004. Fiscal decentralization. In V. Henderson and J. Thisse,
94 Susana Peralta
eds., Handbook of urban and regional economics, forthcoming.

Hindriks, J. and Lockwood, B., 2004. Centralization and political accountability, mimeo.
Oates, W. E., 1972. Fiscal federalism. New York: Harcourt Brace.
OECD, 1999. Taxing powers of state and local government. OECD Tax Policy Studies 1.
Paris: OECD.
OECD, 2001. Tax and the economy: a comparative assessment of OECD countries. OECD
Tax Policy Studies 6. Paris: OECD.
OECD, 2002. Fiscal decentralization in EU applicant states and selected EU member
states. Paris: OECD.
Portugal Diário, 2005/10/25. Governo quer reforçar autonomia financeira das autarquias.
Available online
http://www.portugaldiario.iol.pt/noticia.php?id=601307&dic_id=291
Público, 2005/11/11. PCP insiste na existência de “saco azul” para autarquias. (Page 10)
Ter-Minassian, T., (ed.), 1997. Fiscal federalism in theory and practice. IMF.
Tiebout, C., 1956. A pure theory of local public expenditures. Journal of Political
Economy, 64(5), 416-424.
Wilson, J., 1999. Theories of tax competion. National Tax Journal, 52(2), 269-304.
World Bank, 2000. Entering the 21st century. World Development Report. World Bank
World Bank, 2003. Reforming intergovernmental fiscal relations in the Czech Republic,
PREM (Poverty Reduction and Economic Management) Notes 80.
Sessão 2
Session 2
EQUALITY OF OPPORTUNITY AND EDUCATIONAL ACHIEVEMENT IN
PORTUGAL
Pedro Carneiro∗
University College London, Institute for Fiscal Studies

and Center for Microdata Methods and Practice
Abstract
Portugal has one of the highest levels of income inequality in Europe, and low wages and
unemployment are concentrated among low skill individuals. Education is an important
determinant of inequality. However, there are large differences in the educational attainment
of different individuals in the population, and the sources of these differences emerge early in
the life-cycle when families play a central role in individual development. We estimate that
most of the variance of school achievement at age 15 is explained by family characteristics.
School inputs explain very little of adolescent performance. Children from highly educated
parents benefit of rich cultural environments in the home and become highly educated adults.
Education policy needs to be innovative: 1) it needs to explicitly recognize the fundamental
role of families on child development; 2) it needs to acknowledge the failure of traditional
input based policies.
∗
This paper was prepared for the 2006 Conference on Economic Development in Portugal organized by the Bank
of Portugal. I thank the Bank of Portugal for hospitality and financial and research support. I also thank the
hospitality of the Poverty Unit of the World Bank Research Group. Finally I thank Ana Rute Cardoso, Colm
Harmon, Pedro Martins, Pedro Portugal and Vincent O'Sullivan for useful information on relevant data and
bibliography for Europe and Portugal.
98 Pedro Carneiro
1. Introduction
Portugal has one of the highest levels of income inequality in Europe. Using data
from the OECD, we estimate that in 1993 a worker at the 90th percentile of the earnings
distribution earns 4.05 times more than a worker in the 10th percentile of the earnings
distribution, and 2.47 times more than the worker in the 50th percentile. The ratio between
earnings at the 50th and 10th percentile of the earnings distribution was 1.64 in that year.
The level of inequality in Portugal is comparable to that observed in North America, which
is thought to have the highest level of inequality in the developed world (see OECD,
2005).1
In this paper we examine the role of education as a source of inequality. Education
directly affects individual employment and earnings and therefore it contributes to income
inequality for a given cross section of individuals. Furthermore, children who are born from
better educated parents enjoy a wider range of opportunities than those born from low
educated parents. Parental education is not only associated with higher household income,
but also with better school and home environments. Therefore, education contributes to
intergenerational inequality by naturally creating inequality of opportunity for children
born in different families.
We start by studying the relationship between education and wage inequality. We
review the literature on the returns to schooling and inequality in Portugal and present some
recent results of our own. Education and age together explain 40 to 50% of the total
variance of log wages in the portuguese economy in 2004.2 Loosely speaking, this means
that if everyone in the economy had the same level of education the variance of log wages
would be halved. We can look at this result in two ways: 1) education is an important
source of inequality and education policy can dramatically affect wage disparities in
Portugal; 2) even if we could achieve equal education for all individuals in society there
would still be a lot of inequality left.
1
In Europe, other countries with a comparable (but smaller) level of inequality are Italy, Greece and the United
Kingdom.
2
On average, one additional year of schooling leads to a roughly 1% increase in the employment probability and
a 7% increase in net monthly wages. Age alone explains only 1-3% of the total variance of log wages.
Equality of Opportunity and Educational Achievement in Portugal 99
Then we examine the sources of education inequality. We reproduce a version of

the Coleman report (Coleman, Campbell, Hobson, McPartland, Mood, Weinfeld and York,
1966)3 for Portugal. Our conclusions are similar to those in the original Coleman report: 1)
families play a crucial role in the educational achievement of adolescents (15 years of age)
- student achievement is strongly affected by home and school environments, but the most
important dimension of school environments is the family background of its students; 2)
measured school resources have a weak effect on educational outcomes.
These findings, subject to some caveats (due to both economic and econometric
reasons), raise important questions and implications. First, it is not possible to think of
education policy focusing only on schools or the classroom. Families have to be brought
into the picture more explicitly.4 The question of how to design family-education policy is
both difficult and extremely important.5
Second, our work suggests that an increase in school resources will not generate
large achievement gains. Surprisingly, in the large empirical literature on school quality we
can rarely find cost effective interventions (e.g., Hanushek, 2002). Even when school
resources have some impact on student outcomes the interventions are too costly to be
justified.6 Education policy needs to be innovative: just increasing school resources will not
work.7
Third, our regressions show that there is a strong association between an
individual's achievement in school and the family background of the other students in his
3
The Coleman report was a study of U.S. schools which assessed what factors were behind student achievement.
This study was extraordinarily influential in the policy and academic circles, and stimulated a substantial
amount of research on this topic.
4
This principle has been applied in many developing countries, such as Brazil, Mexico or Colombia, where
families are given incentives to keep their children in school, and for providing adequate nutrition and health
care. Even in developed countries there are some interventions that consider both child and parent.
5
Several recent studies in the US such as Cameron and Heckman (2001) and Carneiro and Heckman (2002,
2003) also emphasize the role of long run family environments in the schooling decisions of adolescents.
Furthermore, they document that the differences of achievement in children from different family backgrounds
emerge very early and if anything they explain. Skill formation is a cumulative process where ''skill begets skill''
(Heckman, 2001), and therefore education policy needs to be considered in a life-cycle perspective, recognizing
that an individual's achievement and ability to learn at a point in time are a result of the history of past
investments (and other events). Once this is recognized, the fundamental importance of early education
interventions for children of disadvantaged backgrounds becomes very clear.
6
One successful exception is ''the literacy hour'', a program evaluated by Machin and McNally (2005). Krueger
and Whitmore (2001) also argue that class size reductions have strong effects on student achievement in
elementary school.
7
A large literature emphasizes the importance of teacher quality on student achievement (e.g., Hanushek, 2002).
Unfortunately, what really constitutes teacher quality is a black box: researchers are generally unable to identify
the attributes that make a good or a bad teacher.
100 Pedro Carneiro
school. Unfortunately, we cannot really argue that this is evidence of peer effects. Peer
effects are hard to identify in a given dataset (e.g., Manski, 2000), and only recently have
economists used research designs suitable for studying this topic,8 even though they have
substantial policy relevance. However, whether peer effects are important for student
achievement or not, our work shows that there is a tendency for individuals with similar
family backgrounds to attend the same schools, generating a somewhat segregated system.
In the dataset we analyze, school effects explain slightly more than 30% of the variance of
test scores of 15 year old adolescents in portuguese schools. In comparison with other
countries, this is a relatively high number, indicating an above average degree of inequality
between schools (OECD, 2002a). Since we do cannot accurately describe the empirical
importance of peer effects, whether such segregation is harmful or not for the achievement
of disadvantaged students is something we cannot really answer. However, segregated
educational systems are often cause for concern since they may have adverse effects on
social cohesion.
Fourth, even though in our data we have available an unusually rich set of
individual, family and school characteristics, a large fraction of achievement inequality
remains unexplained. Our individual test score regressions have at most an R-squared of
40% and the school level regressions have at most an R-squared of 53%. Even when we
allow the effect of home variables to vary across schools (interacting the school dummies
with the home environment variables in the regressions) we get at most an R-squared of
58% in the individual test score regression.9 This raises an important question that needs to
be addressed in future research: why is the explanatory power of these variables so low and
what other variables should we be looking at?10
8
A recent papers by Sacerdote (2000) finds evidence of the importance of peer effects on educational
achievement.
9
Results are available on request.
10
It may also be that our model (linear regression) is too restrictive to fit the data. One other concern is
measurement error: if test scores are a noisy measure of achievement then it is not surprising that we cannot
explain their variance. We attempt to use a less noisy measure of achievement by extractive the first principal
component of reading, math and science scores and using it in our analysis instead of the individual test scores.
The idea is that scores in these different tests are manifestations of the same underlying ability that we can
capture in this first principal component. Our sample size is greatly reduced by this procedure since we only
have scores on the three tests for 25% of the sample. Still, when we redo our analysis using this new measure of
achievement our basic results do not change significantly.
Once we establish the crucial role of family background for student achievement,
we study how unequal are home and school environments of individuals from different
family backgrounds. We show that better educated parents provide home environments
more conducive to learning than less educated parents. However, there are not many
significant differences in the resources of the schools attended by children from low and
high education parents. The largest (observable) difference in school environments for
these two types of children is in the composition of their peer group. Children of highly
educated parents attend school with children who also have highly educated parents, while
children whose parents have low levels of education attend school with similar children. In
summary, the most significant differences between educational opportunities of children of
different family backgrounds are in home environments and peers, not in school resources.
Finally, we document how inequality of opportunity translates into inequality of
educational attainment among adults, using a dataset that collects information on individual
education and parental education for a sample of portuguese adults in 1999. We show that
there is strong intergenerational persistence in educational attainment: differences in adult
education generate differences in educational opportunity for their children and, in
consequence, persistence of educational inequality from generation to generation. For
example, more than 90% of offspring of fathers with an incomplete primary education or
less never finish high school, while 0% of offspring of fathers with a university degree
complete less than high school.
Our paper proceeds as follows. In section 2 we examine the relationship between
inequality in education and inequality in income. In section 3 we study the determinants of
inequality in student achievement. Section 4 documents differences in home and school
environments between children from different family backgrounds, and section 5 presents
estimates of educational mobility in Portugal. Section 6 concludes.
2. Education and Labor Market Outcomes
It is striking how much of wage dispersion in Portugal is due to schooling (and

age). The first two columns of table 1 present the coefficients of a regression of log
monthly wages on age, age squared and years of schooling for males and females aged 25
102 Pedro Carneiro
to 65 (using the Portuguese Labor Force Survey, or LFS,11 for the fourth quarter of 2004).
Schooling accounts for roughly 40% of the total variance of log wages for males and 50%
of the variance of log wages for females, as shown by the R-Squared of the wage
regressions.12 The return to one year of schooling is about 7% for males and 9% for
females.13 Such large values for R-Squared of wage regressions are unusual, especially in
countries with large levels of inequality. A similar regression estimated on US data has an
R-squared of only 15%. Across different countries in Europe (largely with lower levels of
inequality), a set of studies assembled by Harmon, Walker and Westergaard-Nielsen (2001)
show that the R-Squared for similar regressions is below 40% in most cases, and very often
it is below 30%.
Note: The regression is estimated by OLS. Individuals

reporting now wage or reporting wage in intervals are
excluded. Years of Schooling is constructed from the
schooling categories in the survey. Standard errors in
parenthesis.
Most low wage workers in Portugal have very low education levels. The first
column of table 2 shows educational attainment in Portugal for individuals aged 25 to 65.
11
The name of this dataset is Inquerito ao Emprego. The regression we run is the following:
2
ln Yi = α + β Si + γ Ai + θ Ai + ε i
where Yi is monthly wage for individual i , Si is years of schooling and Ai is age.

12
Age also plays a role but a much smaller one which we can ignore for the purpose of the paper. Age alone
explains only 1-3% of the total variance of log wages.
13
However, this estimate does not correct for the heterogeneity and self-selection into schooling (e.g., Card, 1999,
Carneiro, Heckman and Vytlacil, 2005, Carneiro and Lee, 2005). We exclude from the regression individuals
reporting wages in brackets. Including them in the regression and running an interval regression instead of a
standard linear regression yields very similar results.
Less than 10% of the working age population was ever enrolled in university and 64% of
the population has 6 years of schooling or less.14 The second column of table 2 presents the
education composition of the set of workers earning a net wage smaller than 310 euros per
month in 2004 (roughly the net minimum wage), who account for 4.3% of the working
population. It shows that 90% of these workers have at most 6 years of schooling, and 97%
have at most 12 years of schooling. Furthermore, among those workers earning less than
600 euros per month (roughly the median wage), 74% have 6 years of schooling or less and
98% have 12 years of schooling or less. There is a clear link between poverty and lack of
skills.
Finally, education is strongly associated with employment. Table 3 presents the

average derivatives of a regression (probit) of employment on years of schooling, age and
age squared, using individuals aged 25 to 65 from the LFS.15 An increase in one year of
schooling is associated with an increase in the probability of being employed of 1% for
males and 2.7% for females. In 2004, more than 75% of the nonemployed (34% of the
population aged 25 to 65 in this year) have 6 years of education or less and 94% of the
14
Most of the portuguese labor force is relatively low skilled when compared to the rest of Europe (e.g., OECD,
2002a).
15
The regression model is the following:
∗ 2
E = α + βS + γ A +θ A + ε
∗
E = 1 if E >0
where E is employment, S is years of schooling and A is age.
104 Pedro Carneiro
nonemployed have 12 or less years of schooling.16 In summary, education not only explains
a large fraction of the variance of wages, but an overwhelming proportion of low wage and
nonemployed individuals have very low levels of education.17 This suggests that investing
in education is an important mechanism for escaping poverty.
Note: This table presents average marginal derivatives of a probit

of employment on years of schooling and age. Years of
Schooling is constructed from the schooling variable in the
survey. Standard errors are in parenthesis.
Furthermore, to paraphrase Cardoso (1998), inequality in Portugal is high and

rising. Several researchers such as Cardoso (1998), Machado and Mata (2001), Pereira and
Martins (2000) and Hartog, Pereira and Vieira (2001), document an important increase in
inequality during the 1980s and early 1990s (Parente and d'Uva, 2002, suggest that there
was some stabilization after that). This research also shows that most of the recent increase
in inequality has been at the top of the income distribution. Over the same period there has
been a substantial increase in the returns to schooling, which is also documented in the
papers cited above. The increase in the returns to schooling generates an increase in
inequality between individuals with different levels of education. However, there has also
16
Results available on request.
been an increase in inequality within education groups. Economists usually conjecture that
the increases in between and within group inequality have similar causes, in particular, an
increase in the return to (observed and unobserved) skill. In fact, the most standard
explanation for the increase in inequality in the western world is skill biased technical
change.
This increase in inequality occurs at a time of rapid economic growth. The period
comprising the 1980s and the early 1990s was a period of high growth for Portugal. Rich
and poor individuals have benefited very differently from the overall improvement in the
performance of the portuguese economy, both in absolute and in relative terms. Machado
and Mata (2001) document that between 1982 and 1994 wages increased by 20% for the
10th percentile of the wage distribution and by 52% at the top.
In other countries, such as for example the UK, wages at the bottom of the wage
distribution have stagnated in recent years while wages at the top continue to increase, and
in the US low wage individuals have experienced a decline in their real wages in the last 30
years. This is especially worrisome once we think about the problem of poverty and of
poverty alleviation mechanisms. Rebecca Blank (1996) argues that the most effective
poverty alleviation program available to governments is economic growth. In fact, in the
years after the second world war, the poor were doing relatively well across the
industrialized world. Even though most of them were unskilled there were good jobs
available for them, mainly in manufacturing. However, modern economic growth is driven
by technological growth, and access to the benefits of economic growth is restricted to
those individuals who have invested in skills. Due to the overall growth of the economy, as
the demand for services increases, jobs are still available for the poor in these countries but
these are mostly low value added jobs.
Surprisingly, the relative performance of low wage and low skilled individuals in
Portugal has not been as poor as in countries with similar inequality levels, such as the US
17
Obviously, this does not mean that all low skilled individuals have low wages are are not employed. For
example, 91% of those individuals with 6 years of schooling earn more than 310 euros per month, and 33% of
them earn more than 600 euros per month. Similarly, 75% of the individuals in this group are employed in 2004.
What this says is that almost no highly educated individuals are either low wage or nonemployed. Among those
individuals with 12 years of schooling, 98.6% have wages higher than 310 euros per month, and 65 % have
wages above 600 euros per month. Among those with a university degree, these percentages are 98.8% and
93.5% respectively. The fraction of nonemployed individuals is 22% among those with 12 years of schooling
and 15% among those with a university degree.
106 Pedro Carneiro
and the UK. One reason has been that the low skilled jobs in manufacturing are still
available in this country. The other is probably the social safety net which is stronger in
continental Europe than in anglo-saxon countries. However, in the future as the economy
grows, reconverts and adapts to the competitive pressures of the moderns world, the
sizeable low skilled population that exists in Portugal is likely to not only prevent the
development of the economy, but also to see their situation deteriorate, not only in relative
terms, but possibly in absolute terms as well.
Education is not only important for labor market outcomes, but also for many
other dimensions of an individual's life, and therefore the focus on labor market outcomes
is too narrow. For example, education affects criminal behavior (e.g., Lochner and Moretti,
2004), health (e.g., Grossman, 2005, Smith, 2005), family formation and child development
(e.g., the studies summarized in Carneiro and Heckman, 2003), among many other things.
Of particular interest in this paper will be the role of education on parental investments and
child development, an important source of intergenerational mobility. Inequality in
education generates inequality of opportunity among children growing up in different
families, not only because family resources are smaller in families with lower levels of
education, but probably more importantly, because education affects parental behavior
(e.g., Carneiro, Meghir and Parey, 2005).
3. Sources of Inequality in Educational Achievement
The overall picture described above illustrates the crucial (but certainly not
exclusive) role of skill for individual outcomes in a modern economy (and for the aggregate
success of the economy), and the importance of human capital policy in the modern
environment. There are large skill disparities in the portuguese society that dramatically
affect an individual's life chances. However, these skill disparities are likely to emerge well
before individuals enter the labor market. Using US data, Carneiro and Heckman (2003)
document that skill differences among children from different socio-economic groups
emerge very early in life, often as early as ages 1 and 2. These early skill gaps are
substantial and if anything they tend to expand as children grow older. Carneiro and
Heckman (2003) and Cunha, Heckman, Lochner and Masterov (2005) suggest that
complementarity between human capital investments taking place at different periods in
time is a key feature of the technology of skill formation. As a consequence, childhood skill
disparities translate into even larger adult skill disparities. Furthermore, it may also be
difficult and costly to design compensatory programs in adulthood that correct such skill
gaps.18
Carneiro and Heckman (2002, 2003) suggest that skill disparities are mostly a
consequence of home environments, and present a large body of supporting evidence. In
fact, a similar claim was made in what became known as the Coleman Report. Coleman
and his colleagues investigated the determinants of school performance among children in
the US and concluded that the family background of ones peers was the most important
factor affecting individual school performance, while school resources only played a
limited role.19 Although some of the conclusions of this report have been questioned in the
last 40 years, the core of Coleman's argument is still thought to be correct. Not much is
known for Portugal, but the patterns are likely to be similar (as shown in the remaining of
this section).
In this section I examine the sources of inequality in educational achievement in
Portugal using a sample of 15 year old individuals surveyed by the Programme for
International Student Assessment (PISA). I focus on the 2000 wave of this survey. The
PISA is an international assessment of literacy in reading, mathematics and science for 15
year old adolescent in a large set of (mostly developed) countries.20 Its findings have
received wide attention in the media and in the academic world, and were summarized in
OECD (2001, 2003). Portuguese students rank relatively poorly relatively to their
counterparts in other OECD countries (OECD, 2001, 2003). The PISA collects rich
information on cognitive skills, family environments and school environments which
allows us to do an analysis of these different variables on student achievement. We would
18
If early and early and late investments are complementary in the production function of human capital then the
productivity of late investments increases with the amount of early investments. Intuitively, the more I invest
early on the better equiped I am to learn in the future, because learning builds on accumulated skill. This also
implies that late investments are unlikely to be productive if they are not preceded by early investments.
Therefore, compensatory education programs targeted to young disadvantaged adults may not be very effective
because they cannot remedy early neglect. Finally, complementarity also implies that the productivity of early
investments increases with late investments. If early investments are not followed up they do not amount to
anything.
19
Cognitive achievement in the adolescent years is good predictor of final educational attainment, and it also
affects wages in several developed countries (e.g., Carneiro and Heckman, 2002, Heckman and Vytlacil, 2001,
Blundell, Dearden and Sianesi, 2005, Currie and Thomas, 2003).
20
For a detailed description see OECD (2002b).
108 Pedro Carneiro
like to start by examining skill inequality in a younger sample of individuals but no

comparable data exists (in Portugal) for the early childhood years. Even though the spirit of
our work is that of the original Coleman report, our methodology is slightly different. The
original Coleman report was criticized on several methodological aspects, and we try to
address some of these critics here.
I start by estimating the following model
Tij = α + Fi β + S j γ + ε ij
(1)
where Tij is a test score (in reading, mathematics or science) for individual i in school j ,
Fi is a vector of family and home characteristics, S j is a vector of school dummies and
ε ij is the error term (orthogonal to Fi and S j ). Let Φi = Fi β represent family effects
and Ψ j = S j γ represent school effects. The variables in Fi are indices of parental socio-
economic background, cultural communication with parents, social communication with
parents, home educational resources, activities related to classical culture, possessions
related to classical culture in the family home and time spent on homework. These indices
were constructed by the PISA staff from answers to the student questionnaire. A
description of the construction of these indices is provided in the manual for the PISA 2000
database (OECD, 2002b).21 Then we can decompose test scores in family, school and
orthogonal residual effects:
Tij = α + Φ i + Ψ j + ε ij
(this is a restrictive model - it does not allow for interactions between Φ i and Ψ j ,
21
The PISA Index of Socio-Economic Status was derived from students' responses on parental occupation. The
Index of Cultural Communication with parents was derived from students' reports on the frequency with which
their parents engaged with them in the following activities: discussing political or social issues; discussing
books, films or television programmes; and listening to classical music. The index of Social Communication
with parents was derived from students' reports on the frequency with which their parents engaged with them in
the following activities: discussing how well they are doing at school, eating with them around a table, spending
time simply talking to them. The index of Home Education Resources was derived from students' reports on: i)
the availability, in their home, of a dictionary, a quiet place to study, a desk for study and textbooks; and ii) the
number of calculators at home. The index of Activities Related with Classical Culture was derived from the
students' reports on how often they had participated in the following activities during the preceding year: visited
a museum or art gallery; attended an opera, ballet or classical symphony concert; and watched live theatre. The
index of Possessions Related to Classical Culture was derived from the students` reports on the availability of
the following items in the home: classical literature; and books of poetry and works of art. The index of Time
Spent on Homework was derived from the students` reports on the amount of time they devote to homework per
week in reading, mathematics and science.
although these could be included). Finally:
( ) ( ) ( ) ( )
Var Tij = Var Φi + Var Ψ j + Var ε ij + 2 ∗ Cov Φ i , Ψ j ( )
or
1=
( ) + Var ( Ψ j ) + Var (ε ij ) + 2 ∗ Cov ( Φi , Ψ j ) .
Var Φi
( )
Var Tij ( )
Var Tij Var Tij ( ) Var Tij( ) (2)
This model allows me to assess the relative contribution of family background/home
environment and school characteristics to inequality in educational achievement. The R-
Var ( Φi ) Var ⎛⎜ Ψ j ⎞⎟ 2∗Cov⎛⎜ Φi , Ψ j ⎞⎟

⎝ ⎠ ⎝ ⎠
Squared of regression (1) is given by + + .
Var ⎛⎜ Tij ⎞⎟ Var ⎛⎜ Tij ⎞⎟ Var ⎛⎜ Tij ⎞⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
The Coleman Report argued that the main determinant of a child's school success
was the family background of his peers. Therefore, it is instructive to separate school
S
effects ( Ψ j ) into two components: school quality effects ( Ψ j ) and family background
F
effects ( Ψ j ). In order to do that I estimate the following model (using one observation
per school, independently of school size):
Ψ j = θ + S Sj σ + S Fj η + υ j (3)
S
where Ψ j is estimated from equation (1), the variables in S j are school size, hours of
schooling per year, number of computers per student per school, student-teaching staff ratio
F
and the proportion of teachers with a university degree in pedagogy. The variable in S j is
the average index of parental socio-economic background in the school (described above).
Again, we can write:
S F
Ψ j = θ + Ψ j + Ψ j + υij
( ) S
( )F
( ) S F
Var Ψ j = Var Ψ j + Var Ψ j + Var υij + 2 ∗ Cov Ψ j , Ψ j( ) ( )
or (4)
1=
( S
Var Ψ j ) +
Var ( Ψ j ) Var (υij ) 2 ∗ Cov ( Ψ j , Ψ j )
F
+ +
S F
Var ( Ψ j ) Var ( Ψ j ) Var ( Ψ j ) Var ( Ψ j )

110 Pedro Carneiro
Var ⎛⎜ Ψ Sj ⎞⎟ Var ⎛⎜ Ψ Fj ⎞⎟ 2∗Cov⎛⎜ Ψ Sj , Ψ Fj ⎞⎟

⎝ ⎠ ⎝ ⎠ ⎝ ⎠
(the R-Squared of this equation is given by + + .
Var ⎛⎜ Ψ j
⎞
⎟ Var ⎛⎜ Ψ ⎞
⎟
j⎠ Var ⎛⎜ Ψ j ⎞⎟
⎝ ⎠ ⎝ ⎝ ⎠
Table 4a presents the coefficients estimated from equation (1) and table 4b
presents the estimates from equation (3) for three sets of tests: reading, math and science.
All school and home variables appear with the expected sign and are statistically important,
especially in the reading equation. Table 5 presents the variance decompositions
corresponding to equations (2) (panel A) and (4) (panel B). Panel A shows that we can only
explain about 40% of the variance in reading scores (less for the other scores) using school
and observable family variables. School effects alone account for 21% of the total variance
of these scores, which is more than 50% of the explained variance of the model. As shown
in panel B, most of the variance in school effects is due to the family background of the
students in the school (for reading, 38% of the total variance or almost 80% of the
explained variance; similar results hold for math and science). In summary, the main
conclusion of the Coleman report seems to hold for Portugal: the average socio-economic
status of the students in an individual's school is the main observable determinant of
inequality in educational achievement.
Note: Standard errors in parenthesis.

Note: Standard errors in parenthesis.
This may be due to several reasons. Variance decompositions depend not only on
the coefficients in the regressor of interest but also on the variance of the regressors
themselves. For example, a measure of school quality such as the student-teacher ratio may
be an important determinant of student achievement (large coefficient). However, if the
variance of the student-teacher ratio is small then this variable will not explain very much
of the total variance in achievement, even if its coefficient is large. As we will see in detail
112 Pedro Carneiro
in section opportunity, school quality variables do not differ widely between individuals
from different family backgrounds and therefore cannot explain differences in achievement
due to differences in family background. However, school quality variables do vary within
family background groups and their variance is quite considerable in some cases.22
Therefore, the reason for the low contribution of this type of variables to inequality in
achievement does not seem to be a low level of variability of these variables in the
population.
Unfortunately our results cannot be given a straightforward causal interpretation.

For example, in (1), ε ij can be correlated with both Fi and S j , in which case we will not
be able to identify β or γ . Similarly, in equation (3) we cannot identify σ or η if v j is
correlated with s sj or s j . It is quite likely that in both equations the error term is correlated
F
with the regressors. The variance decomposition of equations (2) and (4) will also be
affected, so that it will be difficult to have a structural interpretation for our exercise.
However, let εij be the residual of the projection of ε ij on Fi and S j , and υij be the
S F
residual of the projection of υ j on S j and S j :
ε ij = φ0 + φ1Fi + φ2 S j + εij
(5)
S F
υij = δ 0 + δ1S j + δ 2 S j + υ j
S
where, by construction, εij is orthogonal to Fi and S j , and υij is orthogonal to S j and
F
S j . Substituting in equations (1) and (3):
( ) ( ) (
Tij = α + φ0 + Fi β + φ1 + S j γ + φ2 + εij )
( )
S
( F
) (
Ψ j = θ + δ 0 + S j σ + δ1 + S j η + δ 2 + υ j . ) (6)
It is simple to show that our OLS estimates of equations (1) and (3) converge to the
coefficients of equations (6) (e.g., p lim βÔLS = β + φ1 ). As long as the correlation
between the residuals and the regressors is due to the behavior of families and schools (for
example, better educated parents choose better schools), then we can interpret the OLS
22
Table A1 in the appendix shows mean and dispersion measures for five different school quality variables.
estimates of equations (1) and (3) as the total effect of these variables on achievement.23
The results from this section are consistent with a large literature in the economics
of education which shows very strong family background effects on educational
achievement, and less strong effects of measurable school resources. There are a few school
programs for which significant effects have been found, but they are the exception rather
than the rule (although not many evaluations exist for this type of programs), and their
effects are generally quantitatively small. One important exception has to do with teacher
quality. Teacher quality systematically shows up in studies of achievement as an important
determinant of educational success. However, teacher quality is a black box, generally
measured as a teacher fixed effect in a regression of individual test scores on teacher effects
and other school and family variables. These teacher effects cannot be explained by
observable measures of teacher quality, such as qualification or experience. Researchers
agree that teachers matter very much, but it is not clear what makes a good teacher (e.g.,
Carneiro and Heckman, 2003, or Hanushek, 2001).
Unfortunately, we cannot explain a large fraction of the variance of test scores
neither with family background variables nor with school variables. For example, for
reading scores we can get an R-squared of 40% for equation (1) and an R-squared of 53%
for equation (3). Given that we measure school effects in a flexible way (school dummies),
we can conclude that most of the variance of test scores is within schools and it is likely to
be caused by unobserved family background variables and their correlation with school
variables, which if true will only strengthen our basic result.
23
There are two caveats: 1) we have Ψ j j (
ˆ = S γˆ + φˆ )
2 instead of Ψ j = S j γˆ ; 2) the assumption that all the
ˆ
correlation between regressors and residuals is due to family and school actions may not hold. In principle, 1) is
( )
not serious and merely demands a slight reinterpretation of the estimates of equation (3) - S j γˆ + φˆ2 may be
a more interesting object that S j γˆ to look at. 2) is more serious and may damage our interpretation of the data
(it may happen, for example, if children of better educated parents inherit a genetic endowment that makes them
better students, but that cannot itself be changed by parental education or parental behavior). Unfortunately this
is a problem we cannot address in our study. We conjecture that it is more likely to affect equation (3) than
equation (1) and, if anything, it will lead to an overstatement of the role of the family in this equation. However,
since the dependent variable of equation (3) comes from estimates of equation (1) it is not clear what the final
effects will be.
114 Pedro Carneiro
However, between school variance (equation (3)) may be explained by several

unobservable factors, and which can be either school or family factors.24
Finally, the almost exclusive emphasis that is usually given to cognitive test scores
as a diagnostic of educational success and as a guide to educational policy may be
unbalanced because many other skills matter. Academic achievement is relatively easy to
measure through test scores, but many other unmeasured skills can be equally or more
important for future success in several dimensions of life. For example, Heckman, Hsee
and Rubinstein (2000), Heckman Sixtrud and Urzua (2005) and Carneiro, Crawford and
Goodman (2005) show that several measures of "noncognitive skills" (which aim to capture
traits such as sociability, patience or discipline) are strong determinants of educational
success, labor market outcomes and a variety anti-social behaviors (such as criminal
activity, teenage pregnancy or drug use). Therefore, we replicated the analysis of table 5 for
five measures of noncognitive ability: sense of belonging, engagement in reading, effort
and perseverance, instrumental motivation and interest in reading.25 Table A2 shows our
results. The explained portion of the variance of these variables is even lower than when we
used test scores. Even though the role of an individual's family is relatively more important
than before, surprisingly, the role of the peer's families seems to be relatively weaker than
before.26
24
There are additional school and family variables in the PISA dataset which we have not used. Their inclusion
does not change the conclusions of this paper. More work needs to be done although it may be very difficult to
explain a larger fraction of inequality in achievement than we already do. Our plan is to briefly examine other
databases, namely the TIMSS95 (Trends in Mathematics and Science Study of 1995) and the 2003 wave of the
PISA.
25
These are five indices described in detail in OECD (2002b). The belonging index was derived from students'
reports on their level of agreement with the following statements concerning their school: I feel like an outsider
(or left out of things); I make friends easily; I feel like I belong; I feel awkward and out of place; other students
seem to like me; and, I feel lonely. The engagement in reading index was derived from students' level of
agreement with the following statements: I read only if I have to; reading is one of my favourite hobbies; I like
talking about books with other people; I find it hard to finish books; I feel happy if I receive a book as a present;
for me, reading is a waste of time; I enjoy going to a bookstore or a library; I read only to get information that I
need; and, I cannot sit still and read for more than a few minutes. The effort and perseverance index was derived
from the frequency with which students used the following strategies when studying: I work as hard as possible;
I keep working even if the material is difficult; I try to do my best to acquire the knowledge and skills taught;
and, I put forth my best effort. The index of instrumental motivation was derived from the frequency with which
students study for the following reasons: to increase my job opportunities; to ensure that my future will be
financially secure; and, to get a good job. The index of interest in reading was derived from students' level of
agreement with the following statements: because reading is fun, I wouldn't want to give it up; I read in my
spare time; and, when I read, I sometimes get totally absorbed.
26
This issue can also be further studied using the TIMMS95 and the 2003 wave of the PISA. Furthermore, the
International Association for the Evaluation of Education Achievement (IEA) Civic Education Study (CIVED)
can be used to examine the determinants of civic behavior. We plan to briefly examine these databases.
4. Equality of Opportunity: Families and Schools
In the previous section we showed how schools and families strongly influence the
academic achievement of 15 year old individuals in Portugal. Different individuals are born
in different families, and therefore experience unequal environments at home and in the
school throughout their childhood and adolescence. In this section we document how
inequality in paternal education is associated with differences in home and school
environments. In short, we show how unequal paternal education translates into different
education opportunities, which give rise to inequality in school achievement.27
Table 6 shows that there are large differences in the level of schooling of fathers
of different individuals. These are just a consequence of inequality in education in the
previous generation, which generates inequality of opportunities for the current generation.
Table 7 shows that highly educated fathers show more interest for their children's progress
in school. Table 8 shows that children of highly educated fathers are more likely to be taken
to museums, and table 9 shows that they spend longer hours studying. Table 10 and 11
refer to school environments. Table 10 shows that there are no strong differences in
standard school quality variables between children of fathers with high and low levels of
education. Therefore, differences in school quality variables cannot explain how unequal
parental background translates into unequal test scores. However, table 11 shows that the
education of the father of the child and the education of the fathers of students in the same
school are strongly correlated.
27
We could have picked another measure of family background, such as maternal education. The basic results
would not change.
116 Pedro Carneiro
In summary, families with better educated fathers provide better home and school
environments for their children, which then translate into better test scores. We have
examined only a few dimensions of such environments. It is interesting that children from
better educated fathers do not attend schools with more resources. The most important (and
maybe only significant) difference between schools attended by children from low and high
educated fathers is in the family background of the students in the school. Better educated
fathers enrol their children in schools with similar peers (children from highly educated
parents). This type of segregation is prevalent in a variety of situations. Individuals tend to
associate with others who are alike, for a variety of reasons. Unfortunately, this may result
in strong inequality in outcomes if peer effects are important.
5. Intergenerational Mobility in Education in Portugal
Inequality in school achievement is likely to generate inequality in school

attainment and labor market outcomes. In this section we examine the relationship between
parental education and an individual's final educational attainment and labor market
outcomes. We use the Social Inequalities II dataset collected by the Instituto Nacional de
Estatistica and the Instituto de Ciencias Sociais of the University of Lisbon (as part of a
European network).28
Table 12 is a transition matrix, which gives the probability that an individual
completes a certain educational degree given the educational attainment of his father. It is
obvious that there is a large increase in educational attainment from the parents to the
offspring generation in this dataset, but there is also substantial intergenerational
persistence in educational status. Notice that more than 90% of offspring of fathers with an
28
A detailed description of this dataset can be found in ISSP (1999).
118 Pedro Carneiro
incomplete primary education or less never finish high school, while 0% of offspring of
fathers with a university degree complete less than high school. This disparity is striking,
especially when we consider that 50% of the individuals in this sample have fathers with
less than a complete primary education, while only 1.9% have fathers with a complete
university education.
Table 13 reports employment rates for individuals in different own education -

father's education cells and table 14 reports the proportion of individuals earning more than
500 Euros (net) per month (roughly the median wage in Portugal in 1999) for individuals in
different own education - father's education cells (we compute the mean of the relevant
variable for individuals belonging to each cell). Conditional on own education, the effect of
parental education on labor market outcomes is relatively weak. Inequality of opportunity
generated from being born from parents with different levels of education seems to
translate into unequal labor market outcomes mostly through its effects on educational
attainment. However, this is effect is quite strong and contributes to a persistence of
inequality between individuals from different background for several generations to come.
6. Conclusion
In this paper we examined the role of educational attainment on earnings

inequality. We showed that years of schooling account for 40 to 50% of the variance of log
wages in the portuguese labor market.
Inspired by the Coleman report, we then study the sources of inequality in
educational achievement. The major observable factor driving inequality in test scores
among adolescents in Portugal is family background, especially the family background of
one's peers in school. Measured school resources have a very limited role. However, a lot
of the variance in student achievement remains unexplained by observable variables.
Finally, there is a large degree of persistence in educational status. In such a
system, inequality is likely to persist from generation to generation.
Our study has important policy implications. First human capital policy is a
powerful tool to address the problems of inequality and poverty for future generations.
120 Pedro Carneiro
Second, education policy needs to be creative and recognize the families are the
fundamental education institution in society and that the role of traditional input-based
school policies is very limited. Therefore, improving the life-chances of poor children
requires intervening at early ages when family influences are the most dramatic. Third,
human capital policy has important intergenerational effects: improving the skills of the
current generation not only improves their opportunities to succeed but it also has dramatic
effects of the opportunities of their offspring.
7. Appendix
References
Blank, R. (1996), It Takes a Nation, Princeton University Press.
Blundell, R., L. Dearden and B. Sianesi (2005), “Evaluating the effect of education on
earnings: models, methods and results from the National Child Development
Survey”, Journal of the Royal Statistical Society: Series A, 168(3).
Card, D. (1999), “The Causal Effect of Education on Earnings”, in O. Ashenfelter and D.
Card, Handbook of Labor Economics, Volume 3, North Holland.
Cardoso, A. (1998), “Earnings Inequality in Portugal: High and Rising?”, Review of Income
and Wealth, 44, 3.
Cameron, S. and J. Heckman (2001), “The Dynamics of Educational Attainment for Black,
Hispanic and White Males”, The Journal of Political Economy, 109.
Carneiro, P., C. Crawford and A. Goodman (2005), “The Importance of Cognitive and
Non-Cognitive Skills: Evidence from the National Child Development Survey”,
working paper, University College London.
Carneiro, P. and J. Heckman (2002), “The Evidence on Credit Constraints in Post-
Secondary Schooling”, The Economic Journal, 112.
Carneiro, P. and J. Heckman (2003), “Human Capital Policy”, in J. Heckman and A.
Krueger (eds.), Inequality in America : What Role for Human Capital Policies, MIT
Press.
Carneiro, P., J. Heckman and E. Vytlacil (2005), “Understanding What Instrumental
Variables Estimate: Estimating the Average and Marginal Return to Schooling”,
working paper, The University of Chicago.
Carneiro, P. and S. Lee (2005), “Ability, Sorting and Wage Inequality”, working paper,
University College London.
Carneiro, P., C. Meghir and M. Parey (2005), “Maternal Education and Child Outcomes”,
working paper, University College London.
Coleman, J., E. Campbell, C. Hobson, J. McPartland, A. Mood, F. Weinfeld and R. York
(1966), Equality of Educational Opportunity, Washington, D.C.; U.S. Government
Printing Office.
Cunha, F., J. Heckman, L. Lochner and D. Masterov (2005), “Interpreting the Evidence on
Life-Cycle Skill Formation”, NBER working paper no. 11331.
122 Pedro Carneiro
Currie, J. and D. Thomas (2001), “Early Test Scores, Socioeconomic Status, School
Quality and Future Outcomes”, Research in Labor Economics, 20.
Grossman, M. (2005), “Education and Non-Market Outcomes”, NBER working paper no.
11582.
Hanushek, E. (2002), “Publicly Provided Education”, in A. Auerbach and M. Feldstein,
Handbook of Public Economics, Volume 4, North-Holland.
Harmon, C., I. Walker and N. Westergaard-Nielsen (2001), Education and Earnings in
Europe, Cheltenham, Edward Elgar.
Hartog, J, P. Pereira and J. Vieira (2001), “Changing Returns to Education in Portugal
during the 1980s and Early 1990s: OLS and Quantile Regression Estimators”,
Applied Economics, 33.
Heckman, J. (2001), “Policies to Foster Human Capital”, Research in Economics, 54.
Heckman, J., J. Hsee and Y. Rubinstein (2000), The GED is a Mixed Signal, working
paper, The University of Chicago.
Heckman, J., J. Sixtrud and S. Urzua (2005), “The effect of cognitive and non-cognitive
factors in behavioral and labor outcomes”, working paper, The University of
Chicago.
ISSP (1999), Social Inequalities II, CD-Rom.
Krueger, A. and D. Whitmore (2001), “The Effect of Attending a Small Class in the Early
Grades on College-Test Taking and Middle School Test Results: Evidence from
Project STAR”, The Economic Journal, 11.
Lochner, L. and E. Moretti (2004), “The Effect of Education on Crime: Evidence from
Prison Inmates, Arrests and Self-Reports”, American Economic Review, 94(1).
Machado, J. and J. Mata (2001), “Earning Functions in Portugal 1982-1994: Evidence from
Quantile Regressions”, Empirical Economics, 26.
Machin, S. and S. McNally (2005), “The Literacy Hour”, working paper, University
College London.
Manski, C. (2000), “Economic Analysis of Social Interactions”, The Journal of Economic
Perspectives, 14(3).
OECD (2001), Knowledge and Skills for Life, OECD.
OECD (2002a), Education Policy Analysis, OECD.
OECD (2002b), Manual for the PISA 2000 Database, OECD.

OECD (2003), Literacy Skills for the World of Tomorrow, OECD.
OECD (2005), Society at a Glance, OECD.
Parente, P. and T. B. D'Uva (2002), “Evolucao da Pobreza e da Desigualdade em Portugal
no periodo 1995 a 1997”, Instituto Nacional de Estatistica.
Pereira, P. and P. Martins (2000), “Does Education Reduce Wage Inequality? Quantile
Regressions Evidence from Fifteen European Countries”, IZA Discussion Paper no
120.
Portugal, P. (2004), “Mitos e Factos sobre o Mercado de Trabalho Portugues: a Tragica
Fortuna dos Licenciados”, Boletim Trimestral do Banco de Portugal, Marco.
Sacerdote, B. (2000), Peer Effects with Random Assignment: Results for Dartmouth
Roommates, NBER working paper no. 7469.
Smith, J. (2005), “Unravelling the SES Health Connection”, Aging, Health and Public
Policy, vol. 30.
Solon, G. (1999), “Intergenerational Mobility in the Labor Market”, in O. Ashenfelter and
D. Card, Handbook of Labor Economics, Volume 3, North Holland.
THE INTERNAL RATE OF RETURN TO ON-THE-JOB TRAINING
Rita Almeida∗
The World Bank
Pedro Carneiro
University College London, Institute for Fiscal Studies
and Center for Microdata Methods and Practice
November 2005
Abstract
In this paper we estimate the marginal internal rate of return to training

using firm level data. Our data allows us to adequately account for the costs of
training. We find that (unlike schooling) foregone output accounts for more than 75%
of the total cost of training. This is not surprising given that in our sample firms train
less than 1% of hours worked. We also quantify the effect of training on current and
future output. We use a fixed effects instrumental variables estimator for dynamic
panel data to estimate the production function (e.g., Blundell and Bond, 2000). In our
preferred specification we find that increasing training per employee in 10 hours
(approx. 0.5% of total hours worked/year), leads to an increase in current output of
about 0.5% and to smaller increases in future output since knowledge depreciates
over time. These estimates imply that the private return to training among the firms
providing training is 32%, and it is -9% for those providing no training. The former
firms offer positive but low amounts of training, in spite of large returns. We
conjecture that this is due to coordination problems between management and
employees, and due to uncertainty in the returns to training.
Keywords: On-the-Job Training, Panel Data, Production Function, Rate of Return
JEL Classification codes: C23, D24, J31
∗
We thank conference participants at the 16th European Association of Labor Economists (Lisbon, 2004), 19th
Meeting of the European Economic Association (Madrid, 2004), the IZA/SOLE Meetings (Munich, 2004) the
ZEW Conference on Education and Training (Mannheim, 2005), and the Econometric Society World Congress
2005, especially those of Steve Pischke. We also thank the comments of an anonymous reviewer.
Corresponding author: ralmeida@worldbank.org. Address: 1818 H Street, NW MC 3-348, Washington, DC,
20433 USA.
126 Rita Almeida – Pedro Carneiro
1. Introduction
Individuals invest in human capital over the whole life-cycle. It is estimated that
over one half of lifetime human capital is accumulated through post-school investments on
the firm (Heckman, Lochner and Taber, 1998), either through learning by doing or through
on-the-job training. However, very little is known empirically about investments in on-the-
job training. While there exist many public training programs for which we can find
economic evaluations, most of the job training in the economy is done within firms and the
evidence on their effectiveness is much more scarce.
Furthermore, like any other investment decision, firms will offer training
programs as long as the expected benefits outweighed the total costs of training. A crucial
parameter to evaluate such investments and, therefore, the allocation of resources in the
economy, is the return to training. Surprisingly, there is almost no research estimating (well
defined) returns to training investments (one important exception is Mincer, 1989). Most of
the literature focuses on quantifying the benefits of training for workers or firms with little
or no concern with the costs of training (either with foregone productivity or direct costs).
Most papers estimate the effects of on-the-job training on wages or productivity. In this
paper we estimate the internal rate of return to formal on-the-job training.1 Our data
consists of the census of large firms operating in Portugal between 1995 and 1999. The
coverage, quality and detail of our data allows us to improve the existing literature in
several dimensions. In particular, it allows us to properly account for the costs of training.
Our estimated benefits are on par with or below most of the standard estimates of
the literature. However, when we properly account for the costs of training we conclude
that the implied estimates of the internal rate of return to training for the firms providing
positive amounts of training are quite high. This implies that most of the estimates of the
returns to training in the literature implicitly hide very large estimates of rates of return to
training. Our estimated benefits for firms not currently providing training are generally
quite low or negative. Our estimates are ex-post. We do not know the expected returns the
firm faced at the time the firm made the decision of whether to invest or not invest.
1
We will consider only formal training programs. This is a weakness of most of the literature, since informal
training is very hard to measure.
The Internal Rate of Return to On-the-Job Training 127
Nevertheless, based on our estimates we argue that those firms deciding not to invest in
training behave optimally given that return to training for them is on average so low. On
the other end, those firms who offer training seem to be underinvesting given that their
estimated average marginal return to training is very high. We conjecture that one possible
reason for underinvestment may be coordination difficulties between the management and
the workforce.
Our concept of return will be the marginal internal rate of return of an investment.
Let B jt + s be the flow of marginal benefits of an additional unit of training each period and
let C Tjt be the total marginal cost of the investment in training. Assuming that the cost is all
incurred in one period and that the investment generates benefits for N periods, the
marginal internal rate of return of the investment is given by the rate r that equalizes the
present discounted value of net marginal benefits ( PDV ) to zero:
N Bt + s
PDV = −CtT + ∑ =0 (0.1)
s =1 (1 + r ) s
To estimate the total costs of training, CtT , we need information on the direct cost
of training and on the foregone productivity cost of training. The first is rarely observed in
firm data sets while the second is basically the marginal product of foregone production
time. Our data is unusually rich for this exercise since it contains information on the
duration of training, direct costs of training and training subsidies. In addition, we can also
estimate the marginal productivity of labor and, therefore, the foregone productivity cost of
training.
A major issue is to be able to identify the causal effect of training on output. In
fact, we face the standard problem of how to identify parameters of the production function
when there are missing inputs and inputs are chosen endogenously. Since we have access to
a panel of firms, we will attempt to correct for this problem through a first difference
strategy which uses past inputs as instruments for current differences in inputs, using the
strategies developed by Arellano and Bond (1991), Arellano and Bover (1995) and
Blundell and Bond (2000) for dynamic panel data models. This is also the approach taken
by Dearden, Reed and Van Reenen (2000) in a recent paper on the effects of training on
productivity and wages. From the production function we extract estimates of the marginal
product of training and the marginal product of labor. We then estimate a direct cost of
training function regressing the direct costs of training on polynomials in hours of training
allowing for firm fixed effects. From this function we get the marginal direct cost of
training.
When computing the benefits of the investment in training we recognize that what
determines current output is not the current flow training, but the stock of training.2 This
has implications both for the estimation of the production function and for the computation
of returns. It implies that training increases output contemporaneously but also in future
periods. However, human capital depreciates over time, and in our paper this can occur for
two reasons. The first is genuine depreciation of knowledge. The second is worker
turnover. We assume that workers that leave the firm take with them not only labor but also
human capital, while workers who join the firm have no human capital of the type which is
provided by company training. In our empirical work, we compute a firm specific
depreciation rate accounting for knowledge depreciation and worker turnover.
Well defined rates of return to training are almost never computed in the literature.
Papers such as Bartel (1994), Black and Lynch (1998) or Dearden, Reed and Van Reenen
(2000) report either the effect of training on wages, the effect of training on productivity or
both. In the returns to schooling literature similar quantities are often defined as rates of
return to schooling, because of the assumption that direct costs are irrelevant and that
individuals attend school full time. However, this is not a valid exercise in the context of
our paper because training is usually undertaken simultaneously with work and spend only
a small proportion of their time in training, which means that the cost of foregone work will
be small. Furthermore, direct costs of training are substantial and likely to be an important
component of total costs. Mincer (1989) makes this point very clear and Machin and
Vignoles (2001) emphasize the need to adequately consider costs in this literature.
2
Note that most of the literature estimates the contemporaneous effect of training on output. In other words,
they estimate Bt , neglecting all the future changes in output, Bt + s . Dearden, Reed and Van Reenen (2000)
are one exception since they regress output on a measure of the training stock.
As a first exercise, we compute implicit internal rates of return for some of the
papers in the literature estimating the effects of training on wages or productivity. As best
as we could, we computed the foregone productivity costs of training implicit in these
papers. Since there was no information on the direct costs of training we assumed these
were zero. The numbers we computed are of absurd size, most of them extremely high, due
to the fact that small amounts of training are estimated to generate sizeable benefits in
terms of increases in training or productivity. For example, the implicit internal rate of
return in Dearden, Reed and Van Reenen (2000) is estimated to be, at the very least, 100%;
for Bartel (1994), we estimate rates of return between 50% and 2600%; for Black and
Lynch (1998), we estimate a return of -64%. The details of these calculations are available
in the appendix. These are very simple calculations but they serve to make the point that
most of the papers in this literature have not given adequate consideration to the plausibility
of the size of the effects that they present. We could claim that their estimates are still
plausible because of the neglect of direct costs, but then these need to come into the picture
explicitly in order to assess such a claim. Careful consideration of costs is an essential part
of this whole exercise, and the main concern of this paper.
Several interesting facts emerge from our empirical analysis. The estimated
marginal productivity of training seems to be quite large: an increase in the amount of
training per employee of 10 hours per year, leads to an increase in current productivity of
0.5-1.5%. Increases in future productivity are dampened by the rate of depreciation of
human capital but are still substantial. In a rough comparison, this estimate is below other
estimates of the benefits of training in the literature (see Dearden et al, 2000, and Blundell
et al, 1996). If the marginal productivity of labor were constant (linear technology), an
increase in the amount of training per employee by 10 hours would translate into foregone
productivity costs of 0.5% of output (given that 10 hours corresponds to 0.5% of total hours
worked in a year). However, with decreasing marginal productivity of labor, foregone
productivity is even lower. In our data, marginal direct costs account on average for about
77% of the total marginal costs of training.
We then proceed to estimate the internal marginal rate of return to training
investments. We assume human capital depreciates by 17% a year3 and we estimate

average turnover for each firm. Returns vary across firms with different levels of
productivity. In particular, we estimate that (across specifications) the returns to training for
firms providing no amount of training are generally either negative or quite small in
magnitude. However, the lower bound for our estimates of the returns to training for firms
providing training is 29%. Our estimates of returns are high by any standard. However,
they are implicit in the whole literature on the returns to private training, since the estimates
of benefits are of the same magnitude or higher than the ones we estimate, and estimates of
the marginal product of labor are comparable.4
It is possible that we (and the whole literature) overestimate the benefits of
training because we do not account properly for the fact that training (as well as all other
inputs) is endogenous. However, we allow for time invariant firm effects in the production
function and we instrument current differences in training with lagged values of training
(and current levels of training with lagged differences). We further allow time variant
productivity shocks to be first-order auto correlated (as in Blundell and Bond, 2000). Our
instruments are valid instruments if the error terms in our estimating equation in first
differences are not correlated over time. We test and reject the null of no first order auto-
correlation of the error terms in first differences, but we do not reject the null of no second
order auto-correlation. Therefore, all our instruments are lagged at least two periods. We
have multiple instruments and therefore we can perform a standard overidentification test,
which (for most specifications) we do not reject at conventional levels of significance. Our
estimates of the direct cost of training function also account for firm fixed effects.
We conjecture that those firms not providing training are behaving optimally
given that their returns are so low. However, for those providing training, if returns are as
high as we estimate them to be it is puzzling why they train such a small proportion of the
total hours of work (less than 1%). We conjecture that suboptimal amounts of training may
be the result of a coordination problem. Given that the benefits of training need to be shared
between firms and workers, each party individually only sees part of the total benefit of
3
See Lillard and Tan (1986). Our attempts to estimate this parameter produced very imprecise estimates and
therefore we decided to use a number within the range of reasonable estimates in the literature.
4
Mincer (1989), the only other paper we know that performs a similar calculation of returns, finds estimates
around 17-25%. He computes the return using wages, so this number is a lower bound to the total returns,
since firms and workers should share the rents generated by training.
training. Unless investment decisions are coordinated and decided jointly, inefficient levels
of investment may arise. Furthermore, our estimates of returns are ex post, and there is
considerable heterogeneity. We do not estimate ex ante expected returns faced by firms, nor
do we estimate the risk of human capital investments. Information problems and
uncertainty may also lead firms to invest small amounts in training even though the ex post
average return is substantial. A third hypothesis is that the firm is constrained in its
investment opportunities, for example due to lack of access to credit.
The paper proceeds as follows. In section 2, we present our basic framework for
estimating the production function of the firm, and therefore compute the marginal benefits
of training and the marginal product of labor. We also briefly discuss the direct cost
function. Section 3 describes the data used. In section 4 we present our empirical estimates
of the costs and benefits of training and compute the marginal internal rate of return for
investments in training. We conclude in section 5.
2. Basic Framework
2.1. Estimating Production Functions

The most common approach used to estimate the effect of training on productivity
is to specify a production function which includes training as another input (see e.g. Bartel,
1991).5 One of the problems in this literature is that the functional form and the outcome
variable are often determined by data availability and, therefore, estimates are not directly
comparable across papers. Moreover, the measures of training used also differ across
papers (hours of training, number of trainees, indicators for the presence of training), and so
do the outcome variables for the firm (sales, output, value added, value added per worker).
Most of the literature uses the current or the one-year before level of training to measure
5
An alternative approach to measure the effct of training on productivity is to use wages (see Bartel, 1995, and
Arulampalam et al., 1997). However, changes in the average wages within a firm may be a poor indicator of
changes in the firm's productivity. For example, if training is firm specific then wages may not increase after
training. See Acemoglu and Pishke (1998, 1999) for a model where institutions that foster wage compression
provide an incentive for firms to invest in general training. Wage growth is likely to be a lower bound to
changes in productivity that is likely to be overestimated as more able workers are more likely to have higher
wage growth and are also more likely to receive training. Notice however that the lower bound is not in growth
rates. For example, Dearden et al (2001) find that wages incraese percentualy by less than output but wages
might be a lower bound even if this is not the case as they are measured in different units.
the stock of human capital in the firm.6 The latter procedure is correct only if we assume
that skills depreciate fully in one period. It is usually assumed that the production function
is either log-linear in training (e.g. Barron et al., 1989, and Black and Lynch, 1998), or
semi log-linear (e.g. Bartel, 1991, 1994, Dearden et al., 2000, and Ramirez, 1994),
Table 1 presents a summary of the variables used to measure productivity,
training, the specification and data set used in different papers.7 The diversity of these
measures creates a difficulty in comparing and interpreting the estimates across samples.
However, across studies from the U.S. and Europe there is a robust positive correlation
between a measure of formal on-the-job training and firm level productivity.
In most of the work on this topic the stock of human capital is measured using
current or lagged training. We relax this assumption allowing the stock of human capital to
depend on the past history of human capital investments, and derive the reduced form
equation that will be estimated to quantify the effect of training on productivity. We assume
that the production function is semi-log linear and that firm's stock of human capital
determines the level of current output:
Y jt = At K αjt l βjt1 N βjt2 exp(γ H jt + θ Z jt + η j + ε jt ) (2.1)
where Y jt is a measure of output in firm j at period t , K jt is a measure of capital stock,

L jt is a measure of the labor input, H jt is the aggregate stock of human capital in the firm,
l jt is the hours worked per employee, N jt is the total number of employees in the firm and
Z jt is a vector of firm and workforce characteristics. Given that the production function is
assumed to be identical for all the firms in the sample, η j captures firm heterogeneity and
6
Some papers propose regressing output growth on the period flow of training to overcome the difficulties
associated with measuring the stock of human capital (see e.g. Barret and O'Connell, 1999).
7
Bartel (1991) finds evidence for the US manufacturing firms that if a firm increases the proportion of
occupations with training by 1 percentage point then current productivity increases by 0.17%. Bartel (1994)
shows that firms that provided training to particular occupations, and did not do so before have higher
productivity growth than other firms. And vice-versa, i.e., firms that gave some training and stopped doing so,
registered decreases in productivity growth. The magnitude of the qualitative effect of training depends on the
occupation group: it varies from 17% in managers to 60% in clerical occupations. Black and Lynch (1998)
using data for the US, do not find evidence for a statistically significant effect of the number of trainees on
current or future productivity level. They find that an increase in the proportion of total hours of training off
the job by 1 percentage point increases productivity by 0.002%. Evidence for European countries, as that in
Alba-Ramirez (1994) for Spain and Barrett and O'Connell (1999) for Ireland, has also pointed to a positive
correlation between training and firm productivity. Dearden, Reed and Van Reenen (2000) find also evidence
for the UK, that sectors in manufacturing that offer training receive subsequent increases in output. Their
findings suggest that a 5 percentage point increase in industry training incidence leads to an increase in
industry labor productivity by 4 per cent and a 1.6 percent increase in hourly wages.
ε jt captures time-varying productivity shocks.

∂Y jt ∂H jt
We need to obtain estimates for two components: ∂H jt
and ∂T jt − s
for s = 1,.., N .
∂H jt
To compute ∂T jt − s
we need to make further assumptions on the functional form of the
human capital production function. In particular, we assume that the total human capital in
the firm depreciates for two reasons. On the one hand, skills acquired in the past become
less valuable as knowledge becomes obsolete and workers forget past learning. This
depreciation affects all the workers in the firm. We assume that one unit of knowledge at
the beginning of the period depreciates at rate δ per period. On the other hand, skills
depreciate because each period the worker turnover within the firm implies that, all else
constant, new workers need to acquire skills and those who leave take with them firm
specific knowledge. Therefore, we propose a human capital production function of the
following form (abstracting from j ):
′
H jt +1 = ((1 − δ )h jt + t jt )( N jt − E jt ) + X jt t jt + ω jt
H jt
where h jt is the per capita human capital in period t ( h jt = Njt
), X jt is the number of new
′
employees in period t and E jt the employees who leave the firm in period t .8 ω jt is a
firm time varying productivity shock. At the end of period t , the stock of human capital is
given by the human capital of those N jt − E jt workers that were in the firm in the
beginning of the period (these workers have a stock of human capital and receive some
training on top of that) plus the training of the X jt new workers. This specification implies
that the stock of human capital per employee is given by:
h jt +1 = (1 − δ )h jtφ jt + t jt
N jt − E jt
where φ jt = N jt +1
and 0 ≤ φ jt ≤ 1 . Comparing this equation with the permanent inventory
formula it is straightforward to see that the total skill depreciation in the model is given by
(1 − δ )φ jt . Assume each firm starts out with an initial stock of human capital given by H j 0 .
The human capital of the firm in any future period can be written as:
8
We assume that all entries and exits occur at the beginning of the period. We also ignore the fact that workers
who leave may be of different vintage than those who stay. Instead we assume that they are a random sample
of the existing workers in the firm (who on average have ht units of human capital).
t −1
H jt = (1 − δ )t φ j1 ...φ jt −1 H j 0 + ∑ (1 − δ ) s −1φ jt − s ...φ jt −1t jt − s (2.2)
s =1
H j 0 is the firm's human capital the first period the firm is observed in the sample and it is
unobservable in our data. Plugging (Human capital) into the log-linearized version of
(Production function final) for period t ∈ gives:
t −1
ln Y jt = ln At + α ln K jt + β1 ln l jt + β 2 ln N jt + γ ∑ (1 − δ ) s −1φ jt − s ...φ jt −1t jt − s + θ Z jt + μ jt + ε jt (2.3)
s =1
where μ jt = γ (1 − δ )t φ j1 ...φ jt −1 H j 0 + η j .9
Estimating equation (2.3) by least squares has some problems. There is a possible
misspecification due to the functional form chosen, even though there is no obvious
solution for this (except being non-parametric which is not feasible with so many
explanatory variables). The production function may also be misspecified in the sense that
variables that determine the level of output are omitted and uncorrelated with other
observable inputs but correlated with training. Another obstacle with using this approach is
that the human capital measure, and therefore training, may be correlated with unobserved
productivity shocks. Even though the inclusion of firm time invariant effects mitigates this
problem (see Griliches and Mairesse, 1995), this assumption would be violated if, for
example, transitory productivity shocks determine the decision of offering training.
Moreover, the sign of the bias is indeterminate as it depends on the sign of the correlation
between training and the productivity shock.10 Finding an exogenous variable correlated
with training and uncorrelated to unobserved firm productivity would help identifying the
direction of causality but, in practice, this is a very difficult task due to lack of good
instrumental variables. We take advantage of the panel dimension of our firm sample to
minimize these two problems by following an approach similar to the one that Dearden,
Reed and Van Reenen (2000) use with industry level data. To address the first obstacle we
control for firm time invariant characteristics, unobserved to the econometrician, but that
From this equation in principle it would be possible to estimate the depreciation rate δ . However, our
9
attempts to do so yielded very imprecise estimates and we chose to set δ equal to 17%, a number which is
reasonable in the literature (e.g., Lillard and Tan, 1986).
are potentially correlated with the decision to offer training. To address the second obstacle
we need an instrumental variable that is correlated with training but not with the transitory
productivity shock. We use the Blundell and Bond (1998) GMM estimator to address this
issue. This estimator exploits information in levels and in differences. It's validity is based
on the assumption that the productivity shocks are an AR(1) process. Another advantage of
this approach is that it also corrects for biases generated by measurement error in the
variables in the model.
However, the first differencing approach only works if μ jt is a firm fixed effect,
which only occurs if skills fully depreciate (δ = 1 or φ jt = 0 for all t ), or if there is no
depreciation (δ = 0) and turnover is constant ( φ jt = φ j ). If δ and turnover are positive and
time varying but smaller than 100% , then μ jt − η j is depreciates every period at rate
(1 − δ ) φ jt .11 Given δ , and φ jt it is probably possible to separately identify η j , a time
variant fixed effect, from γ (1 − δ )t φ j1 ...φ jt −1 H j 0 , a firm effect that depreciates every period
at a fixed rate. For computational reasons, we ignore this problem and approximate this
parameter by a time invariant firm effect.12 Therefore, in practice we employ a differencing
procedure that does eliminate the influence of η j , but not of H 0 , in the regression (unless
the latter is well approximated by an invariant firm effect). The extent of the resulting bias
implied by this restriction depends on the correlation between H 0 and the choice of inputs.
In future work we plan to relax this constraint.13
Blundell and Bond (2000) start from the equivalent of the following equation:
ln Y jt = ln At + α ln K jt + β1 ln l jt + β 2 ln N jt + γ H jt + θ Z jt + μ j + ε jt
and assume that

ε jt = ρε jt −1 + ϕ jt

10
If when productivity is high firms allocate more resources to training, γ is overestimated. On the other hand,
it might be when demand is low and labor is idle that firms invest more in training programs (because the

opportunity cost of labor is lower), and γ will be underestimated (Dearden et al, 2000).
= (1 − δ ) φt .
11 μt +1
Just notice that μt
12
We have an unbalanced panel of approximatly 5,500 observations of 1,500 firms. Using this approach would
imply imposing approximately 4,000 restrictions on the regression.
13
Our instrumental variables are not helpful fir this problem unless they are uncorrelated with H 0 , which again
is an unreasonable assumption.
where ϕ jt is i.i.d. Then:
ln Y jt = ln At + α ln K jt + β1 ln l jt + β 2 ln N jt + γ H jt + θ Z jt + μ j +
+ ρ ln Y jt −1 − ρ ln At −1 − ρ a ln K jt −1 − ρβ1 ln l jt − ρβ 2 ln N jt − ργ H jt − ρθ Z jt
− ρμ j − ϕ jt
or
ln Y jt = π 0 + π 1 ln K jt + π 2 ln l jt + π 3 ln N jt + π 4 H jt + π 5 Z jt +
+π 7 ln Y jt −1 + π 8 ln K jt −1 + π 9 ln l jt + π 10 ln N jt + π 11 H jt + π 12 Z jt
+υ j + ϕ jt .
This formulation imposes some restrictions between the regression coefficients,

usually called common factor restrictions. We first estimate the unrestricted model and
them impose (using minimum distance) and test the validity of these restrictions, as is
standard in this literature.
2.2. Computing the Marginal Benefits of Training and Marginal Foregone Productivity
Empirically Y jt is the firm's value added, K jt is the book value of capital, l jt is

hours worked per employee, N jt the total number of employees, Z jt includes time varying
firm and workforce characteristics as the proportion of males in the workforce, the average
age of the workforce, the distribution of the tenure on the job of the workforce,
occupational distribution of the workforce, education of the workforce (measured by the
proportion workers with high education), time, region and sector effects. t jt is measured
with the hours of training per employee.
We assume that depreciation has these two components ( δ and φ ). We assume

that δ = 17% per period14 and we estimate the turnover rate from the data. The average
turnover in the sample is 14%.15 We observe the initial and the end of the period workforce
as well as the number of workers who leave the firm. To compute the number of new hires
we assume that N t +1 = N t + X t − Et . Whenever the initial number of employees in the firm
is unobserved we assumed that φ = 0. Also, because the panel is unbalanced, we assumed
that when the firm is unobserved between any two periods, the training provided is an
14
Lillard and Tan (1986) estimate that depreciation is between 15% and 20% per year.
15
Although there is some sector variation in the turnover rates, all sectors report value above 20% per period.
average of the lead and lagged training values16. Total average skill depreciation in our
sample is 25% per period.17
From the estimates of the production function above we compute the current
marginal product of training and the marginal product of labor. We assume that the future
marginal product of current training investments is equal to current marginal product of
training minus depreciation. We measure training in hours per employee and we measure
labor in number of workers. Therefore in order to get the marginal product of an hour of
training per employee we divide the marginal product of an additional worker by the
number of annual working hours per employee in the firm and we multiply the resulting
quantity by the number of workers in the firm.
3. Direct Costs of Training
As argued in Becker (1962), training involves a direct cost (e.g., the cost of
renting the equipment in training) and a value placed on the time and effort of trainees
(foregone productivity). The latter are costs in the sense that they could have been used in
producing current output if they were not used in raising future output. Let the total training
cost is given by C Tjt = C jt + FPjt . In the previous section we saw how we compute FPjt .
The data used in this paper allows us to compute estimates of the total training
costs because there is information on total direct costs of training in a given year, C jt .
These costs include labor payments to trainers or training institutions, training equipment
as books, or movies and cost related to the depreciation of training equipment (including
buildings and machinery). We the estimate a direct cost function by regressing direct costs
of training on quadratic spline in total hours of training, where the knot point is 2500,
roughly the median number of hours of training in the sample. We also include a time
invariant firm effect in the cost function. Because we measured benefits in terms of hours
of training per employee in order to get a comparable measure of marginal direct costs of
training we multiple the marginal cost of training from the above function by the number of
16
This assumption is likely to have minor implications in the construction of the human capital variables because
there were few of these cases (approx. 3% of the final sample). Moreover, most of the firms are unobserved for
only one period.
17
In results available on request, we perform some sensitivity analysis of the rates of return with respect to the
skill depreciation.
workers in the firm (since we want the marginal cost of giving one extra hour of training to
every worker).18
Once we have marginal benefits and marginal costs we can compute individual
specific marginal internal rates of return, by solving equation (1.1) for each firm.
4. The Data
The main source of data is the “Balanço Social”, an annual survey administered
by the Portuguese Ministry of Employment covering every firm with more than one
hundred employees operating in Portugal. It is based on a mandatory survey and it has
information on different training measures, labor productivity, worker turnover, total wage
bill and direct training costs at the firm level for the period 1995-1999. Details of this
survey are given in appendix. This data set has several advantages relatively to the data
used in other studies. First, it contains employer reports of training activities within the
firm. Having information reported by the employer is better than asking the employee
about past training if the employee recalls less information about on-the-job training.19
Second, training variables are reported for all employees in the firm, not just new hires.
Third, the survey is mandatory for all the large firms in the country, representing 34 % of
the total workforce in 1995. Two problems with the empirical literature on this topic have
been the small sample sizes and the low quality of the available training data. Response
rates on surveys can be very low when surveys are not mandatory.20 Fourth, it has a
longitudinal dimension with time consistent information on training measures, total
18
Full estimates of the production and cost functions are available on request.
19
However, firm level data is likely to be a greater advantage over worker level data for informal training. In this
type of training the worker might consider as working time, the period in which he is actually being trained.
20
Bartel (1989) uses a survey conducted by the Columbia Business School with a 6% response rate. Lynch and
Black (1997) use data on the Educational Quality of the Worforce National Employers survey, which is a
telephone conducted survey with a 64% complete response rate. Barrett and O'Connell (1999) expand an EU
survey and obtain a 33% response rate.
productivity, total wage bill and direct training costs at the firm level.21 More than 50% of
the firms are observed at least twice during the period 1995-1999.22
Table 2 reports the descriptive statistics for the relevant variables in the analysis.
We divide the sample according to whether the firm provides any formal training and, if it
does, whether the yearly total training hours are above the median ( 1, 489 hours) for the
firms that provide training. Firms that offer training programs have a higher value added
per employee and are larger than low training firms and firms that do not offer training.
Total hours on the job per employee, either working or training do not differ significantly
across types of firms. High training firms also invest more in physical capital measured by
the book value depreciation of capital and have higher yearly labor costs per employee
(approximately twice as large as labor costs per employee for firms that do not offer
training).23 In general, the workforce in firms that provide training is more educated and is
older than the workforce in firms that do not offer training (the proportion of workers with
bachelor or college degrees is 8% and 5% in high and low training firms versus 2% in non-
training firms). The workforce in firms that offer training has a higher proportion of male
workers and a higher average tenure.24 These firms also tend to have a higher proportion of
more skilled occupations such as higher managers and middle managers, as well as a lower
proportion of apprentices. High and low training firms differ significantly in their training
intensity. Firms with a small amount of training (defined as being below the median) offer
on average 2.6 hours of training per employee per year while those that offer a large
amount of training offer on average 33 hours of training.25 Even though the difference
between the two groups is large, it is surprising to find such small amounts of training
21
Dearden, Reed and Van Reenen (2000) cover the period 1983-1996 but match the UK Labor Force Survey
with industry level data, which might generate aggregation biases of unknown sign.
22
The major reasons why firms leave the sample are related with firms exiting the market, reduction in
employment below 100 employees, which implies that answering to this survey is not mandatory, and to
changes in the address of the firm headquarter. This implies a change in the firm identifier so that the firm
would still be in the sample but under a different firm identifier.
23
This labor costs excludes the labor costs related to training programs.
24
Arulampalam, Booth and Bryan (2003) also find evidence for European countries that training incidence is
higher among men, and is positively associated with high educational attainment and a high position in the
wage distribution.
25
In our empirical work we use as measure of training the number of hours of training per employee. This
statistic is more informative than the total number of training hours because it controls for the fact that larger
firms have a higher number of trainees. Other useful statistics would be the number of trainees and hours of
training per trainee but the number of trainees is imperfectly measured in this data set.
overall (average yearly hours on the job 1,837 hours), even for the high training firms.
Portuguese firms with more than 100 employees that offer more training than the median,
train at most 1.8% of total time on-the-job, which is a rather small number. High training
firms spend on average almost 5 times more in direct costs of training per employee than
low training firms (41 euros per year and per employee for a low training firm versus 223
euros for a high training firm). As a proportion of value added these costs are 0.3% and
1.4% respectively. This proportion is rather small, but is in line with training such a small
number of hours. In sum, the Portuguese data is in line with surveys from other countries in
Europe and from the U.S., with respect to the profile of the firms that train. There is a lot of
heterogeneity among the firms that offer training, with the low training firms and the high
training firms being very different. Finally, the Portuguese firms train a very small
proportion of time and therefore also spend a small proportion of their value added with
formal training programs.
5. Empirical Results
As our measure of the return to training we use the standard concept of internal
rate of return of an investment. Let B jt + s be the flow of benefits each period and let C Tjt be
the total cost of the investment. Assuming that the cost is all incurred in one period and that
the investment generates benefits for N periods, the internal rate of return of the
investment is given by the rate r that equalizes the present discounted value of net benefits
( PV ) to zero:
N Bt + s
PV = −CtT + ∑ =0
s =1 (1 + r )
s
Panel A of Table 3 presents the coefficients on labor input (number of workers)

and stock of training in the production function (equation (2.3)). The different columns
present the results under different specifications. Column (I) has the lowest number of
controls and column (II) has the highest number of controls.
The point estimates of the effect of training on output are sensitive to the set of
controls that is used. However, the overall pattern is that as the set of controls becomes
richer the point estimates on the training variable decrease, while variation in the point
estimate of the coefficient on labor input does not follow a systematic pattern. The
estimated benefits in all the columns of the table seem to be quite high: a increase in the
amount of training per employee of 10 hours (approximately 0.5% of total hours worked
per year), leads to an increase in current productivity of 0.5-1.5%. This estimate is in line
with other estimates of the benefits of training in the literature (see Dearden et al, 2000, and
Blundell et al, 1996). If marginal productivity of labor is constant (linear technology), an
increase in the amount of training per employee by 10 hours would translate into foregone
productivity costs of 0.5% of output. With decreasing marginal product of labor, foregone
productivity is even lower. Our preferred estimates are the ones of column (VIII) since
those include the full set of controls. We present the remaining columns for completeness
and analysis of the sensitivity of results.
For each column we present four sets of tests. The first one is a test of
overidentifying restrictions (Hansen test). For most specifications (including VIII, our
preferred one) we do not reject these restrictions, although this does not happen for all of
them. The second test is a test of first order autocorrelation of the error term in first
differences. For every specification we reject the null of no first order autocorrelation, and
therefore all our instruments are included with a lag of 2 or higher. Across specifications,
we do not reject the null of no second order autocorrelation. Finally, the last test is a test of
the validity of the common factor restrictions, which we impose (using minimum distance)
to get the estimates in panel A of table 3. Again, for most specifications, we do not reject
the validity of the common factor restrictions.
One important contribution of our paper is to quantify the importance of foregone

costs of training. In the case of education, foregone earnings are a much more important
component of total costs than tuition and other direct costs (full time students forego 100%
of labor earnings to invest in schooling). Not much is known for the on-the-job training.
Median marginal costs of training are reported in panel B of table 3. On average, foregone
productivity accounts for less than 25% of the total costs of training. Notice that marginal
direct costs of training are much higher for firms offering no training than for firms offering
positive amounts of training, suggesting that there are large fixed costs of setting up a
training program in a firm. Panel B of Table 3 also presents, for each specification, the
median estimates of the benefits of giving all employees one extra hour. Benefits are
measured one period after the training is offered.
Our estimates of the median marginal internal rate of return for the whole sample
ranges from 0 to 50%, it ranges from -9 to 23% for the firms who do not provide training,
and it ranges from 29 to 139% for the set of firms offering training. Our preferred
specification is in column VIII, which includes the full set of controls. Notice that the larger
the set of controls included the smaller the return, suggesting that our returns may be
overestimated. With the exception of columns (I) and (IV), the returns to training for firms
not providing training are always below 5% and are negative in some cases. The reason
these firms do not offer training may be precisely because they face low returns and
therefore they may be acting rationally and optimally.
However, the returns for firms providing training are quite high, our lower bound
being of 29% and our preferred estimate being 32%.With such high returns, it is puzzling
why firms train such a small proportion of the total hours of work (less than 1%). Even
though ours may be overestimates of the true internal rate of return (as suggested before),
we believe that if such upward bias exists it should not be as high as to make these numbers
decrease considerably. The reason is that we are being as careful as possible in the
estimation of the production function, by including fixed effects and instrumenting current
inputs with lagged inputs (with a large enough lag so that our results are robust to
autocorrelation in the error terms in differences). Therefore, we conjecture that suboptimal
amounts of training may be the result of a coordination problem. Given that the benefits of
training need to be shared between firms and workers, each party individually only sees
part of the total benefit of training. Unless investment decisions are coordinated and
decided jointly, inefficient levels of investment may arise.
Furthermore, our estimates of returns are ex post, and there is considerable
heterogeneity. We do not estimate ex ante expected returns faced by firms, nor do we
estimate the risk of human capital investments. Information problems and uncertainty may
also lead firms to invest small amounts in training even though the ex post average return is
substantial. In fact, we find significant heterogeneity in the ex post returns to training,
which may indicate also a large amount of uncertainty (although in our current framework
we cannot distinguish ex ante heterogeneity from uncertainty). Focusing on our preferred
specification, the 10th and 90th percentile of the distribution of overall returns to training
are -13% and 86%, for firms providing training these quantities are -15% and 11%, and for
those providing training we have -6% and 94%.
There is a third type of explanation that is usual in this type of situations. It is
possible that firms would like to invest more in their workers but they are constrained (e.g.,
credit constrained). In that case, investments in training are likely to be suboptimal.
6. Conclusion
In this paper we estimate the marginal internal rate of return to training using firm
level data. The literature has neglected the cost side of this investment and focused on
quantifying the benefits of training. The coverage and quality of the firm level survey
allows us to improve the existing literature in several dimensions. We find that (unlike
schooling) foregone output accounts for more than 75% of the total cost of training. This is
not surprising given that in our sample firms train less than 1% of hours worked. We also
quantify the effect of training on current and future output. Even tough disentangling
correlation and causality is difficult, we use a fixed effects instrumental variables estimator
for dynamic panel data to address the problem (e.g., Blundell and Bond, 2000). In our
preferred specification we find that increasing training per employee in 10 hours (approx.
0.5% of total hours worked/year), leads to an increase in current output of about 0.5% and
to smaller increases in future output since knowledge depreciates over time. These
estimates imply that the private return to training among the firms providing training is
32%, and it is -9% for those providing no training. The former firms offer positive but low
amounts of training, in spite of large returns. We conjecture that this is due to coordination
problems between management and employees, and due to uncertainty in the returns to
training.
References
Arulampalam, W., Booth A., and Bryan, M. 2003. “Training in Europe”. University of
Essex working paper.
Arulampalam, W., Booth A., and Elias, P. 1997. “Work-related Training and Earnings
Growth for Young Men in Britain” Research in Labor Economics,16, 119-147,
Fall.
Alba-Ramirez, A.1994. “Formal Training, Temporary Contracts, Productivity and Wages
in Spain”. Oxford Bulletin of Economics and Statistics, vol 56.
Barron, J., Black, D. and Lowenstein, M.1989. “Job Matching and On-The-Job Training”
Journal of Labor Economics. vol 7.
Becker, G. 1962. “Investment in Human Capital: A Theoretical Analysis”. The Journal of
Political Economy, vol. 70, No. 5, Part 2: Investment in Human Beings, pp.9-49.
Black, S. and Lynch, M. 1996. “Human-Capital Investments and Productivity (in
Technology, Human Capital, and the Wage Structure)”. American Economic
Review Papers and Proceedings, vol. 86, No. 2, pp. 263-267.
- -. 1997. “How to Compete: The Impact of Workplace Practices and Information
Technology on Productivity” National Bureau Economic Research Working
Paper No. 6120.
- -. 1998. “Beyond the Incidence of Training: Evidence from a National Employers Survey”
Industrial and Labor Relations Review, Vol.52, no.1.
Bartel, A..1991. “Formal Employee Training Programs and Their Impact on Labor
Productivity: Evidence from a Human Resources Survey”. Market Failure in
Training? New Economic Analysis and Evidence on Trainingof Adult Employees,
ed. David Stern and Jozef Ritzen, Springer-Verlag.
- -.1994. “Productivity Gains From the Implementation of Employee Training Programs”.
Industrial Relations, vol. 33, no. 4.
- -. 1995. “Training, Wage Growth, and Job Performance: Evidence from a Company
Database”. Journal of Labor Economics, Vol. 13, No. 3, pp. 401-425.
Barrett, A. and O'Connell, P. 2001. “Does Training Generally Work? The Returns to In-
Company Training”. Industrial and Labor Relations Review, 54 (3), 647-662.
Blundell, R. and Stephen, B. (1998). “Initial Conditions and Moment Restrictions in

Dynamic Panel Data Models.” Journal of Econometrics 87: 115-143.

Blundell, R. and Stephen, B. (2000), “GMM Estimation with Persistent Panel Data: An
Application to Production Functions”, Econometric Reviews, 19.
Blundell, R., Dearden, L., Goodman, A. and Reed, H. (1997), Higher Education,
Employment and Earnings in Britain, London: Institute of Fiscal Studies.
Blundell, R., Dearden, L. and Meghir, C. (1996), The Determinants of Work Related
Training in Britain, London: Institute of Fiscal Studies.
Blundell, R., Dearden, L., Meghir, C. and Sianesi, B. (1999), “Human capital investment:
the returns from education and training to the individual, the firm and the
economy”, Fiscal Studies, 20, 1, 1-24.
Blundell, R., Dearden, L., Meghir, C. and Sianesi, B. (1999), The Returns from Education
and Training to the Individual, the Firm and the Economy: A Review of the
Theoretical and Empirical Literature, Technical Report to the DfEE Skills Task
Force, Institute for Fiscal Studies: London, 202pp.
Boheim, R. and Booth, A.L. (2001), “The Impact of Bargaining Institutions on Employer-
provided Training”, University of Essex Working Paper.
Booth, A. (1991), “Job-related formal training: who receives it and what is it worth?”,
Oxford Bulletin of Economics and Statistics, vol. 53, pp. 281-94.
Booth, A.L. and Zoega, G. (2000), “Why do Firms Invest in General Training: ‘Good’
Firms and ‘Bad’ Firms as a source of Monopsony Power”, University of Essex
Working Paper.
Dearden, L., Reed, H., Van Reenen, J. 2000. “Who gains when workers train? Training and
corporate productivity in a Panel of British Industries”. Center for Economic
Policy Research, Discussion Paper, No. 2486.
European Commission. 1994. “Continuing Training in Enterprises: Facts and Figures”.
Center for Training Policy Studies. University of Sheffield, England.
Griliches, Zvi. 1977. “Estimating the Returns to Schooling: Some Econometric Problems”.
Econometrica 45, no.1: 1-22.
Griliches, Zvi. 1994. “Productivity, R&D, and the Data Constraint”, The American
Economic Review, Vol. 84, No. 1., pp. 1-23.
Griliches, Zvi and Jacques Mairesse (1995), “Production Functions: The Search for
Identification”, NBER wp. 5067.

Heckman, J., LaLonde, R. and Smith, J.. 1999. “The Economics and Econometrics of
Active Labor Market Programs” in Handbook of Labor Economics, vol. 3A, Orley
Ashenfelter and David Card, eds. Amsterdam: Elsevier.
Heckman, James, Lance Lochner and Christopher Taber (1998), “Explaining Rising Wage
Inequality: Explorations with a Dynamic General Equilibrium Model of Earnings
with Heterogeneous Agents”, Review of Economic Dynamics, 1, no 1, pp 1-58.
Heckman, James, Lance Lochner and Petra Todd (2003), “Fifty Years of Mincer
Regressions”, NBER wp 9732.
Holzer, H., Block, R., Cheatham and Knott, J. 1993. “Are training subsidies for firms
effective? The Michigan experience”. Industrial and Labor Relations, vol 46.
Lalonde, R. (1986), “Evaluating the Econometric Evaluations of Training Programmes
with Experimental Data”, American Economic Review, 76, 4, 604-620.
Lillard, L. and Tan, H..1986. “Training: Who Gets It and What Are Its Effects on
Employment and Earnings?” RAND Corporation, Santa Monica California.
Lynch, L. 1992. “Private-Sector Training and the Earnings of Young Workers”. The
American Economic Review, Vol. 82, No. 1, pp. 299-312.
Machin, S. and Van Reenen, J. (1998) “Technology and Changes in Skill Structure:
Evidence From Seven OECD Countries”, Quarterly Journal of Economics, 113,
1215-44.
Machin, S. and Vignoles, A. (2001), “The economic benefits of training to the individual,
the firm and the economy”. mimeo. Center for the Economics of Education, UK.
Mincer, Jacob (1989), “Job Training: Costs, Returns and Wage Profiles”, NBER wp 3208.
Appendix
A. THE RETURNS TO SCHOOLING AND TRAINING
Similarly to schooling, on the job training is an investment in the individual's

human capital. Therefore, we borrow the framework from the returns to schooling literature
to compute the returns to job training. In the schooling literature the return to a year of
schooling can be estimated from a regression of log wages on years of schooling:
ln Y = α + β S + ε . (6.1)
The reason is that schooling is generally a full time activity and most of the cost of
schooling is foregone earnings. Suppose an individual has s years of schooling and is
considering whether to go to school one more year or not. If he decides to take one more
year of schooling he cannot work during that year and he foregoes the annual wage he
would have earned if had he entered the labor market with s years of schooling. Then the
return to his investment is
Y ( s + 1) − Y ( s )
R= =
Y (s)
= β.
where β is the coefficient on schooling in regression (wage return)26. Assume now that s
is years of job training and that an individual is considering whether to enrol in training one
more year. Unlike schooling, training is a part time activity and, therefore, the foregone
earnings cost of training is not Y ( s ) . If an individual spends a proportion γ of his working
time in training (and if there are no direct costs of training), the returns to training would
be:
Y ( s + 1) − Y ( s )
R= =
γY (s)
(6.2)
β
= .
γ
26
Heckman, Lockner and Todd (2003) discuss under what conditions we can interpret β as an internal rate of
return to schooling.
From the point of view of a firm deciding whether to provide training to its workers,
foregone productivity can be computed as the product of the marginal product of worker's
time and the amount of working time spent in training activities. This simple point, made
by Mincer (1989), has rarely been picked up in the literature. One of the reasons is because
the information on the percentage of time spent training within the working period is rarely
available in surveys. In our sample the average training time that would not be working
time had the training not taken place is approximately 50%. Therefore, assuming that all
the training takes place during working time would lead a significant overestimation of the
costs of training.
There is another important difference between standard analyses of schooling and
our analysis of the returns to on-the-job training. In the estimation of the returns to
schooling direct investment costs are usually neglected (one important exception is
Heckman, Lochner and Todd, 2003). This procedure is justified by the fact that direct costs
are small in absolute value because most of the education is publicly provided.
Furthermore, even when these costs are substantial (such as college tuition in private
universities) they are thought to be much smaller than the foregone earnings cost of
schooling and therefore not worth considering. This assumption is more unattractive for the
case of private training. Although some private training is subsidized, most of it is privately
financed27. Moreover, because foregone productivity is smaller in the case of training,
direct investment costs potentially account for a substantial fraction of total training costs.
In our empirical work the direct costs of training turn out to be quantitatively very
important, both in absolute terms and relative to the foregone productivity cost of training.
Finally, another important difference between the returns to schooling literature

and our paper concerns the assumption on the depreciation rate of human capital. It is
usually assumed that there is no depreciation rate for human capital acquired in school.
However, if depreciation rates are positive then the estimated returns to schooling and
27
In our sample 10% of the firms do not support any cost of training. The average percentage of costs of training
supported by the firm is 75% (100% for the median firm).
β
training are smaller than β and γ respectively.28 In our empirical work we consider
different assumptions on the depreciation rates. In particular, we relate the depreciation rate
with the rate of worker turnover within the firm.
B. IMPLICIT RATES FROM THE LITERATURE
One of the points of our paper is that we need to correctly estimate the costs of
training in order to compute meaningful returns to training. By using the data available in
several papers we can compute estimates of the implicit training costs in those papers. We
can then compute the corresponding rate of return as net benefits divided by the total costs
of training. To exemplify our argument, we compute the implicit returns to training for
three papers in this literature: Dearden, Reed and Van Reenen (2000), Bartel (1994) and
Black and Lynch (1998).
We assume that all costs occur in one initial period. The total benefit is given by
the present discounted value of all future increases in output generated by the training that
is offered in period t . Given that we do not have information on the direct costs of training
for the other samples we compute two different returns. On the first one we focus only on
the forgone output costs and, therefore, understate the total costs of training. We may be
overstating the opportunity costs of training by assuming that all the training takes place
during working hours. On the second return we assume that direct costs are the same
proportion of total costs as in our data.
Table A1 presents estimates of returns to training implicit in the three examples

we chose. The details of the construction of these estimates are available below.29 This
28
If human capital does not depreciate over time the return to training is given by expression (Return intro). An
( ) ( ) ( )
investment of γ Y ( s ) today yelds an increase in output of Y s + 1 − Y s = β Y s in every year in the
βY (s)
( )
future. Therefore: PV = −γ Y s + r , where PV is the present value of the investment in training.
The r that solves PV = 0 (internal rate of return) is given by equation (Return intro). In the extreme case
that depreciation is 100% per year the benefits of training only occur in one year. Then present value of the
β Y ( s)
29
( )
investment is given by: PV = −γ Y s + 1+ r . Solving for r we get that R = γ − 1 .
β
These studies estimate production functions. We use the coefficient on training to compute the benefits of
training and the coefficient on labor to compute the foregone productivity costs of training.
table shows that once we properly account for the costs of training the implied estimates of
returns to training implied from these studies are generally of an implausible magnitude.
B1. Dearden, Reed and Van Reenen (2000)
The specification of the production function in this paper is the following:
⎛ VA ⎞ ⎛T ⎞ ⎛H ⎞
ln ⎜ ⎟ =α +β ⎜ N ⎟ + γ ln ⎜ N ⎟ + Xθ + ε
⎝ N ⎠ ⎝ ⎠ ⎝ ⎠
where VA is value added, T is number of trainees, H is hours worked, N is number of

employees and X are other controls. In their preferred specification: β = 0.8 and
γ = 0.42 . From their table of sample statistics, where they divide firms into high and low
training firms, it is possible to obtain the following figures:
Low T High T
E ⎡⎣ln ( VA
N )⎦
⎤ 2.76 3.19
H 39.1 40.2
E( T
N ) 0.05 0.15
T
Replicating the exercise on their paper (page 49) suppose we increase N
from
10% to 15%. Then the effect on VA
N
is: 0.8 ∗ ( 0.15 − 0.10 ) = 0.04 . This is the benefit of
training.
What about the costs of this exercise? We cannot measure direct costs but we can
say something about foregone productivity. Assume that workers, when they train, they
train full time (this will give us an upper bound on the foregone productivity costs of
training). In table 3.2 of their paper, we interpret “average hours worked” as average hours
worked per employee per week: H
N
= 0.4 . By definition:
H
−ΔH = ΔT ∗
N
because the change in total production hours per week ( ΔH ) has to equal the change in the
number of trainees (individuals who stop working and start training: ΔT ) multiplied by the
H
amount of hours each one worked ( N
). Then:
⎛H⎞ ΔH
−Δ ⎜ ⎟ = −
⎝N⎠ N
ΔT H
= ∗
N N
H
= 0.05 ∗
N
which implies that:
⎛H ⎞ Δ( N )
H
d ln ⎜ ⎟ = = −0.05.
⎝N ⎠ H
N
Using their estimate for γ the effect on VA is:
⎛ VA ⎞ ⎛H⎞
d ln ⎜ ⎟ = γ ∗ d ln ⎜ N ⎟
⎝ N ⎠ ⎝ ⎠
= −0.42 ∗ 0.05
= −0.02.
VA
Foregone productivity is equal to 2% of N
. This is the cost of training.
Finally the return is:
Benefit − Cost
R=
Cost
0.04 − 0.02
=
0.02
= 100%
This estimate of returns assumes that: i) the effects of training last only for one period
(100% depreciation rate); ii) workers who train, full time, and there is no joint production
of training and output; iii) there are no direct costs of training. Relaxing assumptions i) and
ii) leads to an increase in R while relaxing iii) leads to a decrease in R .
B2. Bartel (1994)
The estimated production function that we take from this paper is the following:
⎛Y ⎞
ln ⎜ ⎟ = α + β T + γ ln N + X θ + ε
⎝N⎠
where Y is output, T is an indicator for the presence of training in the firm, N is number
of workers and X are other controls. We can rewrite this equation as
ln Y = α + β T + ( γ + 1) ln N + X θ + ε .
When T corresponds to managerial training the estimates are the following: β = 0.18 and
γ + 1 = 0.12 . T = 1 means that more than 10% of the workforce ( N ) is trained (i.e.,
d ln N = 0.10 )
Assuming that a manager trains full time (no work while in training), the effect of
one manager's work on output is 0.12 ( N is just aggregate labor), training has a 100%
depreciation rate and that the direct costs of training are 90% (average in our data) of the
total training costs, then:
0.18 − 0.012 − 0.1

R= = 50%.
0.012 + 0.1
Here we also make the assumption that teh % of direct costs on total costs is the same
across occupations.
If β Manager > β = 0.12 then R b 2 − 4ac decreases. However, if depreciation rates

are lower than 100% and/or managers train only part time, then all else constant, R should
be higher than 360%. Using other occupations besides managers we estimate:
0.27 − 0.01
RProf = = 2600%
0.01
0.55 − 0.007
Rc = = 7700%
0.007
0.2 − 0.01
RProd = = 1900%.
0.01
B3. Black and Lynch (1998)
The specification estimated in this paper is the following:
⎛O⎞
ln S = α + β ln T + γ ⎜ ⎟ + θ ln H + X η + ε
⎝H⎠
where S is sales, T is number of trainees, O is off-work hours of training, H is hours of

work and X are other controls. They estimate that: β = 0.09 , γ = 0.0002 and θ = 0.47 .
For O we can assume there is no foregone productivity cost. However, we still

need a measure of foregone productivity because not all training is done in off-work hours.
In principle we can assume that:
H = h∗N
where h is hours of work per worker and N is number of workers. From appendix D in
their paper, T
N
= 0.49 in 1993. Assuming that workers work full time (ignore O for now)
then: dT = dN (for a given size of the labor force). Then:
dT dN N
d ln T = = = ( d ln N ) ∗ .
T T T
Suppose that we increase T by 10%: d ln T = 0.1 . Then S increases by

0.1∗ 0.08 = 0.009 . We also know that d ln N = ( d ln T ) TN = 0.1 ∗ 0.49 = 0.049 . Assuming
that h does not vary in this exercise, d ln H = d ln N = 0.049 . Therefore, the reduction in
sales is: 0.47 ∗ 0.05 = 0.025.
Assuming a depreciation rate for training of 100% per period we obtain:
0.009 − 0.025
R= = −0.64 = −64%.
0.025
In this exercise we ignored off-work hours of training so our costs are too high. If we also
used a depreciation rate below 100% the estimated return would be higher. In particular, if
the depreciation rate was 0% then R = 0.009
0.025
= 0.36 = 36% .
C. THE DATA
The main source of data is Balanço Social, an annual survey designed by the
Portuguese Ministry of Employment covering every firm with more than one hundred
employees operating in Portugal30. This is the first mandatory survey collecting
longitudinal firm level data on training practices, productivity, wage bill and direct training
costs. This paper covers the period 1995-1999.
The training information concerns only formal on-the-job training, i.e., structured
training provided by the firm that is offered at the firm or at other location. Examples of
formal training may include seminars, lectures, workshops, audio-visual presentations. In
this survey there is yearly information for the number of training programs provided by the
firm, the number of trainees involved in these programs31, the total training hours and the
total costs of training. Most of this information is available for two different types of
training: internal and external to the firm, depending on whether training is offered inside
the firm or in another location. Other variables available in the Balanço Social and
collected at the firm level include the firm's regional location, ISIC five digit sector codes32,
total sales, value added, number of employees and a measure of capital, given by the book
value of capital depreciation33. Some worker characteristics available at the firm level
include average age and tenure of the workforce and the proportion of males, as well as
several measures of the firm's employment practices (such as number of hires, fires and
proportion of fixed term contracts in the firm).
The original data is composed of 2,923 firms. Due to the well known problems of
estimating productivity in non-manufacturing we restrict the analysis to manufacturing (a
total of 1, 500 firms). Table A 1 reports the sample means for the proportion of firms
providing training programs and for the training hours per employee and per trainee. On
30
Public Administration is not included but state owned firms are. The survey accounts for approximately one
third of the total private employment.
31
There is information for the yearly number of trainees but not for the number of workers enroled. These will
differ as long as the same worker participates in more than one training program per year.
32
ISIC stands for International Sector Industry Classification.
33
It is a function of the book value of the firm's capital stock. It depends on the value of capital and the book
keeping value methods used to depreciate capital in the firm's accounts. If the book value deprecition capital is
linear: BVt = π ∗ K t . It is a very imperfect measure of capital, but the only one available in this data set.
average, 53% of the firms in the sample provide some training34. Conditional on offering
training, large firms, i.e., firms with more than 400 employees, train on average 24 hours
per employee while smaller firms train approximately 15 hours. In general, sectors where
there is a higher proportion of firms providing training also train more hours of training per
employee and have a workforce with higher wages and years of schooling.
34
According to the EU report on continual training programs, the proportion if firms offering training in Portugal
in the early nineties is smaller (13%) than for the EU average (57%). Their survey covers only continual
training (while here I cover also up-front training, i.e., the training the worker receives when he is hired) and
covers both small and large firms. It is a robust finding in the literature that smaller firms are more likely to
offer training.
Table 1 (cont.)
Review of the Literature using the "Production Function Approach"
Authors Survey, Unit of Analysis and Country Training Productivity Specification Point Estimate for the Training Other
Measures Measure Variable (T ratio in brackets) Controls
Dummy if firm implemented Log sales Y(t) on T(t) Managers: 0.06 (.64)
a training program per worker Professionals: 0.02 (.02)
for an occupation group. Clerical: - 0.001 (-.01)
Age of the firm, %
Bartel Columbia Business School Survey. Production: -0.01 (-.09)
workers unionized
dummies for other
(IR, 1994) Firm and Occupation Level Data
personnel policies
Industry two-digit
U.S. Dummy if firm implemented Log sales Y(t+3)-Y(t) on T(t+3)-T(t) Managers: 0.18 (2.4)
ISIC dummies.
a training program per worker Professionals: 0.31 (3.3)
for an occupation group. Clerical: 0.6 (4.2)
Production: 0.25 (2.6)
Log Number Trainees -0.12 (-1.2) Log K, Log Materials

Log Hours, Number
plants,
Age equipment and
Black and
Educational Quality of the Log Number Trainees age firm, Log av.
Lynch
education workforce
Computer, Teamwork
(AER PP,
Workforce National Employers Survey 3 years before Log of sales Y(t) on T(t) and T(t-3) 0.09 (0.99) and Supervisor
1998)
training
Tenure, Dummy if
Firm Level Data
grades/communication
Table 1 (cont.)
skills are important
in recruiting, dummy
U.S. Percentage of Training hours for the use of TQM or
benchmarking
exports dummy,
outside working hours 0.002 (2.1) unianized
establishment.
Industry two-digit
ISIC dummies.
Log Value Added Y(t) on T(t) 0.5 (2.2) OLS

Labor Force Survey and per worker
Dearden, % turnover in t-1, log
Annual Census of Production Trainees per
Reed and capital per employee,
Log hours per
Van
Three-digit SIC Level Data Employees Log Value Added Y(t) on T(t) 0.5 (3.1) WG employee, Log share
Reenen
R&D in Y in t-1
Industry proportion of
(CEPR DP,
U.K. per worker males, age distribution
2000)
in workforce
sector occupational
distributio, % of small
firms in the sector.
Regional, time and
Log Value Added Y(t) on T(t) 1.3 (2.5) GMM
tenure dummies.
per worker
Junior Trainees per -0.01 (-0.03) Log size, Log K

Table 1 (cont.)
Log av hours worked,
Alba -
Collective Bargaining in Large Firms Junior Employees rate capacity
Ramirez
utilization
foreign and public
(OB, 1994) Firm Level Data Log Value Added Y(t) on T(t) dummy variables, tech
change dummy
% Y exported,
Spain Senior Trainees per per worker 0.30 (2.01) fraction temporary
workers
% temporary contracts
Senior Employees in newly hired
workers
% training in newly
hired temporary
contracts
% apprenticeship in
newly hired temporray
workers
% Workers in high
level managers, low
level managers
and clerical workers.
Trainees per Log Sales Y(t+2)-Y(t) on T(t) 0.099(1.8)

Employees per Employee
Change in
Barrett and
Eurostat data for 1993 and extension for 1995. Employment,
O'Connell
Investment,
(ILRR, Broad Sector
Firm Level Data Training Days per Log Sales Y(t+2)-Y(t) on T(t) 0.014 (2.2)
1999) Controls.
Table 1 (cont.)
Ireland Employee per Employee
Training Costs Log Sales Y(t+2)-Y(t) on T(t) 0.005 (0.8)

per Labor Costs per Employee
Table 2 (cont.)
Medians of Some Variables by Training Intensity
No Training Firms Low Training Firms High Training Firms

Value Added 1.934.465 3.460.467 9.313.104
Value Added p.e. 11.113 17.704 26.040
Employees 157 176 308
Total Hours p.e. 1.774 1.796 1.835
Capital Depreciation 248.035 595.769 1.563.233
Labor Costs p.e. 7.232 9.958 13.242
share low educated 0.86 0.77 0.63
share high educated 0.01 0.03 0.06
Av. Age 37 39 41
Share Males 0.4 0.6 0.7
Tenure 6 7 8
Table 2 (cont.)
Medians of Some Variables by Training Intensity
Higher Managers 0.01 0.02 0.03
Middle Managers 0.02 0.02 0.04
Intermediary Staff 0.04 0.05 0.05
High Qual. Prof. 0.41 0.42 0.43
Semi Qual. Prof 0.21 0.20 0.22
Non-Qual. Prof. 0.04 0.05 0.03
Apprenteces 0.03 0.02 0.002
Hours Training p.e. - 2 19
Training Hours / Total Hours - 0.001 0.01
Direct Cost p.e. - 17 158
Direct Cost / VA - 0.001 0.005
Observations 2.578 1.461 1.462
Source: "Balanço Social" .

Note: All nominal variables are in Euros (1995 values). Low training firms are firms offering less than 1,489 total hours of training and High training firms are firms
offering more than 1,489 total hours of training. Employees is the total number of employees in the firm. Total Hours p.e. is the yearly number of hours on-the-job
(working or training) per employee, Capital Depreciation is the book value of capital depreciation, Labor Costs p.e. is total cost of labor supported by the firm
excluding training expenditures per employee, share low educated is the share of workers with at most primary education, share high educated is proportion of
workers with bachelor or college degrees, Av. Age is the average age of the workforce, Share Males is the proportion of male workers in the workforce, Tenure is the
average tenure of workers, Higher Managers-Apprentices are the proportions of each occupational group in the total workforce of the firm, Hours Training p.e. is the
total hours of training provided by the firm per employee.
Training Hours / Total Hours is the proportion of training hours in the total hours spent at work, Direct Cost p.e. is the total direct (monetary) cost of training per
employee and Direct Cost / VA is the total direct cost of training as a proportion of the firm's value added.
Table 3
Returns, Benefits and Costs of an hour of training for all employees
I II III IV V VI VII VIII

A. Production Function Estimates
Human Capital Measure 0.0015 0.0009 0.0010 0.0005 0.0011 0.0007 0.0009 0.0005
(0.0007)** (0.0005)* (0.0004)*** (0.0004) (0.0004)*** (0.0003)** (0.0004)** (0.0003)*
Log Number of Employees 0.7591 0.7185 0.8581 0.7898 0.6142 0.6645 0.6985 0.7140
(0.1873)*** (0.1466)*** (0.1609)*** (0.1301)*** (0.1236)*** (0.1099)*** (0.1288)*** (0.1104)***
Amortization Y Y Y Y Y Y Y Y
Shares in Different Occupations N N N N Y Y Y Y
Average Age, Year, Region and Sector Controls Y Y Y Y Y Y Y Y
Share Highly Educated N Y N Y N Y N Y
Share Male N N Y Y N N Y Y
Observations 3.790 3.790 3.787 3.787 3.790 3.790 3.787 3.787

P-Value Hansen Over-Identification Test 0.393 0.375 0.022 0.024 0.483 0.667 0.358 0.291
P-Value Test for AR(1) in First Differences 0.035 0.038 0.035 0.036 0.032 0.030 0.024 0.025
P-Value Test for AR(2) in First Differences 0.168 0.238 0.115 0.157 0.159 0.163 0.128 0.137
P-Value Common Factor Restrictions 0.128 0.072 0.764 0.278 0.797 0.372 0.382 0.361
Table 3
Returns, Benefits and Costs of an hour of training for all employees
I II III IV V VI VII VIII

B. Yearly Benefits, Total Costs of Training and Average Returns to Training
Median Marginal Benefits (whole sample) 960 576 640 320 704 448 576 320
If Hours of Training = 0 601 360 401 200 441 280 360 200
If Hours of Training > 0 1.548 929 1.032 516 1.135 722 929 516
Median Marginal Direct Cost (whole sample) 863 863 863 863 863 863 863 863
If Hours of Training = 0 1.042 1.042 1.042 1.042 1.042 1.042 1.042 1.042
If Hours of Training > 0 623 623 623 623 623 623 623 623
Median Marginal Foregone Output (whole sample) 279 264 315 290 225 244 256 262
If Hours of Training = 0 132 163 194 179 139 150 158 162
If Hours of Training > 0 439 415 496 457 335 384 404 413
Median Marginal IRR (whole sample) 49% 20% 23% 0% 32% 11% 21% 0%
If Hours of Training = 0 21% 3% 5% -9% 10% -2% 3% -9%
If Hours of Training > 0 139% 76% 79% 29% 106% 56% 77% 32%
Source: author's calculations based on data from "Balanco Social".

*** Significant at 1%, ** Significant at 5%, * Significant at 10%.
Table A1
Simulating the Rates of Return of the Training Investment in the Literature - Some Examples
Authors Productivity Training Training Training Returns

Measure Measure Benefits Costs
Bartel Log sales Dummy if firm implemented 18% increase 0.012 decrease in productivity. 1400%
(IR, 1994) per worker a training program in future productivity.
for managers.
Black and Lynch Log of sales Log number of Trainees Increase of T in 10% An increase of Trainees in 10% -64%
(AER PP, 1998) increases future sales by 0.09%. decreases current sales by 2.5%.
Dearden, Reed and Log Value Added Trainees An increase in T/N from 0.1 to 0.15 An increase in T/N from 0.1 to 0.15 100%
Van Reenen per worker per Employees increases value added by 0.04 increases costs in 0.02 of VA
(CEPR DP, 2000)
Source: Authors calculations. We assume that direct costs of training are the same proportion of total costs as in the Portuguese data.
WILL THE EAST FOLLOW PORTUGAL?
Cátia Batista∗
Department of Economics - University of Oxford
November 2005
Abstract
What is the impact of joining the European Union on a small, less

developed economy? This is the general question driving this research paper. In
particular, the role of factor movements in explaining real wage behavior in Portugal
after its entry in the European Union (EU) is evaluated. Based on these results,
counterfactual exercises are performed to measure the impact of foreign investment
and emigration on skilled and unskilled wages between 1985 and 1999. We find a
small role for labor movements, and a more important one for capital inflows. This
should constitute a good starting point to think about the consequences of the Eastern
enlargement of the EU and of other integration experiences that abolish barriers to
factor mobility.
∗
I wish to thank especially Nancy Stokey, Robert Lucas and Casey Mulligan, for very helpful advice and
support. Suggestions by Fernando Alvarez, Timothy Conley, Steven Davis, Robert Shimer, and Pedro Vicente
are also gratefully acknowledged, as well as those of participants in seminars at Aarhus University, Bank of
Portugal, Catholic University of Louvain, North Carolina State University, Rutgers University, University of
Navarra, University of Oxford, University of Quebec at Montreal and at the University of Chicago. Financial
support was provided by the Portuguese Government Foundation for Science and Technology (BD
1214/2000), the Esther and T. W. Schultz Endowment Fund Dissertation Fellowship, and the Social Sciences
Division at the University of Chicago. Any remaining errors are solely my responsibility. Mailing Address:
Department of Economics, University of Oxford, Manor Road Building, Manor Road, Oxford OX1 3UQ, UK.
Email: catia.batista@economics.ox.ac.uk.
166 Cátia Batista
1. Introduction
While the process of European integration is broadening and deepening at a fast
pace, little attention has been devoted by the economics profession to explaining its real
effects. This paper contributes to the literature by examining the impact of joining the
European Union (EU) on a small, less developed economy.
We consider the case of Portugal, which joined the EU in 19861. At this time, real
wages and human capital in this country were significantly low relative to the average EU
country. In particular, in 1985, Portuguese workers who had not graduated from college
earned only 50% of corresponding French wages in PPP terms (where France may be
regarded as the average EU country), and college graduates earned 72% (also in PPP
terms), although these supplied about 2% of total hours worked. In the period from 1986 to
1999 several interesting phenomena occurred: in 1994, skilled and unskilled real wages had
already converged to 93% and 67% of French wages, respectively; and by 1999 the levels
of both skilled and unskilled wages had increased by about 55%, and educational
attainment of the labor force also increased impressively to 7% of the labor input being
supplied by college graduates.
Our purpose will be to analyze how free factor mobility (emigration and foreign
capital inflows, in particular) contributed to the observed time pattern of real wages of
skilled and unskilled workers in Portugal over the period 1986-1999. This is an important
question on its own, but we believe that its examination can also lead to useful insights on
the consequences of the Eastern enlargement of the EU which started in May 2004, as
Eastern European countries are departing from initial conditions very similar to those of the
Iberian countries in 1986.2 More generally, this paper may contribute to a better
1
Portugal actually joined the European Economic Community (EEC) in 1986, which became the European
Union (EU) only after 1993. For simplicity, the EU designation will be used throughout this dissertation.
2
According to the Economist (2004), entering Eastern European countries and the Iberian countries (Portugal
and Spain) at the time of integration in the EU both shared a fraction of about 20% of the EU population, and
represented a 10% addition to the EU's GDP (in PPP terms). These Eastern countries depart, however, from a
lower per capita GDP relatively to the EU average (46.5% vs. 62.2%, still in PPP terms). An additional
difference is the higher educational attainment of the Eastern labor force.
Will the East Follow Portugal? 167
understanding of other economic integration experiences involving free factor mobility,

such as is expected to eventually happen in the North American Free Trade Agreement
(NAFTA). Unlike trade liberalization, free factor mobility is a distinctive characteristic of
the European integration process that has remained relatively unexplored.
Our theoretical framework consists of a small economy that opens to physical
capital and labor flows, where a representative household optimally decides on human
capital accumulation and migratory movements of its skilled and unskilled members. The
setup of the model builds on the work of Stokey (1996), but includes costly migration flows
since it is also our purpose to understand labor flows in the context of integration
experiences. The technology used in the model allows equipment-specific technological
change as studied by Greenwood et. al. (1997) and distinguishes between skilled and
unskilled labor as in Krusell et. al. (2000). Following this line of research, the aggregate
production function used here has four production factors (skilled and unskilled labor,
capital equipment and structures), and does not display any type of externalities, but
constant returns to scale instead.3
Empirically, the analysis proceeds by using the actual factor inputs in the data to
estimate the parameters in the aggregate production function (and implied elasticities of
substitution) for the Portuguese economy in the period 1985-1999. For this purpose, we use
the two-step Simulated Pseudo Maximum Likelihood Estimation (SPMLE) procedure
proposed and described in detail by Ohanian et. al. (2000). Our results are supportive of the
aggregate technology displaying capital-skill complementarity, which is defined as capital
equipment being more complementary with skilled than with unskilled labor.
The estimation results are used to simulate skilled and unskilled wage series, and
to quantitatively evaluate the impact of the observed increases in emigration and foreign
capital inflows due to Portugal's entry in the EU. The first step constitutes an empirical test
3
Human capital externalities of the type introduced by Lucas (1988) have been used to explain persisting wage
differentials when labor mobility is allowed, as in Lucas (1990) and Giannetti (2003). However, this type of
externalities have generally not been found empirically significant - see Acemoglu and Angrist (2000), for
instance -, although there are some exceptions, as uncovered in local labor markets in Italy by Dalmazzo and
de Blasio (2003).
168 Cátia Batista
to our theoretical framework, which performs well and proves the important role of capital-
skill complementarity in explaining the behavior of real wages in Portugal between 1985
and 1999. In the second step in our simulations, emigration flows are found to be too small
to have a significant impact on labor returns. Foreign capital inflows (especially unilateral
transfers from the EU), however, seem to have a more important role, especially in
accounting for the growth of skilled workers' wages. In light of our results, this should be
interpreted as a result of the strong complementarity between increased capital equipment
stocks and skilled labor.
A final note should be made regarding other empirical studies that have addressed
the real effects of economic integration processes. There is a large literature on the effects
of trade on wages, but the role of free factor mobility, particularly that of migration, has not
been emphasized. Gordon (2003) is an exception, which describes the behavior of Mexican
wages after NAFTA and finds that Mexican regions with higher levels of Foreign Direct
Investment (FDI) and higher rates of migration to the US seem to be associated with greater
increases in wage levels and skill premium, which is qualitatively consistent with the
results in our work.
The absence of literature empirically evaluating the wage impacts of factor flows
is especially true regarding the integration experience of the European Union. More
abundant related literature exists on per capita GDP convergence in Europe. This started
with the work of Barro and Sala-i-Martin (1991), which, in addition to estimating the
European speed of regional convergence, also discusses the small magnitude and role of
migration in this process. It also includes the recent contribution by Caselli and Tenreyro
(2004), who discuss the per worker GDP convergence experiences of the Western countries
in the EU from 1950, and their potential extrapolation to the new Eastern EU members.
This paper is organized as follows. The next section presents the main stylized
facts of the Portuguese economy relevant to our questions and to our theoretical modelling
and empirical strategies to answer them. Section 3 describes the theoretical framework and
its implications. This is followed by section 4, in which the parameters of the technology
used in the model (mainly elasticities of substitution between capital equipment and the
two types of labor) are estimated. In section 5 wage rates and the skill premium are
simulated, and their behavior discussed in light of the capital-skill complementarity
hypothesis. The following section quantitatively evaluates the impact of increases in factor
flows due to entry in the EU on the behavior of real wages and the skill premium. The last
section summarizes the main results of this study and presents directions for future
research.
2. Brief Description of the Main Empirical Facts

This section presents the main facts of interest on the Portuguese labor market,
and also on Portuguese investment flows and capital stocks that motivate our work.
Additional details on data sources, data treatment and alternative indicators are provided in
the Data Appendix.
2.1. The Labor Market

Labor market data are from the Personnel Files (PF) collected by the Portuguese
Ministry of Labor and span the period 1985 to 1999, which includes the periods
immediately before and after Portugal's entry in the EU.
Skilled workers are defined as having completed college. The percentage of the
labor force attaining this level of education is rather small - about 7% in 1999. For this
reason, one may think that a more meaningful definition of skill would be attainment of 12
years of schooling. We perform a sensitivity analysis of the labor market stylized facts
presented in this section and of our empirical findings in the remaining of this work to this
alternative definition of skill. Our results point to minor changes in our results and, hence,
we choose to use the college definition to ease comparison with studies for other countries.
We begin by looking at the evolution of real wages.4 These progressed quite
rapidly, both for skilled and unskilled workers, especially in the period just after Portugal's
integration in the EU and until 1992, when real wage growth started to stagnate for skilled
workers, as shown in the following figure.5
4
Real wages were obtained by deflating nominal wages with a nondurable consumption price index, to avoid
quality adjustment issues.
5
This simultaneous increase of both skilled and unskilled real wages had already been documented for the
1980's and early 1990's by Gouveia and Albuquerque (1994) and Cardoso (1998).
170 Cátia Batista
Evolution of Log Wages from 1985 to 1999

(Skill Defined as College Education)
8.0
7.5
7.0
6.5
Skilled
Unskilled
6.0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Source: Own Calculations based on PF.
The labor quantities corresponding to this pattern of wage behavior emphasize a

significant human capital accumulation process. There is a very substantial increase of the
proportion of skilled (relatively to unskilled) hours worked: it went from a little more than
2% in 1985 to around 7% in 1999, as is displayed in the figure below. This pattern is
mainly due to a significant growth in the supply of skilled labor, which averaged 10% per
year in the period under analysis, while the supply of unskilled labor also increased but
only by an average annual rate of 1.4%.
Proportion of Skilled Hours Worked Relative to Total

8%
7%
6%
5%
4%
3%
2%
1%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Source: Own calculations based on PF.

Because we would like to understand how factor flows caused by joining the EU
have affected real wage behavior, one must naturally present the stylized facts that
correspond to these flows.
Regarding emigration, we consider only permanent movements - defined by a
migration period of no less than one year. This is to avoid dealing with short-term labor
movements, which have increased sharply after Portugal's entry in the EU but are not
necessarily related to typical migration decisions that we are interested in understanding.
Permanent Emigration as a Fraction of Total Workforce

(Permanent Defined as Emigration Period Longer than One Year)
0.6%
0.5%
0.4%
0.3%
0.2%
0.1%
0.0%
1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Source: Permanent Emigration Data from INE, and Estimates by Carrilho and Patricio (2003) for 1989-91; Workforce
calculated from PWT 6.1.
A striking characteristic of these emigration flows is their small magnitude relative

to the total workforce. And, indeed, as the figure illustrates, there was no strong and
immediate jump in emigration after 1986. There was only a gradual increase until 1992, the
year in which emigration attained a peak. This is probably related to the fact that legal
barriers to labor mobility were not immediately abolished when the country entered the EU:
indeed, Portugal joined in January 1st, 1986, but full labor mobility was only granted after
a transition period, which ended in December 31st, 1991.6 Note, however, that after this
period, emigration flows fell sharply to about their pre-integration levels. This is according
6
For more details on the accession clauses the reader is referred to European Commission (2001, pp. 14 - 16).
172 Cátia Batista
additional data from SourceOECD showing that the stock of Portuguese immigrants in EU
countries fell from 18.8% to 18% of the total Portuguese workforce.
From the 90s, the most important migratory phenomenon in the Portuguese labor
market has indeed been very sizable immigration, and no longer emigration. Even though
this phenomenon was not directly created by Portugal's entry in the EU (since most
immigrants came and are coming from outside the EU, and do not, therefore, benefit from
the European open borders), this may be regarded as its indirect consequence. As is clear
from the figure, over our period of interest immigration flows almost offset all emigration
that occurs.
Net Migration as a Fraction of Total Workforce

(Including Permanent and Temporary Migration)
1.00%
0.50%
0.00%
1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.50%
-1.00%
-1.50%
Source: Net Migration Data from INE; Workforce calculated from PWT 6.1.
2.2. Investment Flows and Capital Stocks

We now pay attention to the evolution of capital stocks and associated investment
costs. Following Krusell et. al. (2000), we distinguish between two types of capital:
structures and equipment. The reason to do so is that, especially from around 1980, there
have been enormous quality improvements in equipment capital, in simultaneous with
considerable equipment price decreases. The true magnitude of these decreases can be
captured by performing a quality change adjustment in the equipment price series, in the
spirit of Gordon (1990). The impact of our quality adjustment can be appreciated in the
following table and figure:
Average Annual Change in Price of Equipment

1954-99 1985-99
(Relative to NondurableConsumption)
Quality Adjusted -3.81% -6.37%
NIPA -1.11% -3.59%
Source: Own calculations based on BoP, Long series for the Portuguese Economy, and INE, Quarterly National
Accounts Bulletin 1995-1999.
By any measure, it is clear that the relative price of capital equipment has been
declining over time, which is especially true between 1985 and 1992. This may be
interpreted as technological change specific to the production of capital equipment, as
investigated by Greenwood et. al. (1997), which is corroborated by the approximate
constancy of the relative price of structures over our period of analysis, as the following
figure makes clear.
Annual Growth Rate of Investment Good Prices

(Relative to Nondurable Consumption)
4%
2%
0%
1985 1987 1989 1991 1993 1995 1997 1999
-2%
-4%
-6%
-8%
-10% Equipment - Quality Adjusted

Structures
-12%
-14%
Source: Own calculations based on BoP, Long series for the Portuguese Economy,
and INE, Quarterly National Accounts Bulletin 1995-1999.
Another reason to distinguish between capital equipment and structures is the fact
that equipment stocks have experienced impressive growth rates especially between 1986
and 1993, as is patent in the following figure, where both capital stocks are expressed in per
worker terms, and equipment stock reflects adjustment for quality improvements.
174 Cátia Batista
Annual Growth Rate of Capital Stocks Per Worker

(Skill defined as College Education)
15%
10%
5%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998
-5%
Equipment per Worker

-10% Structures per Worker
Equipment per Skilled Worker
-15%
Source: Own calculations based on PF, BoP - Long series for the Portuguese Economy , and INE - Quarterly
National Accounts Bulletin 1995-1999 .
This figure also displays the behavior of the equipment to skilled labor ratio. This
was, on average, increasing from 1986 to 1993, but started to fall significantly from 1992.
This piece of evidence seems to point in the direction of the capital-skill complementarity
hypothesis (to be formally defined in the following section) as a good candidate to
explaining the behavior of skilled wages, which also grew strongly until around 1992, but
not thereafter.
We now turn to consider foreign capital inflows.7 Foreign Direct Investment (FDI)
displays a rather irregular pattern over time, as can be observed from the figure below.
After an explosive increase, unprecedented in the Portuguese history, FDI fell sharply
around 1992. According to Lopes (1999) this should be attributed to the recession that
affected Europe at that time, as well as to the Portuguese escudo real appreciation.
7
These flows are all measured in net terms: this is especially important for the case of FDI, where it could be
argued that the increase observed after integration was due to the world trend of capital movement
liberalization.
Foreign Direct Investment as a Fraction of GDP
5%
4%
3%
2%
1%
0%
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
Source: Bank of Portugal: Long Series for the Portuguese Economy (1965-1995) and Balance of Payments Statistics.
Note that FDI likely underestimates foreign investment inflows, which may be
included under foreign portfolio investment, for instance. These were, however, extremely
volatile during our period of interest.
In addition to FDI, we consider unilateral transfers from the EU, which started in
1986 and steadily increased also until 1992. From this year, they remained relatively stable
over time, as is depicted in the figure below.
EU Unilateral Transfers as a Fraction of GDP

4%
3%
2%
1%
0%
1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Source: Bank of Portugal: Long Series for the Portuguese Economy (1965-1995) and Balance of Payments Statistics.
176 Cátia Batista
To summarize, we believe the stylized facts for Portugal to be quite interesting

and distinct from those of other studied OECD economies, in terms of fast human capital
accumulation, skill premium behavior8, and physical capital accumulation.
We have unveiled two distinct periods, which are distinct not only in terms of
skilled wage growth (first increasing, then stagnated), but also in terms of approximately
contemporaneous comovements in stocks and prices of equipment, and skilled labor-
equipment ratio. Note that the sharp growth of FDI and transfers from the EU until 1992
helps explaining why total investment growth was so fast in that period, since in addition to
their direct effect, domestic investment (both private and public) likely increased to
complement foreign investment.
3. Model of a Small Economy Open to Factor Flows

The model in this section describes the process of a small economy opening to
capital inflows and labor outflows, as experienced by Portugal when entering the EU. It
aims at reconciling the pieces of empirical evidence presented in the previous section in a
one-country model economy in the spirit of section 2 in Stokey (1996), which opens to
costly migration and capital flows.
The technology in this model displays capital-skill complementarity in the sense
first formalized by Griliches (1969): skilled labor is more complementary with capital than
unskilled labor.9 This property of the aggregate production function is formalized in
Stokey (1996), on which Krusell et. al. (2000) build to develop a more complex functional
form that is empirically tested and found to have strong support in the US data. This is the
functional form adopted in our model.
In particular, the aggregate production function distinguishes between the role of
four different inputs, which agrees with the observed empirical facts The distinct evolution
8
Recall that Davis (1992) studied the skill premium behavior in the 1980s in nine OECD economies, and also in
Brazil, Colombia, South Korea and Venezuela. He found increasing skill premia in seven of the OECD
economies and in Brazil.
9
For more details on how capital-skill complementarity has been modelled in the literature, in particular with
attention to its role in explaining income inequality, the reader is referred to Batista (2002).
of both quantities and prices of capital structures and equipment motivated the inclusion of
these two types of capital. The observed pattern of human capital growth and its associated
pattern of skill premium lead us to differentiate between skilled and unskilled labor.
Stokey (1996) develops a fully specified general equilibrium model, under which
attention is paid to three cases: a benchmark closed economy, capital inflows in a small
open economy, and free trade of goods following economic integration of the small open
economy in a more developed area. This framework allows addressing the issue of the
development of an economy evolving from autarchy to complete economic integration of
the type allowed by NAFTA. However, because integration in the EU also involves
abolishing barriers to labor mobility, this model needs to be extended to allow for labor
movements.
In modelling labor flows, the empirical importance of migration costs10 is matched
by the introduction of a migration cost parameter, which should be interpreted as unofficial
migration barriers, travelling and differential housing costs, psychological and other costs
borne by emigrants related to the loss of local amenities.11 For simplicity, these costs are
assumed to be common for both skilled and unskilled workers.
The following sections set the model up in more detail.
3.1. Technology
In this economy, production of homogeneous output at date t , Y (t ), occurs
according to a homogeneous of degree one aggregate production function with four

production factors, as in Krusell et. al. (2000):
10
This empirical role has been studied in the context of the small magnitude of internal migration within
European countries. Namely, the fact that migration does not seem to respond to shocks (that cause, for
instance, increased regional unemployment) seems to point to the presence of significant costs, which should
be regarded as a lower bound to migration costs of international migration within Europe, given the language
and cultural differences across countries. References are, for Italy, Attanasio and Padoa Schioppa (1991), Faini
et. al. (1997), and Cannari et. al. (2000); and, for Spain, Bentolila (1997), Antolin and Bover (1997), and Bover
and Velilla (1999).
11
Note that these costs are distinct from "assimilation costs" derived from language and cultural differences that
affect a worker's productivity in the host country, relative to nationals of that country. Sjaastad (1962) presents
the seminal analysis of migration as a human capital investment decision, clearly discussing the role of
different types of costs and benefits of migrating.
178 Cátia Batista
[
Y (t ) = A(t ).F K s (t ), K e (t ),U (t ), S (t ) ]
where K e (t ) and K s (t ) denote, respectively, the services of capital equipment and
structures used in production12; U (t ) and S (t ) stand, respectively, for the unskilled and
skilled labor inputs to production at date t 13; and A(t ) is a country-specific Hicks neutral
technological factor, which is introduced to account for the type of cross-country TFP level
differences highlighted by Hall and Jones (1999) and Parente and Prescott (2000).
As noted by Greenwood et. al. (1997) and Krusell et. al. (1997), such an aggregate
production function may be thought of as representing a two-sector economy, where one
sector produces capital equipment goods, and the other produces capital structures and
consumption goods. Assume that there is a common neutral technology factor A(t ) and a
common production function F , but the equipment sector has a specific technology factor
q (t ) , which defines the relative price of equipment in terms of consumption goods
( 1 / q (t )) . Then, as long as F is homogeneous of degree one and factor markets are
perfectly competitive, profit-maximizing firms will allocate the inputs so that input ratios
will be equalized across sectors and the equilibrium relative price between consumption
and equipment will be q (t ) . The two sectors may under these conditions be aggregated,
yielding
X e (t )
Y (t ) = C (t ) + X s ( t ) +
q (t )
where C (t ) represents consumption good output, X e (t ) stands for investment in capital
structures, and X s (t ) denotes output of the equipment sector
12
The services of both types of capital are assumed to be proportional to the original capital stocks, measured in
efficiency units.
13
We assume that U (t ) and S (t ) add to L, the national workforce of the country , which is, for simplicity,
taken as constant over time, i.e. number of retiring people equals number of workers entering the force. This
implies that we are abstracting from the existence of unemployment in this economy.
The choice of the functional form for the aggregate production function, F ,
pursues the objective of compatibility with the capital-skill complementarity hypothesis,
i.e. the elasticity of substitution between capital and unskilled labor being larger than that
between capital and skilled labor, as suggested by the data. We adopt the functional form
used by Krusell et. al. (1997, 2000), which these authors verified to be consistent with
available elasticity estimates and to fit well US data, and use the following specification:14
α
Y (t ) = A(t ).K s (t ) .Y (t )
1−α
where
1
⎡ θ θ
ρ ρ⎤
θ
( ρ
)
Y (t ) = ⎢ μ .U (t ) + (1 − μ ) λ .Ke (t ) + (1 − λ ) .S (t ) ⎥
⎣ ⎦
and
0 < α , μ , λ < 1 and θ , ρ < 1
This is just a Cobb-Douglas production function in capital structures K s and the

aggregate Y , which nests two Constant Elasticity of Substitution (CES) production
functions. This nesting implies that the elasticity of substitution between unskilled labor
and capital equipment is the same as that of unskilled labor and skilled labor, 1 / (1 − θ ) ,
but this potentially differs from the elasticity of substitution between skilled labor and
capital equipment, which is given by 1 / (1 − ρ ) . Capital-skill complementarity therefore
translates in this model simply as θ > ρ , meaning that capital is more complementary with
skilled than with unskilled labor.
14
One possible alternative was Y   1 K,  2 U, S. However, this form does not give rise to capital-skill
complementarity, but to the technology-skill complementarity modelled by Heckman, Lochner and Taber
(1998), as noted by Greenwood and Jovanovic (1999). One other possible alternative would be
Y = Φ ( S , Φ (U , K ) ) , as proposed by Stokey (1996). This implied, however, that the elasticities between
1 2
skilled labor and equipment and between skilled and unskilled labor should be equal, which is contrary to the
empirical observation that the former is smaller than the latter.
180 Cátia Batista
This function may be interpreted, in the spirit of Goldin and Katz (1998), as a two-
stage production function: in a first phase, there is machine installation and maintenance
(skilled workers adopt new technologies and ensure they work efficiently in the
organization); then, there is the machine assembly and production stage (which has
unskilled workers using the product of the skilled labor effort and equipment).
Rewriting the production function in per (national) worker terms, we obtain
α
[ ]
y (t ) = A(t ). f k s (t ), ke (t ), z (t ) = A(t ).k s (t ) .
1
θ
⎡ θ ρ ρ ⎤θ
( ρ
. ⎢ μ (1 − z (t ) ) + (1 − μ ) λ ke (t ) + (1 − λ ) z (t ) ⎥ )
⎣ ⎦
where
Y (t ) S (t ) K i (t )
y (t ) ≡ ; z (t ) ≡ ; ki (t ) ≡ , i = s, e
L L L
Competitive firms' profit maximizing choices determine the following interest
rates and level of skill premium:
[ ]
rs (t ) = A(t ). f1 k s (t ), ke (t ), z (t )
re (t ) = A(t ). f2 [ k s (t ), ke (t ), z (t )]
wz (t ) − wu (t ) = A(t ). f3 [ k s (t ), ke (t ), z (t )]
where rs (t ) and re (t ) denote the rental rates of capital structures and equipment,
respectively, and wz (t ) and wu (t ) stand for the wage rates paid to skilled and unskilled
workers, respectively.
3.2. Introducing Migration

Recall that we are assuming that there are no restrictions to factor mobility in this
country.
∗
If (exogenous) foreign wages, wi (t ), i = z , u , are initially high enough so that the
representative household optimally chooses positive emigration, we can define m z (t ) and
mu (t ) as the fractions of the labor force L that are (stocks of) skilled and unskilled
emigrants, respectively. These are chosen by the representative household at each date. In
their optimization problem, families must consider the cost ψ incurred when migrating.
This cost can be thought of as cultural/psychic separation costs and transportation and other
pecuniary costs. For simplicity, ψ is assumed to be the same regardless of skill level.
The possibility of emigration implies that production will be given by
[
y (t ) = A(t ). f k s (t ), ke (t ), z (t ), mz (t ), mu (t ) ]
1
θ
α ⎡ ρ ρ ⎤θ
θ
(
= A(t ).k s (t ) . ⎢ μ (1 − z (t ) − mu (t ) ) + (1 − μ ) λ ke (t )
⎣⎢
ρ
+ (1 − λ )[ z (t ) − m z (t )] ) ⎥
⎦⎥
where m z (t ) and mu (t ) are taken as given.
This implies the same profit maximizing conditions for competitive firms as
before:
[
rs (t ) = A(t ). f1 k s (t ), ke (t ), z (t ), m z (t ), mu (t )]
re (t ) = A(t ). f2 [ k s (t ), ke (t ), z (t ), m z (t ), mu (t )]
wz (t ) − wu (t ) = A(t ). f3 [ k s (t ), ke (t ), , z (t ), m z (t ), mu (t )]
Given the homogeneity of degree 1 property of the aggregate production function

in both countries, Euler theorem implies that domestic product per worker is equal to:
[ ]
y (t ) = wz (t ). [ z (t ) − m z (t )] + wu (t ) 1 − z (t ) − mu (t ) + r (t ).k (t )
3.3. Preferences
Households are assumed to supply labor inelastically and to privately finance all
investment in human capital. Accumulation of human capital is subject to adjustment costs,
contrary to physical capital accumulation. This is an unrealistic simplifying assumption, but
it should not affect the pattern of transitions, only exaggerate its speed. Adjustment costs
are introduced by the parameters 0 < φ , B < 1 . Human capital depreciates at the rate
0 < η < 1, which includes the effects of retirement. In addition to choosing optimal
emigration paths, the representative households must determine consumption per worker
c(t ) , and investment in human capital z (t ) .
182 Cátia Batista
Free capital mobility implies that the path of both types of physical capital will be
a function of the households' choice for z , m z , and mu , in a way that their marginal
productivity always exactly equals the corresponding depreciation rate ( δ s and δ e ,
respectively  plus the world interest rate ( ρ ) .
A. f1 ⎡⎣ ks ( z , mz , mu ) , ke ( z , mz , mu ) , z , mz , mu ⎤⎦ = δ s + ρ (1)
A. f 2 ⎡⎣ k s ( z , m z , mu ) , ke ( z , m z , mu ) , z , m z , mu ⎤⎦ = δ e + ρ (2)
Because the representative household can borrow and lend at the world interest
rate, its investment and consumption decisions can be thought of as separated. The
investment problem therefore corresponds to the choice of the human capital and migration
levels that maximize the discounted labor income stream net of migration and human
capital investment costs:
⎡ wz (t ). [ z (t ) − mz (t )] ⎤
⎢ ⎥
∞ − ρ t ⎢ + wu (t ).(1 − z (t ) − mu (t )) ⎥
max ∫0 e ∗ ∗ dt
{ I z (t ), mz (t ), mu (t ), t ≥0} ⎢+ wz (t ).mz (t ) + wu (t ).mu (t ) ⎥
⎢ − I z (t ) − ψ . [ mz (t ) + mu (t )] ⎥
⎣ ⎦
φ
s.t . z (t ) = BI z (t ) − η . z (t )
{ ∗ ∗
taking z0 , wz (t ), wu (t ), wz (t ), wu (t ), t ≥ 0 as given. }
Optimality requires the following conditions to hold:
wz (t ) = wz∗ (t ) −ψ if mz (t ) > 0 (3)
∗
wu (t ) = wu (t ) − ψ if mu (t ) > 0 (4)
1 φ
1−φ
z (t ) = B [φμ (t )]1−φ − η.z (t ) (5)
μ (t )
[ ]
1
=η + ρ − . A. f3 k s ( z (t ), m z (t ), mu (t )), ke ( z (t ), m z (t ), mu (t )), z (t ), mz (t ), mu (t ) (6)
μ (t ) μ (t )
where μ (t ) is the co-state for z (t ) .

The consumption problem solved by households is to maximize lifetime utility,

subject to their lifetime earnings:
1−σ
∞ − ρ t c (t ) −1
max ∫0 e dt
{c(t ), t ≥0} 1−σ
⎡ wz (t ).[ z (t ) − mz (t )] ⎤
⎢ ⎥
∞ − ρt ∞ − ρ t ⎢ + wu (t ).(1 − z (t ) − mu (t )) ⎥
s.t. ∫ 0 e c (t ) dt ≤ ∫ 0 e ∗ ∗ dt + k s ( 0 ) + ke ( 0 )
⎢+ wz (t ).mz (t ) + wu (t ).mu (t ) ⎥
⎢ − I z (t ) − ψ .[ m z (t ) + mu (t )] ⎥
⎣ ⎦
given k s ( 0 ) , ke ( 0 ) .
The solution to this problem is given by a constant stream of consumption ( c )

which has a discounted value equal to the maximized value of the objective function of the
investment problem, V ( z ( 0 )) .
c
= V ( z ( 0 )) + k s ( 0 ) + k e ( 0 )
ρ
where t = 0 corresponds to the date of opening the economy to factor flows.

The steady-state values for human and physical capital are the same as before the
economy opened to factor flows, and are given by:
[ ]
A. f1 k s , ke , z , m z , mu = δ s + ρ
A. f 2 [ k s , ke , z , m z , mu ] = δ e + ρ
⎛ −1
φ 1−φ ⎞
−1
[ ]
A. f3 k s , ke , z , m z , mu = (η + ρ ) ⎜ φ
⎜
B [η .z ] φ ⎟
⎟
⎝ ⎠
where an upper bar denotes a steady-state value.
These steady-state levels are set so that the productivity of all types of capital
exactly offsets depreciation and discount rates - adjusted for the impact of adjustment costs
in the case of human capital.
The transitional dynamics for this economy are given by (lomz) and (lommu) for
human capital and its co-state. The two types of physical capital adjust so that their rate of
184 Cátia Batista
return keeps constant given household's choices for human capital and migration, as
described by (loms) and (lome).
At this stage, the theoretical analysis performed qualitatively describes how
opening economies to factor flows is enough to ensure cross-country wage convergence up
to a wedge (created by migration costs ψ ) in an initial stage of transition. This is the result
of a decision of households to migrate immediately after this becomes allowed in order to
equalize wages net of migration costs. During transition to the steady-state, however, as
human capital accumulation leads national wage differentials to fall below ψ , there will be
return migration.
Quantifying these effects is not straightforward provided that the parameters in the
production function (and implied elasticities of substitution) have not been estimated for
Portugal, and that there are no measures available in the literature for ψ . This dictates our
next steps, beginning with the estimation of the technology parameters.
4. Estimating Aggregate Elasticities of Substitution

We estimate the parameter values of the aggregate production function using
actual data for productive inputs, as well as the first order conditions of profit maximizing
firms acting in competitive factor markets, in the tradition of Griliches (1969).15 This
empirical strategy builds on the work of Krusell et. al. (1997, 2000), Ohanian et. al. (2000),
and Greenwood et. al. (1997).
4.1. Econometric Model

We begin by specifying the stochastic nature of the production function.
Labor inputs in our model are defined in efficiency units - such that
t(
S (t ) ≡ ψ s (t ).hs (t ), and U (t ) ≡ ψ u (t ).hu (t ). The efficiency indices ψ s and ψ u
t ) are
15
Griliches (1969) argued that substitution elasticities between different factors of production should be
estimated using derived demand equations instead of the production function.
not observed by the econometrician, and Ohanian et. al. (2000) argue in favor of modelling
⎡ lnψ s (t ) ⎤ as a trend stationary process:
ϕt ≡ ⎢ ⎥
⎣lnψ u (t ) ⎦
ϕt = ϕ0 + γ .t + ϖ t
i.i.d . ⎡γ ⎤
where ϖ t ∼ N (0, Ω ), and γ ≡ ⎢ s ⎥ is the vector of constant growth rates of labor
⎣γ u ⎦
efficiency.16
Given this stochastic specification of the production function, we can now turn to
detailing our econometric model. Our procedure will be to use restrictions implied by the
theoretical model in order to achieve identification of its parameters, which may be
implemented by using the following nonlinear state-space model:
Zt = g ( X t , ϕt ; φ ) + et (Observation equations)
ϕt = ϕ0 + γ .t + ϖ t (State equations)
with observed variables { t t t t }

X t ≡ K s , K e , hs , hu , parameters
⎡0⎤
{ }
i.i.d .
φ ≡ δ s , δ e , α , μ , λ ,θ , ρ ,ηe , ϕ0 , γ , Ω , and shocks et ≡ ⎢⎢ 0 ⎥⎥ , where ε t ∼ N (0,ηε2 ).
⎢ε ⎥
⎣ t⎦
The observation equations summarize the model restrictions: Zt displays the data
counterparts to the expressions derived from the model, g ( X t ,ϕt , et ;φ ) . Since we have
data on hours worked and on corresponding wages, we can use the representative firm's
first-order conditions for skilled and unskilled labor to construct the theoretical counterparts
to the labor share of income and to the wage-bill ratio - corresponding, respectively, to the
first and second observation equations. We do not, however, possess reliable measures for
16
Ohanian et. al. (2000) discuss the details implied by using this trend stationary specification, as well as
potential alternatives and their consequences.
186 Cátia Batista
the rental rates of capital structures and equipment goods. Therefore, the third observation
equation will be a no arbitrage condition, imposed to ensure that the difference between the
model counterparts to these unobserved rates is close to zero.
We will therefore have
⎡ shlt ⎤
Z t ≡ ⎢⎢ wbrt ⎥⎥ (7)
⎢⎣ 0 ⎥⎦
and
⎡ ⎤
⎢
⎢
( )
wst .hst + wut .hut / yt ⎥
⎥
g ( X t , ϕt , et ;φ ) ≡ ⎢
⎢
( )(
wst .hst / wut .hut ) ⎥
⎥
(8)
⎢⎡ ⎡ ⎡ q ⎤ ⎤ ⎥
⎢ ⎢ rs t +1 + (1 − δ s ) ⎥⎤ − ⎢ qt re t +1 + Et ⎢ q t ⎥ (1 − δ e ) ⎥ ⎥
⎣⎣ ⎦ ⎣ ⎣ t +1 ⎦ ⎦⎦
where qt is defined as the price of equipment relative to nondurables.
To simplify, we assume Et ⎡⎢ q t ⎤⎥ (1 − δ e ) = q t (1 − δ e ) + ε t , where ε t is the

q q
⎣ t +1 ⎦ t +1
forecast error. Note that, since the relative price of equipment has been falling over time,
qt
q (1 − δ e ) + ε t corresponds to the expected capital loss on equipment after
t +1
depreciation.17
An additional simplifying assumption will be to abstract from any time trend in
the efficiency levels of skilled and unskilled labor. The state equations in our estimation
will then be given by ϕt = ϕ0 + ϖ t .
17
To express it in this form, we follow KORV and assume that At +1 and ϕt +1 are known when investment
decisions are made, qt +1 being the only unknown at that date.
Moreover, we abstract from the potential existence of a risk premium (which enables us to ignore the
covariance between consumption and capital returns), and take the tax treatment on structures and equipment
investment to be identical.
4.2. Estimation Procedure
4.2.1. Preliminary Parameter Calibration

Since we only have 15 years of observations for all relevant variables, we must
reduce the dimensionality of the parameter vector (15) before proceeding with the
estimation. This is achieved by means of calibrating some parameters and of imposing
additional restrictions (not derived from the model) based on available a priori information.
First, we use the Greenwood et. al. (1997) values for the depreciation rates, which
we have already used to construct the capital stock series: δ s = 0.056 and δ e = 0.124.
We then need to calibrate ηε . For this purpose, we performed some diagnostic

q
tests (Phillip-Perron) and concluded that the hypothesis that q t was covariance
t+1
stationary could not be rejected even at a 1% confidence level. After some more diagnostic
checks (ACF and PACF), we concluded it would be best represented as an ARMA(1,2)
model. From the estimation of this model, we obtained an estimate for the residuals'
standard variation of σˆε = 0.042. Multiplying this standard deviation by 1 −  e  , we
obtained our estimate for ηε = 0.0396. 18
Following Krusell et. al. (1997), one additional simplifying restriction that we
impose is that the shocks to skilled and unskilled labor efficiency are uncorrelated and have
common variance, which is very much in accordance with our results from estimating
ARMA processes for both skilled and unskilled labor inputs. This implies that we can write
2
Ω = ηϖ ∗ I 2 , where ηϖ is the common innovation standard deviation, which takes a
value of 0.2 .
One other restriction is necessary to set the scale of the model (which is not a
model of levels of the variables). We could normalize μ , λ , ϕ s 0 or ϕu 0 . Still following
18
Note that, in order to increase the accuracy of the estimation procedure, we used at this stage our equipment
relative price measure (adjusted for changes in quality) for the longer period 1953-1999.
188 Cátia Batista
Krusell et. al. (2000), we choose to normalize the initial level of skilled labor efficiency,
ϕs 0 , to zero.
Finally, we follow the procedure proposed by Lindquist (2004) to calibrate ,

the average yearly share of national income for capital structures. For this purpose, we
assume a common return for capital equipment and structures, r = re = rs , which we take
to be 6%, consistently with the findings reported by Citron and Walton (2002). This implies
that
1 1999 ⎡ ( r + δ s ) kst ⎤
α= ∑ ⎢ ( 1 − shlt )⎥ = 0.28
15 t =1985 ⎣ ( r + δ s ) k st + ( r + δ e ) ket ⎦
In summary, we will consider the following calibrated values:
α δs δe ηε η w ϕ s 0 γ st γ ut
0.28 0.056 0.124 0.04 0.2 0 0 0
In this manner, we have managed to use our a priori information to reduced the
dimension of our parameter vector to five ( μ , λ , θ , ρ and ϕu 0 ). Although this
calibration of parameters naturally adds uncertainty to the overall empirical exercise, which
we do not take into account when evaluating our estimation results, we believe it to be
advantageous because of the degrees of freedom it also brings.
4.2.2. Simulated Pseudo Maximum Likelihood Estimation (SPMLE)

Ohanian et. al. (2000) show that, when unobserved variables are multivariate
Gaussian trend stationary processes, Simulated Pseudo Maximum Likelihood Estimation
(SPMLE) of our non-linear latent variable model, based on an approximated likelihood
function constructed from the first and second moments of the dependent variables,
produces parameter estimates with negligible bias in small samples, with computational
efficiency. Following their work, we decided to also perform a two-stage version of

SPMLE.19
The two-step version of SPMLE introduced by White (1994), Ch.7, takes into
account potential endogeneity of the labor inputs (which may respond to contemporaneous
shocks in technology and/or labor quality). We assume capital stocks of both equipment
and structures to be predetermined in this sense, given their large magnitude and small
volatility relative to investment (which makes them less likely to respond sizably to i.i.d.
shocks).
A more detailed explanation of the estimation procedures is provided in appendix.
We next present a brief summary.
In the first step, we pay attention to the issue of potential endogeneity of the labor
inputs by predicting skilled and unskilled hours worked based on a simple regression on
current and lagged stocks of equipment and structures, lagged relative price of equipment, a
business cycle indicator and a time trend:
hi = β1 + β 2 K e + β3 K e + β 4 K s + β5 K s + β 6 qt −1 + β7t + β8 BCt −1 + vt , (9)

t t t −1 t t −1
for i = st , ut , t = 1, ..., 15
From these estimations we obtained the fitted labor inputs, which will be used in
the second step, instead of the original labor inputs.
In this second step, we implement SPMLE. We start by performing 500 simulated

i
draws for ε t and ϖ t from their assumed distributions, for each date t. 20 We construct ϕt
for i = 1 t = 1 ...15 , which allows us to compute f ( X t ,ϕti , ei;φ ) for i = 1
19
The SPMLE method is due to Laroque and Salanie (1989). Its basic idea is that instead of using an intractable
original likelihood, one may simulate the associated empirical moments necessary to construct a simpler
function.
20
The number of simulations we performed is the same used by Krusell et al. (2000). Its choice was determined
by the study conducted in Ohanian et al.(2000), which asserted that 50 simulations were enough to ensure that
the mean bias was essentially zero in our model for the key curvature parameters  and  , and also by the
fact that from this number of simulations on the estimated parameters suffered a negligible change due to
simulation uncertainty.
190 Cátia Batista
500, t = 1 ...15. We then use this computation to calculate the first and second sample
moments of the model's counterparts to Z t , which are given by:21
1 500
m500 ( Xt ;φ ) =
i i
∑ f ( X t ,ϕt , et ;φ ) for t = 1 ...15 (10)
500 i =1
1 500 ′
V500 ( Xt ;φ ) = ∑ ⎡⎣ Zt − f ( X t ,ϕt , et ;φ ) ⎤⎦ ⎡⎣ Zt − f ( X t ,ϕt , et ;φ ) ⎤⎦ for t = 1 ...15
i i i i (11)
499 i =1
This allows us to write the log-likelihood function for Zt , using the additive
separability and normality assumed for  t . It is given by
1 15 ′ −1
L500 ( Z ;φ ) = ∑ ⎡⎣ Z t − m500 ( X t ;φ ) ⎤⎦ ⋅ V500 ( Xt ;φ )
30 t =1
( )
⋅ ⎡⎣ Zt − m500 ( X t ;φ ⎤⎦ +
(12)
1 15
+ (
∑ ln det ⎡⎣V500 ( Xt ;φ )⎤⎦
30 t =1
)
Our estimator φˆ PSML is then obtained by maximization of this likelihood
500,15
function.22
The two-step SPML parameter vector estimate and its standard errors are the
following: 23
21
As noted by KORV, these simulated values must be kept throughout our whole exercise, so that the likelihood
function does not become itself a random object.
22
According to the results presented in Ohanian et al.(2000), our estimator has no approximation bias and is
consistent and asymptotically normal, so that as S → ∞, T → ∞ and T → 0, −1 −1 S
( ) (
T φˆST − φ0 → N 0, J 0 ⋅ I 0 ⋅ J 0 )
where ∂L ( Zt ; X t , φ0 ) ∂L ( Zt ; X t , φ0 ) and ⎡ ∂ 2L ( Zt ; X t ,φ0 ) ⎤ .
I =E⎡ ⋅ ⎤
⎣ ∂φ ∂φ ′ ⎦ J =E −
⎢ ⎣
∂φ∂φ ′ ⎥
⎦
Also, the simulation bias is extremely small already at S=20.
−1 −1
Standard errors calculation is based on the asymptotic covariance matrix T J 0  I 0  J 0 , where the sample
23 1
1 ∑15 ⎡ t t 500,15 ⎤ and

counterparts to I and J are given by: ˆ ∂L(Z ; X ,φˆ
t t 500,15 ) ∂L( Z ; X ,φˆ ) ⎡ ∂2L(Zt ; Xt ,φˆ500,15 ) ⎤
I500,15 = 15 t=1 ⎢ ⋅
∂φ ∂φ′ ⎥ Jˆ500,15 = 15 t =1 ⎢
1 ∑15 −
∂φ∂φ′ ⎥.
⎣ ⎦ ⎣ ⎦
P aram eter ρ θ μ λ ϕu 0
E stim ate − 0.9 0 15 0 .16 7 5 0.2 13 6 0 .08 9 7 − 1 3.2 2
(W hite Std E rror ) 0 .4 5 10 0 .16 7 1 0.0 39 3 0 .55 5 4 0 .00 4 8
The estimated values are consistent with the capital-skill complementarity

hypothesis, given that the estimated value for θ is significantly larger than ρ . In addition,
the elasticity estimates implied by our estimates are all sensible in light of the values
reviewed by Hamermesh (1993) and Krusell et. al. (1997).24
Indeed, our estimates are not dramatically different from the ones estimated by
Krusell et. al. (2000) for the US economy, although our values imply lower elasticities of
substitution between equipment and both skilled labor and unskilled labor, which means
there is a lower degree of capital-skill complementarity. The values we obtain are,
however, close to those estimated by Lindquist (2004) for another European country,
Sweden.
Implied Elasticity between Equipment and Labor Skilled Unskilled

Country − Source 1 1
1−
1−
Portugal - own results 0. 53 1. 20
US - Krusell et. al. (2000) 0. 67 1. 67
Sweden - Lindquist (2004) 0. 52 1. 40
The degree of substitutability between equipment and unskilled labor (which is

restricted to be the same as that between skilled and unskilled labor, given the functional
form adopted for the production function) is especially lower than that estimated for the US
and Sweden, which may reflect the poor quality of the Portuguese unskilled labor force.
24
Evidence referred by Krusell et. al. (1997) puts most estimates for the elasticity of substitution between
unskilled labor and equipment capital ranges from 0.5 to 3 (implying  ∈ −1, 0. 67, whereas those for the
elasticity of substitution between skilled labor and equipment vary between near zero and 1. 2 (implying
ρ ∈ [ −∞ , 0.15]).
192 Cátia Batista
5. Simulation of Wage Rates and of the Skill Premium

The objective of this section is to empirically evaluate the technology we proposed
by testing its ability to predict real wage and skill premium behavior. For this purpose, we
simulate these variables using actual input data, the parameters estimated in the previous
section, and our calculations for the TFP series, { At }

1999
.
t=1985
We will abstract from relative efficiency issues by assuming that unobservable
(
efficiency levels of labor ψ s and ψ u
t t ) remained constant over our period of analysis. This
assumption allows us to focus on the extent to which observable variables only can explain
the behavior of our variables of interest.
Since we want to abstract from changes in the (unobservable) efficiency levels of
labor, we normalize ψ s and ψ u to 1 at all periods of time. Under this normalization, our
t t
variables of interest may be written as:
θ −ρ
ρ ⎤ ρ
α ρ −1 ⎡ ⎛ k (t ) ⎞
w z ( t ) = A ( t ).(1 − μ )(1 − α )(1 − λ ). k s ( t ). z ( t ) ⎢λ ⎜ e ⎟ + (1 − λ )⎥
(13)
⎣⎢ ⎝ z ( t ) ⎠ ⎦⎥
1−α −θ
⎛ θ ⎞ θ
⎜ θ ⎡ ⎛ k e (t ) ⎞ ρ ⎤ρ ⎟
. ⎜ μ . (1 − z ( t ) ) + (1 − μ ) ⎢ λ ⎜ ⎟ + (1 − λ )⎥ ⎟
⎜ ⎢⎣ ⎝ z ( t ) ⎠ ⎥⎦ ⎟
⎝ ⎠
α θ −1
wu ( t ) = A ( t ).μ (1 − α ).k s ( t ). (1 − z ( t ) ) . (14)
1−α −θ
⎛ θ ⎞ θ
⎜ θ ⎡ ⎛ k e (t ) ⎞ ρ ⎤ρ ⎟
⎜ μ . (1 − z ( t ) ) + (1 − μ ) ⎢ λ ⎜ z ( t ) ⎟ + (1 − λ )⎥ ⎟
⎜ ⎢⎣ ⎝ ⎠ ⎥⎦ ⎟
⎝ ⎠
θ −ρ
wz (t ) 1− μ ⎡ ⎛ k (t ) ⎞ ρ ⎤ ρ ⎛ z (t ) ⎞θ −1
= . (1 − λ ) . ⎢λ . ⎜ e ⎟ + (1 − λ )⎥ .⎜ ⎟
(15)
wu (t ) μ ⎢⎣ ⎝ z (t ) ⎠ ⎥⎦ ⎝ 1 − z (t ) ⎠
As pointed by Krusell et. al. (2000), the expression for the skill premium
highlights two effects driving the skill premium behavior in this economy: on the one hand,
we have a capital-skill complementarity effect (increasing the stock of capital equipment
per skilled worker tends to increase the skill premium as long as   , i.e., there is
capital-skill complementarity); on the other hand, there is a quantity effect (an increase in
the proportion of skilled labor should decrease the skill premium for the parameter values
we are considering). Note that this does not depend on a country's TFP level.
The results from our simulation are presented in the following figures.
Annual Growth Rate of Skilled Wages
0.15
0.1
0.05
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.05
-0.1
-0.15
-0.2 predicted
actual
-0.25
-0.3
194 Cátia Batista
Annual Growth Rate of Unskilled Wages
0.12
0.1
0.08
0.06
0.04
0.02
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.02
-0.04 predicted
actual
-0.06
Annual Growth Rate of the Skill Premium
10%
5%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-5%
-10%
-15%
predicted
actual
-20%
-25%
The growth in skilled wages is systematically underpredicted during our period of

study, whereas unskilled wage growth is, on average, rather well forecasted. This pattern
results in a simulated behavior of the skill premium that also systematically underpredicted.
Note that the forecast pattern is basically following the growth pattern of capital equipment
per skilled worker, which falls dramatically especially from 1992, as predicted by capital-
skill complementarity. Indeed, this technological characteristic seems to play an important
role in explaining the behavior of the annual growth rates of our variables of interest,
especially their turning points.
Despite these positive findings, our results show persistent gaps between actual
and simulated growth rates. This should indicate that there is either a problem with our
specification/estimation or that there is an important role for factors other than the
observable-driven capital-skill complementarity and quantity effects.
In terms of the specifics of our analysis, an explanation would be to attribute these
gaps to the definition of "skill" used in our exercise: defining skilled workers as those who
completed 12 years of schooling, instead of those who completed college could prove more
meaningful for the concept of capital-skill complementarity in an economy where, in 1999,
only 7% of the labor input corresponded to college graduates, whereas around 25% had
completed 12 years of schooling. It could be that the definition of skill as attaining a
college degree does not capture the significant capital-skill complementarity effect, since
the proportion of college graduates in the labor force is so small. We explore this
hypothesis in the Appendix, but do not find evidence supportive of it.
In addition, simple experimentation of alternative parameter values did not prove
useful in diminishing these gaps, although a more formal test could be performed by
adapting our state-space model to match the skill premium, skilled and unskilled wages,
instead of the three macroeconomic features we used following Krusell et. al. (1997, 2000).
Alternative factors could be related to factors such as institutional labor market
characteristics that are not well captured by our framework, but could be part of the
196 Cátia Batista
"unobservable trend" identified by Katz and Murphy (1992), and that are likely more
relevant in European markets.
In addition, it should be noted that a substantial increase in trade flows has
occurred in the period we are analyzing (the degree of openness25 of the Portuguese
economy more than doubled to about 60% in this period), and that the pattern of
specialization in labor intensive, low technology content exports seems to have deepened,
according to evidence presented by Lopes (1999). This should have strongly contributed to
an increase in real wages (especially unskilled) higher than what the model predicts.
Despite the resulting fall in the skill premium (particularly sharp in the last years of our
sample), this variable and the level of labor earnings (especially unskilled) do not yet seem
close to have been equalized in Northern and Southern EU countries. This will be the
object of further discussion in section 7 of the paper.
6. A Counterfactual Exercise: Measuring the Impact of International Factor

Flows on Wage Rates
This section aims at evaluating how much of the real wage and skill premium
behavior after Portugal's entry in the EU can be quantitatively explained by the
international factor flows occurred at this time. Recall that the basic features of these flows
are described in detail in section 2.
6.1. The Effects of Emigration

There is only a limited amount of data on the educational composition of the
Portuguese permanent emigration flows: survey data from INE26 for 1997 point to the
proportion of skilled workers in emigration flows being approximately the same as that in
the labor force; the same survey for 2001 seems, however, to indicate that only 1.23% of
emigrants were college graduates. These pieces of evidence are consistent with the
observed fact that the skill premium is higher in Portugal than in European host countries27,
25
Measured as the fraction of average exports and imports in GDP.
26
Inquérito aos Movimentos Migratórios de Saída (IMMS). INE. Lisboa. 1993-2002.
27
This type of selection argument in migration was introduced by Borjas (1987), based on the Roy (1950) model.
but do not provide us with reliable information that can be used to simulate the effects of
emigration flows on the behavior of real wages in Portugal.
It is, nevertheless, possible to use this information to establish two benchmarks:
(1) if the proportion of skilled workers who emigrate is the same as that in the overall labor
force, then the small magnitude of emigration flows implies that emigration flows will not
change the fraction of skilled workers employed in production and technology therefore
implying that wages and the skill premium should remain basically unchanged; (2) if, on
the contrary, emigration is essentially unskilled, then effects on wages and the skill
premium should be expected.
We proceed by considering these two benchmark cases: the first provides a lower
bound to the effects of emigration (insignificant); the second provides an upper bound,
which is calculated by simulating the effects on wage rates of totally unskilled emigration.
For this purpose, the counterfactual of no migration above the average level of emigration
in 1983-85 is used. The choice of this threshold level of emigration is intended to filter out
migratory movements that would occur regardless of Portugal's entry in the EU. The results
of the simulations performed are displayed in the following figures.
Counterfactual Effects of Emigration on

0.15
0.1
0.05
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.05
-0.1
-0.15
-0.2 model
counterfactual
-0.25
-0.3
198 Cátia Batista

0.12
0.1
0.08
0.06
0.04
0.02
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.02
-0.04 model
counterfactual
-0.06

10%
5%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-5%
-10%
-15%
model
counterfactual
-20%
-25%
As expected as consequence of a pure quantity effect, a pattern of essentially

unskilled emigration creates upwards pressures in unskilled relative to skilled wages.
Quantitatively, these effects are however pretty small: cumulative effects over the period
1986-1999 on unskilled wages are of a magnitude of 3.3% on unskilled wages, but of less
than 1% on skilled wages and the skill premium. These are displayed in the following table:
6.2. The Effects of Foreign Investment

In order to assess the impact of foreign capital inflows caused by joining the EU, it
is necessary to distinguish what fraction of this investment represents an increase caused by
the integration process. Our approach is a very simple one: consider the difference between
post-1986 FDI and the average FDI level during 1983-85, and consider all EU transfers
(since these only started from 1986 and are a direct result of integration).
An additional simplifying assumption is made regarding the contribution of FDI
and EU transfers to capital equipment and structures, since there is no data available on this
partition. We take each year's proportion of total investment into equipment and structures
and assume it to apply to both FDI and EU transfers as well.
The counterfactual simulation is performed by subtracting the increase in foreign
investment attributed to joining the EU from the original investment series used to
construct capital stocks. This means that the counterfactual effects of no increase foreign
investment are accumulated in the process of capital stock formation.
We begin by considering the effects of total foreign investment (defined as the
sum of relevant FDI and EU transfers), which are displayed in the figures below.
200 Cátia Batista
Counterfactual Effects of Total Foreign Investment on

0.15
0.1
0.05
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.05
-0.1
-0.15
-0.2 model
counterfactual
-0.25
-0.3

0.12
0.1
0.08
0.06
0.04
0.02
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.02
-0.04 model
counterfactual
-0.06

10%
5%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-5%
-10%
-15%
model
counterfactual
-20%
-25%
Increased foreign investment seems to have significantly contributed to an

increase in wages, especially those of more educated workers, which may be attributed to
considerable rises in capital equipment under capital-skill complementarity. This increase
in skilled wages resulted in a sizeable increase in the skill premium. The order of
magnitudes involved is made clear in the table below.
Impact of Total Foreign Investment

Av. Annual Growth Rate Cumulative Effects
Skilled Wages 0.84% 4.85%
Unskilled Wages 0.09% 1.63%
Skill Premium 0.74% 3.43%
Given the quantitative importance of the effects uncovered, an interesting question

is that of distinguishing between the relative importance of FDI and EU transfers in this
exercise. This comparison may be performed by comparing the two tables that follow,
which show that most effects on skilled wages are due to the impact of EU transfers,
202 Cátia Batista
whereas FDI seems to have had a stronger effect on the behavior of the wages of less
educated workers.
Impact of Total Foreign Direct Investment

Impact of Total Foreign Direct Investment

6.3. Summary and Implications

The counterfactual exercises in this section indicate that in the Portuguese
experience of integration in the EU labor flows are quantitatively less important than
capital flows in determining the behavior of real wages.
In particular, depending on the educational composition of the migration flows,
the effects of observed emigration can range from no effects at all (if emigrants are not less
skilled than the labor force in general), to mild cumulative growth effects, up to 3%, over
the period 1986-99 if emigrants are substantially less educated than workers in general. In
addition, if one were to account also for immigration flows (making the reasonable
assumption that their educational composition is likely to be the same as that of emigration
flows), even these small effects on wages would likely vanish away given our previous
discussion on the characteristics of the labor market. Indeed, the approximate magnitudes
of these flows over our period of interest, make them likely to cause offsetting effects over
our period of interest.
This small contribution to the evolution of real wages can be related to evidence
summarized by Barro and Sala-i-Martin (1999) that labor flows seem to contribute little to
regional convergence of per capita income generally, but especially in European regions. A
possible cause for this phenomenon is the small response of migration to persistent real
wage differentials, as those existing between Northern and Southern European countries -
which translates in our case as a small number of emigrants leaving Portugal after its entry
in the EU. This issue deserves further consideration in the next section of the paper.
Regarding the impact of foreign investment on the behavior of real wages, this
seems to be quantitatively more relevant in explaining the behavior of skilled wages (with
an effect of almost 5% over the period 1986-99), which may be attributed to the
technological property of capital-skill complementarity - which should cause skilled wages
and the skilled premium to increase in the presence of a growing capital equipment to
skilled labor ratio.
The results we obtain concerning the effects of foreign investment deserve some
qualification. In particular, some authors, as Lopes (1999), have argued that not all of these
transfers effectively translated into investment. In contrast, it should be noted that increases
in FDI may not fully translate direct foreign investment in Portugal since other types of
foreign financial flows (as portfolio investment) are also likely to have contributed to this
effort. More importantly, it should be emphasized that we are only measuring the direct
effect of foreign capital inflows, and these are in many instances complementary to
domestic investment. This leads us to believe that the EU integration has probably had a
more sizable effect on real investment and, therefore, in real wages to which we can only
provide a lower bound.
7. Concluding Remarks and Directions for Further Research

What was the effect of international factor flows caused by the Portuguese
integration in the EU on the behavior of real wages? Our work answers this question by
showing that emigration played only a small role, whereas foreign capital inflows had a
larger contribution, even if we only take their direct effects into account.
The small impact of migratory movements on wage behavior is related to the
small magnitude of emigration flows after 1986, despite the existing real wage differentials
between Portugal and countries already in the EU. If immigration flows are also
204 Cátia Batista
considered, the wage effects of labor movements over our period of interest are indeed
likely to be insignificant.
The effect of foreign capital inflows associated with EU membership was
especially strong on the growth rates of skilled wages. We interpreted this effect as result of
capital-skill complementarity, a characteristic of our technology that arises from our
estimation of the elasticities of substitution between capital equipment and skilled labor,
and between capital equipment and unskilled labor.
A note should be paid to the fact that the effects of FDI only should be thought of
as a lower bound to the effects of foreign investments in Portugal after the European
integration. This is because foreign portfolio investments, for instance, are also likely to
contribute to increased investment. In addition, the effects of unilateral transfers from the
EU are most likely underestimated since these are associated to complementary increases in
domestic investment (both public and private). As a result we should emphasize that we
only measured the direct effect of foreign capital inflows, which leads us to believe that the
EU integration had probably a more sizable effect on real investment and, therefore, on real
wages to which we can only provide a lower bound.
Nevertheless, our results point to direct effects of factor flows on real wage
growth that are quantitatively not large. This leads us to believe that other aspects of the
European integration process may play an important role in explaining real wage behavior.
Namely, as argued in section 5, increased trade flows with a pattern of specialization in
labor intensive, low-technology content exports may have contributed significantly to
indirect factor flows with important effects on real wages.
In addition, one may argue that important institutional arrangements take place
when a country joins the EU, such as an increased degree of legal uniformity. This type of
effect could be reflected in converging total factor productivity, which could also have a
significant role in explaining the behavior of skilled and unskilled wages.
Further investigating the effects of intensified trade and converging total factor
productivity on real wages should constitute an interesting avenue for future research.
References
Albuquerque, Rui and Miguel Gouveia (1994). Distribuicao dos salarios em Portugal:
1980 e 1990, Bank of Portugal Annual Report.
Attanasio, Orazio P., and Fiorella Padoa Schioppa (1991). "Regional Inequalities,
Migration and Mismatch in Italy", in Fiorella Padoa Schioppa (ed.) (1991).
Mismatch and labour mobility. Cambridge, UK; New York, NY: Cambridge
University Press.
Barro, Robert J., and Xavier Sala-i-Martin (1991). "Convergence across States and
Regions", Brookings Papers on Economic Activity, 1991(1): 107-58.
Batista, Catia (2002). "Skill Premium in Portugal: Some Evidence on the Capital-Skill
Complementarity Hypothesis". Working Paper. University of Chicago.
Bentolila, Samuel (1997). "Sticky Labor in Spanish Regions", EuropeanEconomic Review,
41: 591-598.
Borjas, George (1987). "Self-Selection and the earnings of Immigrants", American
Economic Review, 77(4): 531-553.Bound, John and George Johnson (1992).
Changes in the Structure of Wages in the 1980's: An Evaluation of Alternative
Explanations, The American Economic Review, Vol. 82, No. 3. (Jun., 1992), pp.
371-392.
Bover, Olympia, and Pilar Velilla (1999). "Migration in Spain: Historical Background and
Current Trends", IZA Discussion Papers, No. 88.
Cannari, Luigi, Francesco Nucci and Paolo Sestito (2000). "Geographic Labour Mobility
and the Cost of Housing: Evidence from Italy", Applied Economics, 32: 1899-
1906.
Cardoso, Ana Rute (1998). Earnings Inequality in Portugal: High and Rising?, Review of
Income and Wealth, vol. 44 (3): 325-43.
Caselli, Francesco, and Silvana Tenreyro (2004). "Is Poland the Next Spain?", Working
Paper.
Chiswick, Barry R. (1978). "The Effect of Americanization on the Earnings of Foreign-
Born Men", Journal of Political Economy, 86 (5): 897-921.
European Commission (2001). The Free Movement of Workers in the Context of
Enlargement. http://europa.eu.int/comm/enlargement/docs/pdf/migration_enl.pdf
Faini, Riccardo, Giampaolo Galli, Pietro Gennari, and Fulvio Rossi (1997). "An Empirical
Puzzle: Falling Migration and Growing Unemployment Differentials among
Italian Regions", European Economic Review, 41: 571-579.
Galor, Oded and Omer Moav (2000). Ability-Biased Technological Transition, Wage
Inequality, and Economic Growth, Quarterly Journal of Economics, vol. 115 (2):
469-497.
Giannetti, Mariassunta (2003). "On the Mechanics of Migration Decisions: Skill
Complementarities and Endogenous Price Differentials", Journal of Development
Economics, 71: 329-349.
Goldin, Claudia and Lawrence F. Katz (1998). "The Origins of Technology-Skill
Complementarity", Quarterly Journal of Economics, vol. 113 (3): 693-732.
206 Cátia Batista
Goldin, Claudia and Lawrence F. Katz (1999). The Returns to Skill across the Twentieth
Century United States, NBER Working Paper, No. 7126.
Gordon, Robert J. (1990). The Measurement of Durable Goods Prices. Chicago, IL:
Chicago University Press.
Greenwood, Jeremy, Zvi Hercowitz and Per Krusell (1997). Long run implication of
investment-specific technological change, American Economic Review, vol. 87(3).
Greenwood, Jeremy, Zvi Hercowitz and Per Krusell (2000). The role of investment specific
technological change in the business cycle, European Economic Review, vol. 44:
91-115.
Greenwood, Jeremy and Boyan Jovanovic (2001). Accounting for Growth, in Edwin Dean,
Michael Harper and Charles Hulten ( eds.) NBER Studies in Income and
Wealth: New Directions in Productivity Analysis. Chicago, IL: Chicago
University Press.
Griliches, Zvi (1969). Capital-Skill Complementarity, Review of Economics and Statistics,
vol. 51: 465-468.
Hall, Robert E., and Charles I. Jones (1999). "Why Do Some Countries Produce So Much
More Output Per Worker Than Others?", Quarterly Journal of Economics,
February: 83-116.
Hatton, Timothy J., and Jeffrey G. Williamson (1994). "Latecomers to Mass Migration: The
Latin Experience" in Timothy J. Hatton, and Jeffrey G. Williamson (eds.) (1994).
Migration and the international labour market, 1850-1939. London; New York:
Routledge.
Hamilton, James D. (1994). Time Series Analysis. Princeton, NJ:Princeton University
Press.
Hamermesh, D. S. (1993). Labor Demand. Princeton, NJ: Princeton University Press.
Hanson, Gordon H. (2003). "What Has Happened to Wages in Mexico since NAFTA?
Implications for Hemispheric Free Trade", NBER Working Paper 9563.
Hartog, Joop, Pedro Pereira and Jose Vieira (2001). Changing Returns to Education in
Portugal during the 1980's and early 1990's: OLS and quantile regression
estimators, Applied Economics, vol. (33): 1021-37.
Heckman, James J., Lance Lochner and Christopher Taber (1998). "Explaining Rising
Wage Inequality: Explorations with a Dynamic General Equilibrium Model of
Labor Earnings with Heterogeneous Agents", Review of Economic Dynamics,
1(1): 1-58.
Heston, Alan, Robert Summers and Bettina Aten (2002). Penn World Table Version 6.1.
Center for International Comparisons at the University of Pennsylvania (CICUP).
Hunt, Helen (2000). "Why Do People Still Live in East Germany?", IZA Discussion Paper,
No. 123.
Instituto Nacional de Estatistica. Inquérito aos Movimentos Migratórios de Saída (IMMS).
INE. Lisboa. 1993-2002.
Judd, Kenneth L. (1998). Numerical Methods in Economics. Cambridge, MA: The MIT
Press.
Juhn, Chinui, Kevin M. Murphy, and Brooks Pierce (1993). Wage Inequality and the Rise
in Returns to Skill, Journal of Political Economy, vol. 101(3): 410-442.
Katz, Lawrence F. and Kevin M. Murphy (1992). "Changes in Relative Wages, 1963-1987:
Supply and Demand Factors", Quarterly Journal of Economics, 107(1): 35-78.
Krusell, Per, Lee E. Ohanian, Jose-Victor Rios-Rull and Giovanni L. Violante
(1997).Capital-Skill Complementarity and Inequality: A Macroeconomic Analysis,
Federal Reserve Bank of Minneapolis - Research Department Staff Report 239.
Krusell, Per, Lee E. Ohanian, Jose-Victor Rios-Rull and Giovanni L. Violante (2000).
Capital-Skill Complementarity and Inequality: A Macroeconomic Analysis,
Econometrica, vol. 68 (5): 1029-1053.
Lagarias, J. C., J. A. Reeds, M. H. Wright and P. E. Wright (1998). Convergence
Properties of the Nelder-Mead Simplex Method in Low Dimensions, SIAM Journal
of Optimization, vol. 9 (1): 112-147.
LaLonde, Robert J., and Robert H. Topel (1992) "The assimilation of immigrants in the US
labor market", in George Borjas and Richard B. Freeman (eds.), Immigration and
the workforce:economic consequences for the United states and source areas.
Chicago, IL: Chicago University Press.
Laroque, Guy and Bernard Salanie (1989). Estimation of Multi-Market Fix-Price Models:
An Application of Pseudo Maximum Likelihood Methods, Econometrica, vol.
57(4): 831-860.
Lindquist, Matthew J. (2004). "Capital-Skill Complementarity and Inequality in Sweden",
Scandinavian Journal of Economics, forthcoming.
Luxembourg Income Study (LIS) Microdatabase, (several years). Harmonization of original
surveys conducted by the Luxembourg Income Study, asbl. Luxembourg, periodic
updating.
Lopes, Jose da Silva (1999). A Economia Portuguesa desde 1960. Lisbon, Portugal:
Gradiva.
Lucas, Robert E. (1988). "On the Mechanics of Economic Development", Journal of
Monetary Economics, 22: 3-32.
Lucas, Robert E. Jr. (1990). "Why Doesn't Capital Flow from Rich to Poor Countries?",
American Economic Review, Papers and Proceedings, 80 (2): 92-96.
Machado, Jose A. F. and Jose Mata (2001). Earnings functions in Portugal 1982 - 1994:
Evidence from Quantile Regressions. Empirical Economics, vol. 26 : 115-134.
Parente, Stephen L., and Edward C. Prescott (2000). Barriers to Riches. Cambridge, MA:
MIT Press.
Pedersen, Peder J., Mariola Pytlikova, and Nina Smith (2004). "Selection or Network
Effects? Migration Flows into 27 OECD Countries, 1990-2000", IZA Discussion
Papers, No. 1104.
Ohanian, Lee E., Giovanni L. Violante, Per Krusell and Jose-Victor Rios-Rull (2000).
Simulation-Based Estimation of a Nonlinear, Latent Factor Aggregate Production
Function, in Roberto S. Mariano, Til Schuermann, and Melvyn Weeks eds. 
Simulation-Based Inference in Econometrics: Methods and Applications.
Cambridge, UK: Cambridge University Press.
Organisation for Economic Cooperation and Development (several years). Trends in
international migration: continuous reporting system on migration. Paris:
Organisation for Economic Cooperation and Development.
208 Cátia Batista
Roy, A. D. (1951). Some thoughts on the distribution of earnings, Oxford Economic Papers,
3: 135-146.
Sjaastad, Larry A. (1962). "The Costs and Returns of Human Migration", Journal of
Political Economy, 70 (5-2): 80-93.
Stokey, Nancy L. (1996). Free Trade, Factor Returns, and Factor Accumulation, Journal of
Economic Growth, 1: 421-447.
The Economist (2004). "The Future of Europe: A Club in Need of a New Vision". April
29th.
White, Halbert (1994). Estimation, Inference and Specification Analysis. Econometric
Society Monographs, No. 22. Cambridge, UK: Cambridge University Press
.
8. Appendices
8.1. Labor Market Data

Labor market data are from the Personnel Files (PF) collected by the Portuguese
Ministry of Labor and span the period 1985 to 1999. They were collected in October of
each year from 1994 (March in the previous years), and include full-time wage earners on
the payroll in that month. All firms employing paid labor in Portugal are legally obliged to
report this information, which includes age, highest completed level of education, tenure
with the current firm, and gender.
We used three variables from this data base, each already classified into education
levels: average number of hours worked per year, per worker; number of workers; and
workers' hourly wages.
Regarding average hourly wages, we chose to use the actual gain measure of
wages that includes all money and/or goods paid to workers on a regular basis - which
includes not only the basic wage, but also food and housing subsidies, as well as
productivity supplements. From the available measures, this seemed the most accurate
manner to measure workers' reward for their productivity.
No data were available for any variable for the year 1990 due to an incorrect
sampling experiment performed in that year by the Department of Statistics of the Ministry.
We chose to predict the values corresponding to 1990 by means of simple linear
interpolations, believing that we should use all pieces of available information to
compensate for the data scarceness we face.
The education level taxonomy was changed after 1993. This allowed a more
comprehensive and rigorous classification, at the cost of considering less education levels.
Again, the short span of data available made us opt by using all the data available,
including those for the period between 1985 and 1993. Because of this decision, we had to
make the two taxonomies consistent. For this purpose, we took the latter classification as a
basis and attempted to fit the previous categories into these latter in a manner as rigorous as
possible.
These final categories are:
• No elementary schooling completed (mainly older people who are not
able to read or are able to read without formal education);
210 Cátia Batista
• Elementary schooling completed (4 years of schooling, compulsory from

1956 for boys and from 1960 for both boys and girls);
• Secondary schooling - 1st cycle (6 years of schooling, compulsory from
1964);
• Secondary schooling - 2nd cycle (9 years of schooling, compulsory from
after the mid-1980s);
• Secondary schooling - 3rd cycle (12 years of schooling, has recently been
announced compulsory from 2010);
• Bachelor's degree (15 years of schooling);
• Licentiate degree (most common College degree, averages 17 years of
schooling).
Portugal has a very low level of workers' education in the European context.28 In
1985, only 2.5% of the labor force had a college degree (either Bachelor's or Licentiate)
and 65% had 4 years of schooling or less. These statistics have significantly (given the
restrictions inherent to human capital investment) improved in the period under analysis to
7% and 38.5%, respectively.
Finally a note regarding our decision to ignore some information. The category
`Òthers'' incorporated all workers who could not be fitted into the initial conventional
taxonomy. It ceased to exist from 1994. In the original data provided by the Ministry, there
were also some censored observations. What we did was to sum all the observations in the
category `Òthers'' with those originally ignored and then compute a corrected density for
the number of workers, so that we would not distort the construction of the indicators we
needed for our empirical analysis - mainly averages weighted by this density.
8.2. Investment Data

The Gross Fixed Capital Formation annual time series between 1953 and 1995
were obtained from the Bank of Portugal (BoP)'s study Long Series for the Portuguese
Economy. For the years 1996-99, this information was provided by the National Statistics
Institute (INE), Quarterly National Accounts Bulletin.
We distinguish two categories of investment goods: structures and equipment.
Equipment goods include the original categories ``Machinery and Equipment'' and
``Transportation Materials''. Structures include ``Construction'' and `Òthers''.
The Gross Fixed Capital Formation flows and their deflators were not originally
adjusted for increases in the quality of equipment over time. We decided to use Gordon
(1990)'s equipment quality adjustment index to derive the adjusted equipment price index,
which will also be used to adjust investment series, as described in the next subsection.
Even though there are very substantial differences between the economies of the
United States and Portugal, we believe this usage to be reasonable given nowadays very
rapid technology globalization. Our hypothesis is however most unlikely to hold in the
earlier periods of investment considered. Nevertheless, these flows have little impact on the
stocks of capital from 1985, which is likely to make the impact of this assumption
negligible.
Gordon (1990)'s index was unfortunately only available until 1983. The strategy
followed for the quality adjustment after that date was used by Greenwood, Hercowitz and
Krusell (1997) and suggested to them by Robert Gordon: a fixed 1.5% deduction from the
growth rate of equipment prices.
8.2.1. Computation of Capital Stock Series

Both the structures and (quality adjusted) equipment price indices were deflated
by a non-durable goods and services price index.29 These series are what we designate by
relative (to the price of non-durable consumption) capital price indices.
Capital price indices were then used to deflate the corresponding investment
series, which yielded investment time series in efficiency units.
28
For more details, the reader is referred to Hartog et al. (2001).
29
This is done in order to avoid further problems related to quality adjustment issues.
212 Cátia Batista
The capital stocks were then computed following the Perpetual Inventory Method.
This basically solves the following difference equation
t t t
(1 − δ ) K 0 + ∑ (1 − δ ) I t − j if t ≤ Lmax
j =0
Kt = (1 − δ ) Kt −1 + I t = L
max t
∑ (1 − δ ) I t − j if t > Lmax
j =0
This solution allows us to estimate capital from the investment series only - at the
cost of losing the Lmax initial observations, which is not a problem for us, given that our
data scarcity was driven by the labor data.

The depreciation parameters were chosen following Greenwood, Hercowitz and
Krusell (1997). They were assumed constant across vintages and set at δ e = 0.124 and
δ s = 0.056.
e
We also imposed a maximum length of usage for equipment of Lmax = 15 and
s
for structures of Lmax = 32. This latter was chosen so that investment data would not
become scarce and the former was chosen to be a date at which the remaining percentage
s
value of equipment was the same than that of structures at Lmax = 32 .
These seem sensible numbers according to the study by Böhm et. al. (1998),
which surveyed literature on this estimation procedures in OECD countries.
8.3. National Income and Product Accounts (NIPA) Data

Our main source of information regarding the Portuguese NIPA was again the
Long Series for the Portuguese Economy, provided by the Bank of Portugal. However,
since these only present data until 1995, we again had to make use of information provided
by the (Portuguese) National Institute of Statistics (INE) and also by the BoP Annual
Reports of 1996 through 1999.
In particular, we used the BoP Long Series for data on national labor earnings and
national income (necessary to calculate labor income share) until 1995 and the BoP Annual
Reports from then on.
The GDP series (at 1995 constant prices, in millions of Escudos) used to construct
the business cycle indicator necessary for the first step of the SPML estimation was also
obtained from the BoP until 1995, but after this year we took INE's information.
Finally, the series on private consumption used to calculate the nondurable goods
and services deflator were also made available by the BoP until 1995 and by INE from that
year on.
8.4. Auxiliary Estimation Procedures
8.4.1. First Step of SPMLE

Because SPMLE of the parameters in the production function requires strict
exogeneity of labor and capital inputs, in the first step of SPMLE we pay attention to the
issue of potential endogeneity of the labor inputs. To address this concern, we began by
using the Hodrick-Prescott filter (with smoothing parameter 100, given the annual
frequency of the data) to detrend the Portuguese GDP series (at 1995 prices - please see
Data Appendix for more details). This procedure, as graphed in the figure below, provided
us with a Business Cycle (BC) indicator
Portuguese GDP and its HP Trend from 1980

19500000
18500000
GDP (95 prices)
17500000
HP trend
16500000
15500000
14500000
13500000
12500000
11500000
10500000
9500000
1980 1982 1984 1986 1988 1990 1992 1994 1996 1998
This BC indicator, in addition to a constant term, current and lagged stocks of

equipment and structures, lagged relative price of equipment and a time trend were then
214 Cátia Batista
used as regressors to predict skilled and unskilled hours worked, to prevent potential
endogeneity problems:
hi = β1 + β 2 K e + β3 K e + β 4 K s + β5 K s + β 6 qt −1 + β7 t + β8 BCt −1 + vt ,
t t t −1 t t −1
for i = st , ut , t = 1, ..., 15
From these estimations we obtained the fitted labor inputs, which are used in the
second step instead of the actual inputs.
8.4.2. Second Step of SPMLE Implementation

The econometric estimation described in section 4 was implemented using Matlab
6.0.
A more detailed remark is deserved by our SPMLE procedure. Initially the
function fminunc of Matlab 6.0 was used to perform the likelihood maximization of the
second step of the SPMLE, with options set to use the Medium Scale Optimization. This
uses the numerical gradient BFGS quasi-Newton method, described in Judd (1998),
pp.113-15, with a mixed quadratic and cubic line search procedure. Unfortunately, it was
not capable of achieving convergence, which may be attributed to the highly non-linear
nature of the likelihood function. Therefore, we had to switch to another numerical
optimization method.
The function used to achieve our estimates is “fminsearch” of Matlab 6.0. "optim"
toolbox. This uses the simplex search method of Lagarias et. al. (1998), a direct search
method that does not use numerical or analytic gradients. Numerical gradients and hessians
were then computed using the "Hessian" routine by James P. LeSage, available from
www4.ncsu.edu/~pfackler.
8.5. Sensitivity Analysis to the Definition of Skill

Our classification of data according to skilled and unskilled categories took the
college completion classification used for most studies on the United States labor market.
However, only 7% of the full-time working population in 1999 had attended college. Such
small share of college graduates in the workforce could potentially lead to a biased picture
of the Portuguese skilled workforce that is relevant for the capital-skill complementarity
hypothesis. Therefore, we decided to perform a sensitivity analysis to our definition of skill

of the labor market facts presented in section 2.
For this purpose, let us now define skill as completion of secondary school (12
years of schooling), since this corresponds to approximately 25% of the labor input in
1999, which is about the same proportion of the US labor force completing college.
We obtain that our stylized results would basically remain unchanged: real wage
growth progressed quite rapidly, both for skilled and unskilled workers, especially in the
period immediately after Portugal's integration in the EEC, as shown in the following
figure.
Evolution of Log Wages from 1985 to 1999

7.5
7.0
6.5
Skilled
Unskilled
6.0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Source: Own Calculations based on PF.
Regarding labor quantities, we observe, as expected, a substantial increase of the

proportion of skilled (relatively to unskilled) hours worked, as is displayed in the next
216 Cátia Batista
figure.30 This pattern is mainly due to a significant growth in the supply of skilled labor,
which averaged 6.7% in the period under analysis, while the supply of unskilled labor
remained roughly constant.
Proportion of Skilled Hours Worked Relative to Total

(Skill Defined as 12 Years of Schooling)
25%
20%
15%
10%
5%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Source: Own calculations based on PF.
Using this definition of skill, the two-step SPML parameter vector estimate and its
standard errors are the following:
Parameter ρ θ μ λ ϕu 0
Estimate -0.8441 0.2211 0.1500 0.0021 -20.85
(White Std Error ) 0.4998 0.1608 0.0094 0.0637 0.0067
30
The decline from 1993 to 1994 may be due to changes in the statistical classification of workers' education
level, which we attempted to make consistent over time, as described in more detail in the Data Appendix.
The evolution of the labor force's skill level basically exhibited the same pattern as that of hours worked per
level of skill, the same happening to the relative proportion of skilled vs. unskilled quantities.
The estimated values are very close to those obtained using the "college
definition" of skill. They are also consistent with the capital-skill complementarity
hypothesis, and imply sensible elasticity estimates.
The use of these estimates in simulating the behavior of skilled and unskilled real
wages, as well as of the skill premium shows, however, that the systematic downward gap
in predicting annual growth rates disappears, but no improvements are achieved in terms of
the ability to predict turning points in the behavior of these growth rates. This is clear from
the following figures.
0.15
0.1
0.05
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.05
-0.1
predicted
-0.15 actual
-0.2
218 Cátia Batista
0.12
0.1
0.08
0.06
0.04
0.02
0
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-0.02
-0.04 predicted
actual
-0.06
4%
2%
0%
1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
-2%
-4%
-6%
-8%
predicted
-10%
actual
-12%
-14%
The overall analysis of the effects of the alternative definition of skill as

completing 12 years of schooling leads us to believe that no significant changes or
improvements should occur in terms of the stylized facts and technological characteristics
studied. In particular, it does not seem to be the case that this equating skill to 12 years of
schooling would deepen or make more significant the concept of capital-skill
complementarity. Therefore, we opted to keep the definition of skill as "college
completion", also as a way to ease comparison of this study with others in the literature.
Sessão 3
Session 3
SMALL FIRMS IN PORTUGAL: A SELECTIVE SURVEY OF STYLIZED
FACTS, ECONOMIC ANALYSIS, AND POLICY IMPLICATIONS
Luis M. B. Cabral
New York University and CEPR
December 2005
Abstract
I survey a number of stylized facts pertaining to the dynamics of firm entry,

growth and exit in competitive industries. I focus particularly on data for Portugal,
though I also consider, for comparison purposes, data from other countries. I then
present a series of theoretical models that attempt to explain the stylized facts and
evaluate the welfare impact of market distortions. Finally, I derive a number of policy
implications, all centered around the notion of economic mobility.
224 Luís M. B. Cabral
1. Introduction
Most of the study of industrial organization over the past few decades has focused
on concentrated industries and firms with market power.1 To some extent, the influx of
game theory into industrial organization, for all its benefits, has created a bias away from
industries where market power is of secondary importance.
A second reason why competitive industries have largely been ignored is that,
from our study of the perfect competition model, there would seem to be very little more to
say about industry dynamics and the properties of the long-run equilibrium beyond what we
learn in an economics principles course. In fact, firm behavior under the perfect
competition model is almost trivial: remain active if price is higher than the minimum of
average cost; and, conditional on remaining active, choose output so that price equals
marginal cost. What else is there to study?
Recent empirical studies of competitive industries reveals a wealth of empirical
regularities hitherto ignored. In parallel, a series of theoretical models have attempted to
explain these stylized facts. Altogether, these developments have turned the study of small-
firm dynamics into an exciting area of research within industrial organization — and in fact
a natural bridge between industrial organization and macroeconomics.
In this paper, I present a selective survey of this emergent literature, with an
emphasis on Portugal, both from a data point of view and in terms of policy implications. I
start (Section 2) with a few stylized facts regarding small firm dynamics in a variety of
countries. In fact, one of the striking empirical observations is how regular the main
stylized facts are across countries. I then consider (Section 3) a series of models that
attempt to explain the main stylized facts. One important component of this analysis is the
welfare evaluation of market distortions. This naturally leads into the next section (Section
4), where I consider policy implications of the empirical and theoretical results. The bottom
line of this section is the importance of economic mobility, a broad concept that denotes the
absence of distortions to the activity of small firms.
1
See, for example, Tirole (1988).
Small Firms in Portugal: A Selective Survey of Stylized Facts, Economic Analysis, and Policy Implications 225
2. Industry dynamics: stylized facts

What does the model of perfect competition have to say about entry, exit, and firm
2
size? The long-run equilibrium under perfect competition is a limit point that industries
converge to by means of successive entry and exit. If active firms make positive profits,
then new firms are attracted to the industry. If, on the contrary, active firms make losses,
then some of those firms exit. Finally, in the long-run equilibrium, price equals the
minimum of the long-run average cost. Since technology (i.e., the cost function) is the same
for all firms (because of the equal access assumption), each firm receives zero supra-
normal profits and there is neither entry nor exit.
Concerning the size distribution of firms, the perfect competition model is either
rather extreme or extremely scant: Assuming plant-level cost functions are U-shaped, all
plants must be of the same size in the long run (i.e., there is only one output level that
minimizes average cost). If the managerial costs of owning more than one plant are
positive, then each firm owns one plant only and all firms are of the same size. If, on the
contrary, managerial costs are zero, then there exists a virtually uncountable number of
possible industry configurations, all of which are consistent with the model.3
The empirical evidence from various industries with "many" small firms is widely
at odds with the above view of industry dynamics: in any given period and industry, entry
and exit take place simultaneously; and the size distribution of firms displays a number of
regularities and is not concentrated on a single size. In the following subsections, I present
data on these and related stylized facts. Before proceeding, I address an important
preliminary question: how good is the available data on industry dynamics? Most of these
data are based on official census or registries. This implies that they miss some of activity
in the informal economy. Since much of my analysis compares Portugal to other countries,
the presence of the informal economy is a concern to the extent that it varies greatly across
countries.
2
The following paragraphs are adapted from Cabral (2000).
3
In fact, there may be economies of scale in multi-plant firms-e.g., large-scale purchasing discounts—that counteract
increased managerial costs. But this would takes back to a U-shaped cost curve and the prediction that all firms have
the same number of plants. See Lucas (1978).
Figure 1 plots the values of the share of the informal economy and the level of
Gross National Income (GNI) for all countries with a per-capita GNI greater than
US$5,000. The figure suggests that Portugal is not particularly different from other
countries.
Figure 1: Informal economy and per-capita GNI (countries with per-capita GNI greater
than $5,000). The data point for Portugal is represented by a bullet point.
Source: World Bank.
In fact, correcting for GNI p.c. level, Portugal seems to be right on average.4
Insofar as we compare Portugal to developing countries (as I will do for the most part in
this paper) then we should be aware of the potential bias introduced by the informal
economy, about 23% in Portugal versus 18% for countries with per-capita GNI greater than
US$10,000.
2.1. Entry, growth and exit

One of the most robust stylized facts of competitive industry dynamics is that, in
any given year, entry and exit occur simultaneously. To put it differently, net entry rates are
a small fraction of gross entry rates. Moreover, one finds that entry and exit rates are highly
correlated across industries; that is, industries with higher than average entry rates also
exhibit higher than average exit rates.
4
The line corresponds to the regression of the share of the informal economy with respect to GNI per capita. The
regression has an R 2 of 0.49; the GNI per capita coefficient has a p value of 1.4E-06.
Cable and Schwalbach (1991) developed one of the earliest surveys of studies on
entry and exit in different countries. Table 1 includes some of the results in their Tables
14.1 and 14.2. As can be seen, entry and exit rates are much higher than net entry rates (the
difference between gross entry and gross exit rates). One possible interpretation for this
phenomenon might be that it is an artifact of aggregation: some industries might have high
entry rates while other industries have high exit rates. However, the data on cross-industry
correlation between entry and exit rates suggests that this is not the case Notice moreover
that the results persists at the 3-, 4- and 5-digit industry classification level. Portugal,
interestingly, exhibits some of the higher values of entry and exit rates.
Table 1: Annual gross entry and exit rates

(in %) and correlation.
Source: Cable and Schwalbach (1991).
Table 2 presents data from a more recent source. The second column, turnover
rate, corresponds to the sum of entry and exit rates in term of employment. The entry rate,
for example, is given by the total number of new jobs created by entrants divided by the
total number of workers in the industry. The correlation rates between entry and exit are
also weighted by employment level.
Notice that turnover rates are quite different in Tables 1 and 2. For example, Table
1 implies a turnover rate (in terms of number of firms) of 21.8% for Portugal, whereas
Table 2 shows a much lower value, 9.3%. The main reason for the discrepancy is that Table
2 presents employment-weighted rates, whereas Table 1 refers only to the number of firms.
Table 2:Turnover rate (employment-weighted)

and correlation between entry and exit rates.
Sources: Bartelsman et al (2003); Bartelsman et al (2004).
Table 3: Entrants and exiters relative size

(as a percentage of incumbent firms)
Source: Cable and Schwalbach (1991).
The lower rates in Table 2, compared to Table 1, suggest that entrants and exiters
are of smaller size than incumbent firms. In fact, a second important stylized fact is that
market penetration rates are a small fraction of entry rates. Table 3 includes data on
entrants and exiters relative size in different countries. In all countries, the entrants initial
size is less than one-half the industry average size. In 6 of 8 countries, it is less than one
quarter. Except for the U.S., Portugal has the lowest ratio of entrant and exiter size with
respect to industry average size.
Turning to the dynamics of each individual firm, a third stylized fact is that
survival rates tend to be increasing in firm size and in firm age. For example, Mata and
Portugal (1994) estimate baseline hazard rates (that is, conditional probabilities of exit)
after t years of .19, .14, .12 and .11 for the first four years. Based on the same dataset, Mata
et al. (1995) also show that the probability of survival is increasing in current size. As with
the previous stylized facts, similar results have been obtained for other countries. For the
U.S., Evans (1987) estimates that a 1 percent change in firm size and a 1 percent change in
firm age lead, respectively, to a 7 and a 13 percent change in the probability of survival
over a 5-year period.
To conclude this subsection, I present a stylized fact regarding firm growth:
growth rates are typically decreasing in size, especially for small size levels; and decreasing
with age. Evans (1987) estimates that, for the U.S. and over a 10-year period, a 1 percent
increase in initial size leads to a 0.68 percent increase in ending-period size, that is, growth
is considerably less than proportionate. He also finds that, over the same period, a 1 percent
increase in initial age implies a 1.42% decrease in final size. Hall (1987) derives
qualitatively similar results based on a different sample of U.S. firms. Finally, Dunne and
Hughes (1994) estimate that, over a 5-year period, a 1 percent increase in initial size leads
2
to a 0.93 percent increase in ending-period size (this would correspond to 0.93 ≈ 0.86 over
a period of 10 years; compare with Evans' results).
To summarize, this subsection depicts a typical industry as having many entrants
and exiters each period. A typical entrant is smaller than the industry average and grows
faster than the industry average. Entrants are more likely to exit than older entrants,
especially when they remain small in size. Finally, these facts are fairly robust both across
industries and across countries. In particular, Portugal seems fairly typical.
Figure 2: Firm size distribution in Portugal based on two different datasets.

Source: Cabral and Mata (2003).
2.2. Firm size

The patterns of the firm size distribution depend critically on the type of data
source one considers. Cabral and Mata (2003) show that, when considering the universe of
Portuguese manufacturing firms, the size distribution is skewed to the right. This contrasts
with the distribution of firms from commonly used databases, such as Compustat or
Amadeus, for which the lognormal is a better approximation. Figure 2 illustrates this
contrast. On the left we have the size distribution based on a comprehensive dataset (from
Quadros do Pessoal). On the right, the distribution from a dataset with similar
characteristics to Compustat and Amadeus. One cannot reject that the right-hand
distribution is lognormal, but the one on the left is certainly more right-skewed than a
lognormal.
The lognormal distribution has traditionally played an important role in the study
of the firm size distribution. Empirically, the distribution seems to fit well data from
Compustat and similar databases. Theoretically, the lognormal is the limiting distribution of
a firm growth process such that growth rates are independent of firm size.5 When one looks
at more comprehensive datasets, including small and micro firms, then lognormality is no
longer the rule; nor are growth rates independent of size (see the previous section).
5
See Gibrat (1937), Sutton (1997).
The pattern of right-skewness, with proportionately more small firms than large
firms (with respect to the lognormal distribution), is not unique to Portugal. Figure 3
depicts the distribution of firm size for manufacturing countries in Portugal and the U.S.
Despite the significant differences in size and level of economic development, the
distributions are relatively similar.
In recent research, Bartelsman et al (2003) present some systematic evidence on
firm size and firm size distribution across a series of countries. Table 4 presents the relative
weight of firms with fewer than 20 employees. As can be seen, the number of firms in that
category is virtually identical in Portugal and the U.S. (consistently with Figure 3).
Table 4: Relative importance of firms with fewer than 20 employees:

fraction of total economy (%).
Source: Bartelsman et al (2003).

Table 5: Average firm size (number of employees).
Source: Bartelsman et al (2003).
However, when we consider the fraction of employees accounted for by firms with fewer
than 20 employees, then the number is much lower for the U.S. This reflects the fact that
the right-tail of the U.S. distribution of firm size is much thicker.
Table 5 presents data on average firm size in various countries. Again, Portugal
seems broadly in line with other European countries. Table 5 also presents data on the
subcategories Manufacturing and Services. The cross-country patters of firm size seem
broadly robust to the type of firms considered. In fact, Bartelsman et al (2003) show that
industry composition accounts for a small fraction of the cross-country differences in firm
size.
The results from Table 5 may seem puzzling. After all, Portugal's largest firms
pale in size when compared to their European counterparts; and accordingly one might
expect average firm size to be smaller in Portugal (and the employment share of small firms
to be larger). The solution to this apparent puzzle, at least to some extent, is to be found in
what we might call the extreme-statistic fallacy. Consider two random variables, X and Y.
Let E(X, m, n) be the average value of the m highest values of x out of a sample of n values.
Then, even if the distribution of X and Y were the same, we would inevitably obtain
E(X,m,nX) > E(Y,m,nY) so long as nX > nY. So the fact that the five largest German firms are
greater than the five largest Portuguese firms does not necessarily imply that German firms
are on average greater than Portuguese firms; just like the fact the five tallest German men
are taller than the five tallest Portuguese men does not imply Germans are on average taller.
In summary, Portugal does not seem to be an outlier in terms of firm size and the
firm size distribution. It is often said that there are too many small firms in Portugal, and
that average firm size is too small. However, when looking at SME it is hard to find a
significant difference between Portugal and the average OECD country.
2.3. Productivity
There is extensive evidence, though not extensive systematic evidence, that
productivity in Portugal is substantially lower than the average of all European countries. A
report by McKinsey benchmarks Portugal with reference to a group of four other European
countries and estimates a productivity gap of about 30%.
I am not aware of any systematic estimate of the distribution of productivity levels
by industry in Portugal. Estimates for the U.S. indicate high indices of variability.
Considering how similar different countries are on other dimensions, it seems reasonable to
expect Portuguese industries to exhibit similar levels of dispersion of productivity levels.
An estimate based on a survey of Portuguese innovators indicates a coefficient of
variation greater than 1.6 Such high value of the coefficient of variation has important
implications. In particular, it makes little sense to use the difference in means and
generalize that Portuguese firms are not productive: the fallacy of the average.
A simple calculation will help make the point. Assume that the distribution of
productivity levels is lognormal (which makes sense based on the distributions in other
countries). Assume moreover that the standard deviation is equal to the mean. Then even if
Portuguese firms have an average productivity 50% lower than the European average, 20%
of Portuguese firms have a productivity level higher than the European median; or 11%
above European average.
These calculations are illustrated in Figure 4. In this figure, I normalize units so
that the average productivity level of European firms is 1. I assume that both distributions
are lognormal with a coefficient of variation of 1. As can be seen, about 10% of the mass of
Portuguese firms lies to the right of 1 (the average). Since the median of the lognormal is
6
Pedro Conceição (private communication).
lower than the average, the percentage of Portuguese firms above the median is even
higher.
One of the benefits of competition is that it leads to an efficient allocation of

resources across firms within an industry. In particular, firms with higher productivity
levels will tend to increase in size, whereas firms with lower productivity level will tend to
decrease in size or exit. In this regard, it is useful to measure the gap between weighted and
un-weighted productivity. If this gap is zero, we conclude that there is little correlation
between productivity and size. If this gap is positive, however, we conclude that firms with
higher productivity levels are larger, reflecting the benefits of competitive selection.
Table 6 displays the values of this index. Specifically, I follow Olley and Pakes
(1996) in defining the gap as
where si is firm i's share (in number of employees), Pi is firm i's productivity level, and N is
the number of firms.
Considering the values for other OECD countries, the value for Portugal is fairly
high, though smaller than in the U.S. Two other noticeable differences are the very high
values for southeast Asian countries and the very small (even negative) values for eastern
European transition economies. This suggests that, particularly in these countries, there is
much to gain from reallocating resources across firms within each industry.
At this point, it is worth noting the caveat with which I started my presentation of
empirical evidence: the numbers I present are based on official
Table 6: Gap between weighted and un-weighted labor productivity (1990s)

and share of informal economy.
Source: Bartelsman et al (2005) and World Bank, respectively.
Table 7: Barriers to economic activity.
Source: World Bank Indicators (2003).
statistics, and thus miss much of the informal economy. Insofar as "informal" firms have
lower than average productivity, the real gap is lower than the one reported in Table 6.
Since Portugal exhibits a higher fraction of informal economy activity, the difference
between the productivity gap for Portugal and other European countries may not be so
high—in fact, it may well be negative.
2.4. Distortions to economic activity

It is common wisdom that the Portuguese economy is subject to myriad economic
distortions, and that Portuguese firms must overcome numerous barriers to succeed. In
Table 7, I summarize some of the information from the World Bank's World Economic
Indicators on indices that refer to the economic environment in which firms operate.
One must take these numbers with a grain of salt. For example, the number of
days to enforce a contract in Portugal, 365, suggests that the number was not calculated
with great precision; another suspicious example is that the lowest number of procedures to
enforce a contract is Indonesia at zero (Australia follows with 11); and so forth.
I also include in Table 7, pro memoria, the level of GDP per capita. It is noticeable
that Portugal is an outlier in the sense that the ranking in terms of barrier indices is
significantly lower than that in terms of per capita GDP.
An alternative source of data on entry barriers is the World Bank's database on
doing business. Table 8 presents a few variables of interest, both for Portugal and for the
remaining countries in the sample.7
7
See Djankov et al's (2002) for various notes on this dataset. Note that there are some discrepancies between
Tables 7 and 8 regarding the data for Portugal. First, the values of GDP per capita are different; however, one
must consider that the value is measured for different years and in different units. Second, the time to start a
business is lower in Table 8; but here the measure is in business days, not calendar days, so the difference is not
that great
Cost of starting a business (% of GNI per capita)
Figure 5: Cost of starting a business and per-capita GNI (countries with per-capita GNI
greater than $5,000). The data point for Portugal is represented by a bullet point. Source:
World Bank.
Time to start a business (days)
Figure 6: Time to start a business and per-capita GNI (countries with per-capita GNI
greater than $5,000). The data point Portugal is represented by a bullet point. Source:
World Bank.
Table 8: Entry costs. Source: Djankov et al (2002)
In order to investigate the possibility that Portugal is an outlier in terms of

distortions to economic activity, I consider two graphs, Figures 5 and 6. In these graphs, I
select countries with a Gross National Income (GNI) higher than $5,000 and plot the cost
of starting a new business (Figure 5) and time to start a business (Figure 6), as well as the
level of each country's GNI. I also plot the regression line between the two variables.
Both figures suggest that barriers to entry tend to be smaller in more developed
countries. Regarding the cost of entry, Portugal seems to be right on average.8 However,
regarding time to start a business, Portugal is very much on the outer edge of the
distribution. In summary, bureaucratic barriers to entry are very large in Portugal, even
when controlling for the level of economic development.
Similar calculations can be performed on the sample of 85 countries considered by
Djankov et al (2002). Table presents some summary statistics. Broadly speaking, the
numbers for Portugal don't seem too far off the sample average. What is particularly
striking, however, is the fact the total cost in Portugal is 2.64 times higher than the
monetary cost of entry, whereas the ratio of the averages is only 1.40. In other words, the
8 The lines in each figure are the estimated value from regressing the vertical-axis variable on the horizontal-axis
variable. The values of R2 are 30 and 10%, respectively; the coefficients relating the two variables in each
graph have p values of .003 and .000, respectively.
entry cost in Portugal seems disproportionately due to the cost of time in getting the
necessary approvals for entry.
Table 9: Indeces of Product Market Regulation and

Barriers to Entrepreneur-ship.
Source: Conway et al (2003).
A separate and very different source of data is the OECD data set on product
market regulation.9 Table 9 presents values for the overall index (Product Market
Regulation) as well as a component index, "Barriers to Entrepreneur-ship," which has a
weight of approximately 30% in the overall index. Two facts stand out from the table: first,
the recent trend towards a lower degree of regulation.10 Second, the relatively small
difference between Portugal and the OECD average.11
3. Industry dynamics: economic analysis

Good policy analysis must be based on good knowledge of data and stylized facts;
but it must also be grounded on a coherent model of reality. In this section, I summarize
some of the recent developments in the economic analysis of industry dynamics, especially
as it addresses the stylized facts presented in the previous section: industry turnover, firm
size, productivity. I will conclude with the analysis of how market distortions impact the
dynamics of firm entry and exit.
9
See Conway et al (2003).
10
This tendency is marginally more pronounced in the component "state control."
11
In 2003, the standard deviation of the PCR and BTE indices was 0.43 and 0.42, respectively, so the differences 0.1 and
0.2 are economically and statistically small.
3.1. Explaining industry turnover

In order to explain the stylized facts described in the previous section, we need to
relax some of the assumptions of the model of perfect competition. I will maintain that (i)
firms are price takers; (ii) the product is homogeneous; (iii) information about prices is
perfect. However, in contrast with the perfect competition model, suppose now that: (iv)
firms must pay a sunk cost in order to enter; (v) not all firms have access to the same
technology.
Specifically, suppose that different firms have different degrees of efficiency,
which in turn correspond to different cost functions: more efficient firms have a lower
marginal cost schedule. These differences may result from a variety of factors. For
example, some managers are more efficient in organizing resources than others (more on
this below).
Suppose moreover that each firm is uncertain about its own efficiency. When a
firm first enters an industry, it has only a vague idea of what its efficiency is. As times goes
by, and based on each period's experience, the firm gradually forms a more precise estimate
of its true efficiency. In each period, the firm chooses optimal output based on its current
expectation of efficiency; roughly speaking, the output level such that price is equal to
expected marginal cost.
Given the above elements, we conclude that firms which get a series of bad signals
(high production costs) gradually become "pessimistic" about their efficiency level,
gradually decrease their output, and, eventually, may decide to exit the industry (as variable
profit is insufficient to compensate for the fixed cost). By contrast, firms that receive a
series of good signals (low production costs) remain active and gradually increase their
output.
This model of competitive selection is consistent with several of the stylized facts
described in the previous section. In particular, the model is consistent with the stylized fact
of simultaneous entry and exit in the same industry. Firms which accumulate a series of
very unfavorable productivity signals hold a very unfavorable estimate of their own
efficiency. As a result, their expected value from remaining active is negative, which in
turn leads them to exit. New entrants have no information regarding their efficiency. Their
expected efficiency is therefore much better than the exiting firms': no news is better than
bad news. This justifies that their expected value from being active is positive, in fact,
greater than the entry cost. In summary, it is possible for a firm with no information about
its efficiency to enter while a firm with unfavorable information about efficiency exits.
Efficient firms are firms with a low marginal cost function. Since firms equate
price to (expected) marginal cost, it follows that more efficient firms sell a higher output.
Together with the previous results, this implies that exiters (the active firms with lowest
expected efficiency) are also the firms with lower output. By selection, the firms that
remain active have an efficiency higher than average. In particular, higher than the average
entrant's. It follows that entrants' output is lower than the surviving firms' average output.
In this way, the model is also consistent with the stylized fact that firms that enter and firms
that exit are smaller than average.
Finally, the competitive selection model is also consistent with the empirical
observation that the firm size distribution is neither single-valued nor indeterminate, as the
perfect competition model would imply. In fact, a given population distribution of
efficiency levels implies a particular distribution of firm sizes.12
3.2. Explaining the evolution of new firms

As mentioned in Section 2.2, new firms are characterized by a very right-skewed
distribution. As time moves on, the size distribution of that cohort shifts to the right and
becomes closer to a log-normal. One first natural interpretation of this pattern is that the
smallest of the small firms are exiting in greater numbers, leaving a set of survivors that has
a more symmetric distribution (in particular with fewer very small firms). However, using
Portuguese data Cabral and Mata (2003) reject this interpretation. Figure 7 plots three
curves: the size distribution of 1984 entrants; the 1991 distribution of 1984 entrants who
survived until 1991; and the 1984 size distribution of the 1984 entrants that survived until
1991. As can be seen the first and the third densities are fairly similar, which contradicts
the hypothesis that selection is doing most of the work.
12
It may be worth to point out that the competitive selection model does not depend on firms being asymmetric with respect
to costs. We could alternatively assume that some firms' products are better than others'.
Cabral and Mata (2003) propose an alternative hypothesis, namely that financing
constraints play an important role. Among the set of entrants, some firms start off very
small because they don't have the resources to start off at their efficient level. As they
gradually become less financially constrained, the distribution moves to the right and
becomes more symmetric (as found in the data).
Figure 7: Size distribution of the 1984 cohort of entrants: densities based on 1984 and
1991 data as well 1984 data for the firms that survived through 1991.
Barrios et al (2005) corroborate this explanation for data from Ireland. They show
that the above evolution of firm size skewness is found in Irish firms but not in
multinational firms. This seems consistent with the hypothesis that multinational firms are
less financially constrained and therefore not subject to the above dynamics. However,
Fagiolo and Luzzi (2004), using Italian data, reject the hypothesis that financing constraints
have a significant effect. Based on firm level survey data where respondents indicate if they
are financially constrained, they construct two distributions of new firm size distribution:
that of financially constrained firms and that of not financially constrained firms. The
difference between the two distributions is minimal.
An alternative explanation for the evolution of firm size focuses on the role of
sunk costs. Cabral (1995) provides an theoretical explanation for the negative relation
between firm size and firm growth among new entrants. The idea is that capacity and
technology choices involve some degree of sunkness (that is, investments for which value
is foregone upon exit). Since small entrants are more likely to exit than are large entrants, it
is optimal for small entrants to invest more gradually, and thus experience higher expected
growth rates upon entry than do large entrants.
3.3. Explaining variability in productivity

The characterization of the firm size distribution provided by Jovanovic's model
is, to a great extent, tautological: the distribution of efficiency levels is assumed rather than
derived; a more satisfactory model would also explain the distribution of efficiency levels.
One possibility is to assume that firms invest in R&D and that efficiency levels result from
these R&D investments. This is what the model by Ericson and Pakes (1995) does.
Ericson and Pakes (1995) consider a model where each firm's productivity results
from rationally chosen R&D investments. Specifically, in each period firms decide whether
to remain active and, if so, how much to invest in R&D. Investment leads to a stochastic
improvement in the firm's type, which can be interpreted in a way similar to Jovanovic
(1982).13
Ericson and Pakes (1995) show that there exists a unique rational expectations
Markov equilibrium; moreover, this equilibrium is ergodic. In the long-run, there will be a
distribution of firm types and firm sizes. In other words, instead of assuming a distribution
of types, as in Jovanovic (1982), we now derive such distribution as an equilibrium result.
The key to heterogeneity of firm types is therefore the randomness of the investment
process.
It seems reasonable to assume differences in efficiency and productivity result
from luck of the draw in the R&D process. Still one may wonder how such large
differences in productivity as those reported in the previous section can be sustained in the
long run. After all, one of the main effects of competition is precisely to weed out the
under-performers. One possible explanation, advanced by Syverson (2004a), is that product
differentiation allows firms to survive with below-par productivity levels. Consistently with
this explanation, Syverson (2004a) shows that industries with greater degree of product
differentiation are also industries with greater variability in productivity.
13
Ericson and Pakes (1995) also consider the possibility of non-competitive behavior by firms.
An alternative explanation is that not all industries are that competitive. If the
more productive firms price above cost then less efficient firms may be able to survive (if
with lower margins). Syverson (2004b) tests this hypothesis in the concrete industry. He
shows that in geographical areas where there is less competition there is also greater
variability in productivity levels.
A related test is that of Asplund and Nocke (2005). They show that in
geographical markets where demand density is greater, and thus market competition more
intense, the average life span of an entrant is lower. While they don't directly present
theoretical or empirical results on productivity level, this result is consistent with
Syverson's (2004a,b) on the disciplining effects of competition.
In summary, firm heterogeneity (namely in terms of productivity level) is a
combination of exogenous firm attributes (managerial ability, company culture, etc) and
luck in the investment process.14 Although typical industries exhibit a significant degree of
variation in productivity levels, such dispersion is smaller in more competitive industries
(e.g., industries with lower degree of product differentiation).
3.4. Market distortions and welfare

The fundamental theorem of welfare economics states that, under perfect com-
petition, the market solution is efficient. But, as we've seen in Section 2, the perfect
competition model does not stand up to the facts very well. What then can be said of the
market efficiency result? To my knowledge, Hopen-hayn (1992) provides the most general
extension of the fundamental theorem. Based on an extension of Jovanovic's (1982) model,
he shows that the market equilibrium is efficient if firms are price takers — even if
efficiency varies across firms and across periods. This is a strong result. In particular, if
there are no entry barriers then the equilibrium level of firm turnover (possibly with high
firm turnover compared to the net entry rate) is socially optimal.
14
Pakes and Ericson (1998) test the relative importance of these two sources of heterogeneity. They show that firm type is
ergodic in manufacturing but not in services. This is consistent with the interpretation that Jovanovic's (1982) story does a
better story at explaining the dynamics of firms in the services sector, whereas Ericson and Pakes (1995) is a better
model of firms in manufacturing.
In other words, for all its differences with respect to perfect competition,
competitive selection maintains one important property: efficiency. So long as firms act as
price takers, the equilibrium solution, absent any artificial barriers, is socially efficient.
Each firm's output decision in each period is efficient: price equal to expected marginal cost
is the most efficient output decision, that is, the one that maximizes total surplus.
Moreover, it can be shown that the firm's entry and exit decisions are also optimal from a
social point of view. The basic idea is the same as in the model of perfect competition: a
very small firm has a negligible impact on other firms and on price. It follows that it
internalizes all of the costs and benefits from entering or exiting the industry: what is good
for the firm is good for society.
It might seem inefficient to have firms entering and exiting a given industry
simultaneously. But we must remember that firms are uncertain about their efficiency. The
only way to determine a firm's efficiency is to actually enter the industry. A central planner
who attempted to maximize total surplus would not be able to do better than the market. In
summary, it can be shown that the equilibrium under competitive selection is efficient.
It follows that distortions to the natural workings of the market lead to a lower
level of social welfare. A particularly careful illustration of this point is provided by
Hopenhayn and Rogerson (1993). Based on the Hopenhayn's (1992) model of competitive
selection, Hopenhayn and Rogerson (1993) estimate the impact of a distortion to the
process of firm creation and destruction. Specifically, they consider the impact of firing
costs. The results are staggering: for example, they estimate that a tax on dismissals
equivalent to 1 year's wages reduces steady-state utility by over 2% measured in terms of
consumption. The welfare loss comes about from a 8% reduction in firm turnover. This
implies that less efficient firms are active whereas other, more active firms, remain
inactive. Also, the labor adjustment cost implies that some firms are smaller or greater (in
terms of number of employees) than it would be efficient. Hopenhayn and Rogerson (1993)
estimate that, for more than 90% of the firms, the gap between the marginal productivity of
labor and wage would be greater than 5%.
I suspect that Hopenhayn and Rogerson's (1993) analysis yields a lower bound of
the welfare loss from barriers to entry and mobility. The reason is that they assume all firms
are subject to the same barrier. But casual observation suggests that, just as there is
significant variability in firm productivity, there is also significant variability in the barriers
to entry and mobility that each firm faces.
In order to get an idea of the magnitude of this effect, consider the following
simple model. Suppose there are 1,000 price taking firms each with capacity 1.15 I assume
that marginal cost is constant and normally distributed with mean 100. Consistently with
the empirical evidence, I assume a coefficient of variation of 1, so standard deviation is also
100. Market inverse demand is given by p = 200 — .01Q, where Q is total output (number
of active firms).
I first compute the equilibrium in this economy. This amounts to ordering firms by
marginal cost, thus obtaining market supply; and then finding the supply-demand
equilibrium.
Suppose now that this economy is subject to a series of distortions. Specifically,
firm i's marginal cost is changed by ti , where t i is normally distributed with mean μ and
standard deviation σ . I assume the value ti qi is a transfer, so the only social cost implied
by σ is the distortion it creates.
The social cost of distortion t i can be divided into two terms. First, assuming no
change in costs, t i leads to a gap between price and marginal cost and the corresponding
Harberger excess burden triangle (allocative inefficiency). Second, for a given output level,
t i also implies an increase in production cost compared to the minimum total production
cost (productive inefficiency).
15
The model can easily be extended to firms with different capacities; simply assume that some firms have multiple
establishments, each consisting of one of the "firms" that I consider.
Figure 8: Welfare loss from cost distortions. Each firm's cost is increased by a stochastic
variable with distribution N (μ, σ ).
Figure 8 shows the welfare effects of distortions for various values of μ and σ .
Consider first the case when σ = 0. As we increase the value of μ, the welfare loss
increases. Notice that the relation between μ and welfare loss is convex. If σ is low, then
the welfare loss is limited to allocative inefficiency, which is proportional to t i 2 = μ2. It

follows that a small value of μ implies a very small welfare loss. This is, in essence, the
underlying argument in Harberger's (1954) claim that the welfare loss from market
distortions in the U.S. is very small.
Figure 8 suggests that the welfare loss is also convex with respect to σ. For
example, if σ = 50, then welfare is as large as when μ = 50 and σ = 0. In fact, even if μ is
very small, a high value of σ implies a high welfare loss.
In summary, the welfare cost of distortions depends not only on the size of
distortions (μ) but also on how these vary across firms ( σ ). There are various sources of
cross-firm variation, including tax evasion and more generally the informal economy. I will
return to these in the next section.
4. Policy implications
The theoretical ideal of perfect competition includes, among others, the pre-
sumption that firms are price takers and that there are no barriers to mobility. While the
assumption of price-taking behavior is a reasonable approximation to many industries, the
ideal of a level playing field with no barriers to entry and mobility is far from reality.
Competition policy has made great inroads in the treatment of the classical
"problematic" cases: mergers and acquisitions in concentrated industries, public utilities,
and so forth. But a lot still needs to be done in a variety of so-called "competitive"
industries. Here is where, in my opinion, the main focus of microeconomic policy should
lie.
In this section, I further develop this point. First, I focus on the idea of industry
turnover as a characteristic of the normal behavior of the economy. Next I argue that
turnover is actually a very important channel for the process of industry productivity
growth. Finally, I present the ideal of economic mobility as one of the main goals of
competition policy.
4.1. Turnover and welfare

One of the most common misperceptions regarding micro policy is that it should
protect firms from failing and exiting. It is true that the immediate effect of saving a firm
from exiting is to save a number of jobs equal to that firm's employment. But such policy
would imply a significant welfare cost in terms of resource misallocation. A firm that's
artificially kept alive implies a firm that will not be created, knowing that the latter would
probably be more efficient than the former. In this regard, the welfare cost estimates by
Hopenhayn and Rogerson (1993) are particularly telling.
One variation of the above misperception is that public policy should protect small
firms especially, since their turnover rate is typically higher than average (as documented in
Section 2). But again, a high turnover rate may well be part of the natural process of
experimentation inherent to a healthy competitive industry. In fact, the cost of
experimentation is smaller for small firms, and so it is only natural that in equilibrium we
observe higher entry and exit rates for small firms. So, as a matter of principle one cannot
say that because their turnover rate is higher small firms should be particularly protected.
Having said that, it is true that some market imperfections hit small firms in a
particular way. For example, imperfect credit market conditions are likely to bias the
market against small firms. In general, any imperfections that increase a firm's fixed cost
slant the field against small firms. The solution is then to correct as much as possible for
those credit market imperfections.
4.2. Productivity growth

One of the most important results from the analysis of time series productivity
data is the importance of industry turnover in the process of productivity growth. There are
essentially two ways in which to increase average productivity in a given industry. One is
to increase the productivity level of each firm; the other one is to increase the relative
weight of higher productivity firms. Public policy frequently heralds the former and places
less emphasis on the latter. Governments programs to improve the quality of human and
physical capital inputs have greater political impact; whereas fostering a more fluid process
of firm entry and exit, if anything, carries a political cost. This is unfortunate, for the
empirical evidence is that resource reallocation is the primary source of productivity
growth.
Olley and Pakes (1996) offer an interesting case study of the role of turnover in
productivity growth. They look at the 1980s deregulation process in the U.S.
telecommunications equipment industry. The common wisdom is that deregulation
improves efficiency by forcing incumbent firms to become more efficient. Although
telecommunications equipment firms indeed became more efficient following deregulation,
the greatest source of industry improvement was the process of capital reallocation amongst
incumbent firms.
The exercise of productivity growth accounting leads to different results in
different economies and industries; it also depends on the particular definition of
productivity.16 However, the common message of all of these exercises is the importance of
the process of industry turnover, both in terms of entry and exit and in terms of resource
reallocation among incumbents.
This naturally leads to the question, What can public policy do about it? I next
argue that one of the primary areas of competition policy ought to be guaranteeing the basic
conditions of economic mobility: a level playing field where competition may work
effectively towards the selection of the more efficient firms.
16
See Ahn (2001) for a survey.
4.3. Economic mobility

The Fundamental Theorem of Welfare Economics is remarkably robust. Although
it is usually formulated in the context of perfect competition, which is based on a lot of
assumptions, one only really needs to assume (a) price taking behavior and (b) free entry,
exit and mobility. In particular, as shown by Hopenhayn (1992), market efficiency is
consistent with firm heterogeneity and imperfect information.
Most firms in the Portuguese or any other economy, especially small and medium-
sized firms, face a fairly flat demand curve. It follows that price taking behavior is a fairly
good first-order approximation. We are thus left with condition (b). Unlike price-taking
behavior, the ideal of economic mobility is very far from the reality of the Portuguese and
most other economies.
Contrary to the increasingly discredited advocates of a strong industrial policy
(e.g., government investment in national champions), a growing majority of economics
scholars advocates that the best microeconomic policy is a good competition policy. And
within micro policy I would advocate that promoting economic mobility plays a central
role. Competition policy typically focuses on concentrated industries, industries where
firms have market power. Competitive industries (that is, industries where firms are price
takers) are frequently given less importance. But the term "competitive" can be deceiving:
price taking behavior is an important step towards efficient competition, but it is not the
only one. One must make sure that "competitive" industries are truly competitive. This is
where economic mobility comes in.
I denote by economic mobility the set of conditions ensuring that, in competitive
industries (industries where firms are price takers), competition leads to an efficient
equilibrium. Barriers to entry, such as the bureaucratic costs of creating a new firm, are an
obvious instance of a distortion that drives the market equilibrium away from the efficient
outcome idealized by the Fundamental Theorem. If all firms were equal, barriers to entry
would not be a very big problem: fewer firms might enter than with no barriers, but the loss
would not be that great if the size of the barrier were not that great.17 But all firms are not
equal; and so, more than higher prices, the implication of a barrier to entry is a lower rate of
turnover and thus a less than perfect replacement of less efficient incumbents.
But economic mobility is not just about barriers to entry. The empirical evidence
suggests that much of the reallocation of productive resources takes place among active
firms. Any artificial barrier that encumbers this process has an effect on turnover similar to
a barrier to entry. One example in given by severance payments and, more generally, legal
and economic restrictions to layoffs; but there are more such examples.
Another point suggested by the analysis in the previous section is that a crucial
aspect of economic mobility is not so much the size of distortions but how they differ
across firms. In fact, the loss of productive efficiency is more likely to come from the
variation in the size of distortions than its size. It's bad enough if potential entrants must
pay an extra cost to become active; but it's much worse if some potential entrants must pay
a higher cost than others.
There are several sources of cross-firm variation in barriers to entry and mobility.
For example, when entry entails complicated bureaucratic steps, the worst thing that can
happen is that some firms may be able to evade those bureaucratic costs. Other important
sources include fiscal evasion and more generally avoidance of government imposed
regulations (labor, environmental, and so forth). The point is that fiscal evasion, for
example, is not simply a problem of fairness; it's also an efficiency problem.
In sum, economic mobility is a set of conditions that create a level playing fields,
one where firms can easily enter and exit, grow and contract, according to their relative
efficiency; an environment where market selection leads to efficient selection.
17
This is in essence the point of Harberger's (1954) estimate of the social cost from monopoly: if the distortion is small then
the welfare loss is of second order.
5. Final remarks
Most of microeconomics research in the past few decades has been devoted to
showing when and why markets do not work. The game theory revolution — promptly
taken up by industrial economists — stresses the importance of market power. Other
scholars, from Akerlof to Spence to Stiglitz, focus on asymmetric information and market
failures. In the 1980s, work by Arthur and others highlights the importance of increasing
returns and non-ergodic market outcomes. It would seem that classical and neo-classical
economics are dead.
In fact, the Fundamental Theorem of Welfare Economics is alive and well. True,
there are industries with special problems of market power, asymmetric information,
natural-monopoly structures, and so forth. But for the most part the assumption of price
taking behavior is a fairly good first-order approximation. And the Fundamental Theorem
is then the best guide to public policy: create a level playing field and let the market do the
rest.
References
AHN, SANGHOON (2001), "Firm Dynamics and Productivity Growth: A Review of
Micro Evidence From OECD Countries," OECD Economics Department Working
Paper 297.
ASPLUND, MARCUS, AND VOLKER NOCKE (2005), "Firm Turnover in Imperfectly
Competitive Markets," forthcoming in Review of Economic Studies.
BARRIOS, SALVADOR, HOLGER GORG, AND ERIC STROBL (2005), "The Evolution
of the Firm Size Distribution and Nationality of Ownership," Economics Bulletin
12, 1-11.
BARTELSMAN, ERIC, JOHN HALTIWANGER, AND STEFANO SCARPETTA
(2004), "Microeconomic Evidence of Creative Destruction in Industrial and
Developing Countries," IZA Discussion Paper 1374.
BARTELSMAN, ERIC, JOHN HALTIWANGER, AND STEFANO SCARPETTA
(2005), "Measuring and Analyzing Cross-country Differences in Firm Dynamics,"
Free University Amsterdam, University of Maryland, World Bank.
BARTELSMAN, ERIC, STEFANO SCARPETTA, AND FABIANO SCHIVARDI (2003),
"Comparative Analysis of Firm Demographics and Survival: Micro-Level Ev-
idence for the OECD Countries," OECD Economics Department Working Paper
348.
CABLE, JOHN, AND JOACHIM SCHWALBACH (1991), International Comparisons of
Entry and Exit, in Geroski and Schwalbach (1991).
CABRAL, LUÍS (1995), "Sunk Costs, Firm Size and Firm Growth," Journal of Industrial
Economics 43, 161-172.
CABRAL, LUÍS (2000), Introduction to Industrial Organization, Cambridge, MA: MIT
Press.
CABRAL, LUÍS (2005), "Economic Mobility," New York University.
CABRAL, LUÍS, AND JOSÉ MATA (1996), "On the Evolution of the Firm Size Dis-
tribution: Facts and Theories," American Economic Review 93, 1075-1090.
CONWAY, PAUL, VERONIQUE JANOD, AND GIUSEPPE NICOLETTI (2003),
"Product Market Regulation in OECD Countries: 1998 to 2003," OECD
Economics Department Working Paper 419.
DJANKOV, SIMEON, RAFAEL LA PORTA, FLORENCIO LOPEZ-DE-SILANES,

AND ANDREW SHLEIFER (2002), "The Regulation of Entry," Quarterly
Journal of Economics 117, 1-38..
DUNNE, TIMOTHY AND A. HUGHES (1994), "Age, Size, Growth and Survival: UK
Companies in the 1980s," Journal of Industrial Economics 17, 115-140.
DUNNE, TIMOTHY, MARK ROBERTS, AND LARRY SAMUELSON (1989), "The
Growth and Failure of U.S. Manufacturing Plants," Quarterly Journal of
Economics 104, 671-688.
ERICSON, RICHARD AND ARIEL PAKES (1995), "Markov-Perfect Industry Dynamics:
A Framework for Empirical Work," Review of Economics Studies 62, 53-82.
EVANS, DAVID (1987), "The Relation Between Firm Growth, Size, and Age: Estimates
for 100 Manufacturing Industries," Journal of Industrial Economics 35, 567-581.
FAGIOLO, GIORGIO, AND ALESSANDRA LUZZI (2004), "Do Liquidity Constraints
Matter in Explaining Firm Size and Growth? Some Evidence from the Italian
Manufacturing Industry," LEM Working Paper 2004/08, Sant'Anna School of
Advanced Studies, April.
GEROSKI, PAUL (1991), Market Dynamics and Entry, Oxford: Blackwell.
GEROSKI, PAUL (1996), "What Do We Know About Entry?," International Journal of
Industrial Organization 13, 421-440.
GEROSKI, PAUL, AND J. SCHWALBACH (Eds) (1991), Entry and Market Contesta-
bility, Oxford: Basil Blackwell.
GIBRAT, R. (1931), Les Inegalites Economiques. Applications: Aux Inegalites des
Richesses, a la Concentration des Entreprises, Aux Populations des Villes, Aux
Statistiques des Familles, etc., d'une Loi nouvelle: La Loi de L'ffect Proportionnel,
Paris: Sirey.
HALL, BRONWYN (1987), "The Relationship Between Firm Size and Firm Growth in the
U.S. Manufacturing Sector," Journal of Industrial Economics 35, 583-606.
HARBERGER, ARNOLD C. (1954), "Monopoly and Resource Allocation," American
Economic Review 44, 77-87.
HOPENHAYN, HUGO (1992), "Entry, Exit, and Firm Dynamics in Long Run Equi-
librium," Econometrica 60, 1127-1150.
HOPENHAYN, HUGO, AND RICHARD ROGERSON (1993), "Job Turnover and Policy
Evaluation: A General Equilibrium Analysis," Journal of Political Economy 101,
915-938.
JOVANOVIC, BOYAN (1982), "Selection and Evolution of Industry," Econometrica 50,
649-670.
LUCAS, ROBERT (1978), "On the Size Distribution of Business Firms," Bell Journal of
Economics 9, 508-523.
MATA, JOSÉ AND PEDRO PORTUGAL (1994), "Life Duration of New Firms," Journal
of Industrial Economics 42, 227-246.
MATA, JOSÉ, PEDRO PORTUGAL, AND PAULO GUIMARAES (1994), "The Survival
of New Plants: Start-Up Conditions and Post-Entry Evolution," International
Journal of Industrial Organization 13, 459-481.
OLLEY, STEVEN G., AND ARIEL PAKES (1996), "The Dynamics of Productivity in the
Telecommunications Equipment Industry," Econometrica 64, 1263-1297.
PAKES, ARIEL, AND RICHARD ERICSON (1998), "Empirical Implications of Alter-
native Models of Firm Dynamics," Journal of Economic Theory 79, 1-45.
SUTTON, JOHN (1997), "Gibrat's Legacy," Journal of Economic Literature 35, 40-59.
SYVERSON, CHAD (2004a), "Product Substitutability and Productivity Dispersion,"
Review of Economics and Statistics 86, 534-50.
SYVERSON, CHAD (2004b), "Market Structure and Productivity: A Concrete Example,"
Journal of Political Economy 112, 1181-1222.
TIROLE, JEAN (1988), The Theory of Industrial Organization, Cambridge, MA: MIT
Press.
ASYMMETRIC INFORMATION IN THE STOCK MARKET:
ECONOMIC NEWS AND CO-MOVEMENT BETWEEN US AND PORTUGAL∗
Rui Albuquerque∗∗
Clara Vega∗∗∗
Abstract
We analyze the effect that real-time domestic and foreign news about
fundamentals have on the correlation of stock returns of a small open economy,
Portugal, and a large open economy, the U.S. We also study the role of public and
private information in the price formation process in the U.S. and Portuguese stock
markets. Consistent with our theoretical model, we find that U.S. macroeconomic
news and Portuguese earnings news do not affect the cross-country stock market
correlation, whereas Portuguese macroeconomic news lowers the cross-country stock
market correlation. U.S. public information affects Portuguese stock market returns,
but this effect is diminished when U.S. stock market returns are included in the
regression. This means that part of the co-movement between the U.S. and Portugal is
due to the effect that U.S. macroeconomic fundamentals have on the Portuguese stock
market. Finally, public information news in the U.S. is associated with increased
liquidity, while the effect in Portugal depends on the type of news releases.
Keywords: Private information, public news announcements, information spillovers, international

equity returns, contagion.
∗
Paper prepared for the conference Desenvolvimento Económico Português no Espaço Europeu. We thank
António Antunes for providing us with the Portuguese stock market data, Jośe Mata for providing us with the
macroeconomic news release schedules and the Banco de Portugal for funding this research. The usual
disclaimer applies.
∗∗
Boston University School of Management and CEPR. Address at BU: Finance and Economics, 595
Commonwealth Avenue, Boston, MA 02215. Email: ralbuque@bu.edu. Tel.: 617-353-4614.
∗∗∗
Board of Governors Federal Reserve System and William E. Simon School of Management. Address:
Washington, DC 20551 USA. E-mail: clara.vega@frb.gov. Tel.: 202-452-2379.
258 Rui Albuquerque – Clara Vega
1. Introduction
What drives the cross-country correlation of stock returns? This question, central
to many topics in international finance, is deeply associated with another fundamental
question: what is the role of public and private information in the price formation process?
Under market efficiency, information about future stock prices in one market spreads to
other markets where it can also be used to forecast local prices. Cross-country stock return
correlations thus arise from correlated fundamentals. However, the empirical link between
economic fundamentals and cross-country correlations has by and large eluded researchers,
leading some to explore contagion as the main driver of stock market correlations.
Uncovering the nature of cross-country stock return correlations is especially relevant for
the development and stability of emerging stock markets that are less liquid, more volatile
and have a small and fragile investor base: contagion is likely to increase volatility and
decrease the investor base, forcing smaller investors out of the market.
This paper revisits theoretically and empirically the link between cross-country
correlations and fundamentals and the price discovery process. The standard view in the
literature is that shocks to global factors lead to increases in stock return correlations as
they move the value of firms around the globe in the same direction. Under this view it is
then puzzling that the cross-country correlation of returns is the same on U.S. news
announcement days and non-announcement days. To the extent that shocks to global
factors are shocks to fundamentals, it is also puzzling that U.S. public macroeconomic
news affect returns in other countries only if the U.S. return is not included in the
regression.1
We present a simple model of stock trading, which draws on the work of Kyle
(1985) and King and Wadhwani (1990), that rationalizes the observed, puzzling patterns. In
the model there is one large, foreign country, and one small, local country. The stock price
of the large country is driven by a global factor and the stock price of the small country is
driven by two factors, a global factor and a country-specific factor. For simplicity of
exposition let the large country be the U.S. and the small country be Portugal. Investors in
1
We review the related literature below.
Asymetric Information in the Stock Market: Economic News and Co-movement 259
the U.S. are assumed not to respond to news from Portugal,2 but the converse does not
apply. Each country is independently populated with informed investors, noise traders, and
a competitive market maker. In the model, stock return comovement arises as a result of
market efficiency: investors in Portugal use U.S. returns to infer the private information of
U.S. informed investors about the global factor. In days with high private information
signals about U.S. assets, U.S. net order flow is positive and so are U.S. returns. Since
fundamentals are correlated across countries, this information is good news for the
Portuguese stock market and prices rise locally as well. Thus, Portuguese and U.S. stock
market returns move together even in the absence of public announcements.
The model predicts that this baseline return comovement should not change in
days of U.S. public news announcements. The effect of Portuguese news on comovement
depends on the content of the news. If the news are correlated with U.S. valuations
(Portuguese macroeconomic news), then comovement declines on news days, but if the
news are country-specific (earnings announcements) there is no change in this
comovement.3 The main assumption for these results is the asymmetric response to
Portuguese news by investors: U.S. public news releases generate price discovery in both
U.S. and Portuguese markets, whereas Portuguese news releases generate price discovery
in the local market only. Consider first the effects of macroeconomic Portuguese news
announcements. The absence of price discovery in the U.S. market implies that Portuguese
investors know that trading by U.S. informed investors is not contaminated by the local
news. In the context of our model this means that local macroeconomic news can, among
other things, improve on the information content of U.S. returns leading to a decline in
cross-country correlations. Consider now the effect of Portuguese earnings news on
comovement. The nature of the news assumes that the mechanism just described is absent
2
They might recognize its value, but it is assumed that the costs of processing this information (and that
originating from many other small countries) outweight the benefits, and the information is not used in the
spirit of Grossman and Stiglitz (1980).
3
Portuguese firm-specific earnings announcements are not orthogonal to U.S. valuations. However, to
empirically test our model we need to condition on public announcements that reveal relatively more
information about country-specific factors than about global factors. The PSI-20 equal weighted and value
weighted earnings are not significantly correlated with U.S. GDP growth, while Portuguese GDP growth is
significantly correlated with U.S. GDP growth. Hence, in our empirical test we interpret Portuguese
macroeconomic news as public announcements that are correlated with U.S. valuations, while earnings
announcements predominantly reveal infomation about country-specific factors. In other words, earnings
announcements are nearly country-specific, in Poterba's (1990) terminology.
and comovement is therefore not affected. Finally, consider the effects of U.S. news
announcements. Price discovery in the U.S. market means that informed U.S. investors
‘subtract’ the public news from their private information and trade only on the remaining
portion. Therefore, investors in the Portuguese market can no longer use the (same foreign)
news to improve on the information content of U.S. returns and cross-country correlations
are unchanged. The paper thus provides a theoretical explanation for why cross-country
correlations of returns do not change after U.S. public news announcements and makes
additional predictions on how comovement should change based on local news.
The model also explains why the U.S. return empirically subsumes the
information content of macroeconomic news releases in the U.S. in driving foreign returns.
The reason is that the U.S. return is a sufficient statistic for U.S. public and private
information. Finally, the model predicts that market liquidity should increase in response to
news announcements in the same market. The reason is that in a Kyle (1985) model the
public news ‘destroys’ some of the value of the private information. In turn, the market
maker is less concerned about adverse selection and liquidity increases.
We conduct our empirical analysis studying comovement between Portugal and
the U.S. We focus on Portuguese stock market data for two main reasons. First, the
globalization of the world economy and in particular the process of European integration
means that the Portuguese stock market is increasingly more vulnerable to external shocks.
However, given the relative sizes of the U.S. and Portuguese economies, shocks to the
Portuguese economy are not likely to affect the U.S. This makes it an interesting research
laboratory to study how the small Portuguese stock market responds to economic news at
home and abroad (from the U.S.) and allows us to avoid the endogeneity problem
associated with the comovement literature. Second, a unique feature of the Portuguese data
is that it contains signed trades.4 There is thus no need to rely on artificial algorithms which
add measurement error to estimated order flow imbalances.
In testing the hypothesis of our model we use real-time U.S. and Portuguese
macroeconomic announcements and high frequency stock market returns. Such high
frequency data allows us to probe the workings of the marketplace in powerful ways
4
One notable exception is the TORQ database, which contains signed trades for a sample of 144 NYSE stocks
for the three months November, 1990 through January 1991. The advantage of our dataset is that we have nine
and a half months of data from January 4, 2002 to October 15, 2002.
because: (i) we avoid problems related to the existence of non-synchronous trading periods
between countries (Karolyi and Stulz, 1996); (ii) we measure more accurately the effect
that macroeconomic news announcements have on U.S. and Portuguese prices (Andersen et
al., 2003) by focusing on episodes where the source of price revisions is well identified,
thus leading to a high signal-to-noise ratio; (iii) we are able to test the theoretical
assumption that foreign news are first incorporated into foreign stock prices and then they
are incorporated into local stock prices; (iv) we measure unanticipated order flow which
proxies for private information-based trading and are thus able to analyze the effects of
public news conditional on private information.
Consistent with our assertion that the Portuguese stock market is vulnerable to
global shocks, we find that U.S. public macroeconomic news surprises affect the individual
components of the Portuguese PSI-20 stock market index. As in previous studies, these
announcements lose some of their statistical significance when we include the DJ-30
Industrial value weighted index.5 According to our model, this suggests that part of the
comovement between the U.S. and Portugal is due to the effect that U.S. macroeconomic
fundamentals have on the Portuguese stock market.
While U.S. macroeconomic news affect Portuguese returns, consistent with
previous findings, these news do not affect the correlation of stock returns across the two
countries. The same is true for Portuguese earnings announcements. However, in days of
Portuguese macroeconomic news there is significantly lower comovement of returns. These
last two facts add to our knowledge of comovement of returns as most other studies have
focused on news from the U.S. alone. Also, all three facts are consistent with the model
developed here.
We also empirically show that U.S. macroeconomic news are first incorporated
into U.S. stock market returns and 5 to 15 minutes later they are incorporated into
Portuguese stock market returns. This empirical evidence supports the assumed timing of
the model that Portuguese investors incorporate changes in U.S. returns into their trading,
and to the best of our knowledge it has not been documented before because previous
literature has focused on frequencies lower than five-minute returns. Interestingly, news on
5
The actual DJ-30 Industrial index is a price weighted measure. In this paper, we construct our own index using
individual daily stock returns and taking a value weighted average.
the benchmark interest rate from the European Central Bank−the Portuguese Central Bank
does not have control over its own monetary policy−affect U.S. stock market returns and
price discovery takes place first in the U.S. stock market and then in the Portuguese stock
market.6 One possible explanation is that price discovery takes place first in the most liquid
market, the U.S. stock market (see Hasbrouck (2003)). Another possible explanation is that
the opening of the U.S. market coincides with the end of the press conference of the ECB
and the Reuters release on the ECB’s press conference so the response of the Portuguese
stock market is tied to the later fact not the former.
Finally, public information news in the U.S. increases liquidity in the U.S. market
and the same is true for earnings announcements and liquidity in the Portuguese market.
While these observations are consistent with our model, we also find that Portuguese
macroeconomic news decrease liquidity in the Portuguese market. One explanation for this
unexpected behavior of liquidity is that Portuguese macroeconomic news necessitate more
analysis to be useful, leading to the entry into the market of a different class of informed
investors (see Kim and Verrecchia (1994)).
We proceed as follows. Next we give a brief literature review. In Section 3, we
construct a model of trading to guide our empirical analysis. In Section 5, we describe the
data. In Section 6, we present the empirical results and Section 7 concludes. The appendix
contains the proofs of the results in the main text.
2. Related Literature
Our paper is most closely related to two areas of research. The first examines
international asset market linkages and the second highlights the role of order flow in the
price formation process. In this paper, the two are related because we take the view that
both international information spillovers and order flow are important in the price
discovery process. We briefly discuss these research areas in turn.
One approach to studying international asset market linkages is to assess the effect
of macroeconomic news announcements on asset returns around the world. Becker,
Finnerty and Friedman (1995) use high-frequency futures data from 1986 to 1990 and
6
Ehramann and Fratzscher (2003) show that ECB monetary policy surprises affect U.S. interest rate markets.
Perhaps not surprisingly then we observe that they also affect U.S. stock market returns.
document that U.K. stock market returns react to U.S. and U.K. macroeconomic news,
while the U.S. stock market only reacts to U.S. own news. Wongswan (2005) uses high-
frequency data from 1995 to 2000 to show that there is evidence of transmission of
information from the U.S. and Japan to the equity markets in Korea and Thailand.
Ehramann and Fratzscher (2003) model the degree of interdependence of the U.S. and
European interest rate markets by focusing on the reaction of these markets to
macroeconomic news and monetary policy announcements. They show that the connection
of the Euro area and the U.S. money markets has steadily increased over time, with the
spillover effects from the U.S. to the Euro area being somewhat stronger than in the
opposite direction. Gande and Parsley (2005) document that news ratings on sovereign debt
in one country affect yields in other countries and tie the spillover effects to country
fundamentals. In line with these papers, we find that the Portuguese stock market reacts to
U.S. macroeconomic news, but the U.S. does not react to Portuguese news.
In the context of stock markets some authors argue that the effect highlighted
above is to a large extent subsumed in the foreign return itself. King, Sentana and
Wadhwani (1995) construct a factor model of 16 national stock market monthly returns and
examine the influence of 10 key macroeconomic variables. They show that the surprise
component of these observable variables contributes little to world stock market variation
after conditioning on the common factor that is unrelated to fundamentals. Connolly and
Wang (2003) analyze the U.S., U.K. and Japan equity markets from 1985 to 1996 and
separate the influence of the foreign markets on domestic markets into news about
economic fundamentals and the foreign market return. They find that the macro news effect
is too small to account for any economically sizeable part of the return comovement among
the three national equity markets. While these authors interpret their results as indicative
that contagion, not economic fundamentals, explain stock market linkages, in our rational
asset pricing model, price discovery implies that there is no role for foreign news in
explaining local returns once we condition on foreign returns as foreign returns summarize
both private information and public information. In addition, in our paper we analyze
theoretically and empirically how the correlation across markets changes in the presence of
local and foreign news.
In a seminal contribution, Karolyi and Stulz (1996) model the correlation between
U.S. and Japanese stock markets and test how much of that correlation is explained by the
presence of U.S. news. Recognizing that different shocks can lead to opposite movements
in stock return correlations (thus biasing the results to not finding any association), Karolyi
and Stulz allow for international linkages to be driven by global shocks as well as
competitive shocks: ‘global shocks’ affect the value of all firms (domestic and foreign) in
the same direction, whereas ‘competitive shocks’ increase or decrease the value of firms in
one country relative to the firms in another country.7 In other words, ‘global shocks’
increase cross-country correlations whereas ‘competitive shocks’ lower cross-country
correlations. Relative to their results, we also find that ‘global shocks’−identified in our
paper and theirs as U.S. news−do not change the cross-country correlation of returns.
However, in contrast to their paper, we find that ‘competitive shocks’−identified by
Portuguese macroeconomic news−do affect the correlation of returns. We also provide a
model that can explain these apparently puzzling patterns.8
In our theoretical model there are no contagion effects.9 As in King and
Wadhwani (1990), informed investors in one market use returns in other markets to infer
additional information on fundamentals and hence they will sometimes mistakenly interpret
a high foreign return as evidence of a high private information signal in the foreign market.
However, in contrast to their paper, the equilibrium in our model generates a return
correlation that equals the correlation in the underlying economic fundamentals: on average
investors are inferring just the correct amount of information. Another important difference
between the two setups is that we adopt an equilibrium model of strategic informed trading
à la Kyle (1985) whereas King and Wadhwani (1990) adopt a rational expectations
equilibrium. This allows us to study the role of order flow in the price discovery process
7
Karolyi and Stulz (1996) use U.S. macro announcements and asset prices (foreign exchange rates, U.S.
Treasury bill futures, the Nikkei and Standard and Poor's 500 index returns) to identify global and competitive
shocks in the U.S. and Japan. They consider U.S. macro announcements, shocks to the indexes and interest
rates to be global and shocks to the Yen/Dollar exchange rates to be competitive.
8
The evidence in Karolyi and Stulz (1996) is especially disappointing in light of the papers that have shown
empirically and theoretically the significance of cross-country correlation in economic fundamentals in
explaining the observed size of international return correlations (e.g. Ammer and Mei (1994), Bekaert, Harvey
and Ng (2005), Craig, Dravid, and Richardson (1995), and Dumas, Harvey, and Ruiz (2003)). By studying
contagion effects in extreme market movements, Bae, Karolyi, and Stulz (2003) are able to find some evidence
in favor of the contagion hypothesis, but they also show that modeling international returns with fat tail
distributions can rationalize most of the observed extreme events (see also Campbell, Koedijk, and Kofman
(2002), Campbell, Forbes, Koedijk, and Kofman (2003), and Longin and Solnik (2001)).
9
See Claessens, Dornbusch, and Park (2001) for a comprehensive review of the contagion literature.
following local and foreign news and its effects in international stock return comovement
which had not been previously addressed. Finally, while the literature on contagion focuses
on understanding the mechanisms through which shocks in one country are transmitted to
others in the absence of any correlation in exogenous forcing variables, the focus of our
paper is on how public news in one country spreads to other countries and the subsequent
price discovery process under the premise−that we verify−that fundamentals are correlated.
Several recent studies have highlighted the role of order flow in the price
formation process (e.g., Brandt and Kavajecz (2004) and Evans and Lyons (2004)), while
others have highlighted the role of public information in the price formation process (e.g.,
Fleming and Remolona (1997), Balduzzi, Elton and Green (2001), Green (2004), and
Andersen, Bollerslev, Diebold and Vega (2003, 2004)). Typically, however, the role of
order flow and public information is examined in isolation. Some notable exceptions are
Green (2004), Pasquariello and Vega (2005), and Evans and Lyons (2004) who examine
both the role of public information and order flow in the bond and foreign exchange
markets, respectively. At least two features differentiate our study from theirs. First,
asymmetric information may be more severe in the stock market context than in the bond
and foreign exchange market; since in the later private information is about macroeconomic
factors, while in the former both macroeconomic and firm specific factors matter. Second,
we simultaneously investigate cross-country stock market linkages rather than estimating
the effect of order flow and public news on one market in isolation.
3. Model
There is a local, small economy and a foreign, large economy.10 Each economy
has its own stock market where a single stock is traded (the market index) and a distinct
pool of investors. Investors can also invest in the bond market at a zero interest rate. The
local (foreign) stock market is composed of n ( n∗ ) informed investors, uninformed
investors, and a perfectly competitive market maker. All optimizing agents are risk neutral.
Time is described by t = 1, 2,..., T periods, where T is the time at which the stock pays a
10
We have in mind that the U.S. is the foreign economy and a country such as Portugal the small economy.
liquidating dividend. Foreign economy prices and quantities are identified with an asterisk
‘ ∗ ’.
Stock trading in each period is à la Kyle (1985): investors submit trades given
their information sets; the market maker sets prices given the publicly available information
and the aggregate order flow; and the market clears. Markets are assumed open 24 hours.
At the end of each period t the innovation to the underlying value of the asset, vt +1 , is
realized and becomes public information. Figure 1 shows the timing of the model.
At the start of period t the underlying liquidation value of the local asset is
Vt = V + ∑τt =1vτ known by all investors (there is a similar process for the foreign economy
Vt ∗ ). The asset dividend VT is paid out at the beginning of period T . The variance of the
incremental valuation vt is δ . It is assumed that the local asset’s fundamental value is
affected by a global factor that drives the returns in the foreign, large economy, i.e.
E ⎡⎣vt vt∗ ⎤⎦ = ψ .11
Before trading, informed investors observe a private information signal

st = vt +1 + εt about next period’s incremental valuation, vt +1 , where the variance of εt is
φ . In some periods, and also before trading takes place, market participants receive local or
foreign public news about local or foreign asset values, respectively.12 News are modeled
by the random variable U t = vt +1 + μ t , with the variance of μ t equal to κ . This setup
guarantees that any private or public information received at t is short lived and cannot be
used beyond forecasting time t + 1 valuations. We assume that all variables are normally
distributed and have zero mean and with the exceptions noted above all variables are
assumed to be independently distributed.13
Because informed investors submit their orders without knowing the stock price,
they choose trading based on the private and public information in their information set I It
and their conjecture of the price process {Pτ }τ ≥t to solve
11
This assumption is justified by previous literature and by noting that Portuguese real GDP growth is
significantly positively correlated with U.S. real GDP growth.
12
Whether news come at known dates or not does not affect the equilibrium we study.
13
The model developed here does not allow for time-variation in stock return correlations. This is not critical as
we also do not model any time variation in the volatility of fundamentals (see Forbes and Rigobon (2002)).
⎡ T −1 ⎤
xit
⎣ t =1
( )
max E ⎢ ∑ Pt +1 − Pt xit | I tI ⎥ ,
⎦
(1)
with the convention that PT = VT . Uninformed investors trade the random quantity zt in
period t , where zt has variance ζ . Total order flow in the local stock market is thus
ω t = ∑ in=1xit + zt . For simplicity local and foreign informed investors only trade on their
respective markets. This is not restrictive, but simplifies the analysis considerably. It is
straightforward, but inconsequential for our qualitative results, to allow the private
information of informed investors in different markets to be correlated (even perfectly so)
which would be the case if, for example, foreign investors also traded on the small
economy’s market and vice-versa.
The market maker chooses prices to maximize
⎡T ⎤
E ⎢ ∑ ( Pt +1 − Pt ) ωt | I tM ⎥ . (2)
⎣ t =1 ⎦
Being perfectly competitive maximum profits are zero. By market efficiency, the
market makers’ information set in period t , I tM , includes all available information: the
aggregate order flow ωt and any available public news.
Discussion of Model Assumptions

We describe the behavior of the different investors and the equilibrium prices in
periods with and without public news announcements. We make two main assumptions.
First, investors in the foreign, large economy are assumed to ignore the public news
announced in the local, small economy, but the opposite is not true. The general framework
we have in mind is one where foreign investors are aware of news from many smaller
markets. While these news contain information regarding valuations in their own market,
the costs of having to process all the little pieces of dispersed information outweigh the
benefits (in the spirit of Grossman and Stiglitz (1980)). This assumption leads to what we
believe is a realistic asymmetric treatment of news: while news in the large, foreign
economy lead to price discovery in both markets, news in the local, small economy only
lead to price discovery in the local stock market. It is this asymmetry that is critical to our
results. See section 2 above for evidence of this asymmetric effect when the foreign
economy is the U.S. Also, in our data, U.S. returns do not respond to Portuguese news.
Second, in the same spirit as King and Wadhwani (1990), informed investors and
the market maker in the local economy see the period t stock price in the foreign economy
Pt ∗ before they make their decisions. This second assumption is particularly realistic in our
setting with a large, foreign economy and a small, local economy where local investors
respond to news announced in the foreign economy. Casual observation suggests that
investors in small markets like Portugal wait to see how investors in larger markets,
prominently the U.S., respond to public news in their own countries before they trade on
that information, perhaps trusting foreign investors’ better interpretation of the news. Even
in the absence of news, large upward or downward price movements in the U.S. are
analyzed by local investors trying to learn about the nature of the price change. Confirming
our assumption, section 6.1 below shows evidence of delayed market response in Portugal
to U.S. news announcements. Finally, we believe that the mechanisms we describe here
would still be present in a richer model that does not rely on our timing convention so long
as past price movements contain information that can be useful to forecast future local
returns.
Implicit in our assumption of distinct investors populating each market is the
assumption of no trading across markets. In contrast to King and Wadhwani (1990) this
assumption is unnecessary to our results but again makes the analysis easier. Assuming the
private information signals of informed investors in different markets to be correlated
implicitly allows some of the informed investors to be the same.
Finally, while the model shares many similar features to King and Wadhwani
(1990) (such as the one discussed in the previous paragraph) we differ in the equilibrium
concept. There a rational expectations equilibrium concept is used. Here investors act
strategically on their private information and hence order flow is a noisy measure of that
private information. This feature allows us to determine theoretically and empirically the
effects of public news controlling for private information. It also allows us to study
liquidity effects around announcements as additional predictions from our model in
contrast to their model of contagion which is silent along this dimension of trading.
4. Equilibrium
This section describes the equilibrium prices and trades that result in days with
and without news. Before we start we note that because private information is short lived
the problems (1) and (2) are equivalent to solving the sequence of single period problems
xit ⎣ ( )
max E ⎡ Pt +1 − Pt xit | I tI ⎤ ,
⎦
(3)
and
0 = E ⎡⎣( Pt +1 − Pt ) ωt | I tM ⎤⎦ , (4)
respectively. In the foreign market, informed investors and the market maker solve
identical problems.
4.1. Foreign Stock Market, No Foreign News
Consider first a period t without public news announcements in the foreign

{ }
economy. Informed foreign investors’ information set is ItI ∗ = Vt ∗ , st∗ . In a period of no
news informed investors conjecture that the equilibrium price is
Pt ∗ = Vt ∗ + λ ∗ω t∗ .
They also need to conjecture a price for period t + 1 which depends upon the
existence, or not, of public news in the foreign economy: if there are news announcements
at t + 1 the price displays a different liquidity parameter λ ∗ and depends on the public
news released (see below). In equilibrium, because private information is short lived and
uninformative of future payoffs E ⎡⎣ Pt ∗+1 | ItI ∗ ⎤⎦ = Vt ∗ + E ⎡⎣vt∗+1 | ItI ∗ ⎤⎦ , independently of the
type of news at t  1 . Hence, the solution to the foreign informed investors’ problem is
obtained from
⎡ ⎤
∗
n
max E ⎢vt∗+1 − λ ∗ ∑x ∗jt | I tI ∗ ⎥ xit∗ .
⎣⎢ ⎦⎥
∗
xit
j =1
{ }
Similarly, knowing I tM ∗ = Vt ∗ , ωt∗ , the foreign market maker solves
0 = E ⎡⎣vt∗+1 | ωt∗ ⎤⎦ − λ ∗ωt∗ ,
where λ ∗ measures the information content of order flow. It is well known (see Kyle
(1985)) that the solution to this static problem entails
δ∗ n∗
λ∗ = , (5)
(
n∗ + 1 ζ ∗ δ ∗ + φ ∗ )
and
ζ∗
xit∗ = st∗ . (6)
n∗ ( δ ∗ + φ ∗ )
From (6), informed investors are more aggressive in exploring their private
information when they can better use their information monopoly: the signal is very
informative ( φ ∗ is low); there are fewer of them ( n∗ is low); there are many uninformed
investors ( ζ ∗ is large). In contrast, the market maker, fearing more informed trading,
responds by reducing liquidity (i.e., λ ∗ increases) when the signal is very informative ( φ ∗
is low) or there are fewer informed investors ( n∗ is low) (see 5).
4.2. Foreign Stock Market, Foreign News
Let t be a period with a public news announcement in the foreign economy,

U t∗ = vt∗+1 + μt∗ , with E ⎡⎣ μ t∗2 ⎤⎦ = κ ∗ . Foreign informed investors correctly conjecture that the
equilibrium stock price is,14
( )
Pt ∗ = Vt ∗ + λ ∗ ω t∗ − β ∗U t∗ + σ ∗U t∗ .
Note that it is the unexpected part of the order flow, ω t∗ − β ∗U t∗ , that is relevant to
the market maker to learn about informed investors’ private information (with β ∗
determined in equilibrium). With this price conjecture, the information set
I∗
{ ∗
I t = Vt , s ,U ∗
t t
∗
} , and the equilibrium property that
E ⎡⎣ P | I t ⎤⎦ = Vt + E ⎡⎣v | I ⎤⎦ , ∗
t +1
I∗ ∗ ∗
t +1 t
I∗
we solve the problem faced by informed investors (see the appendix). The market maker
{
observes the aggregate order flow, i.e., I tM ∗ = Vt ∗ , ωt∗ ,U t∗ . The equilibrium price }
parameters are:
The parameter λ here is not equal to that when there are no news. Indeed all equilibrium price parameters
14 ∗
used in the model should carry a time subscript to reflect whether news are announced or not in a particular
period. We omit the time subscript to avoid cluttering the notation.
λ =
∗
(
var vt∗+1 | U t∗ ) n∗
,
n +1 ∗
(
ζ ∗ var st∗ | U t∗ ) (7)
δ ∗
σˆ ∗ = ,
δ +κ∗
∗
(
with σ ∗ = σˆ ∗ + λ ∗ β ∗ ,15 var vt∗+1 | U t∗ = ) δ ∗κ ∗
δ ∗ +κ ∗ (
and var st∗ | U t∗ = ) δ ∗κ ∗ +φ ∗δ ∗ +φ ∗κ ∗
δ ∗ +κ ∗
and the asset
demand is
ζ∗
xit∗ = (s ∗
)
− σˆ ∗U t∗ . (8)
n var st∗ | U t∗
∗
( ) t
The coefficient σ̂ ∗ associated with the public news is given by the ratio of the
covariance of the public news with the private information st to the variance of the public
news and describes the part of private information that can be inferred from public news.
Adding public information has two conflicting implications for trading by
informed investors (see (8) and (6)). On the one hand, the ‘amount’ of private information
is reduced to the unforecastable part of informed investors’ private information,
( )
st∗ − E st∗ | U t∗ = st∗ − σˆ ∗U t∗ , making informed investors trade less. On the other hand, the
public news reduces the conditional variance of the private information to
( )
var st∗ | U t∗ = δ ∗κ ∗ +φ ∗δ ∗ +φ ∗κ ∗
δ ∗ +κ ∗
( )
from var st∗ = δ ∗ + φ ∗ . This makes informed investors more
aggressive in acting upon their private information. Both of the forces affecting the order
flow of informed investors also affect the liquidity parameter λ ∗ in the pricing function in
opposing ways. However, it can be show that the former effect is stronger and that:
Proposition 1 In the foreign market liquidity always increases in news days relative to no
news days.
15
We also get that in the price function,
n∗ζ ∗
β ∗ = −δ ∗ .
(δ κ ∗ ∗
)(
+ φ ∗δ ∗ + φ ∗κ ∗ δ ∗ + κ ∗ )
4.3. Local Market, No Local or Foreign News
Investors in the local stock market observe the time t price of the foreign stock
∗
asset Pt and use this knowledge to infer the private information of informed investors in
that market:
ωt∗ = λ ∗−1 ( Pt∗ − Vt ∗ ) .
Because stock fundamentals are correlated ωt∗ is also informative about the local
stock market. With this additional information, the local economy’s informed investors
conjecture a pricing function for the current period t of
Pt = Vt + λω t + ηωt∗
( ( ))
(9)
= Vt + λω t + η λ ∗−1 Pt ∗ − Vt ∗ .
This pricing function implies that returns in both markets are correlated, with the
conditional correlation given by ηλ ∗−1 . Cross-country correlation of returns arises because
fundamental asset valuations are correlated and foreign returns carry the private
information of foreign informed investors about their own asset’s valuation. Intuitively,
after large price changes in the foreign market, local investors try to infer whether such
move was motivated by informed or uninformed trading and use this information to
improve their forecasts of local valuations.
Local informed investors solve (3) knowing E ⎣⎡ Pt +1 | I tI ⎦⎤ = Vt + E ⎡⎣ vt∗+1 | I tI ⎤⎦ , and

{ }
I tI = Vt , st , ωt∗ . The solution to this problem, described in the appendix, enables us to
compute the aggregate order flow. Given the local and foreign aggregate order flow and
{ }
I tM = Vt , ωt , ωt∗ , the local market maker sets prices to meet the zero profit condition (4).
The time t equilibrium is described by:
λ=
(
var vt +1 | ωt∗ ) n
,
n +1 ζ var ( st | ωt∗ )
(10)
ψ n∗
η= ,
n∗ + 1 ζ ∗ δ ∗ + φ ∗( )
where ( )
var vt +1 | ωt∗ = δ − n∗
( n +1)(δ
∗ ∗
+φ ∗ )
ψ2 and ( )
var st | ωt∗ = δ + φ − n∗
( n +1)(δ
∗ ∗
+φ ∗ )
ψ 2 . The
liquidity parameter thus has a similar form and interpretation to those in (5) and (7). The
parameter η describes how much of the foreign order flow can be used to forecast the local
( )
asset, i.e. E st | ωt∗ = ηωt∗ . The equilibrium asset demand is:
ζ
xit = ( s − ηω ) .
∗
(11)
(
n var st | ωt∗ ) t t
As the order flow from the foreign stock market is public information to all
investors (via knowledge of the foreign return), its effects on the asset demand of informed
investors are identical to those discussed above regarding the informed asset demand in the
foreign country in the presence of public news (8).
We can now state the time t price function:
Proposition 2 The time t equilibrium price when there are no public news in either market
is
ψ
(
Pt − Vt = λω t + ∗ Pt ∗ − Vt ∗ .
δ
)
The slope coefficient of regressing local returns on foreign returns on no news
days is the ratio of the covariance of valuations ψ to the variance of the foreign asset value
δ ∗ . On average the correlation of stock returns reflects the correlation on fundamentals.
4.4. Local Market with Local News, No Foreign News
Suppose now that period t has a public news announcement in the small, local
economy and recall that local news are assumed not to affect investor behavior in the
foreign economy. News in the local economy are U t = vt +1 + μ t , with E ⎡⎣ μ t2 ⎤⎦ = κ . The
market maker is conjectured to choose the following price function (where again only the
unexpected local order flow contains useful information to the market maker):
Pt = Vt + λω t + ηω t∗ + σ U t .
The appendix gives the solution to the informed investors’ problem given the
{ }
information set I tI = Vt , st ,U t , ωt∗ . With the solution to this problem and the uninformed
trades we construct the local aggregate order flow and solve the market maker’s problem of
finding the Pt which ensures zero profits, given I tM = Vt ,U t , ωt , ωt∗ . The equilibrium { }
trading at t is characterized by:
λ=
(
var vt +1 | ω t∗ , U t ) n
,
n +1 (
ζ var st | ω t∗ ,U t )
η=
κ ( )
var vt∗+1 n∗ ψ
,
(
var U t | ω ∗
t ) ∗
n +1 ζ var ( s
∗
t
∗
) δ∗
σ=
(
var vt +1 | ω ∗
t ).
(
var U t | ω t∗ )
The equilibrium asset trades of informed investors are given by
ζ
xit = ⎡⎣ st − ηωt∗ − σ U t ⎤⎦ ,
( ∗
n var s | ω ,U
)
⎡ ⎤ ⎡ ⎤
( )
where var s | ω ∗ , U = ⎢(δκ + φδ + φκ ) − (κ + φ ) ∗ n ∗ ∗ ψ 2 ⎥ / ⎢(δ + κ ) − ∗ n ∗ ∗ ψ 2 ⎥ .
⎣ (
∗
n +1)(δ + φ )
⎦ ⎣ (
∗
n +1)(δ +φ )
⎦
Again η and σ are the components of the linear projection E ⎡⎣ st | ω t∗ , U t ⎤⎦ associated with
ωt∗ and U t , respectively: informed investors trade on the unexpected component of their
private information st − E ⎡⎣ st | ωt∗ , U t ⎤⎦ = st − ηωt∗ − σ U t . Note that adding ωt does not
affect the values of η and σ because the private information that goes into the local net
order flow already ‘subtracts’ the information contained in ω ∗ , and U . The price t t
coefficient  contains three parts. One is an adjustment for how useful the foreign net order
( )
flow is given the local public news ( κ / var U t | ω t∗ ). The second accounts for liquidity in
the foreign market (recall that λ = ∗ ( ) ∗
var v t +1
n∗
) and, finally, the last part measures ex-
n∗ +1 ( )
ζ ∗ var st∗
ante covariation between incremental dividends, ψ / δ ∗ . We are now ready to give our
second result:
Proposition 3 The local equilibrium price when there are public news in the local market
is
(
Pt − Vt = λω t + ηλ ∗−1 Pt ∗ − Vt ∗ + σ U t , )
with
κ ψ
ηλ ∗−1 = . (12)
( )
var U t | ω t∗ δ
∗
ψ
The slope ηλ ∗−1 increases with the noise in the public news, κ , and ηλ ∗−1 → δ∗
if κ → ∞
Note that if the local public news are uninformative and κ → ∞ , we recover a
ψ
correlation of stock markets of ηλ ∗−1 = δ∗
. Also, because the slope in the price equation
ηλ ∗−1
increases with the uninformativeness of public information κ , it must be that for
ψ
any informative piece of news κ < ∞ , ηλ ∗−1 < δ∗
. Thus,
Corollary 4 Days with news in the small, local market display lower correlation of returns
between countries than days without news.
In days of news in the local market, local prices respond less to foreign prices
because some of the information contained in the foreign order flow is now provided via
the local public news. This is true unless the public news are uninformative and κ = ∞ .
The next result indicates the effect of local news on liquidity:
Proposition 5 In the local market liquidity always increases in days of local news relative
to no news days.
Recall that public information has two opposing effects. First, it increases the
precision of the private information of informed investors, which leads to an increase in
adverse selection costs to the market maker and a higher λ . This effect is outweighed by
the fact that public news reduces the amount of private information that informed investors
trade on (see 8).
4.5. Local Market with Foreign News, No Local News
When there are news in the foreign economy, but not in the local economy, the
foreign order flow incorporates these news. Hence, the local market maker cares not about
the total order flow, but only that component of the order flow that is unrelated to news,
wt∗ = ω t∗ − β U t∗ (and similarly for the local order flow). Informed investors conjecture that
the time t equilibrium price function is:
Pt = Vt + λω t + η w t∗ + σ U t∗ .
{ }
The relevant information sets are I tI = Vt , st ,U t∗ , ωt∗ for informed investors and
I t
M
{ ∗
t
∗
= Vt , U , ωt , ω
t } for the market maker. The equilibrium parameters λ , η and σ are
the coefficients on the linear projection of vt +1 on ωt , w ,U , respectively and are given in
∗
t
∗
t
the appendix. The next proposition characterizes the equilibrium pricing rule:
Proposition 6 The local equilibrium price when there are public news in the foreign
market is
ψ
(
Pt − Vt = λω t + ∗ Pt ∗ − Vt ∗ .
δ
)
The proposition shows that in days of foreign news the return correlation does not
change from that in days without news, and is higher than that for days with news in the
local economy. The intuition for this result is that as news are released in the foreign
market, foreign informed investors chose to trade on the orthogonal component of their
private information (see (8)). Therefore, the public news U t∗ are uninformative about the
‘net-private information’ used by foreign investors, st∗ − σˆ ∗U t∗ , which is captured in the
foreign return: the information content of U t∗ does not affect the information content of the
foreign return and the return correlation is unchanged in the presence of foreign news.
Furthermore, because of the linearity built in the model the foreign return is a sufficient
statistic for both foreign private and public news. This means that a regression of local
returns on local order flow, foreign returns and foreign news should produce a zero
coefficient on foreign news.
The proposition can explain the evidence in Karolyi and Stulz (1996) that
correlations between stock returns in U.S. and Japan do not vary with news announcements
in the U.S., if, as it is reasonable, one takes the Japanese market as a follower to the U.S.
market in the presence of U.S. news. While such lack of connection between fundamentals
and international return correlations has often been viewed as evidence of contagion
through the form of investor sentiment or noise trading, in our model it occurs because
investors in the ‘follower’ market know that they cannot use the U.S. public news to help
them filter the private information of U.S. investors contained in the U.S. return.
Another important result from Proposition 6 is that after controlling for the foreign
price, Pt ∗ − Vt ∗ , foreign news are no longer relevant to forecast local prices. Foreign prices
already include the impact of foreign news and local investors’ inference accounts for that
as well. Obviously, if we omit the foreign return as an explanatory variable of local prices,
foreign news will have explanatory power, but as the model suggests foreign returns are an
all encompassing variable for foreign private and public information. Hence, Proposition
can be used to explain the results of King, Sentana and Wadhwani (1995), Connolly and
Wang (2003) and ours−see below−where foreign news have no explanatory power of local
returns over and above that implied by foreign returns.
4.6. Country-Specific News

So far we have treated news in the small, local economy as also being informative
of the foreign valuations. Consider now that valuations in the small economy are given by
v t1  v 1t1  v 2t1 ,
with E ⎡⎣v1t +1vt∗+1 ⎤⎦ = ψ and E ⎡⎣v2t +1vt∗+1 ⎤⎦ = 0 . We maintain the normalization that
E ⎡⎣vt2+1 ⎤⎦ = δ so that E ⎡⎣v12t +1 ⎤⎦ = δ1 , E ⎡⎣ v22t +1 ⎤⎦ = δ 2 and δ = δ1 + δ 2 . We wish to characterize
the equilibrium in the presence of local country-specific news:
U t = v2t +1 + μ t ,
where, abusing notation, E ⎡⎣ μ t2 ⎤⎦ = κ , and E ⎡⎣U t vt∗+1 ⎤⎦ = 0 . The signal that informed
investors in the local economy get is on the common component, st = v1t +1 + εt . The main
result below is not affected by this assumption.
The market maker is assumed to choose the following price function
Pt = Vt + λω t + ηω t∗ + σ U t .
The appendix shows the equilibrium price parameters and asset trades of informed
investors.
Proposition 7 The local equilibrium price when there are local country-specific public
news is
ψ δ2
δ
( )
Pt − Vt = λω t + ∗ Pt ∗ − Vt ∗ +
δ2 + κ
Ut .
Therefore, if all news in the local, small economy are country-specific, then there
should not be any change in correlation. The intuition is quite straightforward. The foreign
return contains information on the common valuation component v1t +1 , whereas the
country-specific news contain information about v2t +1 . As these components are
orthogonal, so are the news, and the information content of the foreign return is not
affected. In our empirical implementation we test this differential treatment of news by
looking at macroeconomic news in Portugal and contrasting its effects with firm-specific
news as given by earnings announcements.
Finally we have:
Proposition 8 In the local market liquidity always increases in days of local country-
specific news relative to no news days.
5. Data Description
We test the implications of the model presented in the previous section using
Portuguese and U.S. stock market data and Portuguese and U.S. earnings and
macroeconomic announcements. As mentioned in the Introduction, this choice is motivated
not only by the quality and availability of high frequency Portuguese stock market data, but
because Portugal is a small emerging market economy vulnerable to U.S. news
announcements, while there is relative immunity of the U.S. economy to Portuguese
shocks. The data are novel in several respects, such as the simultaneous high frequency
data in the U.S. and Portuguese stock markets and the availability of signed trades in the
Portuguese stock market. Here we describe them in detail.
5.1. Portuguese and U.S. Stock Market Data

We analyze the individual components of the main Portuguese stock market index,
the PSI-20 Index, and the individual components of the DJ 30 Industrial Index listed in
Table 1A and Table 1B respectively. Our sample period is determined by the availability of
high frequency stock market data from Euronext Lisbon.16 This database contains all time-
stamped transactions, signed trades and bid/ask quotes from January 4, 2002 to October 15,
2002. In total there are 195 trading days and 2,441,490 orders. This allow us to observe the
number of buyer-initiated and seller-initiated trades, the number of orders placed right after
public announcements, whether or not orders are cancelled, changed, executed, or if they
have simply expired. As it is shown in Table 1A, the majority of the orders (between 84 to
96 percent) are limit orders and the liquidity of the market, measured by the number of
transactions, is highly correlated with market capitalization. When we analyze daily
comovement between the U.S. and Portuguese stocks we only use Portuguese limit orders
that were placed and executed on the same day (see, Antão, Antunes and Martins 2004).
For the U.S. we use Trades and Quotes (TAQ) data, which contains bid quotes,
ask quotes and transaction prices from stocks traded in different U.S. stock markets. To
calculate the daily number of buys and sells we use the Lee and Ready (1991) algorithm for
NYSE listed stocks and the Ellis, Michaely and O’Hara (2000) suggested variation of the
Lee and Ready algorithm for Nasdaq listed stocks.17 We only use trades and quotes from
the exchange they are most frequently traded in, which in our case coincides with the
exchange they are listed in.
16
On February 6, 2002 the Bolsa de Valores de Lisboa e Porto (BVLP) changed its name to Euronext Lisbon.
17
Odders-White (2000), Lee and Radhakrishna (2000) and Ellis, Michaely, and O'Hara (2000) evaluate how well
the Lee and Ready algorithm performs and they find that the algorithm is 81% to 93% accurate, depending on
the sample and stocks studied.
One of the main problems of using daily returns to measure asset market
comovement is the existence of non-synchronous trading periods around the globe (e.g.,
Karolyi and Stulz 1996). In general, studies use closing market prices to estimate daily
returns. Since stock markets around the world close at different times, these studies are not
measuring contemporaneous stock market correlations. In this paper, we avoid this problem
by using high frequency data and estimating daily returns using 11:30 EST prices, which
correspond to Portuguese stock market closing prices as shown in Table 2. An advantage of
using 11:30 EST prices is that the U.S. stock market has been open for two hours and most
U.S. macroeconomic news announcements are released at 8:30 and 10:00 EST. Hence,
using prices close to these announcement release times allows us to measure more
accurately the news effect on U.S. and Portuguese prices (Andersen et al. 2003). While we
note that our qualitative results do not depend on using 11:30 EST prices, the results do
depend on measuring synchronous returns between the two countries.
In Table 3A and 3B we show descriptive statistics for the Portuguese and U.S.
stock market returns for trading days common to both countries from January 4, 2002 to
October 15, 2002. In total there are 189 trading days. The tables show that the average
market capitalization of the components of the DJ 30 Index (100 billion dollars) is
approximately 50 times larger than the average market capitalization of the individual
components of the Portuguese PSI-20 (2 billion dollars). Liquidity is twenty six times
higher, measured by the bid-ask spread divided by the average daily price in the sample, in
the U.S. market compared to the Portuguese market, and return volatility is 35% larger in
the Portuguese market than in the U.S. market.
5.2. Macroeconomic and Earnings Data
To estimate news surprises for the U.S. and Portugal, we require data from four
different sources: Bloomberg, International Money Market Services (MMS), Reuters and
IBES. We use Bloomberg to collect real-time data on the realizations of 9 of the most
relevant Portuguese macroeconomic announcements, including the European Central Bank
(ECB) benchmark refinancing interest rate. We use one-month Euribor yield data from
Bloomberg to estimate the ECB interest rate market expectation, following the method
described in Kuttner (2001), Cochrane and Piazzesi (2002), Rigobon and Sack (2002),
among others, who use the current month federal funds futures contract, one-month
Eurodollar deposit rate and three-month Eurodollar futures rate, respectively, to estimate
the market expectation of the federal funds target rate. The Portuguese Central Bank does
not control the ECB benchmark interest rate, yet it is the Portuguese monetary policy
instrument. Hence, we do not consider it a Portuguese macroeconomic announcement (as
we consider the federal funds rate a U.S. macroeconomic announcement), but we include
this variable in all of our estimates. Foreshadowing subsequent results, the U.S. stock
market does not react to any of the Portuguese announcements, except for the ECB
announcement.
Since Bloomberg only reports professional forecasts for the Portuguese CPI, we
construct our own forecast for the remaining series listed in Table 4A. In untabulated
results, we show that a seasonal random walk describes the data well.18 In other words, the
optimal forecast of this month’s or quarter’s announcement is last year’s announcement
about the corresponding month or quarter, and this is the proxy we use for the market
expectation.
We use MMS real-time data on the realizations of 25 of the most relevant U.S.
macroeconomic announcements. Consistent with our estimate of the ECB interest rate
market expectation, we use one-month Eurodollar deposit rate data from Bloomberg to
estimate the market expectation of the federal funds target rate.19 For the remaining 24
U.S. macroeconomic announcements listed in Table 4B, we use the MMS median forecast,
which is better than our own univariate forecasts.20 Table 4A and Table 4B provide a brief
description of the most salient characteristics of these U.S. and Portuguese macroeconomic
news announcements: the total number of observations in our sample, the agency reporting
18
We estimate the optimal forecast of these series using the Bloomberg sample period from July 7, 1997 to
September 8, 2005. The in-sample period starts in July 7, 1997 and ends in December 31, 2004 and the out-of-
sample period is from January 1, 2005 to September 8, 2005. Since we need historical data to calculate these
optimal forecasts we are unable to use the following Portuguese macroeconomic announcements which started
to be reported in Bloomberg in 2002 or in later years: Government Budget Deficit, Total Construction
Licenses, Consumer Confidence, Labor Costs, Manufacturing Production and New Car Sales.
19
We could have also used the current month federal funds futures contract, since Gürkaynak, Sack and Swanson
(2002) find that this instrument outperforms the eurodollar deposit rate's one-month ahead forecasting power.
We note, that our conclusions do not depend on whether we use the eurodollar deposit rate or the current
month federal funds futures contract.
20
For a more detailed description of the MMS data we refer the reader to Andersen, Bollerslev, Diebold, and
Vega (2003).
each announcement, and the time of the announcement release. The only announcement
with uncertain release time is the ECB’s announcement. The Governing Council of the
ECB announces the interest rate level at 7:45 EST every month of the year, and the
president of the council holds a press conference that ends anytime between 8:50 EST and
10:00 EST. Since ECB interest rate announcement surprises are very few, the market
focuses on the president’s statement and the future trajectory of interest rates rather than on
the actual announcement. The evidence in Section 6.1 further supports this view.
Finally, we use the individual stock’s earnings announcements to analyze the

effect Portuguese country specific news have on the US-Portuguese stock market
comovement (Section 4.6) and to control for non-global U.S. public announcements. We
use IBES and Reuters data to measure the U.S. and Portuguese earnings realizations and
expectations, respectively. Since Reuters does not collect data for all Portuguese stocks, we
use the previous year’s earnings when the forecast is not available.21
We define announcement surprises as the difference between announcement

realizations and their corresponding expectations. More specifically, since units of
measurement vary across variables, we standardize the resulting surprises by dividing each
of them by their sample standard deviation. The standardized news associated with
announcement indicator j at time t is therefore computed as
Ajt − E jt
S jt = , (13)
σl j
where Ajt is the announced value of indicator j , E jt is the MMS median forecast for U.S.
macroeconomic data, the IBES median forecast for U.S. earnings data, the Eurodollar
implied forecast for the federal funds rate, Bloomberg median forecast for Portuguese CPI,
last year’s announcement for the remaining Portuguese macroeconomic announcements,
Reuters data for Portuguese earnings, and the Euribor implied forecast for the ECB
21
Reuters collects data from investment bank professional forecasters, e.g. Caixa Valores, Deutsche Bank, Banco
Espíritu Santo, Banco Finantia, Banco Santander Central Hispano, BPI, Lehman Brothers, etc. Reuters does
not collect data for the following Portuguese stocks: COFINA, IBERSOL, IMPRESA, NB, PARAREDE,
PORTUCEL, PTM, SAG, SEMAPA, SONAE and TD.
benchmark interest rate.22 The denominator, σl j , is the sample standard deviation of

Ajt − E jt estimated using the full sample period of expectations and announcements.
Equation (13) facilitates the aggregation of news described below and it facilitates
meaningful comparisons of responses of different stock price changes to different pieces of
news. Operationally, we estimate the responses by regressing stock price changes on news.
Because σl j is constant for any indicator j , the standardization affects neither the
statistical significance of the response estimates nor the fit of the regressions. In Table 4A
and 4B we show the sample mean and standard deviation of each news announcement
surprise for the United States and Portugal between January 4, 2002 and October 15, 2002,
respectively. Note that the standardized variables show a standard deviation that is not
equal to one, because the standard deviation we use was obtained in the full sample, with
the purpose of more accurately measuring the standard deviation of the surprises in a larger
sample.
To keep the analysis parsimonious, we aggregate the macroeconomic
announcements into seven groups as shown in Table 4A and Table 4B: real activity, each
of the GDP components (i.e., consumption, investment, government purchases and net
exports), prices, and forward-looking announcements. For example, U.S. real activity
surprises are defined as the sum of GDP Advance, GDP Preliminary, GDP Final, Nonfarm
Payroll, Retail Sales, Industrial Production, Capacity Utilization, Personal Income and
Consumer Credit standardized surprises (according to equation (13)); while Portuguese real
activity surprises are defined as the sum of GDP, the Employment Report, Industrial
Production and Industrial Sales standardized surprises. The benchmark interest rate
announcements and the weekly announcements are not aggregated because interest rate
announcements do not fall into any of the above described categories and weekly
announcements are more volatile than quarterly and monthly announcements. The
aggregation does not affect the conclusions of the paper, as we show in Section 6.7, but it is
solely done for expositional purposes. Summary statistics for these aggregated
macroeconomic surprises are shown in Table 5 and will be used to calculate the economic
22
As mentionde before, since Reuters does not collect data for all Portuguese stocks, we use the previous year's
earnings when the forecast is not available.
significance of these announcements.
6. Empirical Analysis
The model of Section 3 generates several implications that we now test. In the
database described in Section , we are able to directly observe local price changes, Pt − Pt −1 ,
foreign price changes, Pt ∗ − Pt ∗−1 , Portuguese macroeconomic news, S p por t , Portuguese firm-
specific earnings, S pepor t , U.S. macroeconomic news, S pus t , U.S. firm-specific earnings,
S peus t , ECB public news surprises, S pecb t , local aggregate order flow, ωt , and foreign
aggregate order flow, ωt∗ . With our availability of intraday data we avoid non-synchronous
trading biases by defining price changes from Portuguese market closing time 11:30 EST to
closing time the next day 11:30 EST (see Table 2). In our setting, it is only the unexpected
portion of aggregate order flow that affects the equilibrium prices. Furthermore, ωt and ωt∗
are assumed to depend only on informed and liquidity trading. Yet, in reality, many
additional microstructure imperfections can cause lagged effects in the observed order flow
(see Hasbrouck, 2004). Therefore, to implement our model, we estimate Ωt and Ω∗t , the
unanticipated portion of aggregate local and foreign order flow. For that purpose, we use
the linear autoregressive model of Hasbrouck (1991),
xi = ax + b( L)ri + c( L) xi + v( x)i ,
(14)
xi∗ = ax∗ + b( L)ri∗ + c( L) xi∗ + v∗ ( x)i ,
where xi ( xi∗ ) is the transaction by transaction net order flow for local (foreign) asset i
(purchases take a +1 and sales take a −1 ), ri ( ri∗ ) is the quote revision changes for local
(foreign) asset i , and b( L) and c( L) are polynomials in the lag operator. We choose 25
lags for all assets because this lag structure is sufficient to eliminate all the serial
correlation in the data. However, the results that follow do not rely on this particular lag
structure.
As previously mentioned, we focus on daily horizons, for broader intervals lead to

less powerful tests of market comovement (Karolyi and Stulz 1996) and the influence of
macroeconomic fundamentals (Andersen et al. 2003, and 2005). Therefore, we compute
aggregate unanticipated (or abnormal) net order flow over each day t , Ωt = ∑ in=t 1v( x)it
∗
( Ω∗t = ∑ in=t 1v∗ ( x)it ), where nt ( nt∗ ) is the number of transactions taking place on day t in
the local (foreign) market. Consistent with the daily return definition we aggregate order
flow from 11:30 EST the previous day to 11:30 EST today.
6.1. Model’s Timing Assumption
One of the model’s main assumption, in the same spirit as King and Wadhwani
(1990), is that Portuguese informed investors and the market maker observe the period t
U.S. stock price Pt ∗ before they make their decisions about foreign public news (U.S. and
ECB announcements).23 To investigate the plausibility of this assumption, we calculate
cumulative 5-minute responses of U.S. and Portuguese stock market returns to foreign
public news in three different scenarios: ( i ) the Portuguese stock market is open, but the
U.S. stock market is not (Figure 2A and 2B), ( ii ) the Portuguese and U.S. stock markets
are both open (Figure 3A and 3B), and ( iii ) the U.S. stock market is open, but the
Portuguese stock market is not (Figure 4A and 4B). In particular, we estimate the following
contemporaneous stock market response:
15
ritpor = a por + ∑λ
pus = 7
por
spus S pus t + λspporecb S pecb t + ε it ,
15
(15)
r =a+
us
it ∑λ
pus = 7
spus S pus t + λspecb S pecb t + ε it ,
where ritpor = (ln( Pitpor ) − ln( Pitpor

−1 )) × 100 is the individual stock return for Portuguese firm
i = 1,..., 20 , ritus = (ln( Pitus ) − ln( Pitus−1 )) ×100 is the individual stock return for U.S. firm
i = 1,...,30 , S pus t is the standardized news corresponding to announcement pus
( pus = 7,...,15 ) listed in Table 5, and S pecb t is the standardized ECB benchmark interest rate
news surprise made at time t . We estimate equation (15) using only those observations
( ritpor , ritus , S pt ) such that an announcement was made at time t , where p = pecb ,7,...,15 .
Focusing first on the response to public announcements when the Portuguese stock
market is open, but the U.S. stock market is not, we estimate the reaction to the 8:30 and
9:15 EST U.S. macroeconomic announcements and the 7:45 EST ECB announcement. In
23
Figure 1 shows the timing of the model.
each panel of Figure 2A we plot λsppor , the cumulative return response following the
announcement. The left hand side of the x-axis in each plot coincides with the time
indicated in the title of the plot. Each tick advances time by 5 minutes. For example, in the
top left hand corner plot the first tick indicates 7:50 EST and the return is measured from
just before 7:45 EST to 7:50 EST. The second tick indicates 7:55 EST and the return is
measured from just before 7:45 EST to 7:55 EST. In Figure 2A the vertical line indicates
9:30 EST when the U.S. market opens.24
The first panel of Figure 2A shows that the Portuguese stock market does not react
to the ECB announcement until 9:30 EST, after the President of the ECB finished
delivering the monetary policy statement and Reuters puts out a report on the ECB’s press
conference, but also as the U.S. stock market opens. For the U.S. real activity,
consumption, investment and price announcements, the Portuguese stock market reacts
from 15 minutes before 9:30 (45 minutes after the announcement) to 20 minutes after 9:30
( 1 hour and 20 minutes after the announcement). For the remaining 8:30 and 9:15 EST
announcements, there is no significant response. To the best of our knowledge, this is the
first paper that finds direct evidence of delayed response to U.S. macroeconomic
announcements in foreign markets, which coincides with the U.S. stock market open.25
This evidence suggests that Portugal reacts more to New York’s assessment of U.S.
announcements than to the news themselves. In all panels of Figure 2A, it is clear that
volatility increases around 9:30 EST, consistent with the empirical evidence in King and
Wadhwani (1990), who show that volatility in the FTSE-100 index is higher from 9:45 to
10:15 EST just after New York’s stock market opens. In contrast to their paper, we do not
interpret this result as evidence of contagion, but as evidence of price discovery taking
place in the U.S. first and then in Portugal. This interpretation is also consistent with
Hasbrouck (2003), who shows that price discovery first takes place in the most liquid
market, the U.S. stock market, and then in the least liquid market, the Portuguese stock
24
The fact that the GLOBEX market began on the CME trading the S&P 500 futures contract after regular
trading does not show up in our data.
25
King and Wadhawani (1990) document that volatility in the FTSE-100 index is lower from 8:30 to 9:30 EST
than from 9:45 to 10:15 EST, which they interpret as evidence that the U.K. reacts more strongly to New
York's interpretation of the news, than to the news themselves. The difference between their results and ours, is
that (i) we focus on the conditional mean, not the conditional variance, and (ii) we estimate the stock market
response to news surprises, not the response to announcements.
market.
In Figure 2B we show the U.S. response to the same 8:30 and 9:15 EST U.S.
macroeconomic announcements and the 7:45 EST ECB announcement. Since the U.S.
stock market is not yet open, the first interval in all panels on the x-axis captures the U.S.
stock market response from the previous day’s close to 9:35 EST, the second interval
captures the cumulative response from the previous day’s close to 9:40 EST, and so on.
The last interval captures the response from the previous day’s close to 11:30 EST. In the
first panel, we can observe that the U.S. market response to ECB announcements is
immediate. There is some overshooting, since the immediate effect of -0.5 is reversed 20
minutes later to +0.1, but settles 30 minutes later to -0.2 (a typical unanticipated rate hike
of 25 basis points in the ECB rate is associated with a decrease of roughly 0.2 percent in
the level of stock market prices, compared with Kuttner and Bernanke’s (2005) estimate of
1 percent for an unanticipated hike of 25 basis points in the Fed funds rate). For the real
activity, consumption, forward-looking and initial unemployment claims announcements,
the effect is immediate, with a tendency of overshooting, and becoming stable within one
hour of the announcement’s release. The effect of the investment announcements are fully
reversed by the end of the day, the net exports announcement only becomes statistically
significant towards 11:30 EST and the price announcements are insignificant.
In Figure 3A and 3B, we report the contemporaneous 5-minute cumulative
response to the 10:00 EST announcements, when both markets are open. Though these
results are not fully satisfying because the most important U.S. announcements are at 8:30
EST rather than 10:00 EST, they have the advantage of being generated when both markets
are open. The Portuguese stock market significantly reacts to the consumption
announcement 15 minutes after they are released (panel 1 in Figure 3A), this reaction
temporarily reverses itself, and 45 minutes later it becomes stable. In contrast, the U.S.
stock market also reacts 15 minutes later to this announcement (panel 1 in Figure 3B), and
40 minutes later it becomes stable. The Portuguese stock market does not react to the
investment announcement (panel 2 in Figure 3A), while the U.S. stock market (panel 2 in
Figure 3B) has an immediate reaction which becomes insignificant by 11:30 EST. Finally,
the Portuguese stock market reacts to forward-looking announcements 15 minutes after
they are released (panel 3 in Figure 3A), whereas the U.S. stock market (panel 3 in Figure
3B) incorporates the information in these announcements immediately. This evidence,

though weaker than that presented in Figure 2A and Figure 2B, supports the view that
Portugal waits for New York’s assessment of U.S. announcements.
Finally, Figure 4A and 4B reports the local and foreign stock market response to
U.S. news when the U.S. market is open, but the Portuguese market is not (14:00, 14:15
and 15:00 EST macroeconomic announcements). The first panel in Figure 4A shows that
the Portuguese reaction to the previous day’s U.S. government budget deficit surprise is
incorporated into Portuguese stock market returns within 20 minutes of the market open in
Portugal (at 3:30 EST the next day). However, this news becomes statistically insignificant
by 11:30 EST, so it does not have a permanent effect on Portuguese stock market prices.
This is not surprising, because the U.S. stock market (panel 1 Figure 4B) does not react
strongly to this announcement. In contrast, the federal funds rate surprise is important to
both markets. It is incorporated into the Portuguese stock price as soon as the market opens
(panel 2 Figure 4A), though it experiences a small reversal in the next 10 minutes, and it
becomes stable within two hours of market open. This quick reaction is in contrast to those
in Figure 3A where responses in the Portuguese market to U.S. announcements were never
immediate. Surprisingly, the federal funds rate affects Portuguese stock market prices more
than the ECB interest rate, since a typical unanticipated rate hike of 25 basis points in the
federal funds rate is associated with a decrease of roughly 0.8 percent in the level of
Portuguese stock market prices, compared with a decrease of 0.2 percent associated with
the ECB rate announcement. The U.S. stock market (panel 2 Figure 4B) reacts within 5-
minutes of the announcement and it has a permanent effect; an unanticipated rate hike of 25
basis points in the federal funds rate is associated with a decrease of roughly 2.5 percent in
the level of U.S. stock market prices, compared with 1 percent estimated by Bernanke and
Kuttner (2005) using the NYSE-AMEX-NASDAQ value weighted index and data from
June 1989 to December 2002. Finally, the U.S. consumer credit surprise is incorporated
into Portuguese stock market prices (panel 3 Figure 4A) within one hour and 15 minutes of
market open.
Taken together, the empirical evidence presented in Figures 2, 3 and 4 supports
the model’s timing assumption that Portuguese investors wait for news to be incorporated
in the U.S. before acting on them. We next turn to estimate the impact of unanticipated
U.S. order flow and U.S. public news surprises on daily U.S. price changes, to test the
predictions of the model outlined in Section 4.2.
6.2. Impact of U.S. News on U.S. Returns

We translate the equilibrium prices in Section 4.2 into the following estimable
equation:
15 J
ritus = a + λspeus Sipeus t + λspecb S pecb t + ∑λ
pus = 7
spus S pus t + ∑λ j Ωit − j (1 − Dtus ) +
j =0
J J J
(16)
∑λ
j =0
pj Ωit − j D + ∑β r
t
us
j =1
us
j it − j (1 − D ) + ∑β r
t
us
j =1
us
pj it − j D + ε it ,
t
us
where Dtus is an indicator function for U.S. public (earnings or macroeconomic)

announcement release dates, ritus = (ln( Pitus ) − ln( Pitus−1 )) ×100 is the individual stock return
for firm i = 1,...,30 , in the DJ 30 Index, S pus t is the standardized news corresponding to
announcement pus ( pus = 7,...,15 ) listed in Table 5, and S pecb t is the standardized ECB
benchmark interest rate news surprise made at time t . We include the ECB benchmark
refinancing interest rate and none of the Portuguese macroeconomic news announcements,
because we do not expect the Portuguese macroeconomic announcements to affect the U.S.
economy, however the ECB Governing Council decision has started to influence the U.S.
economy since the advent of the European Monetary Union (Ehrmann and Fratzscher,
2003).26 Lagged unanticipated order flow values and lagged price changes are included in
equation (16) to differentiate between our informed-trading hypothesis from the equally
sensible inventory-model alternative (first formalized by Garman, 1976).27 Price changes
may react to net order flow imbalances to compensate market participants for providing
liquidity, even when the order flow has no information content. To assess the relevance of
this alternative hypothesis, we follow Hasbrouck (1991) and include lagged values of
unanticipated order flow and price changes in all of our equations. As in Hasbrouck (1991),
26
Consistent with our expectation, the U.S. stock market does not respond to Portuguese macroeconomic
surprises. Previous studies have found similar results with, for example, United Kingdom macro surprises
(Becker, Finnerty and Friedman, 1995) and German surprises (Andersen, Bollerslev, Diebold, Vega, 2003).
27
The order of the lagged polynomial, J , is set to asses the permanence of order flow rather than setting it
optimally using the Akaike and Schwarz information criteria. In this study we define the impact of order flow
on yield changes as permanent (i.e., driven by information effects) when lasting for at least five trading days.
Hence, we set J = 5 in all equations herein.
we assume the permanent impact of trades is due to information shocks and the transitory
impact is due to non-information (e.g., liquidity) shocks. Hence, positive and significant
contemporaneous estimates for λ0 and λ p 0 are driven by transitory inventory control
effects when accompanied by a negative and significant impact of lagged unanticipated net
order flow on price changes. In other words, significant contemporaneous order flow
effects are transitory if they are later reversed. On the other hand, positive and significant
estimates for λ0 and λ p 0 are driven by permanent information effects (consistent with our
model) when accompanied by positive and significant, or statistically insignificant impact
of lagged unanticipated net order flow on yield changes.
We use a GARCH(1,1)-X model to control for heteroskedasticity in the data.
Specifically, we model the conditional variance of ε it as follows:
15
σ it2 = ω + βσ σ it2−1 + βε ε it2−1 + φTit + ψ p Dtp +ψ p Dtp +
eus
eus
ecb
ecb
∑ψ
pus = 7
pus Dtpus , (17)
where Tit is the number of transactions on day t and stock i , Dtpeus is an indicator function
equal to one when a U.S. earnings announcement is released, Dtpecb is an indicator function
equal to one when the ECB benchmark interest rate is announced and Dtpus is equal to one
when U.S. macroeconomic announcement indicator pus is released.
In Table 6 we show that news on U.S. macroeconomic fundamentals exert a
significant influence in the U.S. stock market. The sign of the coefficients indicate that an
increase in real economic activity during economic recessions is good news for the stock
market. While our data is too short to separate between expansion periods and recession
periods, the signs associated with U.S. macroeconomic news are consistent with Andersen
et al. (2005) and Boyd, Jagannathan and Hu (2001) if one takes the period we study in 2002
to be a recession.28 The economic explanation advanced in Andersen et al. (2005) and
Boyd, Jagannathan and Hu (2001) is that the discount rate effect dominates during
28
According to the NBER the starting date for the closest recession to 2002 in the U.S. is 03/2001 and the end
date of the same recession is 11/2001. There is however wide disagreement about the recession's end date as
explained in Andersen et al. (2005) and the NBER's website http://www.nber.org/cycles/recessions.html,
because employment continued to decline through June 2003. Andersen et al. (2005) define the recession end
date as 12/2002, which would define our sample period as a recessionary period in the U.S. Under this
ambiguity it is reasonable to believe that the market did not know whether the recession was over during the
11/2001-12/2002 period (note that the end of the 03/2001 recession was declared on July 17, 2003).
economic expansions while the cash flow effect dominates during economic contractions,
because the Federal Reserve Bank is less likely to increase interest rates during recessions.
This claim is further supported by the statistical insignificance of the inflationary shocks
(PPI and CPI surprises). A one standard deviation unexpected increase in the federal funds
target rate decreases stock market returns by 0.76 percent, consistent with Bernanke and
Kuttner (2005), who find a 1 percent decrease in the value weighted CRSP index.
Similarly, a one standard deviation unexpected increase in the ECB benchmark rate
decreases stock market returns by 0.16 percent.
The results in Table 6 provide evidence that strongly supports the informed-
trading hypothesis and some evidence supporting small inventory control effects. The
estimated contemporaneous correlation between unanticipated order flow and price changes
( λ p 0 and λ0 ) are positive and significant (at the 1% level). First order lagged unanticipated
order flow ( λ p1 and λ1 ) is negative and statistically significant at the 1% significance
level, however its magnitude is about 4 times smaller than the contemporaneous
coefficient. Furthermore, the third and higher order (not shown) lagged unanticipated order
flow is statistically insignificant at all significance levels. In other words, we find evidence
that the contemporaneous impact of unanticipated U.S. stock market order flow is reversed
the next day, but the magnitude of the reversal is small enough to be consistent with a small
inventory control effect and a larger informational order flow effect studied in the model.
According to our model, we also expect λ0 − λ p 0 > 0 , holding the inventory effect constant
on announcement and non-announcement days and as long as the private information
which agents receive is sufficiently precise. Table 6 shows that we fail to reject this null
hypothesis at five percent significance level. In other words, liquidity is statistically
significantly higher in announcement days versus non-announcement days.
In untabulated results and consistent with Andersen and Bollerslev (1998) we find
that volatility is higher during announcement days. Previous literature found that daily
stock market volatility is positively correlated with the number of transactions (e.g., Jones
et al., 1994). However, we found that the number of transactions becomes insignificant
when including the announcement dummies. Since the number of transactions is a proxy
for investors’ dispersion of beliefs, the results suggest that the announcement dummies
might more accurately measure this dispersion of beliefs.
6.3. Impact of Portuguese Macroeconomic News on Portuguese and U.S. Comovement

We next turn to examine the impact of unanticipated Portuguese order flow and
Portuguese public news surprises on Portuguese daily price changes and U.S.-Portuguese
stock market correlations. Similar to the previous specification, we translate the equilibrium
prices into the following estimable equation:
4 15
ritpor = a por + λspporepor Sipepor t + λspporecb S pecb t + ∑ λsp S p
p por = 2
por
por por t
+ ∑λ
pus = 7
por
spus S pus t +
J J J
∑λ
j =0
por
j Ω∗it − j (1 − Dtpor ) + ∑λ pjpor Ω∗it − j Dtpor + ∑β jpor rtus− j (1 − Dtpor )
j =0 j =0
(18)
J J
+ ∑β pjpor rtus− j Dtpor + ∑β por j rit − j + ε it ,
por por por
j =0 j =1
where Dtpor is an indicator function for Portuguese macroeconomic announcement release

dates, ritpor = (ln( Pitpor ) − ln( Pitpor
−1 )) × 100 is the individual stock return for firm i = 1,..., 20 , in
the PSI-20 Index, rtus = (ln( Pt us ) − ln( Pt us−1 )) ×100 is the value weighted DJ 30 Industrial
Index return, S ip epor t is the standardized news corresponding to Portuguese earnings
announcement for firm i = 1,..., 20 , S p por t is the standardized news corresponding to
Portuguese announcement p por ( p por = 2,3, 4 ) listed in Table 5, S pecb t is the standardized
ECB benchmark interest rate news surprise, and S p us t is the standardized news
corresponding to announcement pus ( pus = 7,...,15 ) listed in Table 5. We use a similar
GARCH(1,1)-X model to control for heteroskedasticity in the data. Specifically, we model
the conditional variance of ε itpor as follows:
σ itpor 2 = ω por + βσporσ itpor

−1 + β ε ε it −1 + φ Tit + ψ p Dt
2 por por 2 por por pepor
epor
4 15 (19)
+ψ pporecb Dtpecb + ∑ ψ p Dt + ∑ψ Dtpus + βεporus rt us−12 ,
por p por por
por pus
p por = 2 pus = 7
where rtus−12 in equation (19) controls for volatility spill over effects (e.g., Karolyi and Stulz,
1996) from the U.S. market to the Portuguese market.
According to the model, we expect that stock return correlations fall on public
macroeconomic news announcement days, β 0por − β ppor
0 > 0 . In the first two columns of
Table 7 we find evidence to strongly support this claim. One standard deviation shock to
the value weighted DJ 30 Industrial index return (1.83%) increases Portuguese returns by
0.46% during non-announcement days and only by 0.15% on announcement days. This
difference is statistically and economically significant (.
The signs and significance of the coefficients associated with the Portuguese
macroeconomic news announcements are consistent with the U.S. market results presented
above. The only exception is that Portuguese inflationary shocks have a strong negative
effect on Portuguese stock market returns, while U.S. inflationary shocks had no significant
effect on U.S. stock market returns during the U.S. recession period. Interestingly, the U.S.
macro news announcements slightly gain statistical significance when we drop the value
weighted DJ 30 Industrial Index return from equation (18). This means that part of the
comovement between the U.S. and Portugal is due to the effect U.S. macroeconomic
fundamentals have on the Portuguese stock market.
Consistent with the U.S. market results in Table 6, Table 7 also provides evidence
that strongly supports the informed-trading hypothesis and to a lesser extent evidence in
favor of small inventory effects. The estimated contemporaneous correlation between
unanticipated order flow and price changes ( λ ppor
0 and λ0
por
) are positive and significant (at
the 1% level). First order lagged unanticipated order flow ( λ ppor
1 and λ1
por
) is negative and
statistically significant at the 1% significance level, however its magnitude is 9.7 times
smaller on announcement days ( λ ppor
1 ) and 4.36 times smaller on non-announcements days
( λ1por ) than the contemporaneous coefficient. Furthermore, the third and higher order
lagged unanticipated order flow effects are statistically insignificant at all significance
levels for announcement and non-announcement days (not shown).
Opposite to the U.S. market, liquidity during macroeconomic news announcement
days is statistically significantly lower than liquidity on non-announcement days. One
explanation for this unexpected behavior of liquidity is that Portuguese macroeconomic
news necessitate more analysis to be useful, leading to the entry into the market of a
different class of informed investors (see Kim and Verrecchia, 1994).
In untabulated results we note that, in contrast with the U.S., Portuguese stock
market volatility is lower during announcement days of real activity. However, and in
contrast with the evidence for the U.S., daily stock market volatility is strongly positively
correlated with the number of transactions as in Jones et al. (1994).
6.4. Impact of Portuguese Earnings Announcements on Portuguese and U.S. Comovement
Now, we examine the impact of country specific news (section 4.6) on the U.S.-
Portuguese stock market correlations. Similar to the previous specification, we replace
Dtpor in equation (18) with Dtepor , an indicator function for Portuguese earnings
announcement release dates. According to the model, we expect that cross-country return
correlations are unchanged with earnings announcements, β 0epor − β pepor
0 = 0 . In column 3
and 4 of Table 7 we find evidence to strongly support this claim. One standard deviation
shock to the value weighted DJ 30 Industrial index return (1.83%) increases Portuguese
returns by 0.42% during non-announcement days and only by 0.43% on announcement
days and this difference is not statistically significant.
Otherwise, the signs and significance of the coefficients associated with the
Portuguese macroeconomic news announcements, earnings news announcements, ECB
benchmark interest rates and U.S. macroeconomic news announcements are very similar to
those in the previous section.
6.5. Impact of U.S. News on Portuguese and U.S. Comovement
Finally, we examine the impact of U.S. macroeconomic public news surprises on

the U.S.-Portuguese stock market comovement. We replace Dtpor in equation (18) with
Dtus , an indicator function for U.S. macroeconomic announcement release dates. We
expect that cross-country return correlations remain unchanged across U.S. macroeconomic
announcement days and non-announcement days, i.e., β 0us − β pus0 = 0 . In the last two
columns of Table 7 we find evidence to strongly support this claim. In particular, one
standard deviation shock to the value weighted DJ 30 Industrial index return increases
Portuguese returns by about 0.4% in announcement as well as non-announcement days.
Consistent with the model the sign of λ0 − λ p 0 > 0 meaning that at announcements
liquidity increases, but this difference is not statistically significant. There appears to be no
impact of U.S. macroeconomic news announcements on Portuguese adverse selection costs
at the daily frequency. Finally, the results in the last two columns are broadly consistent
with those in columns one through four for Portuguese and U.S. news announcements.
6.6. Robustness Checks

Since some of the macroeconomic announcements listed in Table 4A and Table
4B do not significantly affect stock market returns, our definition of announcement days
might bias the results presented above towards not finding any change in the cross-country
stock market correlation across days. In other words, we may be comparing non-
announcements days with essentially non-announcement days (days when insignificant
macroeconomic news are released). To explore this possibility we re-estimate equation (18)
using only significant U.S. and Portuguese macroeconomic news announcements. For
comparison, we report in Table 8 the coefficient estimates of equation (16), when we don’t
aggregate the U.S. announcements. We can observe that only 50% of the macroeconomic
news announcements listed in Table 4B are statistically significant, hence our previous
results could be biased. In the first two columns of Table 9 we report the coefficient
estimates of equation (18) redefining our indicator function for announcement days to be
equal to days when a statistically significant Portuguese macroeconomic announcement is
released (6 out of the 8 announcements listed in Table 4A are statistically significant). The
coefficient estimates for the Portuguese-U.S. stock market correlations are qualitatively the
same as those in Table 7 (column 1 and 2). Interestingly, all the U.S. announcements that
are significant in Table 9 (column 1, 3, and 5) when the U.S. value weighted return is not
included in equation (18) remain significant in Table 8 (with the exception of the business
inventories). This means that the Portuguese stock market only reacts to those U.S.
macroeconomic announcements that affect the U.S. stock market. As previously shown,
many of the U.S. macroeconomic announcements become insignificant when the value
weighted U.S. return is included in the regression. Similarly, the results in Table 9, column
4 and 6, report that the Portuguese-U.S. stock market correlation is the same during U.S.
significant announcement days and non-announcement days (including days when
insignificant announcements are released). Hence, our results are robust to different
announcement day definitions.
Finally, we test for predictable patterns in the Portuguese-U.S. stock market

correlation in the same spirit as Karolyi and Stulz (1996). In particular, we replace Dtpor in
equation (18) with DtLARGE , an indicator function equal to one if the U.S. value weighted DJ
30 index experiences a top 10% jump on day t , and with day-of-the week indicator
functions {Dt
Monday
}
, DtTuesday , DtWednesday , DtThursday . In the first specification we test the
hypothesis that the Portuguese stock market overreacts to any U.S. stock market shock. In
the second specification, we consider the possibility of predictable patterns in the cross-
country stock market correlation, that cannot be explained with our model. In Table 10 we
show that, in contrast to Karolyi and Stulz (1996), the Portuguese stock market does not
react significantly to large shocks to the U.S. stock market (i.e., β 0por
LARGE is insignificant)
and there is no evidence of day-of-the week effects (only the Thursday correlation is
slightly significantly lower at 10% significance level). Thus, we conclude our model
explains the time-varying properties of the U.S.-Portuguese stock market correlation better
than these two patterns previously found in the literature.
7. Conclusion
The main goal of this paper is to deepen our understanding of cross-country
correlations. To that end, we theoretically identify and empirically document important
public news and order flow effects in the U.S. and Portuguese stock markets. To guide our
analysis, we develop a parsimonious model of speculative trading in the presence of short-
lived private and public information about future dividends. We then test its equilibrium
implications by studying the relation between daily U.S. and Portuguese stock market
returns, order flow and real-time macroeconomic news releases.
Our evidence suggests that cross-country stock market return correlations are
unchanged when U.S. news is released, however this correlation decreases when
Portuguese news is released, so long as the news are not country-specific. In addition, we
show that adverse selection costs decrease in days of news announcements in the U.S.
market only. Also, U.S. public information affects Portuguese stock market returns, but this
effect is diminished when U.S. stock market returns are included in the regression. This
means that part of the comovement between the U.S. and Portugal is due to the effect U.S.
macroeconomic fundamentals have on the Portuguese stock market. Consistently, we

observe that only those macroeconomic announcements that affect the U.S. stock market
affect the Portuguese stock market.
Appendix
The appendix gives the full details for the results presented in the paper. We start
with the large economy and compute the equilibrium in days of no news. The market maker
sets a price Pt ∗ = Vt ∗ + λ ∗ω t∗ , and the informed investors’ problem is
⎡ n ⎤
max E ⎡⎣ Pt ∗+1 − Pt ∗ ⎤⎦ xit∗ = max E ⎢ vt∗+1 − λ ∗ ∑x ∗jt ⎥ xit∗ ,
⎣ j =1 ⎦
with first order necessary and sufficient condition (with conjecture that x ∗jt = xit∗ ):
⎡ n ⎤
0 = E ⎢ vt∗+1 − λ ∗ ∑x ∗jt | st∗ ⎥ − λ ∗ xit∗ .
⎣ j =1 ⎦
Because
⎡vt∗+1 ⎤ ⎛ ⎡δ ∗ δ∗ ⎤⎞
⎢ ∗⎥ → ⎜ ⎥ ⎟,
⎜ ⎢
N 0,
⎣ st ⎦ ⎝ ⎣ δ ∗ + φ ∗ ⎦ ⎟⎠
we have that E ⎡⎣vt∗+1 | st∗ ⎤⎦ = δ∗

δ ∗ +φ ∗
st∗ and
1 δ∗
xit∗ = st∗ .
λ ∗ ( n∗ + 1) δ + φ
∗ ∗
The market maker’s problem is
0 = E ⎣⎡ Pt ∗+1 − Pt ∗ ⎦⎤ ωt∗ = E ⎣⎡vt∗+1 | ωt∗ ⎦⎤ − λ ∗ωt∗ .
n∗ δ∗
With ω t∗ = st∗ + zt∗ = β ∗ st∗ + zt∗ we obtain:
( ) ∗
λ ∗ n∗ +1 δ +φ
∗
⎡vt∗+1 ⎤ ⎛ ⎡δ ∗ β ∗δ ∗ ⎤⎞
⎢ ∗⎥ → N ⎜ 0, ⎢ ⎥ ⎟.
⎣ ω t ⎦ ⎜ ⎢ β ∗2 (δ ∗ + φ ∗ ) +ζ ∗⎥⎟
⎝ ⎣ ⎦⎠
β ∗δ ∗
Therefore E ⎡⎣vt∗+1 | ωt∗ ⎤⎦ = ωt∗ . Replacing in the market maker’s optimality
( )
β ∗2 δ ∗ +φ ∗ +ζ ∗
condition we get
δ∗ n∗
λ∗ = ,
(
n∗ + 1 ζ ∗ δ ∗ + φ ∗ )
and
ζ∗
xit∗ = st∗ .
n∗ (δ ∗ + φ ∗ )
Consider a day of news in the foreign economy. Prices are

Pt ∗ = Vt ∗ + λ ∗ω t∗ + σ ∗U t∗ . The informed investors’ problem yields:
⎡ n ⎤
max E ⎡⎣ Pt ∗+1 − Pt ∗ ⎤⎦ xit∗ = max E ⎢ vt∗+1 − λ ∗ ∑x ∗jt − σ ∗U t∗ ⎥ xit∗ ,
⎣ j =1 ⎦
with FONC (with conjecture that x ∗jt = xit∗ ):
⎡ n ⎤
0 = E ⎢ vt∗+1 − λ ∗ ∑x ∗jt | st∗ ,U t∗ ⎥ − σ ∗U t∗ − λ xit .
⎣ j =1 ⎦
Because
⎡vt∗+1 ⎤ ⎛ ⎡δ ∗ δ∗ δ∗ ⎤⎞
⎢ ∗⎥ ⎜ ⎢ ⎥⎟
⎢ st ⎥ → N ⎜ 0, ⎢ δ ∗ + φ∗ δ ∗ ⎥ ⎟,
⎢U t∗ ⎥ ⎜ ⎢ δ ∗ + κ ∗ ⎥⎦ ⎟⎠
⎣ ⎦ ⎝ ⎣
we obtain
1 ⎡ δ ∗κ ∗ ∗ ⎛ δ ∗φ ∗ ⎞ ⎤
xit∗ = ⎢ st + ⎜ − σ ∗ ⎟ U t∗ ⎥ ,
∗
(
λ n + 1 ⎣⎢ Δ 5
∗
) ⎝ Δ5 ⎠ ⎦⎥
with Δ 5 = δ ∗κ ∗ + φ ∗δ ∗ + φ ∗κ ∗ . The market maker’s problem is to find λ ∗ and σ ∗ such

that:
0 = E ⎡⎣ Pt ∗+1 − Pt ∗ ⎤⎦ ωt∗ = E ⎡⎣ vt∗+1 | ωt∗ ,U t∗ ⎤⎦ − λ ∗ωt∗ − σ ∗U t∗ .
Because total order flow is

n∗ δ ∗κ ∗ ⎛ δ ∗φ ∗
n∗ ⎞
ω t∗ = st∗ + ⎜ − σ ∗ ⎟ U t∗ + zt ,
(
λ ∗ n∗ + 1 Δ 5

) ∗
(
λ n + 1 ⎝ Δ5
∗

)
⎠
β0 β1
we get
⎡vt∗+1 ⎤ ⎛ ⎡δ ∗
⎜
( β0 + β1 ) δ ∗ δ∗ ⎤⎞
⎢ ∗⎥ ⎢ ⎥⎟
⎢ ωt ⎥ → N ⎜ 0, ⎢ β 02 ( δ ∗ + φ ∗ ) + β12 (δ ∗ + κ ∗ ) + 2 β 0 β1δ ∗ + ζ ∗ ( )
β 0δ ∗ + β1 δ ∗ + κ ∗ ⎥ ⎟ .
⎢U t∗ ⎥ ⎜ ⎢ ⎥⎟
⎣ ⎦ ⎜ δ∗ +κ∗ ⎥⎦ ⎠⎟
⎝ ⎢⎣
With the expectations E ⎡⎣vt∗+1 | ωt∗ ,U t∗ ⎤⎦ we can then obtain after much algebra:
n∗  ∗  ∗  2
 ∗2  ∗ ∗ ∗ ∗ ∗ ∗
 ∗  ∗   ∗ n ∗  1 2         
∗
∗  ∗  ∗ .
 
Proof of Proposition 1. In the foreign market news lead to an increase in liquidity

as the adverse selection is reduced λA∗ < λNA
∗
:
δ ∗κ ∗ n∗ δ∗ n∗
λA∗ = < ∗
= λNA ,
∗
( )(
n + 1 ζ ∗ δ ∗ + κ ∗ ⎡ δ ∗ + κ ∗ φ ∗ + δ ∗ − δ ∗2 ⎤
⎣ ⎦ )( ) ∗
(
n +1 ζ ∗ δ ∗ +φ∗ )
because
(
0 < δ ∗κ ∗ + δ ∗ + κ ∗ φ ∗ + φ ∗κ ∗ .)
■
On a day of no news, equilibrium in the small market requires Pt = Vt + λω t + ηωt∗
because the foreign order flow contains useful information for local investors. The
informed investors’ problem is
⎡ n ⎤
max E ⎡⎣ Pt +1 − Pt ⎤⎦ xit = max E ⎢vt +1 − λ ∑x jt − λ zt − ηω t∗ ⎥ xit ,
⎣ j =1 ⎦
with FONC
λ ( n + 1) xit = E ⎡⎣ vt +1 | st , ω t∗ ⎤⎦ − ηω t∗ .
It can be checked that
⎛ ⎡ ⎤⎞
⎡vt +1 ⎤ ⎜ δ δ β ∗ψ ⎥⎟
⎢ s ⎥ → N ⎜ 0, ⎢⎢ δ +φ β ∗ψ ⎥⎟,
⎢ t ⎥ ⎜ ⎢ ⎥⎟
⎢⎣ ω t∗ ⎥⎦ ⎜
⎝ ⎢⎣ ( )
n +1 ζ ⎥ ⎟
∗ ∗
⎦⎠
so that E ⎡⎣vt +1 | st , ω t∗ ⎤⎦ = a0 st + a1ω t∗ with
δ ( n∗ + 1) ζ ∗ − β ∗2ψ 2
a0 =
(n ∗
)
+ 1 ζ ∗ (δ + φ ) − β ∗2ψ 2
β ∗ψφ
a1 = ,
( n∗ + 1) ζ ∗ (δ + φ ) − β ∗2ψ 2
and
1
xit = ⎡ a0 st + ( a1 − η ) ω t∗ ⎤⎦ .
λ ( n + 1) ⎣
The market maker solves
0 = E ⎣⎡ Pt +1 − Pt ⎦⎤ ωt = E ⎣⎡vt +1 | ωt , ωt∗ ⎦⎤ − λωt − ηωt∗ .
It can be shown that
⎛ ⎡δ n
⎡⎣ a0δ + ( a1 − η ) β ∗ψ ⎤⎦ β ∗ψ ⎤⎞
⎡vt +1 ⎤ ⎜ ⎢ λ ( n +1)
⎥⎟
⎢ ω ⎥ → N ⎜ 0, ⎢ ⎡ a β ψ + ( a1 − η ) n + 1 ζ ⎤ ⎥ ⎟
⎢ t⎥ ⎜ ⎢ E ⎣⎡ω t2 ⎦⎤ n
λ ( n +1) ⎣ 0
∗ ∗
( ∗
)
⎦⎥ ⎟
⎢⎣ ω t∗ ⎥⎦ ⎜ ⎢ ⎥⎟
⎜
⎝ ⎢⎣
(
n +1 ζ
∗
)
∗
⎥⎦ ⎠⎟
with
2
⎛ ⎞
t
2 n
1 0 (1 )
⎡ a 2 δ + φ ) + ( a − η ) 2 n∗ + 1 ζ ∗ + 2a ( a − η ) β ∗ψ ⎤ + ζ .
⎜ λ ( n + 1) ⎟⎟ ⎣ 0 (
E ⎡⎣ω ⎤⎦ = ⎜
⎦
⎝ ⎠
Out of which we solve for E ⎡⎣vt +1 | ωt , ω t∗ ⎤⎦ . Tedious algebra then shows that the market
maker’s efficiency condition is satisfied with
( )
2
⎡δ n∗ + 1 ζ ∗ − β ∗2ψ 2 ⎤
λ = 2 1 n ⎣ ⎦
ζ ( n∗ + 1) ζ ∗ ( n + 1) ( )
n + 1 ζ (δ + φ ) − β ψ
∗ ∗ ∗2 2
2
βψ ∗
η= .
(n ∗
)
+1 ζ ∗
Proof of Proposition 2.The equilibrium return regression when there are no

public announcements in either economy is
ψ
(
Pt − Vt = λω t + ∗ Pt ∗ − Vt ∗
δ
)
ψ
because ηλ ∗−1 = δ∗
:
n ∗ ∗

∗ ∗
1 
 ∗−1   .
n ∗  1 ∗ ∗ n∗ ∗
n ∗1  ∗ ∗ ∗ 
Consider now a day of news in the small, local economy U t = vt +1 + μ t , with
E ⎡⎣ μ t2 ⎤⎦ = κ . The market maker is assumed to choose the following price function
Pt = Vt + λω t + ηω t∗ + σ U t . The informed investors’ problem is
⎡ n ⎤
max E ⎡⎣ Pt +1 − Pt ⎤⎦ xit = max E ⎢ vt +1 − λ ∑x jt − λ zt − ηωt∗ − σ U t ⎥ xit ,
⎣ j =1 ⎦
with FONC
λ ( n + 1) xit = E ⎡⎣ vt +1 | st , ωt∗ , U t ⎤⎦ − ηωt∗ − σ U t .
To compute expectations we can show that
⎡vt +1 ⎤ ⎛ ⎡δ δ β ∗ψ δ ⎤⎞
⎢s ⎥ ⎜ ⎢ ⎥⎟
δ +φ β ∗ψ δ ⎥⎟
⎢ t ⎥ → N ⎜ 0, ⎢
⎜ ⎢ ⎟
⎢ ω t∗ ⎥ ⎜ ⎢ β ∗ψ (n ∗
)
+1 ζ ∗ β ∗ψ ⎥⎥ ⎟
⎢ ⎥ ⎜ ⎢
⎣ Ut ⎦ ⎝ ⎣ δ β ∗ψ δ + κ ⎥⎦ ⎟⎠
which implies
1
xit = ⎡ a0 st + ( a1 − η ) ωt∗ + ( a2 − σ )U t ⎤⎦
λ ( n + 1) ⎣
with (abusing notation)
κ ⎡⎣δ ( n∗ + 1) ζ ∗ − β ∗2ψ 2 ⎤⎦ κ
a0 = = a2 ,
(δκ + φδ + φκ ) ( n ∗
)
+ 1 ζ − (φ + κ ) β ψ
∗ ∗2 2
φ
β ∗ψφκ
a1 = .
(δκ + φδ + φκ ) ( n∗ + 1) ζ ∗ − (φ + κ ) β ∗2ψ 2
0 = E ⎡⎣ Pt +1 − Pt ⎤⎦ ωt = E ⎡⎣vt +1 | ωt , ωt∗ , U t ⎤⎦ − λωt − ηωt∗ − σ U t .
It is easy to show that
⎡vt +1 ⎤
⎛ ⎡δ E [ vt +1ωt ] β ∗ψ δ ⎤⎞
⎜ ⎢ ⎥⎟
⎢ ω ⎥ ⎜ ⎢ E ⎡⎣ω ⎤⎦
2
E ⎡⎣ω t ω ⎤⎦ ∗
E [ω tU t ]⎥ ⎟
⎢ ⎥ → N ⎜ 0, ⎢
t t t
⎥⎟
⎢ ω t ⎥
(n )
∗
⎜ ⎢ ∗
+1 ζ ∗ β ∗ψ ⎥ ⎟
⎢ ⎥ ⎜ ⎢ ⎥⎟
⎣ t⎦
U ⎜ ⎢ δ + κ ⎦⎥ ⎟⎠
⎝ ⎣
with
n
E [ω tU t ] = ⎡ a δ + ( a1 − η ) β ∗ψ + ( a2 − σ )(δ + κ ) ⎤⎦
λ ( n + 1) ⎣ 0
n
E [ vt +1ω t ] = ⎡ a0δ + ( a1 − η ) β ∗ψ + ( a2 − σ ) δ ⎤⎦
λ ( n + 1) ⎣
E ⎡⎣ω t ω t∗ ⎤⎦ =
n
λ ( n + 1) ⎣
( )
⎡ a0 β ∗ψ + ( a1 − η ) n∗ + 1 ζ ∗ + ( a2 − σ ) β ∗ψ ⎤
⎦
and
2
⎛ ⎞
E ⎡⎣ω t2 ⎤⎦ = ⎜
n
1 ( 2 )
⎡ a 2 δ + φ ) + ( a − η )2 n∗ + 1 ζ ∗ + ( a − σ )2 (δ + κ ) + 2a ( a − η ) β ∗ψ +
⎜ λ ( n + 1) ⎟⎟ ⎣ 0 ( 0 1
⎝ ⎠
2a0 ( a2 − σ ) δ + 2 ( a1 − η )( a2 − σ ) β ∗ψ ⎤⎦ + ζ .
We can then solve for E ⎡⎣vt +1 | ωt , ωt∗ , U t ⎤⎦ and, after much algebra, find the parameters that
satisfy the market maker’s zero profit condition:
κ 2 ⎡⎣δ ( n∗ + 1) ζ ∗ − β ∗2ψ 2 ⎤⎦
2
n 1
λ = 2
( n + 1) (δκ + φδ + φκ ) ( n∗ + 1) ζ ∗ − (φ + κ ) β ∗2ψ 2 ζ ⎡⎣( n∗ + 1) ζ ∗ (δ + κ ) − β ∗2ψ 2 ⎤⎦

2
ζ ∗ n∗
κ ψ
(δ ∗
+φ ∗ )
η=
(δ + κ ) ( n∗ + 1) ζ ∗ − (δζ +nφ ) ψ 2
∗ ∗
∗ ∗
δ ( n∗ + 1) ζ ∗ − β ∗2ψ 2
σ= .
(n ∗
)
+ 1 ζ ∗ (δ + κ ) − β ∗2ψ 2
Proof of Proposition 3. Price correlation across markets is given by
 ∗ 
 ∗−1   ∗−1
n  1    −  ∗2  2
∗ ∗
   
 ∗  ∗.
∗
   1 − Corr̃ t , Ũ t 
2  ∗
varŨ t |̃ t  
∗−1
The slope 
d log ηλ ∗−1 ( )= (
var νt +1 | ωt∗ ) > 0.
dκ (
κ var U t | ωt∗ )
Noting that as κ → ∞ , ηλ ∗−1 → ψ
δ∗ ( )
and because d log ηλ ∗−1 / d κ > 0 it must be that
∗−1 ψ
ηλ < δ∗
for finite κ .■
Proof of Proposition 5.What are the liquidity effects in days of news relative to
no news? Compare
( )
2
⎡δ n∗ + 1 ζ ∗ − β ∗2ψ 2 ⎤
λNo
1 n ⎣ ⎦
_A =
2
ζ ( n∗ + 1) ζ ∗ ( n + 1) ( )
n∗ + 1 ζ ∗ (δ + φ ) − β ∗2ψ 2
2
with
κ 2 ⎡δ ( n∗ + 1) ζ ∗ − β ∗2ψ 2 ⎤
2
λA2, POR =
n ⎣ ⎦ 1
.
( n + 1) (δκ + φδ + φκ ) ( n∗ + 1) ζ ∗ − (φ + κ ) β ∗2ψ 2 ζ ⎡⎣( n∗ + 1) ζ ∗ (δ + κ ) − β ∗2ψ 2 ⎤⎦
2
If liquidity increases in news days then λA2, POR < λNo

2
_ A or
n∗
⎡ ⎤ δ + κ − ( n∗ +1)(δ ∗ +φ ∗ ) ψ
2
( φ + κ ) n∗
κ 2 < ⎢(δκ + φδ + φκ ) − ∗ ψ 2
⎥ .
⎢ ∗
n +1 δ +φ
∗
( )
⎥ δ + φ − n∗ +1 nδ ∗ +φ ∗ ψ 2
∗
⎣ ⎦ ( )( )
The right hand side of this inequality can be written as
⎡ ⎤ ( )
2
φ var vt +1 | ω t∗
2
(
κ + κ var vt +1 | ω ⎢1 + ∗
) ⎥ +φ .
⎢ var st | ω t∗
⎣
t
( ) ⎥
⎦ (
var st | ω t∗ )
Therefore we get that λA2, POR < λNo
2
_ A .■
There is also price discovery in the small country after news in the large economy.
The market maker is assumed to choose the following price function:
Pt = Vt + λω t + η ⎡⎣ω t∗ − β1∗U t∗ ⎤⎦ + σ U t∗ .

wt∗ = β 0∗ st∗ + zt∗
The informed investors’ problem is
⎡ n ⎤
max E ⎡⎣ Pt +1 − Pt ⎤⎦ xit = max E ⎢ vt +1 − λ ∑x jt − λ zt − η wt∗ − σ U t∗ ⎥ xit ,
⎣ j =1 ⎦
with FONC
λ ( n + 1) xit = E ⎡⎣ vt +1 | st , wt∗ ,U t∗ ⎤⎦ − η wt∗ − σ U t∗ .
It is possible to show that
⎡vt +1 ⎤ ⎛ ⎡ δ δ β 0∗ψ ⎤⎞ ψ
⎢s ⎥ ⎜ ⎢ ⎥⎟
δ δ +φ β 0∗ψ ψ
⎢ t ⎥ → N ⎜ 0, ⎢ ⎥⎟
⎢ w t∗ ⎥ ⎜ ⎢ ∗ ∗ ∗ ⎥⎟
⎜ ⎢ β 0ψ β 0∗ψ E ⎡⎣ w t ⎤⎦ β 0 δ
∗2
⎢ ∗⎥ ⎥⎟
⎣U t ⎦ ⎜ ⎢ β 0∗δ ∗ δ ∗ + κ ∗ ⎥⎦ ⎠⎟
⎝ ⎣ ψ ψ
with
ζ ∗ (δ ∗ + κ ∗ ) n∗
E ⎡⎣ w t∗2 ⎤⎦ =
δ ∗κ ∗ + φ ∗κ ∗ + φ ∗δ ∗
(δ ∗
)
+ φ ∗ + ζ ∗.
Tedious algebra shows that

1
xit = ⎡ a0 st + ( a1 − η ) wt∗ + ( a2 − σ ) U t∗ ⎤⎦ ,
λ ( n + 1) ⎣
with
a0 = ⎣⎢ ( ( ) ⎣ t ⎦ )
⎡ δ δ ∗ + κ ∗ −ψ 2 E ⎡ w ∗2 ⎤ − δβ ∗2δ ∗2 − β ∗2ψ 2 κ ∗ − δ ∗ ⎤
0 0
⎦⎥
( )
( 2
) 2 ∗2 2 ∗ ∗ ∗2
(
(δ + φ ) ⎡⎣ δ + κ E ⎡⎣ w t ⎤⎦ − β0 δ ⎤⎦ +ψ ⎡⎣ δ − κ β 0 − E ⎡⎣ w t∗2 ⎤⎦ ⎤⎦
∗ ∗ ∗ ∗
)
φκ ∗ β 0∗ψ
a1 =
(δ + φ ) ⎡⎣(δ ∗ + κ ∗ ) E ⎡⎣ w t∗2 ⎤⎦ − β0∗2δ ∗2 ⎤⎦ +ψ 2 ⎡⎣(δ ∗ − κ ∗ ) β0∗2 − E ⎡⎣ w t∗2 ⎤⎦ ⎤⎦
⎡ − β 0∗2δ ∗ + E ⎡ w t∗2 ⎤ ⎤ψφ
⎣ ⎣ ⎦⎦
a2 = .
(δ + φ ) ⎡⎣(δ ∗ + κ ∗ ) ∗2 ∗2 ∗2 ⎤
⎦ ⎣ ( )
E ⎡⎣ w t ⎤⎦ − β 0 δ +ψ 2 ⎡ δ ∗ − κ ∗ β 0∗2 − E ⎡⎣ w t∗2 ⎤⎦ ⎤
⎦
0 = E ⎡⎣ Pt +1 − Pt ⎤⎦ ωt = E ⎡⎣vt +1 | ωt , wt∗ , U t∗ ⎤⎦ − λωt − η wt∗ − σ U t∗ .
The market maker considers
⎡vt +1 ⎤
⎛ ⎡δ E [ vt +1ωt ] β 0∗ψ ψ ⎤⎞
⎜ ⎢ ⎥⎟
⎢ ω ⎥ ⎜ ⎢ E ⎡⎣ω t2 ⎤⎦ E ⎡⎣ω t w t∗ ⎤⎦ E ⎡⎣ω tU t∗ ⎤⎦ ⎥ ⎟
⎢ t ⎥ → N ⎜ 0, ⎢ ⎥⎟
⎢ w t ⎥
∗
⎜ ⎢ E ⎡⎣ w t∗2 ⎤⎦ β 0∗δ ∗ ⎥ ⎟
⎢ ∗⎥ ⎜ ⎢ ⎥⎟
⎣U t ⎦ ⎜ ⎢ δ ∗ + κ ∗ ⎥⎦ ⎟⎠
⎝ ⎣
where
E ⎣⎡ω tU t∗ ⎦⎤ =
n
λ ( n + 1) ⎣
⎡ a0ψ + ( a1 − η ) β 0∗δ ∗ + ( a2 − σ ) δ ∗ + κ ∗ ⎤
⎦( )
n
E [ vt +1ω t ] = ⎡ a0δ + ( a1 − η ) β 0∗ψ + ( a2 − σ )ψ ⎤⎦
λ ( n + 1) ⎣
n ⎡ a0 β 0∗ψ + ( a1 − η ) E ⎡ w t∗2 ⎤ + ( a2 − σ ) β 0∗δ ∗ ⎤
E ⎣⎡ω t wt∗ ⎦⎤ = ⎣ ⎦
λ ( n + 1) ⎣ ⎦
and
2
⎛ ⎞
n ⎡ a 2 δ + φ ) + ( a − η )2 E ⎡ w ∗2 ⎤ + ( a − σ )2 δ ∗ + κ ∗ + 2a ( a − η ) β ∗ψ
⎜ λ ( n + 1) ⎟⎟ ⎣ 0 (
E ⎡⎣ w t2 ⎤⎦ = ⎜ 1 ⎣ t ⎦ 2 0 1 0 ( )
⎝ ⎠
+2a0 ( a2 − σ )ψ + 2 ( a1 − η )( a2 − σ ) β 0∗δ ∗ ⎤⎦ + ζ .
Computing E ⎡⎣vt +1 | ωt , wt∗ ,U t∗ ⎤⎦ we can then solve for the price parameters:
( ( ) ) ( ( ) )
2
n ⎡ β ∗2 ψ 2 δ ∗ − κ ∗ − δδ ∗2 + E ⎡ w ∗2 ⎤ δ δ ∗ + κ ∗ −ψ 2 ⎤
ζ ( n +1) ⎣⎢ 0
2
⎣ t ⎦ ⎦⎥
λ2 =
⎣ ⎣ ⎦ (
⎡ E ⎡ w t∗2 ⎤ δ ∗ + κ ∗ ) ⎦ ⎣⎣
∗ ∗
(⎦
∗2
)
− β 0 δ ⎢ (δ + φ ) δ + κ −ψ E ⎣⎡ w t ⎦⎤ + β 0 ψ δ − κ ∗ − δ ∗2 (δ + φ )⎤ ⎤⎥
∗2 ∗2 ⎤ ⎡ ⎡ 2⎤ ∗2 ⎡ 2
⎣
∗
⎦⎦ ( )
β ψκ ∗ ∗
η= 0
E ⎡⎣ w ⎤⎦ (δ + κ ∗ ) − β 0∗2δ ∗2
∗2
t
∗
ψ ⎣⎡ E ⎣⎡ w t∗2 ⎦⎤ − β 0∗2δ ∗ ⎦⎤
σ= .
E ⎡⎣ w t∗2 ⎤⎦ (δ ∗ + κ ∗ ) − β 0∗2δ ∗2
Proof of Proposition 6 The regression of returns in the small economy on returns

in the large economy conditional on local order flow and the available public information
yields a slope on the foreign return of
ψ ζ ∗ (δ ∗ + κ ∗ )( n∗ + 1) ψ
ηλ ∗−1 = =
t (
E ⎡⎣ w ⎤⎦ δ + κ
∗2 ∗ ∗
)−β ∗2
0δ ∗2
δ ∗
δ∗
(
n∗ζ ∗ δ ∗ + κ ∗ )
since β 0∗ = δ ∗κ ∗ + φ ∗δ ∗ +φ ∗κ ∗
. We also need to show that the coefficient on foreign news is
zero. To see this note that
Pt − Vt = λω t + η ⎡⎣ω t∗ − β1U t∗ ⎤⎦ + σ U t∗
( )
= λω t + ηλ ∗−1 Pt ∗ − Vt ∗ − ⎡η λ ∗−1σ ∗ + β1∗ − σ ⎤ U t∗ .
⎣ ⎦ ( )
But,
ψ
ηλ ∗−1σ ∗ = ,
δ ∗ +κ∗
and
⎛n∗ δ ∗φ ∗ ⎞
ηβ1∗ = η ⎜ −σ ∗ ⎟
∗ ∗
( ∗ ∗
) ∗ ∗ ∗ ∗
λ n +1 ⎝ δ κ + φ δ + φ κ ⎠
ψ n∗ δ ∗κ ∗
=− .
δ ∗ + κ ∗ ( n∗ + 1) δ ∗κ ∗ + φ ∗δ ∗ + φ ∗κ ∗
Therefore we obtain:
ψ δ ∗κ ∗ + ( n∗ + 1) φ ∗ (δ ∗ + κ ∗ ) ψ ⎣⎡ E ⎣⎡ w t∗2 ⎦⎤ − β 0∗2δ ∗ ⎦⎤
η ( λ ∗−1σ ∗ + β1∗ ) − σ = − = 0.
δ ∗ + κ ∗ ( n∗ + 1)(δ ∗κ ∗ + φ ∗δ ∗ + φ ∗κ ∗ ) E ⎡⎣ w t∗2 ⎤⎦ (δ ∗ + κ ∗ ) − β 0∗2δ ∗2
Finally, we consider country-specific news in the local, small economy. The price
function is Pt = Vt + λω t + ηω t∗ + σ U t . Informed investor’s problem is:
⎡ n ⎤
max E ⎡⎣ Pt +1 − Pt ⎤⎦ xit = max E ⎢ vt +1 − λ ∑x jt − λ zt − ηωt∗ − σ U t ⎥ xit ,
⎣ j =1 ⎦
with FONC given by
λ ( n + 1) xit = E ⎡⎣ vt +1 | st , ωt∗ , U t ⎤⎦ − ηωt∗ − σ U t .
Knowing that
⎡vt +1 ⎤ ⎛ ⎡δ δ1 β ∗ψ δ2 ⎤ ⎞
⎢s ⎥ ⎜ ⎢ ⎥⎟
δ1 + φ β ∗ψ
⎢ t ⎥ → N ⎜⎜ 0, ⎢ ⎥⎟
0
⎢ ⎟
⎢ ω t∗ ⎥ ⎜ ⎢ β ∗ψ (n ∗
)
+1 ζ ∗ 0 ⎥⎥ ⎟
⎢ ⎥ ⎜ ⎢
⎣ Ut ⎦ ⎝ ⎣ 0 0 δ 2 + κ ⎥⎦ ⎟⎠
we can write
1
xit = ⎡ a0 st + ( a1 − η ) ωt∗ + ( a2 − σ )U t ⎤⎦
λ ( n + 1) ⎣
with
δ1 ( n∗ + 1) ζ ∗ − β ∗2ψ 2 δ1 ( n∗ + 1) ζ ∗ − β ∗2ψ 2
a0 = = a1
(n ∗
)
+ 1 ζ ∗ (δ1 + φ ) − β ∗2ψ 2 β ∗ψφ
δ2
a2 = .
δ2 + κ
0 = E ⎡⎣ Pt +1 − Pt ⎤⎦ ωt = E ⎡⎣vt +1 | ωt , ωt∗ , U t ⎤⎦ − λωt − ηωt∗ − σ U t .
and because
⎛ ⎡ δ E [ vt +1ω t ] β ∗ψ δ2 ⎤⎞
⎡vt +1 ⎤ ⎜ ⎢ ⎥⎟
⎢ ω ⎥ ⎜ ⎢ E [ vt +1ω t ] E ⎡⎣ω t2 ⎤⎦ E ⎡⎣ω t ω t∗ ⎤⎦ E [ω tU t ]⎥ ⎟
⎢ t ⎥ → N ⎜ 0, ⎢ ⎥⎟
⎢ ω t∗ ⎥
⎢ ⎥
⎜ ⎢ β ∗ψ
⎜ ⎢
E ⎡⎣ω t ω t∗ ⎤⎦ (n ∗
+1 ζ ∗ ) 0 ⎥⎟
⎥⎟
⎣ Ut ⎦ ⎜ ⎢ δ2 E [ω tU t ] 0 δ 2 + κ ⎥⎦ ⎟⎠
⎝ ⎣
with
n
E [ω tU t ] = ( a2 − σ )(δ 2 + κ )
λ ( n + 1)
n
E [ vt +1ω t ] = ⎡ a0δ1 + ( a1 − η ) β ∗ψ + ( a2 − σ ) δ 2 ⎤⎦
λ ( n + 1) ⎣
E ⎣⎡ω t ω t∗ ⎦⎤ =
n
(
⎡ a0 β ∗ψ + ( a1 − η ) n∗ + 1 ζ ∗ ⎤
λ ( n + 1) ⎣ ⎦ )
and
2
⎛ ⎞
n
1 ( )
⎡ a 2 δ + φ ) + ( a − η )2 n∗ + 1 ζ ∗ + ( a − σ )2 (δ + κ ) + 2a ( a − η ) β ∗ψ ⎤ + ζ
⎜ λ ( n + 1) ⎟⎟ ⎣ 0 ( 1
E ⎡⎣ω t2 ⎤⎦ = ⎜ 2 2 0 1
⎦
⎝ ⎠
we obtain
2
⎡δ n∗ + 1 ζ ∗ − ψ 2 ⎤⎥
1 n ⎣⎢
1 ( ) n∗ζ ∗
δ ∗ +φ ∗
⎦
λ2 = ,
ζ ( n + 1) ζ
∗ ∗
( n + 1) (n ) (δ
2
∗
+1 ζ ∗
1 +φ ) − δ +φ
n∗ζ ∗
∗ ∗ ψ2
δ ∗
n ∗
ψ
η= ∗
n +1 ζ ∗
(δ ∗
+φ ∗
) δ∗
δ2
σ= .
δ2 + κ
Proof of Proposition 7. Regressing prices in the local economy on local order

flow, local country-specific news and foreign returns yields a coefficient on foreign returns
of
ψ
ηλ ∗−1 = .
δ∗
Country-specific news in the local economy do not change conditional correlations.■

References
Ammer, J., Mei J. (1996), Measuring International Economic Linkages with Stock Market
Data, Journal of Finance, 51, pp. 1743-1763.
Andersen, T.G., Bollerslev, T. (1998), Deutsche Mark-Dollar Volatility: Intraday Activity
Patterns, Macroeconomic Announcements, and Longer Run Dependencies,
Journal of Finance, 53, pp. 219-265.
Andersen, T., Bollerslev, T., Diebold, F., Vega, C. (2003), Micro Effects of Macro
Announcements: Real-Time Price Discovery in Foreign Exchange, American
Economic Review, 93, pp. 38-62.
Andersen, T., Bollerslev, T., Diebold, F., Vega, C. (2005), Real-Time Price Discovery in
Stock, Bond, and Foreign Exchange Markets, Working Paper, Northwestern
University.
Antão, P., Antunes, A., Martins, N. (2004), Liquidity and Order Flow in the Portuguese
Stock Market, Banco de Portugal Economic Bulletin, 10, pp. 115-126.
Bae, K. H., Karolyi, G. A., and Stulz, R. M. (2003), A New Approach to Measuring
Financial Contagion, Review of Financial Studies, 16, pp. 717-763.
Balduzzi, P., Elton, E., Green, C. (2001), Economic News and Bond Prices: Evidence From
the U.S. Treasury Market, Journal of Financial and Quantitative Analysis, 36, pp.
523-543.
Becker, K. G., Finnerty, J. E., Friedman, J. (1995), Economic News and Equity Market
Linkages between the U.S. and U.K., Journal of Banking and Finance, 19, pp.
1191-1210.
Bekaert, G., Harvey, C.R., and Ng, A., (2005), Market Integration and Contagion, Journal
of Business, 78, pp.39-69.
Bernanke, B. S., and Kuttner, K. N. (2005), What Explains the Stock Market’s Reaction to
Federal Reserve Policy?, Journal of Finance, 60, pp.1221-1257.
Boyd, J. H., Jagannathan, R. and Hu, J. (2001), The Stock Market’s Reaction to
Unemployment News: Why Bad News Is Usually Good For Stocks, Journal of
Finance, 60, pp. 649-672.
Brandt, M., Kavajecz, K. (2004), Price Discovery in the U.S. Treasury Market: The Impact
of Order Flow and Liquidity on the Yield Curve, Journal of Finance, 59, pp.
2623-2654.
Campbell, R., Koedijk, K., and Kofman, P., (2002), Increased Correlation in Bear Markets,
Financial Analysts Journal, 58, pp. 87-94.
Campbell, R., Forbes, C. S., Koedijk, K., and Kofman, P., (2003), Diversification
Meltdown or the Impact of Fat Tails on Conditional Correlation?, LIFE Working
Paper 03-027.
Claessens, S., Dornbusch, R. , and Park, Y. C., (2001), Contagion: Why Crises Spread and
How This Can Be Stopped, in S. Claessens, and K. J. Forbes (ed.), International
Financial Contagion. Kluwer Academic Publishers.
Cochrane, John H. and Monika Piazzesi (2002), The Fed and Interest Rates: A High-
Frequency Identification, American Economic Review Papers and Proceedings,
92, 90-95.
Connolly, R., Wang, F. A. (2003), International Equity Market Co-movements: Economic
Fundamentals or Contagion? Pacific-Basin Finance Journal, 11, pp. 23-43.
Craig, A., Dravid, A., and Richardson, M., (1995), Market Efficiency Around the Clock:
Some Supporting Evidence Using Foreign-Based Derivatives, Journal of
Financial Economics, 39, pp. 161-180.
Dumas, B., Harvey, C.R., Ruiz, P., (2003), Are Correlations of Stock Returns Justified by
Subsequent Changes in National Outputs?, Journal of International Money and
Finance, 22, pp. 777-811.
Ehrmann, M., Fratzscher, M. (2003), Interdependence Between The Euro Area and the
U.S.: What Role for EMU? European Central Bank Working Paper No 200.
Evans, M., Lyons, R. (2002), Order Flow and Exchange Rate Dynamics, Journal of
Political Economy, 110, pp.170-180.
Evans, M., Lyons, R. (2004), Exchange Rate Fundamentals and Order Flow, Working
Paper, Haas School of Business, University of California at Berkeley.
Fleming, M., Remolona, E., (1997), What Moves the Bond Market? Economic Policy
Review, Federal Reserve Bank of New York, pp. 31-50.
Gande, A, Parsley, D. C., (2005), News Spillovers in the Sovereign Debt Market, Journal
of Financial Economics, 75, pp. 691-734.
Garman, Mark B., (1976), Market Microstructure, Journal of Financial Economics, 3, pp.
257-275.
Green, C. (2004), Economic News and the Impact of Trading on Bond Prices, Journal of
Finance, 59, pp. 1201-1233.
Grossman, Sanford J., Stiglitz, Joseph E. (1980), On the Impossibility of Informationally
Efficient Markets, The American Economic Review 70, pp. 393-408.
Gürkaynak, Refet S., Sack, Brian P., Swanson, Eric (2002) Market-Based Measures of
Monetary Policy Expectations, Unpublished manuscript, Board of Governors of the
Federal Reserve System.
Hasbrouck, J., (1991), Measuring The Information Content of Stock Trades, Journal of
Finance, 46, pp. 179-207.
Hasbrouck, J., (2003), Intraday Price Formation in U.S. Equity Index Markets, Journal of
Finance, 58, pp. 2375-2399.
Hasbrouck, J., (2004), Economic and Statistical Perspectives on the Dynamics of Trade in
Securities Markets, Teaching Notes, Stern School of Business, New York University.
Jones, C. M., Kaul, G., Lipson, M. L. (1994), Transactions, Volume, and Volatility, Review
of Financial Studies, 7, pp. 631-351.
Karolyi, G. Andrew, and Rene M. Stulz (1996), Why Do Markets Move Together? An
Investigation of U.S.-Japan Stock Return Co-movements, The Journal of Finance, 51,
pp. 951-985.
King, M., Wadhwani S. (1990), Transmission of Volatility Between Stock Markets, Review
of Financial Studies, 3, pp. 5-33.
King, M., Sentana E., Wadhwani S. (1994), Volatility and the Link Between National
Stock Markets, Econometrica, 62, pp. 901-33.
Kim. O., Verrecchia, R. (1994), Market Liquidity and Volume around Earnings
Announcements, Journal of Accounting and Economics, 17, pp. 41-67.
Kim. O., Verrecchia, R. (1997), Pre-Announcement and Event-Period Private Information,
Journal of Accounting and Economics, 24, pp. 395-419.
Kodres, L. E., Pritsker, M. (2002), A Rational Expectations Model of Financial Contagion,
Kuttner, Kenneth N. (2001), Monetary Policy Surprises and Interest Rates: Evidence from
the Fed Funds Futures Market, Journal of Monetary Economics, 47, 523-544.
Kyle, A. S. (1985), Continuous Trading and Insider Trading, Econometrica, 53, pp. 1315-
1335.
Longin, F., Solnik, B., (2001), Extreme Correlation of International Equity Markets,
Pasquariello, P. (2004), Imperfect Competition, Information Heterogeneity, and Financial
Contagion, Working Paper, University of Michigan.
Poterba, James (1990), Transmission of Volatility Between Stock Markets: Discussion,
Review of Financial Studies, 3, pp. 34-35.
Forbes, K., Rigobon, R., (2002), No Contagion, Only Interdependence: Measuring Stock
Market Co-movements, Journal of Finance 57, pp 2223-2261.
Rigobon, R., Sack, B. (2002), The Impact of Monetary Policy on Asset Prices, Finance and
Economics Discussion Series. Board of Governors of the Federal Reserve System.
Stulz, R. (1999), International Portfolio Flows and Security Markets, in Feldstein, M. (ed.),
International Capital Flows. Chicago: The University of Chicago Press.
Wongswan, J. (2005), Transmission of Information Across International Equity Markets,
Review of Financial Studies, forthcoming.
Table 1A
Individual Stock Components of the PSI-20
This table lists the Portuguese stocks used in our empirical analysis, the ticker symbol,
company name, the exchange they are more frequently traded in, the average number of transactions
per day, the average number of limit orders placed per day and the percentage of orders that are limit
orders. The stocks are ordered from highest market capitalization to smallest market capitalization.
The sample period is from January 4, 2002 to October 15, 2002.
Ticker Symbol Company Name Exchange Transactions Limit Percent

PT Portugal Telecom, SGPS - Nom. Euronext Lisbon 1503 2816 0.93
BCP Banco Comercial Português Euronext Lisbon 462 879 0.86
BES Banco Espírito Santo Euronext Lisbon 79 161 0.87
EDP EDP Euronext Lisbon 471 1097 0.84
BRISA Brisa (Priv.) Euronext Lisbon 358 738 0.87
CIM POR CIM POR, SGPS Euronext Lisbon 104 297 0.89
BPI BPI, SGPS Euronext Lisbon 189 444 0.93
PTM PT Multimédia, SGPS - Nom. Euronext Lisbon 175 358 0.91
SONAE Sonae, SGPS Euronext Lisbon 193 564 0.94
JM Jerónimo M artins, SGPS Euronext Lisbon 126 337 0.94
PORTUCEL Portucel Euronext Lisbon 64 183 0.91
SEM APA Semapa, SGPS Euronext Lisbon 27 101 0.91
SNC Sonae.com, SGPS Euronext Lisbon 211 501 0.92
TD Teixeira Duarte, SA Euronext Lisbon 17 67 0.88
SAG SAG - Gest, SGPS Euronext Lisbon 19 73 0.85
NB Novabase, SGPS Euronext Lisbon 40 120 0.87
IM PRESA Impresa, SGPS Euronext Lisbon 184 427 0.93
COFINA Cofina, SGPS Euronext Lisbon 26 88 0.91
IBERSOL Ibersol, SGPS Euronext Lisbon 15 67 0.96
PARAREDE Pararede, SGPS Euronext Lisbon 101 334 0.93

Table 1B
Individual Stock Components of the DJ 30 Industrial Index
This table lists the U.S. stocks used in our empirical analysis, the ticker symbol, company
name, the exchange they are more frequently traded in, the average number of transactions per day
and the average number of quotes per day. The stocks are ordered from highest market capitalization
to smallest market capitalization. The sample period is from January 4, 2002 to October 15, 2002.
Ticker Symbol Company Name Exchange Trades Quotes

GE General Electric Corporation NYSE 5005 6826
MSFT M icrosoft Corp. NASDAQ 68213 76353
XOM Exxon M obil Corporation NYSE 3427 5570
WMT Wal M art Stores Inc. NYSE 2768 4828
C Citigroup Inc. NYSE 3889 6244
JNJ Johnson & Johnson NYSE 2375 3425
INTC Intel Corp NASDAQ 65596 72147
IBM International Business M achines Corp. NYSE 3277 8828
KO Coca Cola Co NYSE 2180 3612
MRK M erck CO Inc. NYSE 2393 3711
PG Procter & Gamble Co NYSE 2588 5754
MO Philip M orris Companies Inc. NYSE 2976 4812
SBC S B C Communications Inc. NYSE 2935 4534
HD Home Dep ot Inc. NYSE 3433 4853
JPM J P M organ Co NYSE 2949 4832
T AT&T Corp. NYSE 2094 3232
HPQ Hewlett Packard Co NYSE 2537 3703
AXP American Express Company NYSE 2602 4178
MMM M innesota M ining & M fg Co (3M Co) NYSE 1782 3841
DD Du Pont De Nemours E I & Co NYSE 1972 3903
DIS Walt Disney Company NYSE 2878 5633
BA Boeing Company NYSE 2178 3990
MCD M cdonalds Corp NYSE 2174 3450
UTX United Technologies Corp NYSE 1830 3476
GM General M otors Corp NYSE 2318 5284
HON Honeywell International Inc NYSE 1769 3505
AA Alcoa Inc NYSE 1809 3138
IP International Paper Co NYSE 1812 3553
CAT Caterpillar Inc NYSE 1464 3168
EK Eastman Kodak CO NYSE 1222 2519

Table 2
Portuguese and U.S. Stock Market Trading Hours
This table presents the Portuguese and U.S. stock market open and close.
Portuguese Portuguese
M arket Open M arket Close
PST 8:30  14:30  16:30  21:00
U.S. U.S.
M arket Open M arket Close
EST 3:30  9:30  11:30  16:00

Table 3A
PSI-20 Summary Statistics
This table presents daily descriptive statistics for the individual components of the
Portuguese PSI-20 stock market index. The first two columns report the daily mean and standard
deviation of firm-specific stock returns measured as the difference in log prices from 11:30 EST to
11:30 EST. The third column reports the daily standard deviation of unanticipated order flow
estimated using Hasbrouck (1991) VAR model. The last three columns report the mean percentage
spread, market capitalization and order execution time. The percentage spread is the ask price minus
the bid price divided by the average daily price, market capitalization (size) is measured in millions of
euros, and the order execution time is defined as the time it takes for a limit order to be executed
conditional on the order being executed on the same day it is placed.
Ticker Symbol M ean Return Std. Return Std. Order Flow Mean % Spread M ean Size M ean Time
PT -0.2564 2.10 181.87 0.109 8215.563 13.959
BCP -0.3916 1.81 67.07 0.288 5304.910 23.529
BES -0.1602 1.90 13.03 0.300 3750.000 23.123
EDP -0.2171 1.52 66.71 0.467 3298.312 32.208
BRISA 0.0400 1.42 64.25 0.227 2842.007 20.424
CIMPOR -0.0814 1.46 19.43 0.442 1934.312 25.722
BPI -0.1125 1.66 43.83 0.431 1656.800 28.744
PTM -0.1523 1.91 24.63 0.372 1571.596 22.906
SONAE -0.3696 1.99 26.96 1.504 800.000 34.870
JM -0.2341 2.02 18.76 0.522 666.218 27.729
PORTUCEL -0.0020 1.36 11.98 0.783 394.188 38.139
SEM APA -0.1387 1.39 8.40 1.035 390.497 38.373
SNC -0.4162 2.30 28.71 0.486 355.213 27.987
TD -0.2712 1.22 6.88 1.185 273.000 38.180
SAG -0.0836 0.93 7.33 0.742 222.000 31.298
NB -0.3576 1.42 9.12 0.714 163.885 31.481
IM PRESA -0.2168 2.20 26.45 0.588 136.800 28.044
COFINA -0.0761 1.22 8.64 1.072 116.500 40.170
IBERSOL -0.1118 1.21 4.21 1.351 70.000 34.125
PARAREDE -0.6867 3.90 20.22 3.326 25.018 30.327

Table 3B
DJ 30 Industrial Index Summary Statistics
This table presents daily descriptive statistics for the individual components of the U.S. DJ
30 Industrial index. The first two columns report the daily mean and standard deviation of firm-
specific stock returns measured as the difference in log prices from 11:30 EST to 11:30 EST. The
third column reports the daily standard deviation of unanticipated order flow estimated using
Hasbrouck (1991) VAR model. The last two columns report the mean percentage spread and market
capitalization. The percentage spread is the ask price minus the bid price divided by the average daily
price and market capitalization (size) is measured in billions of dollars.
M ean Return Std. Return Std. Order Flow M ean % Spread M ean Size
GE -0.267 2.87 106.44 0.065 309.804
M SFT -0.186 2.41 301.38 0.006 295.425
XOM -0.07 2.06 77.85 0.068 257.870
WM T -0.019 1.97 70.01 0.062 246.777
C -0.236 3.29 74.53 0.084 205.566
JNJ -0.02 2.38 80.53 0.068 176.798
INTC -0.431 3.63 351.17 0.011 156.148
IBM -0.326 2.74 91.77 0.063 145.065
KO 0.052 1.61 69.38 0.059 123.754
M RK -0.128 2.17 70.28 0.063 122.944
PG 0.058 1.42 75.28 0.047 113.952
MO -0.091 2.06 95.75 0.055 103.439
SBC -0.296 3.08 80.45 0.096 102.786
HD -0.242 2.71 88.96 0.077 88.688
JPM -0.361 3.76 77.95 0.108 57.642
T -0.227 3.15 79.33 0.129 57.396
HPQ -0.333 3.10 88.59 0.114 52.534
AXP -0.058 2.97 80.85 0.094 49.408
MMM 0.014 1.77 51.23 0.055 47.732
DD -0.036 2.37 65.76 0.078 44.984
DIS -0.186 2.88 101.11 0.113 40.387
BA -0.13 2.45 77.87 0.099 31.746
M CD -0.219 (2.04) (76.52) 0.095 31.122
UTX -0.068 (2.34) (60.1) 0.079 30.978
GM -0.173 (2.43) (78.49) 0.065 27.539
HON -0.252 (2.95) (68.07) 0.117 25.841
AA -0.267 (2.77) (65.57) 0.100 25.269
IP -0.066 (2.07) (71.9) 0.076 19.143
CAT -0.162 (2.23) (59.35) 0.078 16.530
EK 0.018 (2.40) (52.28) 0.101 9.127

Table 4A
Portuguese Macroeconomic News Announcements
Portuguese news announcements are partitioned into three groups: real activity, net-exports
and prices. Within each group, we list Portuguese news announcements in chronological order of
their release. The first column lists the total number of observations in our sample period from
January 4, 2002 to October 15, 2002, the second column lists the agency that releases the public
announcement, Instituto Nacional de Estatística (INE) and the European Central Bank (ECB), the
third column lists the pre-scheduled release time (Eastern Standard Time), the last two columns report
the mean and standard deviation of the surprise estimated using equation (13).
Footnotes:
1. Since we need historical data to estimate optimal forecasts we were unable to analyze the following
Portuguese announcements which started to be released in 2002 or later: Government Budget
Deficit, Total Construction Licenses, Consumer Confidence, Labor Costs, Manufacturing
Production and New Car Sales.
2. Total number of observations in our sample period from January 4, 2002 to October 15, 2002.
3. Release times vary in Portugal.
4. Day of the week announcement is usually released in.
5. Industrial Production was released at 10:00 EST on January 8, 2002, at 11:00 EST on March 13,
2002 and at 5:00 EST on April 5, 2002.
6. Industrial Sales was released at 10:00 EST on January 16, 2002.
7. Retail Sales was released at 10:00 EST on June 11, 2002.
8. Producer Price Index was released at 10:00 EST on January 25, 2002.
9. The ECB announces the benchmark interest rate at 7:45 EST, and the President of the ECB,
Willem F. Duisenberg, holds a press conference that ends anytime between 8:50 to 10:00 EST.
Table 4B
U.S. Macroeconomic News Announcements
Announcements Obs 1 Source 2 Time 3 Mean Std. Dev. DOW 4

Real Activity
Quarterly Announcements
1- GDP Advance 3 BEA 8:30 0.342 1.649 Varies
2- GDP Preliminary 3 BEA 8:30 0.497 1.386 Varies
3- GDP Final 3 BEA 8:30 1.108 0.508 Varies
Monthly Announcements
4- Nonfarm Payroll 10 BLS 8:30 -0.069 0.349 Friday 5
5- Retail Sales 9 BC 8:30 0.143 0.975 Varies
6- Industrial Production 8 FRB 9:15 0.074 1.010 Varies
7- Capacity Utilization 8 FRB 9:15 0.175 0.856 Varies
8- Personal Income 8 BEA 8:30 0.049 0.680 Varies
9- Consumer Credit 9 FRB 15:00 0.352 2.174 Varies
Consumption
10- New Home Sales 9 BC 10:00 0.262 1.038 Varies
11- Personal Consumption Expenditures 8 BEA 8:30 -0.146 0.669 Varies
Investment
12- Durable Goods Orders 9 BC 8:30 0.284 1.020 Varies
13- Factory Orders 10 BC 10:00 -0.147 0.726 Varies
14- Construction Spending 7 BC 10:00 -0.188 1.081 Varies
15- Business Inventories 10 BC 8:30 0.171 0.886 Varies
Government Purchases
16- Government Budget 9 FM S 14:00 -0.129 0.510 Varies
Net Exports
17- Trade Balance 9 BEA 8:30 -0.168 1.118 Varies
Prices
18- Producer Price Index 10 BLS 8:30 -0.546 1.076 Varies
19- Consumer Price Index 9 BLS 8:30 -0.377 0.768 Varies
Forward-Looking
20- Consumer Confidence Index 9 CB 10:00 0.083 1.062 Tuesday 6
21- NAPM Index 8 NAPM 10:00 0.113 1.330 Varies
22- Housing Starts 9 BC 8:30 0.103 0.572 Varies
23- Index of Leading Indicators 9 CB 8:30 0.302 1.126 Varies
Six-Week Announcements
24- Target Federal Funds Rate 6 FRB 14:15 -0.525 0.605 Tuesday 7
Weekly Announcements
25- Initial Unemployment Claims 40 ETA 8:30 0.139 1.045 Thursday
Notes to Table 4B
U.S. news announcements are partitioned into seven groups: real activity, each of the GDP
components (consumption, investment, government purchases and net exports), prices, and forward-
looking. Within each group, we list U.S. news announcements in chronological order of their release.
The last two columns list the mean and standard deviation of the surprise estimated using equation
(13).
Footnotes:
1. Total number of observations in our sample period from January 4, 2002 to October 15, 2002.
2. Bureau of Labor Statistics (BLS), Bureau of the Census (BC), Bureau of Economic Analysis
(BEA), Federal Reserve Board (FRB), National Association of Purchasing Managers (NAPM),
Conference Board (CB), Financial Management Office (FMO), Employment and Training
Administration (ETA).
3. Eastern Standard Time. Daylight savings time starts on the first Sunday of April and ends on the
last Sunday of October.
4. Day of the week announcement is usually released in.
5. Nonfarm Payroll is released the first Friday of the month.
4. The Conference Board’s Consumer Confidence Index is released the last Tuesday of the month.
5. The FOMC has eight scheduled meetings per year. Since March 22, 1994 these meetings are
usually scheduled on Tuesday, except for the first meeting of the year, which is a two-day meeting
starting on Tuesday and ending on Wednesday when the announcement is released to the public.
Table 5
Aggregate Portuguese and U.S. Macroeconomic News Announcements
This table reports estimates of the aggregate Portuguese and U.S. News. We aggregate
announcements into seven groups as shown in Table 4A and Table 4B: real activity, each of the GDP
components (i.e., consumption, investment, government purchases and net exports), prices, and
forward-looking announcements. The six-week and weekly announcements are not aggregated. For
example, U.S. real activity surprises are defined as the sum of GDP Advance, GDP Preliminary, GDP
Final, Nonfarm Payroll, Retail Sales, Industrial Production, Capacity Utilization, Personal Income
and Consumer Credit standardized surprises (according to equation (13)), while Portuguese real
activity surprises are defined as the sum of GDP, the Employment Report, Industrial Production and
Industrial Sales standardized surprises.
Table 6
Influence of U.S. Macroeconomic Announcements on U.S. Returns
This table reports coefficient estimates of the following equation:
15 J
us
rit = a + λsp S ip
eus eus t
+ λsp S p
ecb ecb t
+ ∑λ
pus = 7
spus S pus t + ∑λ Ω
j =0
j it − j (1 −
us
Dt ) +
J J J
∑λ
j =0
us
pj Ω it − j Dt + ∑β r
j =1
us
j it − j (1 −
us
Dt ) + ∑β
j =1
us us
pj rit − j Dt + ε it ,
where ritus is the daily stock return for i = 1,...,30 DJ 30 individual stocks, Ω∗it is the unanticipated
order flow (defined in Section 5), Dtus is an indicator function equal to one if a U.S. public
announcement (earnings or macroeconomic news) is released on date t , Sipeus t is the standardized
earnings news surprise for stock return i , S pecb t is the standardized ECB benchmark refinancing rate
news surprise, and S pus t is the aggregate standardized U.S. macroeconomic news surprise for
pus = 7,...,15 (defined in Section 4.2 and listed in Table 5). The t-statistics are estimated using the
GARCH(1,1)-X model (equation (17)) to correct for heteroskedasticity and autocorrelation. We mark
the coefficients and F-statistics with a “ ∗ ” , “ ∗∗ ” , or “ ∗∗∗ ” to indicate significance at the 10%, 5%,
or 1% level, respectively.
Coefficient t-statistic
Contemporaneous Order Flow
Non-Announcement,  0 0.0102 19.39 ∗∗∗
Announcement,  p0 0.0084 23.08 ∗∗∗
H 0 :  0 − p0  0 (F-statistic) 0.0018 8.59 ∗∗∗
Lagged Order Flow

Non-Announcement,  1 -0.0029 -5.11 ∗∗∗
Non-Announcement,  2 -0.0010 -1.61
Announcement,  p1 -0.0016 -3.86 ∗∗∗
Announcement,  p2 -0.0011 -2.78 ∗∗∗
European Announcement,  sp ecb
ECB benchmark refinancing rate -0.1703 -3.44 ∗∗∗
U.S. Announcements,  sp eus ,  sp us
Earnings 0.5595 4.58 ∗∗∗
Real Activity 0.1680 3.38 ∗∗∗
Consumption 0.8889 7.85 ∗∗∗
Investment 0.1258 1.77 ∗
Government Purchases 0.0050 0.02
Net Exports 0.3678 3.46 ∗∗∗
Prices -0.1570 -1.86 ∗
Forward-Looking 0.2351 3.20 ∗∗∗
Target Federal Funds Rate -0.7300 -4.21 ∗∗∗
Initial Unemployment Claims -0.1010 -2.26 ∗∗
13.42 ∗∗∗
por por
H 0 :  s7 . . .   s15  0 (F-statistic)
Adjusted-R 2 0.1665
Table 7
Influence of News Announcements on US-Portuguese Stock Market Comovement
Portuguese Portuguese U.S.

Macroeconomic Ann. Earnings Ann. Macroeconomic Ann.
por
Non-Announcement,  0 0.0182 ∗∗∗ 0.0186 ∗∗∗ 0.0193 ∗∗∗ 0.0141 ∗∗∗ 0.0235 ∗∗∗ 0.0165 ∗∗∗
por
Announcement,  p0 0.0240 ∗∗∗ 0.0244 ∗∗∗ 0.0097 0.0101 ∗ 0.0187 ∗∗∗ 0.0131 ∗∗∗
por por
H0 :  0 −  p0  0, (F-statistic) -0.0058 ∗∗∗ -0.0058 ∗∗∗ 0.0096 0.0039 0.0048 ∗∗ 0.0034 ∗∗∗
Lagged Order Flow

por
Non-Announcement,  1 -0.0038 ∗∗∗
-0.0048 ∗∗∗
-0.0036 ∗∗∗ -0.0033 ∗∗∗ -0.0060 ∗∗ -0.0045 ∗∗∗
por
Non-Announcement,  2 -0.0027 ∗∗ -0.0020 ∗ -0.0023 ∗∗ -0.0015 ∗∗ -0.0050 ∗∗ -0.0016
por
Announcement,  p1 -0.0051 ∗ -0.0044 -0.0025 -0.0034 -0.0036 ∗∗ -0.0033 ∗∗∗
por
Announcement,  p2 -0.0012 -0.0001 -0.0050 ∗ -0.0022 ∗ -0.0013 -0.0013 ∗∗
U.S. Value Weighted Return

por
Non-Announcement,  0 0.2515 ∗∗∗ 0.2289 ∗∗∗ 0.2278 ∗∗∗
por
Announcement,  0p 0.0867 ∗∗∗
0.2367 ∗
0.2370 ∗∗∗
por por
H0 :  0 −  0p  0, (F-statistic) 0.1648 ∗∗∗ -0.0078 -0.0092
por por
Portuguese Announcements,  sp epor ,  sp por
Earnings 0.8201 ∗∗∗ 0.7349 ∗∗∗ 0.8411 ∗∗∗ 0.7701 ∗∗∗ 0.7741 ∗∗∗ 0.7558 ∗∗∗
∗∗∗ ∗∗∗
Real Activity 0.0058 0.1546 0.0217 0.2623 0.0215 0.2443 ∗∗∗
Net Exports 0.5737 ∗∗∗ 0.4920 ∗∗∗ 0.5768 ∗∗∗ 0.4833 ∗∗∗ 0.6150 ∗∗∗ 0.4489 ∗∗∗
Prices -0.2975 ∗∗∗ -0.2289 ∗∗∗ -0.3013 ∗∗∗ -0.1515 ∗∗∗ -0.3108 ∗∗∗ -0.1417 ∗∗∗
10.53 ∗∗∗ 12.58 ∗∗∗ 11.32 ∗∗∗ 13.32 ∗∗∗ 12.82 ∗∗∗ 11.92 ∗∗∗
por por por
H0:  s2   s3   s4  0 (F-statistic)
por
European Announcement,  sp ecb
∗∗
ECB benchmark refinancing rate -0.0910 -0.0163 -0.0685 ∗∗ -0.0560 -0.0765 ∗∗ -0.0539
por
U.S. Announcements,  sp us
Real Activity 0.0399 -0.0292 0.0399 -0.0225 0.0411 -0.0219
∗ ∗∗ ∗ ∗ ∗
Consumption 0.1967 -0.2320 0.1935 -0.1863 0.1940 -0.1827 ∗
Investment 0.1429 ∗ 0.0188 0.1372 ∗ 0.0232 0.1402 ∗ 0.0278
Government Purchases 0.3094 0.0499 0.2902 0.0457 0.2928 0.0549
Net Exports 0.0431 -0.1774 0.0516 -0.1355 0.0402 -0.1420
Prices -0.0578 0.1256 -0.0385 0.0948 -0.0369 0.0974

∗∗ ∗ ∗∗ ∗∗ ∗∗
Forward-Looking 0.1615 0.1132 0.1655 0.1331 0.1647 0.1317 ∗∗
Target Federal Funds Rate -1.3058 ∗∗∗ -0.8551 ∗∗∗ -1.3209 ∗∗∗ -0.8765 ∗∗∗ -1.3085 ∗∗∗ -0.8664 ∗∗∗
Initial Unemployment Claims -0.1216 ∗∗ -0.0300 -0.1248 ∗∗ -0.0429 -0.1177 ∗∗ -0.0383

∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗
3.07 ∗∗∗
por por
H0:  s7 . . .   s15  0 (F-statistic) 6.93 3.30 6.98 3.12 6.86
Adjusted-R2 0.1652 0.2270 0.1670 0.2335 0.1621 0.2328

Notes to Table 7
4 15 J
por
rit =a
por
+ λsp
por
epor
Sip
epor t
+ λspporecb S pecb t + ∑λ
p por = 2
por
sp por Sp
por t
+ ∑λ
pus = 7
por
spus S pus t + ∑λ
j =0
por ∗
j Ω it − j (1 − Dt ) +
J J J J
∑λ
j =0
por ∗
pj Ω it − j Dt + ∑β
j =0
por us
j rt − j (1 − Dt ) + ∑β
j =0
por us
pj rt − j Dt + ∑β
j =1
por por
porj rit − j + ε it
por
,
where ritpor is the daily return for stock i = 1,..., 20 of the PSI-20 Index, Ω∗it is the unanticipated
order flow (defined in Section 5), Dt = Dt { por

, Dtepor , Dtus } is an indicator function equal to one if a
Portuguese macroeconomic announcement, Portuguese earnings announcement, and a U.S.
macroeconomic annoucement is released, respectively, on date t , S pecb t is the standardized ECB
benchmark refinancing rate news surprise, Sipepor t is the Portuguese standardized earnings news
surprise for stock i , and S pus t is the aggregate standardized U.S. macroeconomic news surprise for
pus = 7,...,15 and S p por t is the aggregate Portuguese standardized macroeconomic news surprise for
p por = 2,3, 4 (both of which are defined in Section 4.2 and listed in Table 5). The coefficients in
column one and two correspond to the equation estimates when Dt = Dtpor , the coefficients in
column three and four correspond to the equation estimates when Dt = Dtepor , and the coefficients in
the last two columns correspond to the equation estimates when Dt = Dtus . The t-statistics are
estimated using the GARCH(1,1)-X model (equation (19)) to correct for heteroskedasticity and
autocorrelation. We mark the coefficients and F-statistics with a “ ∗ ”, “ ∗∗ ”, or “ ∗∗∗ ” to indicate
significance at the 10%, 5%, or 1% level, respectively.

Table 8. Influence of U.S. Disaggregate Macroeconomic Announcements on U.S. Returns
Coefficient t-statistic
Non-Announcement,  0 0.0105 20.65 ∗∗∗
Announcement,  p0 0.0084 22.25 ∗∗∗
H0 :  0 −  p0  0 (F-statistic) 0.0021 11.18 ∗∗∗
European Announcement
ECB benchmark refinancing rate -0.1613 -2.92 ∗∗∗
U.S. Announcements
Earnings 0.4906 3.89 ∗∗∗
GDP Advance -0.5452 -3.71 ∗∗∗
GDP Preliminary -0.0370 -0.25
GDP Final 0.1142 0.47
Nonfarm Payroll 0.1573 0.32
Retail Sales -0.0227 -0.18
Industrial Production -0.1628 -0.48
Capacity Utilization 1.1817 3.67 ∗∗∗
Personal Income 1.2411 4.28 ∗∗∗
Consumer Credit 0.1778 2.01 ∗∗
New Home Sales -0.6252 -4.8 ∗∗∗
Personal Consumption Expenditures 1.9291 10.7 ∗∗∗
Durable Goods Orders 0.3076 2.51 ∗∗
Factory Orders 0.0839 0.64
Construction Spending 0.2049 0.57
Business Inventories -0.1039 -0.55
Government Budget 0.1917 0.88
Trade Balance 0.4005 3.61 ∗∗∗
Producer Price Index 0.0090 0.06
Consumer Price Index -0.1917 -1.33
Consumer Confidence Index 0.0064 0.06
NAPM Index -0.5410 -1.88 ∗
Housing Starts -1.1441 -6.17 ∗∗∗
Index of Leading Indicators 0.4112 3.48 ∗∗∗
Initial Unemployment Claims -0.1128 -2.56 ∗∗
Target Federal Funds Rate -0.9267 -4.73 ∗∗∗

2
Adjusted-R 0.1763
Notes to Table 8
25 J
us
rit = a + λsp Sip
eus eus t
+ λsp S p
ecb ecb t
+ ∑λ
pus =1
spus Sp
us t
+ ∑λ Ω
j =0
j it − j (1 −
us
Dt ) +
J J J
∑λ
j =0
us
pj Ω it − j Dt + ∑β r
j =1
us
j it − j (1 −
us
Dt ) + ∑β
j =1
us us
pj rit − j Dt + ε it ,
where ritus is the daily stock return for i = 1,...,30 DJ 30 individual stocks, Ω∗it is the unanticipated
order flow (defined in Section 5), Dtus is an indicator function equal to one if a U.S. public
announcement (earnings or macroeconomic news) is released on date t , Sipeus t is the standardized
earnings news surprise for stock return i , S pecb t is the standardized ECB benchmark refinancing rate
news surprise, and S pus t is the disaggregate standardized U.S. macroeconomic news surprise for
pus = 1,..., 25 (defined in Section 4.2 and listed in Table 4B). The t-statistics are estimated using the
the coefficients and F-statistics with a “ ∗ ”, “ ∗∗ ”, or “ ∗∗∗ ” to indicate significance at the 10%, 5%, or
1% level, respectively.
Table 9
Influence of Statistically Significant Disaggregate Announcements on US-Portuguese stock
Market Comovement
Portuguese Portuguese U.S.

Macroeconomic Ann. Earnings Ann. Macroeconomic Ann.
por
Non-Announcement,  0 0.0151 ∗∗∗
0.0139 ∗∗∗
0.0150 ∗∗∗ 0.0144 ∗∗∗ 0.0144 ∗∗∗ 0.0137 ∗∗∗
por
Announcement,  p0 0.0148 ∗∗∗ 0.0156 ∗∗∗ 0.0112 ∗∗ 0.0112 ∗∗ 0.0153 ∗∗∗ 0.0150 ∗∗∗
por por
H0 :  0 −  p0  0, (F-statistic) 0.0003 -0.0017 ∗∗ 0.0038 0.0032 -0.0009 ∗ -0.0013 ∗∗

por
Non-Announcement,  0 0.2287 ∗∗∗
0.2065 ∗∗∗ 0.2151 ∗∗∗
por
Announcement,  0p 0.0889 ∗∗ 0.2077 0.1962 ∗∗∗
por por
H0 :  0 −  0p  0, (F-statistic) 0.1398 ∗∗∗ -0.0012 0.0189
Portuguese Announcements
∗∗∗ ∗∗∗
Earnings 0.7721 0.7461 0.7783 ∗∗∗ 0.7512 ∗∗∗ 0.7742 ∗∗∗ 0.7530 ∗∗∗
GDP 0.0090 ∗∗∗ 0.0101 ∗∗∗ 0.0087 ∗∗∗ 0.0102 ∗∗∗ 0.0093 ∗∗∗ 0.0109 ∗∗∗
Industrial Sales 0.1964 ∗ -0.01823 0.2052 ∗∗ -0.1897 ∗∗∗ 0.1859 ∗ -0.1788 ∗∗
Trade Balance 0.4853 ∗∗∗ 0.4724 ∗∗∗ 0.5017 ∗∗∗ 0.4377 ∗∗∗ 0.4956 ∗∗∗ 0.4297 ∗∗∗
Consumer Price Index -0.5303 ∗∗∗ -0.3630 ∗∗∗ -0.5301 ∗∗∗ -0.2791 ∗∗ -0.5284 ∗∗∗ -0.2727 ∗∗
Producer Price Index -0.2089 ∗∗∗ -0.1400 ∗∗ -0.2125 ∗∗∗ -0.1100 -0.2147 ∗∗∗ -0.1092
European Announcement
∗∗ ∗
ECB benchmark refinancing rate -0.1048 -0.0662 -0.1062 ∗∗∗ -0.0733 ∗∗ -0.1045 ∗∗∗ -0.0714 ∗∗
U.S. Announcements
Personal Income 0.9951 ∗∗∗ 0.3847 ∗ 0.9938 ∗∗∗ 0.4684 ∗∗ 0.9854 ∗∗∗ 0.4614 ∗∗
Consumer Credit -0.1270 ∗∗ -0.2304 ∗∗∗ -0.1226 ∗∗ -0.2020 ∗∗∗ -0.1217 ∗∗ -0.1995 ∗∗∗
Personal Consumption Expenditures 0.9212 ∗∗∗ 0.0817 0.9129 ∗∗∗ 0.0867 0.9158 ∗∗∗ 0.0941
Business Inventories -0.7147 ∗∗∗ -0.3485 ∗∗ -0.7179 ∗∗∗ -0.3221 ∗∗ -0.7115 ∗∗∗ -0.3178 ∗∗
Housing Starts -0.5527 ∗∗ -0.2369 -0.5487 ∗∗ -0.2370 -0.5331 ∗∗ -0.2274

∗∗ ∗∗ ∗∗
Indexof Leading Indicators 0.2867 0.1698 0.2896 0.1830 0.2873 0.1817
∗∗ ∗∗ ∗∗
Initial Unemployment Claims -0.1259 -0.0211 -0.1289 -0.0256 -0.1284 -0.0269
Target Federal Funds Rate -1.3169 ∗∗∗ -0.8166 ∗∗∗ -1.3280 ∗∗∗ -0.8282 ∗∗∗ -1.3267 ∗∗∗ -0.8353 ∗∗∗
Adjusted-R2 0.1887 0.2393 0.1857 0.2355 0.1861 0.2367

Notes to Table 9

Ppor Pus J
por
rit =a
por
+ λsp
por
epor
S ip
epor t
+ λspporecb S pecb t + ∑λ
p por =1
por
sp por Sp
por t
+ ∑λ
pus =1
por
spus S pus t + ∑λ
j =0
por ∗
j Ω it − j (1 − Dt ) +
J J J J
∑λ
j =0
por ∗
pj Ω it − j Dt + ∑β
j =0
por us
j rt − j (1 − Dt ) + ∑β
j =0
por us
pj rt − j Dt + ∑β j =1
por por
por
,
where ritpor is the daily return for stock i = 1,..., 20 of the PSI-20 Index, Ω∗it is the unanticipated
order flow (defined in Section 5), Dt = Dt { por

, Dtepor , Dtus } is an indicator function equal to one if a
statistically significant Portuguese macroeconomic announcement, Portuguese earnings
announcement, and a statistically significant U.S. macroeconomic annoucement is released,
respectively, on date t , S pecb t is the standardized ECB benchmark refinancing rate news surprise,
Sipepor t is the Portuguese standardized earnings news surprise for stock i , and S pus t is the
disaggregate standardized U.S. macroeconomic news surprise for statistically significant news
pus = 1,..., Pus and S p por t is the disaggregate Portuguese standardized macroeconomic news
surprise for statistically significant news p por = 1,..., Ppor . The coefficients in column one and two
correspond to the equation estimates when Dt = Dtpor , the coefficients in column three and four
correspond to the equation estimates when Dt = Dtepor , and the coefficients in the last two columns
correspond to the equation estimates when Dt = Dtus . The t-statistics are estimated using the
the coefficients and F-statistics with a “ ∗ ”, “ ∗∗ ”, or “ ∗∗∗ ” to indicate significance at the 10%, 5%, or
1% level, respectively.
Table 10
Day of the Week Effects and Large U.S. Returns
Coefficient t-statistic Coefficient t-statistic

por
0 0.0136 43.45 ∗∗∗ 0.0134 40.90 ∗∗∗
Lagged Order Flow

por
1 -0.0032 -6.70 ∗∗∗ -0.0030 -5.83 ∗∗∗
por
2 -0.0016 -3.24 ∗∗∗
-0.0014 -2.58 ∗∗∗
por
3 -0.0014 -2.62 ∗∗∗ -0.0016 -2.65 ∗∗∗

por
0 0.2200 10.58 ∗∗∗ 0.2551 10.44 ∗∗∗
por
 0LARGE 0.0251 0.99
por
Monday,  0 0.0114 0.30
por
Tuesday,  0 -0.0413 -1.11
por
Wednesday,  0 -0.0170 -0.37
por
Thursday,  0 -0.0647 -1.72 ∗
por por
Portuguese Announcements,  e ,  sp
Earnings 0.7457 5.92 ∗∗∗ 0.7920 7.06 ∗∗∗
Real Activity 0.2277 4.05 ∗∗∗ 0.2156 4.17 ∗∗∗
Net Exports 0.4785 3.87 ∗∗∗ 0.4859 2.63 ∗∗∗
Prices -0.1799 -3.26 ∗∗∗ -0.1258 -2.23 ∗∗

∗∗∗
F-statistic 10.12 13.43 ∗∗∗
European Announcement,  us
ecb
ECB benchmark refinancing rate -0.0578 -1.72 ∗ -0.0503 -1.37

por
U.S. Announcements,  sp
Real Activity -0.0117 -0.3 -0.0214 -0.54
∗
Consumption -0.1953 -1.81 -0.1780 -1.56
Investment 0.0334 0.47 0.0307 0.43
Government Purchases 0.1057 0.41 0.0500 0.19
Net Exports -0.1132 -0.97 -0.1520 -1.26
Prices 0.0803 0.93 0.0869 0.99

∗∗
Forward-Looking 0.1383 2.07 0.1254 1.87 ∗
Target Federal Funds Rate -0.8698 -4.26 ∗∗∗ -0.8777 -4.24 ∗∗∗
Initial Unemployment Claims -0.0350 -0.62 -0.0409 -0.71

∗∗∗
F-statistic 2.62 2.60 ∗∗∗
Adjusted-R2 0.2334 0.2321

Notes to Table 10
Ppor Pus J
por
rit =a
por
+ λsp
por
epor
S ip
epor t
+ λspporecb S pecb t + ∑
p por =1
λsp por S p por t +
por
∑
pus =1
λspus S pus t +
por
∑λ
j =0
por ∗
j Ω it − j (1 − Dt ) +
J J J J
∑
j =0
por ∗
λ pj Ω it − j Dt + ∑
j =0
por us
β j rt − j (1 − Dt ) + ∑
j =0
por us
β pj rt − j Dt + ∑β j =1
por por
por
,
where ritpor is the daily return for stock i = 1,..., 20 of the PSI-20 Index, Ω∗it lim is the
δ x →0
unanticipated order flow (defined in Section 5),
Dt = D { t
LARGE
{
, D t
Monday
,D Tuesday
t ,D Wednesday
t ,D Thrusday
t }} where DtLARGE is an indicator function
equal to one if the U.S. value weighted DJ 30 index experiences a top 10% jump on day t , and
{D t
Monday
, DtTuesday , DtWednesday , DtThrusday } are day-of-the week indicator functions, S pecb t is the
standardized ECB benchmark refinancing rate news surprise, Sipepor t is the Portuguese standardized
earnings news surprise for stock i , and S pus t is the aggregate standardized U.S. macroeconomic
news surprise pus = 2,..., 7 and S p por t is the aggregate Portuguese standardized macroeconomic
news surprise p por = 2,3, 4 . The t-statistics are estimated using the GARCH(1,1)-X model (equation
(19)) to correct for heteroskedasticity and autocorrelation. We mark the coefficients and F-statistics
with a “ ∗ ”, “ ∗∗ ”, or “ ∗∗∗ ” to indicate significance at the 10%, 5%, or 1% level, respectively
Figure 1
Timeline of Events
Foreign Economy Local Economy
Market makers observe

∗
Public news U arrives, t aggregate order flow Public news U t and Pt ∗ arrives, Market makers observe
t informed traders learn s ∗
t ω =∑ x +z
∗
t
n ∗
i =1 it
∗
t informed traders learn st ωt = ∑ in=1xit + zt t +1
↔ ↔
Orders are placed: xit∗ , zt
∗
Price is chosen Pt ∗ . Orders are placed: xit , zt Price is chosen Pt .
Figure 2A Portugues Response to 7:45 ECB announcement, 8:30 and 9:15 EST U.S.
Macroeconomic Announcements
We plot λsp
por
(equation (15)), the cumulative Portuguese response to 8:30 and 9:15 EST
U.S. macroeconomic announcements, and the 7:45 EST ECB announcement. The left hand side of the
x-axis in each plot coincides with the time indicated in the title of the plot. Each tick advances time
by 5 minutes. For example, in the top left hand corner plot the first tick indicates 7:50 EST and the
return is measured from just before 7:45 EST to 7:50 EST. The second tick indicates 7:55 EST and
the return is measured from just before 7:45 EST to 7:55 EST. The vertical line corresponds to 9:30
EST when the U.S. stock market opens. The last tick captures the response from the time indicated in
the title of the plot to 11:30 EST.
ECB Benc hm ark Rate 7:45 Real Ac tivity 8:30 Consum ption 8:30
0.4 0.4 0.4
0.2 0.2 0.2

Response
Response
Response
0.0 0.0 0.0
-0.2 -0.2 -0.2
-0.4 -0.4 -0.4

5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 5 10 15 20 25 30 35
Time Time Time
Inve stm ent 8:30 Net Exports 8:30 Prices 8:30
0.4 0.4 0.4
0.2 0.2 0.2

Response
Response
Response
0.0 0.0 0.0
-0.2 -0.2 -0.2
-0.4 -0.4 -0.4

5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25 30 35
Time Time Time
Forward-Looking 8:30 Initial Une m ploy m ent Claim s 8:30 Real Ac tivity 9:15
0.4 0.4 0.4
0.2 0.2 0.2

Response
Response
Response
0.0 0.0 0.0
-0.2 -0.2 -0.2
-0.4 -0.4 -0.4

5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25
Time Time Time

Figure 2B. U.S. Response to 7:45 ECB announcement, 8:30 and 9:15 EST U.S.
Macroeconomic Announcements
We plot λsp (equation (15)), the cumulative U.S. response to 8:30 and 9:15 EST U.S.
macroeconomic announcements, and the 7:45 EST ECB announcement. Since the U.S. stock market
is not yet open, the first tick in all panels on the x-axis captures the U.S. stock market response from
the previous day’s close to 9:35 EST, the second tick captures the cumulative response from the
previous day’s close to 9:40 EST, and so on. The last tick captures the response from the previous
day’s close to 11:30 EST.
ECB Benchmark Rate 7:45 Real Activity 8:30 Consumption 8:30

0.8 0.8 2.0
0.6 0.6
1.5
0.4 0.4
0.2 0.2 1.0
Response
Response
Response
0.0 0.0 0.5
-0.2 -0.2
-0.4 -0.4 0.0
-0.6 -0.6 -0.5
-0.8 -0.8
5 10 15 20 5 10 15 20 5 10 15 20
Time Tim e Time
Investment 8:30 Net Exports 8:30 Prices 8:30

0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
Response
Response
Response
0.0 0.0 0.0

-0.2 -0.2 -0.2
-0.4 -0.4 -0.4
-0.6 -0.6 -0.6
-0.8 -0.8 -0.8
5 10 15 20 5 10 15 20 5 10 15 20
Time Tim e Time
Forward-Looking 8:30 Initial Unemployment Claims 8:30 Real Activity 9:15

0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
Response
Response
Response
0.0 0.0 0.0

-0.2 -0.2 -0.2
-0.4 -0.4 -0.4
-0.6 -0.6 -0.6
-0.8 -0.8 -0.8
5 10 15 20 5 10 15 20 5 10 15 20
Time Tim e Time

Figure 3A. Portuguese Response to 10:00 EST U.S. Macroeconomic Announcements
We plot λsp
por
(equation (15)), the cumulative Portuguese response to 10:00 EST U.S.
macroeconomic announcements. The first tick in all panels on the x-axis captures the U.S. stock
market response from 10:00 to 10:05 EST, the second tick captures the cumulative response from
10:00 to 10:10 EST, and so on. The last tick captures the response from 10:00 to 11:30 EST.
Consumption 10:00
0.5
0.4
0.3
Response
0.2
0.1
0.0
-0.1
-0.2
2 4 6 8 10 12 14 16 18
T ime
Investment 10:00
0.5
0.4
0.3
Response
0.2
0.1
0.0
-0.1
-0.2
2 4 6 8 10 12 14 16 18
T ime
Forward-Looking 10:00
0.15
0.10
Response
0.05
0.00
-0.05
2 4 6 8 10 12 14 16 18
T ime
Figure 3B. U.S. Response to 10:00 EST U.S. Macroeconomic Announcements
We plot λsp (equation (15)), the cumulative U.S. response to 10:00 EST U.S.
macroeconomic announcements. The first tick in all panels on the x-axis captures the U.S. stock
market response from 10:00 to 10:05 EST, the second tick captures the cumulative response from
10:00 to 10:10 EST, and so on. The last tick captures the response from 10:00 to 11:30 EST.
Consumption 10:00
0.5
0.4
0.3
Response
0.2
0.1
0.0
-0.1
-0.2
2 4 6 8 10 12 14 16 18
Time
Investment 10:00
0.5
0.4
0.3
Response
0.2
0.1
0.0
-0.1
-0.2
2 4 6 8 10 12 14 16 18
Time
Forward-Looking 10:00
0.5
0.4
0.3
Response
0.2
0.1
0.0
-0.1
-0.2
2 4 6 8 10 12 14 16 18
Time
Figure 4A. Portuguese Response to 14:00, 14:15 and 15:00 EST U.S. Macroeconomic
Announcements
We plot λsp
por
(equation (15)), the cumulative Portuguese response to 14:00, 14:05 and
15:00 EST U.S. macroeconomic announcements. Since the Portuguese stock market is not open, the
first tick in all panels on the x-axis captures the Portuguese stock market response from this day’s
close to 3:35 EST the next day, the second tick captures the cumulative response from this day’s close
to 3:40 EST the next day, and so on. The last tick captures the response from this day’s close to 11:30
EST.
Government Budget 14:00

1.0
0.5
Response
0.0
-0.5
-1.0
-1.5
10 20 30 40 50 60 70 80 90
Time
Fed Funds Target Rate 14:15

0.2
0.0
-0.2
-0.4
Response
-0.6
-0.8
-1.0
-1.2
-1.4
10 20 30 40 50 60 70 80 90
Time
Consumer Credit 15:00

0.4
0.3
0.2
Response
0.1
0.0
-0.1
-0.2
10 20 30 40 50 60 70 80 90
Time
Figure 4B. U.S. Response to 14:00, 14:15 and 15:00 EST U.S. Macroeconomic
Announcements
We plot λsp (equation (15)), the cumulative U.S. response to 14:00, 14:15 and 15:00 EST
U.S. macroeconomic announcements. The left hand side of the x-axis in each plot coincides with the
time indicated in the title of the plot. Each tick advances time by 5 minutes. For example, in the top
plot the first tick indicates 14:00 EST and the return is measured from just before 14:00 EST to 14:05
EST. The second tick indicates 14:10 EST and the return is measured from just before 14:00 EST to
14:10 EST, and so on. The last tick captures the response from the time indicated in the title of the
plot to 11:30 EST.
Government Budget 14:00

1.0
0.5
Response
0.0
-0.5
-1.0
-1.5
5 10 15 20 25 30 35 40 45
Time
Fed Funds Target Rate 14:15

1
0
Response
-1
-2
-3
-4
5 10 15 20 25 30 35 40 45
Time
Consumer Credit 15:00

0.4
0.3
0.2
Response
0.1
0.0
-0.1
-0.2
5 10 15 20 25 30 35
Time

Desenvolvimento Económico Português No Espaço Europeu - 2006

Enviado por

Dados do documento

Título original

Direitos autorais

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Desenvolvimento Económico Português No Espaço Europeu - 2006

Enviado por

Direitos autorais:

Conferência do Banco de Portugal

Economic Research Department

Number of copies printed

Legal Deposit no. 236983/06

Com a realização desta 3ª Conferência sobre o «Desenvolvimento Económico

economia portuguesa pode atrair projectos de investimento que integrem estas

With this third Conference on “Portuguese Economic Development in the

Superior de Economia e Gestão), Isabel Horta Correia (Banco de Portugal and

Nota de Apresentação / Foreword

Conferência do Banco de Portugal sobre “Desenvolvimento Económico Português no

Third Conference on “Portuguese Economic Developments in the European Context:

Impostos e Decisão de Trabalho: Portugal, Europa e Estados Unidos ............................... 41

Budget Setting Autonomy and Political Accountability..................................................... 65

The Internal Rate of return to On-the-Job Training............................................................125

Will the East Follow Portugal?...........................................................................................165

Asymmetric Information in the Stock Market: Economic News and Co-movement

Third Conference on "Portuguese Economic Development in

José A. Ferreira Machado

A nation’s standard of living is determined by the productivity of its workers that

given state of technological knowledge, this productivity depends essentially on the

the services sector and the efficiency costs of taxes

productivity in manufacturing since productivity growth in agriculture and services lagged

The previous analysis is probably too aggregate to draw specific policy

recommendations on how to increase productivity. Nevertheless, the focus on the service

A significant part of the differences observed in workers’ productivity (measured

by their wages) is explained by diverse levels of education in the workforce (Pedro

Carneiro’s contribution). Educational policy is thus critical to increase productivity.

performance is the socio-economic family background and, consequently, traditional

Cátia Batista remarks the aggregate technology in Portugal is characterized by the

will Portugal be to new capital.

Improving the productivity of workers and firms is certainly a way of improving

efficiency in the sense of adequate reaction to fundamentals.

Governments’ interventions have also major effects on the incentives to an

market into less productive activities and, thus, lower productivity.

As a very personal summation, four major lessons may, in my view, be drawn

from these communications to the Bank of Portugal 3rd Conference:

▪ Services are critical to continued convergence of productivity and standards of

▪ It is not enough to invest additional resources in education in order to improve

▪ Sound competition policies may provide a significant boost to aggregate

In this paper we document the structural transformation - the reallocation

Keywords: productivity, structural transformation, agriculture, manufacturing, services.

JEL Classification: O1,O4.

We address the long-run economic performance of Portugal and its process of

2. Transformation and Long-run Performance

In this section we document the economic performance and the process of

2.1. The Behavior of Aggregate Labor Productivity

In this expression, L represents the number of workers employed, h represents average

Equation (1) can be re-written in terms of output per worker Y / L as:

2.2. The Process of Structural Transformation

referred to as the structural transformation of the economy in the development literature.4

Portugal has experienced a process of structural transformation that is also broadly

We develop a simple model of the structural transformation of an economy where

3.1. Description of Economic Environment

Production At each date there are three goods produced: agriculture ( a ),

Yi = Ai Li , i ∈ {a, m, s}, (3)

where Yi is output in sector i , Li is labor allocated to production in sector i , and Ai is a

max { pi Ai Li − wLi } , (4)

where Li is the demand of labor in sector i .

Population The economy is populated by an infinitely-lived representative household of

one.) The household has preferences over consumption goods as follows:

where ca , t is the consumption of agricultural goods at date t and ct is the consumption of

u (ct , ca , t ) = log(ct ) + V (ca , t ),

where V (ca , t ) is non-homothetic, i.e., there is a subsistence level of agriculture a below

max log ⎡⎣bcmρ + (1 − b)(cs + s ) ρ ⎤⎦ + V (ca ) ,

In what follows we normalize the wage rate to one.

3.2. Equilibrium of the Model