Available online at www.sciencedirect.

com

Marine Pollution Bulletin 55 (2007) 591–602

www.elsevier.com/locate/marpolbul

An application of contingent valuation and decision tree analysis

to water quality improvements
Jonathan P. Atkins a, Daryl Burdon b,*
, James H. Allen b

a
Centre for Economic Policy, The Business School, University of Hull, Hull HU6 7RX, UK
b
Institute of Estuarine and Coastal Studies (IECS), University of Hull, Hull HU6 7RX, UK

Abstract

This paper applies contingent valuation and decision tree analysis to investigate public preferences for water quality improvements,
and in particular reduced eutrophication. Such preferences are important given that the development of EU water quality legislation is
imposing significant costs on European economies. Results are reported of a survey undertaken of residents of Århus County, Denmark
for water quality improvements in the Randers Fjord. Results demonstrate strong public support for reduced eutrophication and identify
key determinants of such support.
Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Contingent valuation; Decision tree analysis; Water quality; Eutrophication; Randers Fjord

1. Introduction
The European Union’s Water Framework Directive
(2000/60/EC) has provided the legislative framework for
water quality improvements within its region, seeking to
restore all water bodies to ‘good ecological status’. In
embracing the Urban Waste Treatment Directive (91/
271/EEC) and the Nitrates Directive (91/676/EEC), a prin-
ciple concern of this legislation has been the impact of
eutrophication on water quality. Eutrophication is the
accelerated production of organic matter in a water body
and is linked to an increase in nutrients, in particular nitro-
gen and phosphorus, being discharged into aquatic ecosys-
tems (Bricker et al., 1999). It has been recognised by the
scientific community as constituting a serious threat to
the ecology of and services provided by freshwater systems
since the mid-1900s and to estuarine and coastal systems
since the 1980s (Elliott and de Jonge, 2002). Furthermore,
there are plans to extend the scope of legislation within the
European Union (see Apitz et al., 2006) with the introduc-
tion of the EC Marine Strategy (EC, 2005a) and the pro-
posed Marine Strategy Directive (EC, 2005b), which seek
the restoration of marine ecosystems and require ‘good
environmental status’. Despite these developments, pub-
lished evidence on public preferences for such measures is
scarce, and this is despite the high costs borne by society
in the form of ‘costs of treatment, mitigation and/or com-
pensation’ (Elliott and de Jonge, 2002) in implementing
and/or meeting the requirements of this legislation.
In this paper, an evaluation is undertaken of public pref-
erences for reduced eutrophication, particularly with
respect to its implications for water quality improvements.
The focus of the evaluation is the Randers Fjord, Den-
mark,1 and the preferences of residents of the neighbouring
Århus County for a less eutrophied Fjord, where costs of
implementing the Urban Waste Water Treatment Directive
alone in the catchment have been estimated to be in
the region of €300 million (Atkins and Burdon, 2006).
While the approach adopted here follows the widely used

*
Corresponding author. Fax: +44 (0) 1482 464130.
E-mail address: d.burdon@hull.ac.uk (D. Burdon).
1
The Randers Fjord was one of four chosen study sites for the EU-
funded project EUROTROPH, which investigated nutrient cycling and
the trophic status of European coastal ecosystems.

0025-326X/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.marpolbul.2007.09.018
592 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
contingent valuation methodology (CVM), the findings
update and extend preliminary results in Atkins and Bur-
don (2006), further contribute to a small empirical literature
which reports investigations into eutrophication at other
locations (Söderqvist, 1996; Markowska and Źylicz, 1999;
Gren et al., 2000; Bateman et al., 2006; Frykblom et al.,
unpublished) and may constitute the first survey evidence
of this kind for an estuarine system. In addition, this paper
advances the use of decision tree analysis as a basis for the
investigation of the CVM survey respondents’ behaviour.
Classification and regression (decision) tree analysis is a
recursive partitioning technique which predicts the mem-
bership in the classes of a categorical dependent variable
(or predicts values of a continuous dependent variable)
from one or more explanatory variables (Breiman et al.,
1984). This technique here is used to examine relationships
between both CVM voting and bid decisions, and identified
characteristics of individual survey respondents.
The remainder of the paper is structured as follows. In
Section 2, the theory and empirical literature on goods
and services provided by estuarine and coastal systems is
reviewed, with a particular emphasis on the valuation of
public preferences and eutrophication. In Section 3, eutro-
phication in the Randers Fjord is described and the CVM
survey and findings presented. In Section 4, decision tree
analysis is used to evaluate the CVM bidding decisions.
Conclusions are drawn in Section 5.
2. Background
There is a large theoretical and empirical literature on
the goods and services provided by the natural environ-
ment (see De Groot et al., 2002; Eftec, 2005), and a small
literature on the goods and services provided by the estua-
rine and coastal environment (Turner et al., 2001; Beau-
mont et al., 2007). Much of this literature takes as its
starting point the framework established by Boyle and
Bishop, 1985 in their analysis of the total economic value
of wildlife resources, whereby a range of use and non-use
values can be identified for an estuarine ecosystem. Use
values are derived from consumptive and non-consumptive
human interaction with the ecosystem, and might include
activities associated with recreation (fishing, boating and
other water sports), commercial fisheries, agriculture,
industry, and tourism along with biodiversity, research
and education. Retaining an option to use the ecosystem
in an uncertain future may constitute an important addi-
tional use value. In contrast, non-use values can take vari-
ous forms including existence, altruistic and bequest values
(Bateman et al., 2002). Existence values often reflect a sense
of well being and are derived from knowledge of an ecosys-
tem’s existence, but are not associated with actual or
planned use. Altruistic values are an expression of concern
for the availability of the ecosystem for others in the cur-
rent generation. Bequest values are associated with the cur-
rent generation placing value on ensuring the availability of
the ecosystem to future generations.
Estuarine goods and services can be diminished by dete-
riorating water quality associated with increased levels of
eutrophication. While the impacts of increased nutrient
concentrations can vary greatly with the individual features
of the system in question, Bricker et al. (1999) identify gen-
eric ‘estuarine use impairments’ (p. 7) as primary and sec-
ondary symptoms of eutrophication and goods and
service flows, presented schematically in Fig. 1.
The fact that many of these effects and impairments
impact on goods and service flows which are either not
marketable or marketed at prices that do not reflect their
social worth, necessitates the adoption of non-market val-
uation procedures to assess the values placed by society on
any changes in the flows. The contingent valuation meth-
odology is such a non-market valuation procedure as it
relies on direct questioning of consumers to obtain willing-
ness-to-pay bids for a scenario created by the analyst. In
effect, where an actual market place in an environmental
good or service flow may not exist, a hypothetical market
can be created, and agents requested to reveal the value
that they place on a proposed change in the flow. In prac-
tice, this methodology has the potential to reflect a range of
ecosystem goods and services if an appropriate scenario
and bidding process can be established (see Turner et al.,
2001; Bateman et al., 2002 for a more detailed explanation
of the technique). An alternative approach is choice model-
ling, which also relies on stated preferences; this is particu-
larly applicable where relative values for different attributes
of an environmental resource are being sought (Bateman
et al., 2002). The application of CVM to investigate public
preferences and to elicit willingness-to-pay for reduced
eutrophication in the Randers Fjord is discussed in the fol-
lowing section.
Turning to empirical applications, the authors are una-
ware of any previous applications of CVM investigating
eutrophication in the Randers Fjord. However, the tech-
nique has been used to examine water quality improve-
ments, including eutrophication, in the US (Bockstael
et al., 1989), Europe (Söderqvist, 1996; Markowska and
Źylicz, 1999; Gren et al., 2000; Hanley et al., 2003; Bat-
eman et al., 2006; Murato et al., 2006; Frykblom et al.,
unpublished) and Asia (Yaping, 1997; Tapvong and Krua-
van, 2000). The studies undertaken within Europe include
two related CVM surveys in Sweden and Poland which
were part of a larger project that sought to evaluate the
costs and benefits of a less eutrophied Baltic Sea (see Gren
et al., 2000 for further details). This Baltic study obtained
contingent willingness-to-pay valuations in both countries
using a mailed questionnaire as the survey instrument, with
the results extrapolated to obtain values for other market
and transition economies which border the Baltic Sea.
The studies found that the perceived benefits of reducing
eutrophication in the Baltic Sea by 50% through a notional
‘international action plan’ designed to abate such pollution
to be substantial – estimated to be SEK 31,527 million
ð€3720Þ million per year when extrapolated to all nine
coastal states. Of particular interest to the current paper
 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602 593

External Nutrient Inputs Primary Secondary Potential Effects and Use

and Susceptibility Symptoms Symptoms Impairment

Loss of Habitat
Loss of Commercial fishing
Submerged Recreational fishing
Decreased Aquatic Tourism
Light Vegetation
Influence of Availability Increase in Algal Toxins
Physical Commercial Fishing
and Nuisance / Recreational Fishing
Toxic Algal Human Health Problems
Biological
N Blooms and Swimming
Processes
Algal Macroalgal Tourism
and i.e. FW
P Inflow, Dominance Mats
Fish Kills
Wetlands Changes Commercial Fishing
Uptake, Recreational Fishing
Filter Benthic Aesthetic Values
Feeders. community Tourism
Increased change
Organic Loss of Habitat
Matter Commercial Fishing
Recreational Fishing
Production Low Tourism
Dissolved
Oxygen Offensive Odours
Aesthetic Values
Tourism

Fig. 1. Primary and secondary symptoms of eutrophication (adapted from Bricker et al., 1999).

is the finding reported in Gren et al. (2000) that the esti-
mated willingness-to-pay in Denmark for a less eutrophied
Baltic Sea was SEK 3515 ð€415Þ per person per year.
3. Evaluation of reduced eutrophication in the Randers Fjord
3.1. The Randers Fjord and eutrophication
The Randers Fjord is situated in the County of Århus,
on the east coast of Jutland, Denmark. It is approximately
27 km long from its head at the city of Randers to its
mouth at Udbyhøj on the Kattegat coast. The Fjord is a
shallow estuary with an average depth of 1.6 m (Sømod
et al., 1999) and a tidal amplitude of about 0.5 m. Within
the Fjord, the penetration of saline intrusion is dependent
on the strength and direction of the wind (Nielsen et al.,
1993). The Fjord is a popular location for recreation and
tourism and is regularly used for angling, bathing, boating,
and water sports. There is little evidence of commercial
fisheries or industry within the Fjord.
The Randers Fjord is located within the Gudenå catch-
ment. This catchment drains an area of 3260 km2, with the
main freshwater input to the system from the River
Gudenå, which drains approximately 6% of Denmark
(European Environment Agency, 1995). In addition, there
is a smaller freshwater input from the Alling Å. The major-
ity of the land lying within the catchment is allocated to
agricultural activities (approximately 8% of the total Dan-
ish agricultural land area (Sømod et al., 1999)), which are a
major diffuse source of nutrient inputs. In addition to agri-
cultural run-offs, industrial and household sewage consti-
tutes important point sources for nutrient inputs into the
system; Andersen (1999) has calculated the amount of
waste water treated within the catchment which originates
from a combination of both industry and urban sewage to
be 600,000 population equivalent.2 Anderson states that as
a result of both increased agricultural activity and waste
water treatment discharges there is evidence of a 10-fold
increase in nutrient transport when compared to natural
levels in the Fjord.
The increased nutrient loading makes the Fjord suscep-
tible to seasonal eutrophication, particularly during the
spring and summer months. For example, evidence of
greatly reduced eelgrass beds within the Fjord over the last
Century is consistent with this finding. Nielsen et al. (2003)
provide evidence that eelgrass beds within the Fjord were
abundant in the early 1900s then disappeared following a
wasting disease in the 1930s. By the mid-1950s eelgrass
beds were partly re-established within the Fjord, only to
decrease greatly again over the next 50 years due to a
reduction in light penetration caused by increased microal-
gal growth and sedimentation of organic matter.
3.2. Contingent valuation survey
The Baltic Sea study discussed in Gren et al. (2000) has
demonstrated that there is scope to use CVM in assessing
public preferences for reduced eutrophication. The study
has informed the design of the Randers survey which has
2
A population equivalent is defined by the EU Urban Waste Water
Treatment Directive as the organic biodegradable load having a 5-day
biochemical oxygen demand of 60 g of oxygen per day (Council of the
European Communities, 1991).
594 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
also involved a mailed contingent valuation questionnaire
as the survey instrument. The design of the questionnaire,
including the contingency scenario, was adapted from
Söderqvist (1996). The development of the survey materi-
als, including the findings of the pilot survey, is outlined
below and discussed more fully in Burdon et al. (2004).
The questionnaire, translated into Danish, was divided
into four sections. Section A sought information on the
respondent’s various occupational, vacational and recrea-
tional uses of and links with the Randers Fjord. Section
B introduced the topic of eutrophication to the respondent,
with detailed descriptions and colour photographic evi-
dence of six signs and symptoms: microalgal blooms; mac-
roalgal mats; foaming; fish kills; reduced water clarity; and
toxic shellfish. The respondents were asked whether they
had heard of eutrophication before receiving the question-
naire, and to identify any signs and symptoms of eutrophi-
cation of which they were familiar and/or had observed,
both generally and at the Randers Fjord. Respondents
were also asked to indicate the seriousness of eutrophica-
tion against other environmental and socio-economic prob-
lems faced by Denmark. Whilst this section was scientific in
nature, the pilot survey did not identify any issues sur-
rounding respondents’ understanding of the survey materi-
als. Section C asked respondents to consider whether they
would support an action plan to reduce nutrient inputs and
thereby improve the water quality in the Randers Fjord.
The maximum willingness-to-pay by the respondent for
the action plan was then elicited, without providing
guidance as to what might be an appropriate amount. If
they would not support the action plan they were asked
to provide reasons for this. Finally, Section D of the ques-
tionnaire sought information on the respondent’s back-
ground; this provided information that could be used to
explain the bids offered by the respondents in Section C
of the questionnaire. The respondent’s postcode also
enabled them to be placed into one of three broad geo-
graphical zones according to distance from the Fjord.
The questionnaire in Danish, along with its English trans-
lation, is reproduced in Burdon et al. (2004).
Of central concern to the design of the CVM scenario,
which provided the context for which bids were elicited,
was the Water Framework Directive which focuses atten-
tion on restoring all water bodies to a ‘good ecological sta-
tus’. Nielsen et al. (2003) suggest that this status should be
as close as possible to reference conditions. In the case of
the Randers Fjord such reference conditions are defined
as ‘pristine conditions’; those conditions prevailing in the
Fjord pre-1915/16 before anthropogenic influence. Nielsen
et al. argue that nutrient reduction programmes could be
implemented to create such an improved ecological status
that it may be expected that Secchi depth measurements3
in the inner Fjord (Randers) will increase from 1.5 m to
3
Secchi depth is a measurement taken by lowering a patterned ‘Secchi
disk’ into water and noting the depth at which the pattern is no longer
visible.
3.7 m and in the outer Fjord (Udbyhøj) from 2.5 m to
5.5 m. Since the average depth of the Fjord is only 1.6 m,
this decrease in turbidity might result in the Fjord bottom
being visible whilst swimming, fishing or boating through-
out its length. It is likely that a return to pristine conditions
would also include the re-development of eelgrass beds and
their associated communities, reduced fish kills and
thus improvements in catch rates for local recreational
anglers, reduced macroalgal mats in the outer Fjord and
a reduction of foaming on local coastal shores. In the ques-
tionnaire it was proposed that an action plan would
achieve well defined outcomes consistent with these pristine
conditions if there were sufficient support for its
implementation.
In the questionnaire the respondent was asked whether
they would, possibly would, or would not accept to pay
an increase in taxes, in order to make it possible to imple-
ment the action plan and receive the associated water qual-
ity improvements and its benefits. If the respondent
accepted the nutrient reduction action plan (either voting
probably or definitely in favour) they were asked to offer
their maximum increase in taxes (in DKK) that they would
be willing to pay per month, over a 10-year period in order
to fund the action plan and receive the perceived benefits.
While an increase in taxes was the preferred payment
mechanism due to its simplicity and credibility, and its con-
sistency with the procedure followed by Söderqvist (1996),
other mechanisms might have been used. For example,
Frykblom et al. (unpublished) have proposed bidding
involving an increase in food prices due to higher agricul-
tural product costs and/or increased water bills to finance
improved waste water treatment, as consistent with the
hypothetical action plan involving reduced eutrophication.
In addition, this mechanism follows an open-ended format
and was chosen as it is informative of an individual’s max-
imum willingness-to-pay and does not lead to anchoring
bias. However, it is recognised that this mechanism does
have weaknesses including its impact on response rates
and protest answers/outliers (Bateman et al., 2002).
The main postal survey was undertaken between Octo-
ber and December 2003, following a successful pilot survey
in September of the same year. A random sample was
drawn from households located in Århus County, with
each receiving a mailed questionnaire addressed to a
named individual. The survey population comprised the
entries in the EUROPE INFO database (Kapitol, 2001),
which identified 477,201 names and addresses within the
County. The sample selection procedure was weighted
against the population size in each of the postal code areas.
A total of 1510 questionnaires were sent out of which 226
were completed and returned. In 39 cases, questionnaires
were returned-to-sender – in these cases questionnaires
were sent out again to the same address marked for the
attention of ‘the occupant’. The achieved response rate
was 15%. Not all returned questionnaires provided com-
plete information, with 13 failing to complete Section C
as the results in Section 3.3 indicate.
 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
3.3. Survey results
The questionnaire provided evidence of the characteris-
tics of the sample, on the role of the Randers Fjord within
the local economy, and on the public perception of eutro-
phication.4 Possibly reflecting the sample selection proce-
dure, over half of the respondents were male (63%) and
the majority fell within the age range of 30–60 years
(68%). The annual household income was distributed
across all the interval categories presented, thus showing
that the survey sample covered all economic backgrounds
[household incomes from less than DKK 100,000
ð€13; 400Þ to incomes of more than DKK 1,000,000
ð€134; 000Þ per year]. Almost all respondents (99%) nor-
mally live within the County of Århus, with 15% consider-
ing themselves to live within the vicinity of the Randers
Fjord itself. The respondent population included individu-
als living in cities (33%), towns (36%), villages (25%) and
isolated rural dwellings (5%). Only 4% of the respondents
had an occupation directly linked to the Randers Fjord,
and these comprised employment in the tourism industry
and occupations related to the environment and shipping.
Respondents used the Fjord for recreation and tourist
activities, including recreational hunting/fishing (11%),
walking, jogging and running (12%), and bird/wildlife
watching (9%), in addition to the more popular activity
of enjoying the views and the sounds of the water (22%).
A small number of respondents, or their families, own a
summer cottage near the Fjord (5%) whilst 9% had taken
a vacation around the Fjord since January 2001.
On knowledge of eutrophication, nearly half of the
respondents (43%) had heard of this issue before they
received the questionnaire. Macroalgal mats, foaming on
local beaches, fish kills and reduced water clarity were iden-
tified as the most common symptoms of eutrophication. A
lower proportion of the respondents had heard of or seen
the effects of eutrophication within the Randers Fjord itself
(36%), with macroalgal mats, foaming on local beaches,
fish kills, reduced water clarity and dead sea bottoms being
heard of or seen by at least a quarter of the respondents.
Only 8% of respondents had observed or had any direct
experience of changes in the Randers Fjord (either positive
or negative) which they considered to be wholly or partially
due to eutrophication. In these cases the occurrence of fish
kills and algal blooms during summer months within the
Fjord were cited. When asked to assess various general
environmental problems faced by Denmark, the typical
respondent considered eutrophication to be a more impor-
tant issue than noise from traffic but less of a problem than
the acidification of lakes and soils. It was also considered
more important than a range of social and economic issues
faced by Denmark such as inequality between men and
women.
4
For a more detailed discussion of the survey results see Burdon et al.
(2004).
595
1100
1000
Max Willingness-To-Pay (DKK)
900
800
700
600
500
400
300
200
100
0
-100
N= 129 122 119
A B C
Responses
Fig. 2. Boxplots for the willingness-to-pay responses (DKK 1 ¼ €0:134).
Support for the proposed action plan to reduce nutrient
inputs so as to improve the water quality of the Fjord was
strong among the sample. Some 69% of respondents indi-
cated a willingness to support the action plan, voting either
probably or definitely in favour of it.5 The bids offered by
these respondents in support of the action plan ranged up
to DKK 1000 ð€134Þ per month over a ten year period,
with a mean bid of DKK 93 ð€12:46Þ and median bid of
DKK 50 ð€6:70Þ [N ¼ 129, standard deviation = DKK
146 ð€19:56Þ]. The box plot shown in Fig. 2 further illus-
trates these results and identifies three ‘outliers’ [DKK
250 ð€33:50Þ and two DKK 300 ð€40:20Þ bids] and seven
‘extreme values’ [DKK 1000 ð€134Þ, DKK 800 ð€107:20Þ
and five DKK 500 ð€67Þ bids] within Plot A6 which is the
boxplot for the full sample. If the extreme and outlier bid
values are removed from the sample (shown progressively
in Plots B and C), the mean bid falls to DKK 57 ð€7:64Þ
[N ¼ 119, standard deviation = DKK 51 ð€6:83Þ]. The
findings suggest that the presence of outliers and extreme
values has a strong influence on the willingness-to-pay val-
ues and were used to justify the removal of these data
observations from the later analysis. With these observa-
tions there was concern that they represent high protest
bids being associated with respondents who were not char-
acterised by unusually high incomes.
Given the questionnaire was addressed to an individual
within each sample household, there was a potential for
bias associated with the respondent’s gender and age. For
5
The respondents who voted against the action plan most commonly
cited ‘I think I pay too many taxes already’ (38 cases) and ‘I am not
responsible for the poor water quality, let the polluters pay’ (33 cases) as
reasons against the action plan. Less frequently cited reasons included ‘the
money would be better spent elsewhere’ and ‘I do not believe my
contribution would lead to any improvement’. The latter response, in
particular, may be indicative of a protest vote.
6
A bid was defined as an ‘outlier’ if it was 1.5–3 box lengths from the
upper and lower edge of the box, and an ‘extreme value’ if it was greater
than 3 box lengths away.
596 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
example, Dupont (2001) has found that gender can have a
significant impact on voting behaviour and willingness-to-
pay in a study involving water quality improvements. The
current survey provides partial support for this finding,
with 66% of males and 74% of females found to vote in
favour of the action plan (although this finding is not sta-
tistically significant using a Pearson Chi-square test
ðv2 ¼ 1:864; p > 0:05Þ), and a mean bid value higher for
males when compared to females by DKK 21 ð€2:81Þ per
month (a result that is statistically significant using a
Mann–Whitney U-test ðZ ¼ 2:077; p < 0:05Þ. These
results suggest that gender has an effect on the amount
bid by the population and therefore gender should be
included within any aggregation of the results. An equiva-
lent analysis relating to the four age categories used in the
survey (<30 years, 30–44 years, 45–60 years and >60 years)
did not find support for age being a significant determinant
of either voting or bid decisions, so age was omitted from
the aggregation process.
With a total adult population (over 16 years of age) of
Århus County in 2003 of 579,885 (Statistics Denmark,
2003), aggregating the sample evidence to the population
of the County as a whole suggests that the value of the
potential benefit received from improved water quality, as
implied by the scenario, in the Randers Fjord may be in
the region of DKK 22.8 million ð€3:06Þ million per month
over the 10-year period of the proposed action plan. This
aggregation treats the population as homogenous and
assumes that 69% of the County would vote in favour of
the action plan. If the aggregation allows for gender, the
value of the potential benefit falls to DKK 22 million
ð€2:95Þ million per month given an adult male and female
population of approximately 286,000 and 294,000, respec-
tively. Such findings can be interpreted as strong support
for reduced eutrophication in the Randers Fjord by resi-
dents of Århus County at the time that the survey was
undertaken.
4. A decision tree analysis
4.1. Methodology
To investigate the contingent voting and the bid deci-
sions, classification and regression (decision) tree analysis
provides valuable exploratory tools. Such tools, often
referred to as recursive partitioning techniques, chart the
relationships between a predictor variable (which may be
nominal, ordinal or continuous) and a number of explana-
tory variables. As such, these methods have some parallels
with standard statistical techniques (Venables and Ripley,
2002; De’ath and Fabricius, 2000), but are not hypothesis
testing in the traditional sense (although their output can
be used in formal statistical tests and probability distribu-
tions may be used to assist the development of the model).
Consequently, such models are less strict with regard their
assumptions, and are ideal for determining patterns and
deriving models from large noisey datasets typically
obtained from contingent valuation surveys. These models
are particularly useful where there is little a priori knowl-
edge or theories regarding which variables are related and
how. Decision tree analysis has been applied to a variety
of research problems including segmentation, stratification,
prediction and data reduction/screening (see for example
Eisenberg and McKone, 1998; Podgorelec et al., 2002;
Hautaniemi et al., 2005).
Rather than implicitly fit the data to a pre-defined rela-
tionship (e.g. linear) they progressively sort through the
explanatory variables (e.g. gender, age and income) and
extract the variable which gives the best classification or
prediction by splitting the data into sub-groups or nodes.
For example, for a categorical target variable such as vot-
ing ‘yes’ or ‘no’ the model would pick the best explanatory
variable e.g. income, which best splits the cases into ‘yes’
voters on one side of the tree and ‘no’ voters on the other.
The degree of success of the split is defined by the impurity
level of the sub-groups. In addition to picking the best var-
iable to split the cases, the tree also defines cut-off levels or
rules, which define the split. This process is applied recur-
sively to the subgroups (to define sub-subgroups) and so
on until the decision tree is finished, as defined by various
stopping criteria.
The decision tree models used in the current study were
primarily based on the Classification and Regression Trees
(C&RT) methodology7 as defined by Breiman et al. (1984),
although other techniques such as Chi-squared Automatic
Interaction Detector (CHAID) have also been examined
(see Kass, 1980). C&RT is a binary tree algorithm which
partitions data into two subsets such that the samples
within each subset are more homogenous, with regard to
the target variable, than the previous subset. This is a
recursive process as it repeats until a pre-defined measure
of homogeneity/purity (or other measure of completion)
is satisfied. It should be noted that the same predictor var-
iable may be used a number of times at different levels of
the tree.
The detailed algorithms upon which the decision tree is
defined may be found in Breiman et al. (1984) and SPSS
(2001). For the current study 10-fold cross-validation was
used to validate the models. Cross-validation involves split-
ting the sample into a number of smaller groups. Trees are
then generated with the data from each sub-sample
excluded in turn. For example, with 10-fold cross-valida-
tion, the data are split into 10 sub-samples (sample folds)
and then 10 trees are generated. The first tree is calculated
using all of the cases except those in the first sample fold;
the second tree is calculated using all the cases except those
in the second sample fold, and so on. For each tree the mis-
classification risk is estimated by applying the tree to the
sub-sample excluded in generating it.
7
A number of decision tree methods are currently available and in the
current study the software package AnswerTreeTM was used to apply these
methods to the data.
 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602 597

In terms of stopping rules, the maximum tree depth was
set to five and as the dataset was relatively small, the min-
imum number of cases for parent nodes was set to five and
the minimum number of cases for child nodes was two. In
C&RT the growth of the tree is also determined by an
impurity measure which in the current study was the Gini
impurity measure for categorical target variables (voting
yes or no) and the least squared deviation (LSD) measure
used for continuous target variable (actual bid). The Gini
measure is based on the squared probabilities of member-
ship for each target category in the node. It reaches its min-
imum (zero) when all cases in the node fall into a single

Vote Y/N

Node 0
Category % n
Y 70.05 145
N 29.95 62
Total (100.00) 207
—
Resident in…
Improvement=0.0120

1;3;2 4

Node 1 Node 2
Category % n Category % n
Y 71.50 143 Y2 8.57 2
N 28.50 57 N7 1.43 5
Total (96.62) 200 Total (3.38) 7
—
DISTANCE
Improvement=0.0111

<=2 >2

Node 3 Node 4
Category % n Category % n
Y 76.64 105 Y 60.32 38
N 23.36 32 N 39.68 25
Total (66.18) 137 Total (30.43) 63
— —
READERS INCOME
Improvement=0.0090 Improvement=0.0085

1 0 <=150000 >150000

Node 7 Node 8 Node 9 Node 10

Category % n Category % n Category % n Category % n
Y 79.20 99 Y5 0.00 6 Y2 0.00 1 Y 63.79 37
N 20.80 26 N5 0.00 6 N8 0.00 4 N 36.21 21
Total (60.39) 125 Total (5.80) 12 Total (2.42) 5 Total (28.02) 58
— —
GENDER INCOME
Improvement=0.0145 Improvement=0.0172

1 2 <=350000 >350000

Node 15 Node 16 Node 19 Node 20

Category % n Category % n Category % n Category % n
Y2 5.00 2 Y 100.00 4 Y 100.00 11 Y 55.32 26
N7 5.00 6 N0 .00 0 N0 .00 0 N 44.68 21
Total (3.86) 8 Total (1.93) 4 Total (5.31) 11 Total (22.71) 47
—
INCOME
Improvement=0.0160

<=550000 >550000

Node 27 Node 28
Category % n Category % n
Y 38.46 10 Y 76.19 16
N 61.54 16 N 23.81 5
Total (12.56) 26 Total (10.14) 21

Fig. 3. Decision tree analysis for votes in favour and against the proposed action plan.
598 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
target category. The LSD impurity measure is computed as
the within-node variance, adjusted for frequency or case
weights.
4.2. Application and results
4.2.1. Analysis of voting behaviour
The first application of the decision tree analysis was to
voting behaviour. This was based on a sample of 207
respondents, for whom a complete set of data was avail-
able, and is presented in Fig. 3. The significance of the split
in the data into sub-groups or nodes was tested using the
Pearson Chi-Square test, with the results reported in Table
1. The first division of the decision tree model with respect
to voting behaviour shows that those respondents who are
resident in an ‘isolated rural dwelling’ were less likely to
vote in favour of supporting the action plan, than those
who reside in either a town, a city or a village ðp < 0:05Þ.
Those resident in an isolated rural dwelling were removed
from the next branch (Nodes 3 and 4), where a significant
difference ðp < 0:05Þ was identified between those respon-
dents who live closer to the Fjord (DIST1 and DIST2)
and those that live further away (DIST3). Those that live
closest are more likely to vote in favour of supporting such
an action plan. Other factors, such as gender and the read-
ing habits of respondents (the latter being used as an indi-
cator of possible awareness of such issues) also influenced
the direction of the decision tree however the influence of
these factors was not shown to be significant ðp > 0:05Þ.
The final two divisions on the right hand side of the tree
both showed that annual household income had a signifi-
cant effect ðp < 0:05Þ on the voting behaviour of the
respondents. It is clear from the decision tree model that
income has an impact on the voting behaviour when the
main group is sub-divided into smaller sub-categories with
respect to income. Fig. 3 also shows that there is not a
straightforward relationship between voting behaviour
and income. Those that have a higher income are not nec-
essarily more likely to vote in favour of (or against) the
action plan, as this has been found to vary between sub-
divisions of income. For example those that earn more
than DKK 350,000 ð€46; 900Þ were significantly more likely
to vote against the action plan (Nodes 19 and 20) and,
Table 1
Pearson Chi-square test results for voting behaviour
Factor Difference between Chi value
Resident in. . . Node 1 and Node 2 4.071
Distance Node 3 and Node 4 4.871
Readers Node 7 and Node 8 3.711
Income Node 9 and Node 10 2.086
Gender Node 15 and Node 16 3.375
Income Node 19 and Node 20 5.891
Income Node 27 and Node 28 5.251
within this category, those that earned in excess of DKK
550,000 ð€73; 700Þ were significantly more likely to vote
in favour of the action plan (Nodes 27 and 28).
Therefore, in this examination of voting behaviour, the
decision tree analysis has been effective in identifying statis-
tically significant relationships between the respondent’s
decision and a reduced number of variables, in this applica-
tion namely whether the respondent is resident in an iso-
lated rural dwelling, how far the respondent lives from
the Fjord, and what is the respondent’s annual household
income. A strength of the procedure is clearly identified
in the relationship identified between the voting decision
and the income variable which is shown to be complex,
and might be missed or found to be insignificant using
more standard modelling procedures such as regression
and probit analysis.
4.2.2. Analysis of willingness-to-pay bids
The second application of decision tree analysis is to the
willingness-to-pay bids. Here a sample comprises those 179
respondents who expressed a willingness-to-pay having
voted in favour of supporting the action plan. The decision
tree is depicted in Fig. 4 and the significance of the data
splits are tested using the Mann–Whitney U-test, with the
results reported in Table 2. The decision tree model for
the bids shows that distance from the Fjord has the most
influence on the willingness-to-pay of respondents, with
those living closest to the Fjord (DIST1) being willing to
pay DKK 22 ð€2:95Þ more per month over the suggested
10-year period than those respondents that live further
away from the Fjord (DIST2 and DIST3). This result
was shown to be statistically significant ðp < 0:05Þ. ‘DIST1’
respondents were then divided into those who had prior
knowledge of eutrophication in the Randers Fjord and
those who did not. The results show that those with prior
knowledge of eutrophication in the Fjord were willing to
pay more to fund the action plan than those with no prior
knowledge. However, this finding was not statistically sig-
nificant at a 95% probability level. The respondents that
lived further away from the Fjord (DIST2 and DIST3)
were then sub-divided into those whose annual household
income is less than or equal to DKK 950,000 ð€127; 300Þ
and those whose income was greater than this (Nodes 5
and 6). Those whose annual household income was greater
than DKK 950,000 ð€127; 300Þ were willing to pay DKK
33 ð€4:42Þ per month more, although this finding was not
shown to be significant at the 95% level. This segregation
of income at the highest level removes those respondents
p Value
with the greatest income (Node 6). One factor that may
0.044
cause a lack of significance at this level was the highest
0.027
0.054 income category (Band 11) on the questionnaire being
0.149 open-ended (that is greater than DKK 1,000,000
0.066 ð€134; 000Þ) thus the behaviour of respondents with income
0.015 in this band may differ greatly. For the purpose of the anal-
0.022
ysis, all of the respondents in this income band had to be
 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602 599

Bid (DKK)
Node 0
Mean 37.7430
Std. Dev. 49.6894
n 179
% 100.00
Predicted 37.7430
—
DISTANCE
Improvement=86.9448

<=1 >1

Node 1 Node 2
Mean 55.1250 Mean 32.7410
Std. Dev. 61.4930 Std. Dev. 44.7579
n 40 n 139
% 22.35 % 77.65
Predicted 55.1250 Predicted 32.7410
— —
Randers Know. INCOME
Improvement=57.4766 Improvement=52.1576

0 1 <=950000 >950000

Node 3 Node 4 Node 5 Node 6

Mean 36.4706 Mean 68.9130 Mean 30.5846 Mean 63.8889
Std. Dev. 36.6898 Std. Dev. 72.5032 Std. Dev. 42.2915 Std. Dev. 67.4434
n 17 n 23 n 130 n 9
% 9.50 % 12.85 % 72.63 % 5.03
Predicted 36.4706 Predicted 68.9130 Predicted 30.5846 Predicted 63.8889
—
DISTANCE
Improvement=29.5214

<=2 >2

Node 7 Node 8
Mean 35.6250 Mean 22.5200
Std. Dev. 43.6358 Std. Dev. 39.1279
n 80 n 50
% 44.69 % 27.93
Predicted 35.6250 Predicted 22.5200
—
INCOME
Improvement=22.6146

<=450000 >450000

Node 9 Node 10
Mean 29.8667 Mean 11.5000
Std. Dev. 45.3969 Std. Dev. 24.2845
n 30 n 20
% 16.76 % 11.17
Predicted 29.8667 Predicted 11.5000
—
AGE INCOME
Improvement=37.6852 Improvement=8.0726

<=3 >3 <=550000 >550000

Node 11 Node 12 Node 13 Node 14

Mean 38.9091 Mean 5.0000 Mean 3.0000 Mean 20.0000
Std. Dev. 50.0570 Std. Dev. 7.5593 Std. Dev. 6.7495 Std. Dev. 32.2318
n 22 n 8 n 10 n 10
% 12.29 % 4.47 % 5.59 % 5.59
Predicted 38.9091 Predicted 5.0000 Predicted 3.0000 Predicted 20.0000

Fig. 4. Decision tree analysis for willingness-to-pay bids for the proposed action plan.
600 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602

Table 2 the European Marine Strategy and proposed Marine

Mann–Whitney U-test results for willingness-to-pay bids Framework Directive relating to the marine environment,
Factor Difference between Z Value p Value published evidence on public preferences for such measures
Distance Node 1 and Node 2 2.308 0.021 is scarce. When this policy intervention and associated
Randers knowledge Node 3 and Node 4 1.207 0.227 management action imposes substantial costs on the local
Income Node 5 and Node 6 1.358 0.174 community their support for such measures is imperative.
Distance Node 7 and Node 8 1.732a 0.086
Income Node 9 and Node 10 1.687 0.092
With this in mind, the focus of this paper has been on
Age Node 11 and Node 12 1.977 0.048 the determination and evaluation of public preferences
Income Node 13 and Node 14 4.075 0.000 for reduced eutrophication with its implications for water
a
Data showed a similar variance and therefore a T-test was used (Z quality improvement, relating specifically to the Randers
value is therefore a T value). Fjord, Denmark. The analysis was based on responses
obtained from a contingent valuation survey of residents
of Århus County carried out in 2003. It has identified
treated the same, although they may have differing levels of
implicit public support for the legislation.
income above DKK 1,000,000 ð€134; 000Þ.
Water quality and the particular issue of eutrophication
Those in the lower income band (less than or equal to
are clearly in the Århus County public conscience. A large
DKK 950,000 ð€127; 300Þ) were then segregated by dis-
proportion of survey respondents reveal an appreciation
tance, with those living closest to the Fjord (DIST1 and
and/or experience of the signs and symptoms of eutrophi-
DIST2) being willing to pay an extra DKK 13 ð€1:74Þ
cation; macroalgal mats, foaming on local beaches, fish
per month, although this split was not significant
kills, reduced water clarity and dead sea bottoms being
ðp > 0:05Þ. Those respondents who live furthest away from
heard of or seen in the Randers Fjord by at least a quarter
the Fjord (DIST3) were then sub-divided with respect to
of the respondents. When asked to assess various problems
annual household income (Nodes 9 and 10). The respon-
faced by Denmark, the issues associated with eutrophica-
dents with a household income less than or equal to
tion were deemed to be more important than noise from
DKK 450,000 ð€60; 300Þ were willing to pay DKK 18
traffic but less of a problem than the acidification of lakes
ð€2:41Þ per month more than those in the higher annual
and soils, and a greater problem than a number of social
household income category. This result was not shown to
and economic issues.
be significant ðp > 0:05Þ. The final divisions of the analysis
Moreover, there appears to be strong public support for
divided the lower household income category by age
the introduction of an action plan to reduce nutrient inputs
(Nodes 11 and 12) and the upper household income cate-
from water treatment plants and agricultural sources in this
gory by a different annual household income classification
estuarine system. With the successful application of CVM
(Nodes 13 and 14). With respect to age, it was shown that
to the examination of the issue, it was established that a
the respondents over 60 years old are willing to pay signif-
majority of respondents were in favour of the action plan,
icantly less (DKK 33 ð€4:42Þ per month) than those who
with a mean willingness-to-pay (in terms of a local tax) of
are 60 years old or less ðp < 0:05Þ. Finally, the upper
DKK 57 ð€7:64Þ per person per month over a ten year per-
household income group (Node 10) was separated with
iod. Aggregating the sample evidence to the adult popula-
respect to income. It was shown that those with an income
tion of the County as a whole suggests that the estimated
greater than DKK 550,000 ð€73; 700Þ but less than or equal
value of the benefit received from improved water quality,
to DKK 950,000 ð€127; 300Þ were willing to pay DKK 17
as implied by the scenario, in the Randers Fjord is DKK
ð€2:28Þ per month more than those that earned between
22.8 million (€3:06 million) per month over the ten year
DKK 450,000 ð€60; 300Þ and DKK 550,000
period of the action plan. When compared to the findings
ð€73; 700Þðp < 0:05Þ.
of Gren et al. (2000), the willingness-to-pay for a less eutro-
Therefore, the decision tree analysis has been shown to
phied Randers Fjord is approximately a quarter of that
be effective, in this case, in identifying and clarifying the
offered in support of a 50% reduction in eutrophication
relationship between the value of the bid offered in support
for the Baltic Sea (DKK 2785 ð€373:19Þ per person per
of the proposed plan and a reduced number of variables.
year).
Significant relationships were identified between bidding
In the current study classification and regression (deci-
behaviour and the distance the respondent lives from the
sion) tree analysis has been used as an exploratory tool
Fjord, their annual household income and their age. In
to investigate the contingent voting and bid decisions. This
the case of the income variable, the relationship with the
analysis has revealed the complexity of the relationships
bid value is found to be complicated.
surrounding such decisions. In particular, the analysis has
highlighted the significance of the respondents’ distance
5. Concluding comments from the Fjord and their annual household income in
determining their voting behaviour, and the importance
Despite recent developments in EU water quality legisla- of the respondents’ distance from the Fjord, their annual
tion, namely the introduction of the Water Framework household income and their age in determining the size
Directive, and prospects for further legislation such as of their bid. These relationships may have been missed or
 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
found to be insignificant using more standard modelling
procedures such as regression and probit analysis owing,
for example, to noise within the data set. If a larger sample
was obtained, this would allow the analysis to progress
beyond an exploratory role, as presented in this paper, to
one of prediction whereby voting and bidding behaviour
could be predicted given information on individual profiles.
Thus, the focus of this paper has been to report the find-
ings of a CVM study and decision tree analysis on public
preferences for reduced eutrophication in the Randers
Fjord, Denmark; the findings may have wider relevance.
The Randers Fjord has a relatively small catchment area
and population owing to topographic controls, whereas
most estuarine systems across Europe drain much larger
catchments where identifying the cause and effect of eutro-
phication controls would be much more difficult, as would
engaging the wider community. This may restrict the appli-
cation of this methodology and the transferability of the
findings to similar catchments across Northern Europe.
Nevertheless, as more studies are undertaken, common fea-
tures of public preferences may become apparent, allowing
for a greater role for prediction to inform policy makers
about the extent and nature of public support for their
actions in this field.
Acknowledgements
The authors would like to thank the European Commis-
sion for financial support of this work as part of the multi-
national EUROTROPH project, Contract Number EVK3-
CT-2000-00040, of which the Institute of Estuarine and
Coastal Studies (IECS), University of Hull was a partici-
pant. The authors benefited from comments received from
participants at the ECSA 41st International Conference,
Venice, 16–20 October 2006. The authors would also like
to thank three anonymous referees for helpful comments
on an earlier draft.
References
Andersen, J.M., 1999. 25 Years of Pollution Abatement and Environ-
mental Improvement in the Gudenaa – a Danish Lake-River-Estuary
System. Publication of the Gudenaa Committee, Højbjerg, Denmark.
Apitz, S.E., Elliott, M., Fountain, M., Galloway, T.S., 2006. European
environmental management: moving to an ecosystem approach.
Integrated Environmental Assessment and Management 2 (1), 80–85.
Atkins, J.P., Burdon, D., 2006. An initial economic evaluation of water
quality improvements in the Randers Fjord, Denmark. Marine
Pollution Bulletin 53 (1–4), 195–204.
Bateman, I.J., Carson, R.T., Day, B., Hanemann, M., Hanley, N., Hett,
T., Jones-Lee, M., Loomes, G., Maurato, S., Özdemiroglu, E., Pearce,
D.W., Sugden, R., Swanson, J., 2002. Economic Valuation with Stated
Preference Techniques: A Manual. Edward Elgar, Cheltenham, UK.
Bateman, I.J., Day, B., Dupont, D., Georgiou, S., Matias, N.G.,
Morimoto, S., Subramanian, L., 2006. Cost-benefit analysis and the
prevention of eutrophication. In: Pearce, D.W. (Ed.), Environmental
Valuation in Developed Countries. Edward Elgar, Cheltenham, UK.
Beaumont, N.J., Austen, M.C., Atkins, J.P., Burdon, D., Degraer, S.,
Dentinho, T.P., Derous, S., Holm, P., Horton, T., Van Ierland, E.,
601
Marboe, Starkey, D.J., Townsend, M., Zarzycki, T., 2007. Identifica-
tion, definition and quantification of goods and services provided by
marine biodiversity: implications for the ecosystem approach. Marine
Pollution Bulletin 54 (3), 253–265.
Bockstael, N.E., McConnell, K.E., Strand, I.E., 1989. Measuring the
benefits of improvements in water quality: the Chesapeake Bay.
Marine Resource Economics 6, 1–18.
Boyle, K.J., Bishop, R.C., 1985. The total value of wildlife resources:
conceptual and empirical issues. Invited paper, Association of Envi-
ronmental and Resource Economists Workshop on Recreational
Demand Modelling, Boulder, Colorado, 17–18 May 1985.
Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J., 1984. Classifica-
tion and Regression Trees. Belmont, California, Wadsworth.
Bricker, S.B., Clement, C.G., Pirhalla, D.E., Orlando, S.P., Farrow,
D.R.G., 1999. National estuarine eutrophication assessment: effects of
nutrient enrichment in the nation’s estuaries. NOAA, National Ocean
Service, Special Projects Office and the National Centres for Coastal
Ocean Science. Silver Spring, MD, p. 71.
Burdon, D., Atkins, J.P., Mutemererwa, A.M., Allen, J.H., 2004. An
economic evaluation of water quality improvements in European
estuarine and coastal ecosystems. Report 3 of 4. Institute of Estuarine
and Coastal Studies (IECS), University of Hull, Hull, UK. Report to
the European Commission (Contract No. EVK3-CT-2000-00040).
Council of the European Communities, 1991. Urban Waste Water
Treatment Directive 91/271/EEC. Official Journal of the European
Communities, L135, Brussels, 30 May 1991, pp. 40–52.
De’ath, G., Fabricius, K.E., 2000. Classification and regression trees: a
powerful yet simple technique for ecological data analysis. Ecology 81
(11), 3178–3192.
De Groot, R.S., Wilson, M.A., Boumans, R.M.J., 2002. A typology for
the classification, description and valuation of ecosystem functions,
goods and services. Ecological Economics 41, 393–408.
Dupont, D.P., 2001. Gender and willingness-to-pay for recreational
benefits from water quality improvements. In: Johnston, R.S., Shriver,
A.L. (Eds.), Proceedings of the Tenth Biennial Conference of the
International Institute for Fisheries Economics and Trade (IIFET),
10–14 July 2000, Corvallis, Oregon, USA.
EC, 2005a. Thematic strategy on the protection and conservation of the
marine environment. Communication from the Commission to the
Council and European Parliament, European Commission, Brussels,
24.10.2005, COM(2005)504 final.
EC, 2005b. Framework for Community Action in the field of marine
environmental policy (Marine Strategy Directive). Proposal for a
Directive from the European Commission, Brussels, 24.10.2005,
COM(2005)505 final.
Eftec, 2005. The economic, social and ecological value of ecosystem
services: a literature review. Report produced by Economics for the
Environment Consultancy (Eftec) for the Department for Environ-
ment, Food and Rural Affairs (Defra), January 2005.
Eisenberg, J.N.S., McKone, T.E., 1998. Decision tree method for
classification of chemical pollutants: incorporation of across-chemical
variability and within-chemical uncertainty. Environmental Science &
Technology 32 (21), 3396–3404.
Elliott, M., de Jonge, V.N., 2002. The management of nutrients and
potential eutrophication in estuaries and other restricted water bodies.
Hydrobiologia (475/76), 513–524.
European Environment Agency, 1995. Europe’s environment. In: Stan-
ners, D., Bourdeau, P. (Eds.), The Dobris Assessment: The Report on
the State of Pan-European Environment Requested by the Ministers
for the Whole of Europe at the Ministerial Conference Held at Dobris
Castle, Czechoslovakia, June, 1991. Office of Official Publications of
the European Communities, Luxembourg.
Frykblom, P., Scharin, H., Söderqvist, T., Helgesson, A., unpublished.
Cost-benefit analysis in Sweden: general use and an application to
coastal water quality management.
Gren, I.-G., Turner, K., Wulff, F. (Eds.), 2000. Managing a Sea – The
Ecological Economics of the Baltic. Earthscan Publications Ltd.,
London, UK.
602 J.P. Atkins et al. / Marine Pollution Bulletin 55 (2007) 591–602
Hanley, N., Bell, D., Alvarez-Farizo, B., 2003. Valuing the benefits of
coastal water quality improvements using contingent and real behav-
iour. Environmental and Resource Economics 24 (3), 273–285.
Hautaniemi, S., Kharait, S., Iwabu, A., Wells, A., Lauffenburger, D.A.,
2005. Modelling of signal-response cascades using decision tree
analysis. Bioinformatics 21 (9), 2027–2035.
KAPITOL, 2001. EUROPE INFO address database (DVD-Rom 1995–
2001). Kapitol, Chaussée de Saint-Job 506 Sint-Jobsesteenweg, B-
1180, Brussels, Belgium.
Kass, G., 1980. An exploratory technique for investigating large quantities
of categorical data. Applied Statistics 29 (2), 119–127.
Markowska, A., Źylicz, T., 1999. Costing an international public good:
the case of the Baltic Sea. Ecological Economics 30, 301–316.
Murato, S., Provins, A., Özdemiroglu, Georgiou, S., Newcombe, J., 2006.
Is it worth revising the European Bathing Water Directive? A choice
experiment. In: Pearce, D.W. (Ed.), Environmental Valuation in
Developed Countries. Edward Elgar, Cheltenham, UK.
Nielsen, K., Sømod, B., Hansen, D.F., 1993. Eutrophication of Randers
Fjord Estuary, Denmark – deteriorations and improvements (1900–
1991). European Water Pollution Control 3 (4), 44–51.
Nielsen, K., Sømod, B., Ellegaard, C., Krause-Jensen, D., 2003. Assessing
reference conditions according to the European Water Framework
Directive using modelling and analysis of historical data: an example
from Randers Fjord, Denmark. Ambio, vol. 32(4), June 2003, pp. 287–
294.
Podgorelec, V., Kokol, P., Stiglic, B., Rozman, I., 2002. Decision trees: an
overview and their use in medicine. Journal of Medical Systems 26 (5),
445–463.
Söderqvist, T., 1996. Contingent valuation of a less eutrophied Baltic Sea.
Beijer Discussion Paper Series, No. 88. Beijer International Institute of
Ecological Economics, Stockholm, Sweden.
Sømod, B., Larsen, J.E., Hansen, D.F., Düwel, L., Andersen, P., 1999.
Vandmiljøet i Randers Fjord 1997. Århus Amt, Højberg, Danmark.
SPSS, 2001. AnswerTreeÒ 3.0 User’s Guide. SPSS Inc., USA.
Statistics Denmark, 2003. Statistical Yearbook 2003, 107th Ed. Published
by Statistics Denmark, June 2003.
Tapvong, C., Kruavan, J., 2000. Water quality improvements: a contin-
gent valuation study of the Chao Phraya river. EEPSEA Research
Reports, Int. Dev. Res. Centre, Ottawa, Canada.
Turner, R.K., Bateman, I.J., Adger, W.N., 2001. Economics of Coastal
and Water Resources: Valuing Environmental Functions. Kluwer
Academic Publishers, The Netherlands.
Venables, W., Ripley, B., 2002. Modern Applied Statistics with S.
Springer-Verlag, Berlin, p. 495.
Yaping, D., 1997. The value of improved water quality for recreation in
East Lake, Wuhan, China: an application of contingent valuation and
travel cost methods. EEPSEA Research Reports, Int. Dev. Res.
Centre, Ottawa, Canada.