See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/261222909 Ecosystem Service Valuation from Floodplain Restoration in the Danube River Basin: An International Choice Experiment Application Conference Paper · June 2009 DOI: 10.13140/2.1.2884.7044 CITATIONS READS 12 1,118 10 authors, including: Markus Bliem Roy Brouwer Regional Government of Carinthia University of Waterloo 22 PUBLICATIONS 225 CITATIONS 253 PUBLICATIONS 9,182 CITATIONS SEE PROFILE S. Kerekes Corvinus University of Budapest 41 PUBLICATIONS 327 CITATIONS SEE PROFILE Some of the authors of this publication are also working on these related projects: Biosphere Reserve Integrated Monitoring View project Agricultural Water Futures View project All content following this page was uploaded by Markus Bliem on 01 April 2014. The user has requested enhancement of the downloaded file. SEE PROFILE Ecosystem Service Valuation from Floodplain Restoration in the Danube River Basin: An International Choice Experiment Application Roy Brouwer1, Markus Bliem2, Zsuzsanna Flachner3, Michael Getzner4, Sandor Kerekes5, Simon Milton5, Teodora Palarie6, Zsuzsanna Szerényi5, Angheluta Vadineanu6, Alfred Wagtendonk1 1 Institute for Environmental Studies (IVM), VU University, Amsterdam, The Netherlands. 2 Institute for Advanced 3 4 5 Studies Carinthia (IHSK), Austria. RISSAC, Budapest, Hungary. University of Klagenfurt, Austria. Corvinus 6 University Budapest, Hungary. University of Bucharest, Romania. Abstract The objective of this paper is to estimate the welfare impacts of ecological floodplain restoration and test their transferability across the second largest river basin in Europe. Floodplain restoration reduces flood risk and improves water quality, the welfare impacts of which are measured in an identical choice experiment conducted in three countries. The estimated choice models differ in preference and scale parameters due to preference heterogeneity. Although transfer errors are reduced by up to almost 75% if preference heterogeneity is controlled for, the estimated welfare measures are not transferable when imposing a 20% error tolerance level. Similarly, substantial errors are found if average values are aggregated without controlling for preference heterogeneity and socio-economic conditions that are unevenly distributed in space. Key words: Choice Experiment, Ecological Restoration, Water Quality, Flood Risk, Benefits Transfer, GIS, Water Framework Directive 1 1. Introduction Integrated river basin management and the economic valuation of water resources have been at the forefront of the European environmental economics research agenda since the introduction of the European Water Framework Directive (2000/60/EC). In particular, payments for watershed ecosystem services have received a lot of attention (e.g. Porras et al., 2008). This paper addresses the economic valuation of the ecosystem services associated with floodplain restoration. More specifically, the non-market benefits of ecological restoration of heavily modified river stretches in three different countries in the international Danube river basin are estimated. With a length of 2,780 km the Danube is the second largest river in Europe. The river has been subject to anthropogenic modification and environmental pressures over the past centuries, including canalization, construction of embankments, navigation, use of hydropower and pollution (ICPDR, 2007). The shape of the river has changed and large parts of the floodplains have been drained for agricultural purposes. The hydrological connectivity of the river and its tributaries to the surrounding floodplains has been reduced substantially (e.g. Hohensinner et al., 2004). Only in small areas, such as the Danube National Park in Austria, are the original Danube ecosystem and adjacent wetlands still intact. Restoring the river to a ‘good ecological status’, as required by the European Water Framework Directive (WFD), is only feasible if parts of the river are transformed back as close as possible to their original state with regard to natural hydro-morphological conditions. Through the application of a common choice experiment design, the economic values of river restoration in three river basin countries (Austria, Hungary and Romania) are compared. The main objective of the study presented here is to estimate public willingness to pay (WTP) for the ecosystem services associated with river restoration in terms of reduced flood risk and 2 water quality improvements and to test the transferability of the non-market benefits of river restoration policy scenarios across the three river basin countries, whilst accounting for preference heterogeneity. Geographical information systems (GIS) are used to aggregate economic welfare implications of these policy scenarios across the relevant population of beneficiaries in the three countries, controlling for spatially correlated choices through distance-decay functions and differences in income distribution across the international river basin. Distance-decay refers to the expected negative relationship between WTP for river restoration and distance of the respondent’s place of residence from the Danube river (e.g. Bateman et al., 2006). Respondents living closer to the river are expected to assign more value to the benefits of river restoration than respondents living further away. Estimated non-market values can be used, together with potential market benefits such as the avoided costs of flood damage or water purification, to justify investments in Danube river restoration projects to achieve the ecological objectives of the WFD based on economic welfare considerations. These projects are expected to create a double dividend from the necessary investments to transform river stretches back in their natural conditions, as they restore the natural floodwater storage capacity of the river (hence reduce flood risk) and also improve water quality at the same time (e.g. Schiemer et al., 1999). In view of the public nature of the ecosystem services provided by the river restoration projects, stated preference methods, such as choice experiments, are considered important tools to inform policy and decision-making regarding these projects (Brouwer, 2008). Choice models are considered superior to contingent valuation methods for the purpose of value transfer because of their ability to value marginal changes in good and site characteristics (Morrison and Bergland, 2006). However, the application of choice models to water quality improvements in the context of the WFD is limited (e.g. Hanley et al., 2006). Although various economic valuation 3 methods and techniques, including stated preference methods such as contingent valuation, have been applied over the past decades to value the non-priced public goods and services provided by wetlands and floodplains (e.g. Brouwer et al., 1999; Brander et al., 2006), a common valuation design for international comparison is missing. 2. Choice experiment design A common choice experiment (CE) was developed to assess the non-market value of the ecosystem services associated with river restoration. The design consists of two exclusive categories of benefits: the impact of river restoration on floodwater storage and a corresponding reduction in flood risk, and the river’s nutrient retention capacity and hence water quality. These benefits make up two of the three attributes used to evaluate policy alternatives in the CE. A monetary cost price was included as a third attribute to enable monetization of the benefits of different river restoration projects. The alternatives describe different end states created through river restoration measures. The link between these end states and river restoration projects is explained in the introduction to the CE. Respondents are told that river restoration measures will affect both flood risk and water quality as a result of an improvement of the river’s floodwater storage and nutrient retention capacity. Variations in end states are caused by different degrees of river restoration and corresponding scale effects. In this way, respondents are not asked to value the river restoration measures per se, but rather their outcomes in order to avoid correlation due to causality. To increase the realism of the presented alternatives, respondents were shown existing river restoration plans on a map. The same map was used in the three case study applications: a 2000 CORINE land cover map at 1:100,000 scale, displaying the main level 3 4 ecosystem types (human settlements, agriculture, forests and meadows, wetlands and freshwater ecosystems). A main effects fractional factorial design was used to generate 32 different choice sets blocked in 8 different versions of 4 cards. Each respondent was thus shown 4 cards. The allocation of card sets across respondents was random. Each choice card consists of two unlabelled restoration alternatives (A or B) and a description of the status quo. Respondents were asked to choose between the two river restoration alternatives or the status quo situation (see the example card in Figure 1). The trade-off here is the price respondents pay as a private household for the presented public river restoration benefits on top of their current annual water bill. Hence, the estimated welfare measure is individual WTP to secure the river restoration benefits (compensating surplus). If respondents choose the current situation, they obviously forego restoration benefits and the cost price is zero. An overview of the attributes utilised and their corresponding levels is presented in Table 1. Table 1: Overview of attribute levels Level Flood return period Water quality Cost price (€/household/year) Baseline Once every 5 years Moderate 0 1 Once every 25 years Good 3 2 Once every 50 years Very good 10 3 Once every 100 years 30 4 50 5 Figure 1: Example of a choice card Flood frequency Option A Option B Current situation Once every 25 years Once every 25 years Once every 5 years Good Very good Moderate €3 € 10 €0 Water quality Increase in water bill I prefer: Option A Option B Neither Flood risk is defined as the flood return period. Currently, the study areas face severe highimpact flooding approximately once every 5 years (as opposed to regular low-impact annual floods). As a result of river restoration, this return period can be reduced to once every 100 years. Water quality is described in categorical terms and explained with the help of the water quality ladder created by Resources for the Future (RFF) in the United States (Carson and Mitchell, 1993). The ladder categorises water quality into recreational uses such as ‘swimmable’, ‘boatable’ and ‘fishable’ on the one hand and illustrates levels of biological diversity of aquatic life on the other. Coloured pictograms were used to visualise the water quality levels (where red reflects poor water quality conditions, yellow moderate, blue good and green very good water quality conditions). Based on expert consultation, moderate water quality levels were chosen as the baseline category (i.e. the current situation). The monetary 6 attribute was specified as an increase in the household water bill. Payment levels varied from 3 to 50 Euros per year (with the corresponding monthly amounts also presented between brackets). In the case of Hungary and Romania these values were converted into national currencies (Hungarian Forint and Romanian Lei). 3. Econometric model The choice model used here has its roots in random utility theory (e.g. Ben-Akiva and Lerman, 1985). The Multinomial Logit Model (MNL) is the most commonly used structure for choice models. The MNL model assumes that the random components of the utility of the alternatives are independently and identically (i.i.d.) Gumbel distributed with a type I extreme value (EV) distribution (Train, 2003). The random parameters or mixed logit model is more flexible than the MNL model and relaxes the assumption of independence of irrelevant alternatives (IIA) as a result of the iid property, and allows – among other aspects – for (random) preference heterogeneity. Any random utility model can be approximated by a mixed logit model (McFadden and Train, 2000). The model is defined based on the functional form for its choice probabilities. The standard indirect utility function underlying the mixed logit model is: (1) U ij = Vij + ε ij = β i X ij + ε ij where Uij refers to the utility of individual i obtained from choice alternative j, Vij is the measurable component of utility measured through a vector of utility coefficients β associated with a vector of observed attribute and individual characteristics Xij, and εij captures the unobserved influences on an individual’s choice with an iid extreme value distribution. The utility coefficients β vary according to individual (hence βi) with density f(β). This density 7 can be a function of any set of parameters, and represents in this case the mean and covariance of β in the sample population: (2) U ij = β X ij + f ( β ) X ij + ε ij In equation 2, β is a vector of fixed parameters and f(β) the vector of standard deviation parameters, reflecting the deviation from mean β. Thus instead of being fixed across individuals (as in the MNL model), β represents a random variable. Based on the choice design used in this case study, equation (1) can be rewritten as (ignoring the notation of random terms for the sake of simplicity): Vijl = β 0l + β 1l Flood ijl + β 2l Quality ijl + β 3l Pr iceijl + ε ijl (3) where β0 is the alternative specific constant (ASC) and β1 to β3 refer to the vector of coefficients related to the attributes flood return period (Flood), water quality (Quality) and cost price (Price). The subscript l represents the specific country in which the survey was conducted (l=1,2,3). The null hypothesis tested here is that the part-worth utilities associated with the different attributes are uniform across the three different river basins. Choice behaviour is expected to be related negatively to the flood return variable (i.e. the lower the flood return period, the higher the probability of choosing the river restoration alternative), and positively to the water quality variable (the higher the quality level, the higher the probability of choosing the river restoration alternative). The cost price is expected to have a negative effect on choice behaviour (the higher the price, the lower the probability of choosing the river restoration alternative). 8 The conditional choice probability L that individual i prefers choice alternative j (if ε is i.i.d. EV distributed) can be expressed in terms of the following logistic distribution (McFadden, 1974): Lij = e λβ i X ij ∑e (4) λβ i X j j∈C Because βi is not observed, the unconditional choice probability P equals the integral of equation 4 over all possible variables βi (Train, 2003): λβ X e i ij Pij = ∫ λβ X ∑ e i j j∈C f ( β ) dβ (5) In equations 4 and 5, λ is a scale parameter, typically assumed to be 1, implying constant error variance, and C is the choice set. For the purpose of facilitating international comparison of the estimated choice models, we explicitly test the equality of scale parameters across countries using a Full Information Maximum Likelihood (FIML) estimation procedure (Hensher, 1986). The scale factor and model parameters are confounded (Vij=λ[β0+ β0X1 + ...+βnXn]); that is, the model estimates equal β/σε , making it impossible to attribute model differences to differences in parameter estimates or scale (Swait and Louviere, 1993). Even if the model parameters from the three countries are truly identical, this may be obscured by variance inequality, and if the scale parameter varies by respondent and/or country, this may significantly affect predicted probabilities (Louviere et al., 2002). 9 The inclusion of a monetary attribute in the choice model allows for the estimation of monetary Hicksian welfare measures for different river restoration policy scenarios and changes in individual components of these scenarios (e.g. Hensher et al., 2005). These monetary welfare measures form the basis for the international comparison and transferability tests (e.g. Morrison et al., 2002). The marginal rate of substitution (MRS) for a change in one of the river restoration attributes, for example flood return period, is calculated as the negative ratio of the first derivatives of the flood return period (β1 in equation 3) to the cost price (β3 in equation3). The price attribute in the denominator is interpreted here as the marginal utility of income. The MRS hence equals in this example marginal WTP for a reduction in the flood return period. Besides testing the equality of these marginal (implicit) prices across river basin countries, we also compare and test the equality of the estimated compensating surplus (CS) of a number of policy scenarios, i.e. foreseen discrete changes in flood risk and water quality as a result of river restoration in Austria, Hungary and Romania, accounting for demographic and socio-economic factors. 4. Case study description The Danube is the second largest river in Europe. It originates in the Black Forest in Germany and flows through 10 Central and Eastern European counties before emptying into the Black Sea. The river has a length of 2,780 km, and a catchment area of more than 800,000 km2. Although some parts of the river still are in a near-natural state, most river stretches have been classified as heavily modified due to embankment and regulation works, and intensive navigation. The shape of the river has been drastically changed and large parts of the associated formerly-waterlogged area have been drained for agricultural purposes, reducing the connectivity between the Danube and the surrounding area and its tributaries to small patches. The structure and state of the riparian zone directly influence the biological and 10 hydro-morphological quality elements of the river. Various river restoration projects have been identified in the Danube river basin. Three of them were included in this study (Figure 2): the Donau-Auen National Park in Austria, Által-ér in Hungary, and the Islands of Braila in Romania. These case studies will be briefly described below. Figure 2: Location of the Danube river restoration projects The Austrian Danube National Park is located east of Vienna. It is a green ribbon floodplain area with a length of 35 km, linking Vienna and Bratislava. The national park covers an area of 93 km2 and is a complex ecosystem with an enormous diversity of habitats, plants and animals. In some parts, the floodplain is still to a high degree ecologically intact, displaying the characteristics of a large stream. Within and around the floodplain, the main economic activities include agriculture, forestry and fishing. Overall, water quality can be classified as moderate to good. The main source of pollution is the wastewater treatment system of 11 municipalities upstream (notably the City of Vienna) although the latter has been improved substantially over the past decades. Parts of this stretch of the Danube are classified as heavily modified water bodies, especially since the river bed was canalized and the construction of the hydropower station of Freudenau (within the city limits of Vienna downstream to the east) further changed the free-flowing character of the river. Achieving the WFD objectives along the Danube and in particular around Vienna is uncertain due primarily to non-point source pollution (Institute for Water Quality, 2008). The Hungarian Által-ér is a tributary of the Danube which has been partly disconnected from the main river over the past centuries. The watershed covers an area of 520 km2 and the Általér river has a length of approximately 50 km. The river’s watercourses are highly regulated and run in artificial riverbeds. A series of dams were constructed in the past to assure water supply. The main sources of diffuse pollution are wastewater from small-scale domestic animal husbandry and manure and fertilizer runoff from arable land. The most serious water quality problems are found in and around the industrialized cities of Tatabánya, Vértesszőlős and Tata due to emissions of heavy metals and nutrients. Most of the (often illegal) municipal landfills located in this area also contain dangerous substances, which leak into surface and groundwater bodies. Although monitoring data in the watershed are inadequate to assess current water quality levels, overall water quality in the watershed is rated as ‘moderate’ based on expert judgment. Finally, the Romanian Braila Islands are part of the former Inland Danube Delta. Braila City is the most important urban centre in this area and has approximately 215 thousand inhabitants. Land use is predominantly agricultural (approximately three quarters of the total area), while industry is concentrated in Braila City. The Danube river has been classified in this section (195 km in length) as heavily modified due to hydro-technical works and dredging of more than 20 percent of the river bed for the purpose of navigation. The wetlands from the Small Island of Braila Natural 12 Park (210 km2) and the floodplains between the riverbanks and dikes are the main remnants of the natural floodplains. Almost 80 percent of the whole area has been drained for agricultural purposes. As a consequence, connectivity between the Danube and the floodplains is very limited. Water quality in this stretch of the Danube river is rated as moderate. The main pollution sources are agriculture, industry, navigation, and domestic households. On the other hand, the Small Island of Braila is especially rich in bird species: 136 different species have been identified, of which 47 are listed in the annex of the EU-Bird Directive. Together with the coastal Danube Delta, the wetland system is an important stepping stone for bird migration routes in SouthEastern Europe. 4. Survey design and implementation The questionnaire was developed during several meetings of the international project group over a six month time period and subsequently pre-tested in each of the three river basin countries. After each pre-test, the results were shared to improve the common valuation design. Although the questionnaire was identical in all three countries, survey administration differed. Interviewing took place face-to-face in Hungary and Romania, while a web-based survey was conducted in Austria. For the pre-test in Austria, 526 people were recruited randomly in Vienna, of whom 109 completed the web-based questionnaire (a response rate of 21%). In addition, 15 questionnaires were sent to water experts with whom face-to-face interviews were conducted to test the structure and wording of the questionnaire. In Hungary, the pre-test consisted of 32 face-to-face interviews in Tata, one of the largest towns in the Által-ér catchment, where approximately 20 percent of the catchment population live. The questionnaire was pre-tested in two rounds in Romania, targeting a total of almost 100 local residents living in different villages throughout the case study area and the city of Braila. After each pre-test, minor changes were introduced to the structure and wording of the 13 questionnaire. Special attention was paid to respondents’ understanding of the choice experiment and the credibility of combination of alternatives. The final questionnaire consists of 37 questions, most of which are close-ended (multiple choice), and is divided into three main parts. The first part of the questionnaire contains questions about respondents’ general perception of water related issues, including water quality and flood experience. People were also asked about their recreational activities, e.g. how often they visit the case study area and whether they would visit more often if water quality was improved. The second part consists of the CE, while the third part of the questionnaire is designed to collect information on standard respondent demographic and socio-economic characteristics. The main survey was carried out simultaneously in all three countries in November 2007. In Austria, the main survey targeted a random sample of 1,977 households from a representative household market panel in Vienna and Lower Austria. The response rate was 26 percent (n=506). Stratified sampling procedures were followed in Hungary and Romania, based on gender, age and representative shares of the rural and urban population living along both sides of the Danube river. In Hungary, 892 people were asked to participate in the survey, of whom 471 agreed (a response rate of 53%), while in Romania 519 of the 850 respondents who were asked to participate completed the questionnaire (a response rate of 61%). 14 5. Results 5.1. Sample characteristics Sample characteristics across the three subsamples are summarized in Table 2. Generally, the samples are fairly representative compared to the underlying national population from which they were drawn. Men are slightly overrepresented in Hungary (the share of the male population at national level is 47%), but overall an equal number of men and women are included. The average age is also more or less the same across countries (41 in Austria, 43 in Hungary and 44 in Romania), although minor differences are found in age structure and distribution. Some degree of self-selection may have played a role in Hungary and Romania in view of the relatively high share of respondents with higher (secondary and higher) education in these two sub-samples (26%) when compared to Austria (10%). As expected, pronounced differences are found when comparing disposable household income between the three sub-samples. Mean annual household income in the Austrian sample (€22,025) is two and a half times higher than in Hungary (€8,925) and four times higher than in Romania (€5,590). As can be seen from Table 2, most Romanian households earn less than 500 Euros per month, while around 30 percent of respondents in Hungary and less than 10 percent in Austria fall into the lowest income category. Most Hungarian respondents earn between 500 and 1,000 Euros per month, and most Austrian respondents between one and two thousand Euros per month. Average household income in the Austrian sample is slightly higher than the national average, while average household income in Hungary and Romania conforms more or less to the national average in these respective countries (Eurostat, 2008). Almost 10 percent of the Austrian sample are members or donate 15 to an environmental organization. This share is substantially lower in Hungary (5%) and Romania (<1%). Table 2: Respondent demographic and socio-economic characteristics Relative distribution (%) Austria Hungary Romania 48 55 51 =< 19 year 10 6 3 20-29 18 20 16 30-39 21 19 22 40-49 22 17 21 50-59 18 20 21 >= 60 year 11 18 17 €0-500/month 8 29 68 €500-1000 13 46 25 €1000-2000 38 23 6 €2000-3000 24 1 1 > €3000/month 17 1 0 Share male respondents Age groups Income groups Most respondents visited the study area where the river restoration measures are to take place. This share is lowest in Romania (47%) and highest in Hungary (88%). The average distance respondents live from the Danube river is lowest in the Hungarian sample (4 km), and highest in the Austrian sample (47 km). Romanian respondents live, on average, about 15 km from the Danube. In all three samples, a majority of respondents (80%) walk regularly along the 16 river. Fewer people fish or swim in open waters. The share of respondents who swim in open waters is highest in Austria (69%) and lowest in Hungary (22%). On the other hand, the share of respondents fishing in the Danube is lowest in Austria (9%), while respectively 23 to 40 percent of the Hungarian and Romanian sample fish in the river. Investigating flood experiences and water quality perception, Romanian respondents have least experience with floods. Less than 10 percent have ever experienced a flood during their lifetimes. This share is higher in Austria and Hungary (respectively 16 and 19%). Significant differences are found when examining the public perception of water quality (Figure 3). Figure 3: Public perception of Danube water quality in the three samples Based on the WFD water quality classification, almost three quarters of all Austrian respondents believe that current water quality is good to very good. Only 5 percent perceive water quality to be poor. In Hungary and Romania almost half of the sample classify current water quality as poor. Just over 15 percent think it is good and less than one percent rate it as very good. Hence, in the latter two countries a discrepancy exists between the expert 17 classification of current water quality (and thus the corresponding status quo description of water quality in the CE) and public perception. Around a quarter (Austria) to a third (Hungary and Romania) of the respondents perceive current water quality to be at the level it is described in the CE. We now turn to the CE results. 5.2. Marginal WTP for floodplain restoration and tests of international transferability Out of the more than 17,000 choice occasions in all three samples (6,000 in Austria, 5,445 in Hungary and 5,988 in Romania), the status quo was chosen in 22 percent of all cases. This percentage was lowest in Austria (17%) and highest in Romania (28%), presumably due to a lack of monetary resources to pay for the restoration benefits. No major differences were found between the three countries in terms of choices for option A or B. Option A was chosen in 36 percent of all the choice occasions and option B in 42 percent. Marginal WTP values for the attributes can be derived from the estimated choice model (see section 3). For this, a simple model including the design attributes only is estimated. The null hypothesis of equal model parameters across the three river basin countries, whilst permitting scale factors to differ between countries, is tested by pooling the datasets and jointly estimating the model and scale parameters in a nested logit model (using the FIML procedure in NLOGIT 3.0). The Likelihood Ratio (LR) test rejects the null hypothesis of equality of model parameters at the one percent confidence level (chi-squared (5) = 151.382). Following Swait and Louviere (1993), this also implies that the hypothesis of equality of scale parameters is rejected. So, we expect the marginal WTP values to be different due to the fact that both model parameters and variances differ between the three country specific models. The marginal WTP values and their standard errors, estimated using the delta method (Greene, 2003), are presented in Table 3. For the purpose of international comparison, the 18 values in Hungary and Romania were adjusted using the purchasing power of the Euro in Austria (e.g. Ready and Navrud, 2006)1. The flood risk attribute is only significantly different from zero in Austria, and indicates what a household is willing to pay, ceteris paribus, for a reduction of the flood return period by one year. The values for good and very good water quality equal what households are willing to pay for a change in water quality from moderate to good and very good conditions. In order to test for possible non-linearity, dummy coding is used for the categorical water quality levels (moderate water quality being the baseline category). Equality of dummy variables for the different water quality levels (within each country) was tested with the help of the Wald test. The null hypothesis of equal parameter estimates is convincingly rejected at the one percent confidence level in each country, reflecting sensitivity to scope2. Table 3: Attribute implicit prices in the three samples (€/household/year) Marginal change in: Austria Hungary Romania 0.20 0 0 (0.05) (not significant) (not significant) 44.5 21.2 23.0 (6.5) (3.1) (10.7) 75.3 42.5 36.8 (8.4) (4.2) (14.1) Flood risk Water quality conditions moderate → good moderate → very good Standard errors between brackets. 1 Purchasing power parities (PPP) were taken from the World Bank 2008 World Development Indicators. 2 Chi-square values (1 degree of freedom) are respectively 66.62 (p<0.001), 108.87 (p<0.001) and 6.25 (p<0.012) for Austria, Hungary and Romania. 19 Using the two one-sided t-test (Kristofersson and Navrud, 2005) to test whether the marginal values are equivalent between river basin countries at a significance level of α=0.05 and a transfer error tolerance level of 20%3, the null hypothesis of inequality is rejected for Hungary and Romania (test results are available from the authors). So, the PPP adjusted estimated marginal WTP values for both good and very good water quality are transferable between Hungary and Romania, but not between Austria and Hungary and Austria and Romania. In the latter case, the transfer errors range (depending on transfer between countries and the extent of the water quality improvement) between 44 and 110 percent. 5.3. Modelling preference heterogeneity for floodplain restoration Preference heterogeneity is picked up in the fixed and stochastic part of the estimated utility functions. The statistically significant results for the country specific random parameters logit models are presented in Table 4. For efficiency purposes, the model is estimated (in NLOGIT 3.0) using the Halton sequence in a quasi-Monte Carlo maximum likelihood simulation method (Bhat, 2001). Random effects were detected for the highest water quality attribute in each country. The standard deviation is highly significant in all three samples, suggesting that preferences are indeed heterogeneous and taste variations are partly random. The flood risk attribute is (as before) only statistically significant in Austria. Flood risk was included in this case as number of years (flood return period). The expected positive coefficient hence implies that an alternative is more likely to be chosen if the flood return period is lower (e.g. 100 instead of 50 years). The water quality attributes are statistically significant and positive in all countries (the impact of very good quality again being significantly higher than that of good quality based on the Wald test), except in Romania, where only very good water quality is 3 Following Kristofersson and Navrud (2005). Other levels are also possible. The tolerance level is the transfer error the researcher or policymaker is willing to accept. 20 significant. As expected, the cost attribute has a significant negative impact on choices in all the three countries. Table 4: Estimated random parameters choice models for the three samples Austria Hungary Romania Variable Coef. est. S.e. Coef. est. S.e. Coef. est. S.e. *** ASC 0.010 0.140 -0.685 0.202 0.234 0.097** Design attributes FLOODRISK 0.005 0.002*** -0.0004 0.002 0.009 0.006 *** *** GOOD QUALITY 1.353 0.171 0.837 0.191 0.198 0.148 VERY GOOD QUALITY 1.814 0.180*** 2.209 0.184*** 0.457 0.162*** COST -0.035 0.006*** -0.067 0.007*** -0.046 0.005*** Preference heterogeneity HIGHEDUC x ASC 0.484 0.178*** * VISITOR x ASC 1.184 0.627 0.905 0.195*** INCOME x COST 0.002 0.001* 0.010 0.001*** 0.014 0.002*** *** *** QPERCEP x GQ -0.011 0.003 -0.011 0.004 QPERCEP x VGQ -0.010 0.003** -0.006 0.003* -0.009 0.003*** FUTVISIT x GQ 0.792 0.307*** 0.325 0.197* 0.503 0.170*** *** FUTVISIT x VGQ 1.321 0.320 0.565 0.183*** DIST x GQ -0.005 0.002*** DIST x VISITOR -0.006 0.003* DIST x FLOODAFFECT -0.047 0.020** Standard deviation VERY GOOD QUALITY 1.211 0.437*** 1.591 0.325*** 1.623 0.334*** VISITOR 2.134 0.817*** Model fit LL -1884.105 -1499.563 -1801.288 LR TEST 238.172 p<0.001 613.111 p<0.001 626.2394 p<0.001 ADJ R2 0.139 0.167 0.059 N 2000 1815 1996 *** 1% significance ** * 5% significance 10% significance Preference heterogeneity was accounted for in the fixed part of the estimated utility functions through interactions with the ASC and the design attributes. In the former case, whether or not a respondent had ever visited the area where the river restoration is planned (VISITOR) and whether or not a respondent has a higher secondary education degree (HIGHEDUC) have a significant positive impact on choices in Austria and Hungary. These results show which 21 characteristics of respondents increase the likelihood of being in favour of river restoration compared to maintenance of the status quo or no restoration policy. VISITOR also appeared to have a significant standard deviation, suggesting random taste variation around this variable. Significant interaction terms with the design attributes include disposable household income (INCOME) and cost price. As expected, in all countries higher income groups are more likely to choose one of the two river restoration alternatives at a higher price than lower income groups. Other interaction terms are respondent perception of current water quality (QPERCEP) and the value they attach to water quality improvements (good quality GQ and very good quality VGQ) and whether a respondent would visit the case study sites more often in the future (FUTVISIT) if water quality were improved. In the first case, the negative sign indicates that respondents who already perceive water quality as good value a water quality improvement less. In the latter case, the positive sign tells us that those respondents who said they would visit the area more often if water quality were improved are more likely to pay for river restoration than respondents who said their visiting frequency would not change as a result of any water quality changes. Additionally, in all three samples significant distance-decay (DIST) effects were found4, albeit through different interaction terms. In Austria, a significant distance-decay effect is found if water quality is improved to good status (GQ), while distance-decay is only significant in Romania for users (VISITOR), not for non-users. In Hungary, distance-decay 4 Distance is defined as the one-way distance reported by respondents from their home to the Danube River due to lack of sufficient variation in the available postal codes in the surveyed rural areas in this case study needed to calculate the distance from a respondent’s home to the river using GIS). 22 plays a significant role for respondents who have been affected by flooding (FLOODAFFECT). Those affected by flooding, but living further away from the river are less likely to choose one of the river restoration alternatives. The distance-decay functions demonstrate spatial correlation in choice behaviour. We return to these spatial relationships in section 5.5 when aggregating the welfare effects of the policy scenarios presented in the next section across the population of beneficiaries using GIS. 5.4. Compensating surplus for floodplain restoration and tests of international transferability Based on the statistically best fit models presented in the previous section, a number of policy scenarios were simulated and their welfare implications estimated, changing flood frequency and water quality simultaneously. The compensating surplus (CS) is estimated using the standard Hanemann formula (e.g. Bennett and Blamey, 2001): -1/βcost (V0-V1) where V0 and V1 are linear combinations of attribute levels in the estimated utility functions in the current situation and a new policy scenario respectively. To what extent these CS measures are transferable across the three river basin countries is tested again using the two one-sided t-test (α=0.05; tolerance level=20%). The estimated welfare measures for five different policy scenarios in each river basin country are presented in Table 5. As before, standard errors needed to calculate the 95 percent confidence intervals are estimated using the delta method and the values are adjusted for differences in purchasing power. Two policy scenarios involve the improvement of water quality to a good ecological status, with flood risk variations of once every 25 and 50 years, and three policy scenarios involve water quality improvements up to very good ecological status with flood risk reductions varying from once every 25 years to once every 100 years. Average values were used for respondent characteristics, including income and distance, 23 assuming that the samples are more or less representative for the entire river basin in each country. Table 5: Compensating surplus welfare measures for different policy scenarios (€/household/year)1 Policy scenario 1 2 3 4 5 1 CS Flood risk Water quality Once every 25 yrs Good Once every 50 yrs Once every 25 yrs Once every 50 yrs Once every 100 yrs Austria Hungary Romania 69.6 20.2 4.8 (37.5-101.8) (15.2-25.3) (4.0-5.5) 73.5 20.2 4.8 (40.8-106.2) (15.2-25.3) (4.0-5.5) 85.0 32.1 9.5 (50.2-119.8) (26.3-37.9) (6.8-12.2) 88.8 32.1 9.5 (53.4-124.2) (26.3-37.9) (6.8-12.2) 96.4 32.1 9.5 (59.6-133.3) (26.3-37.9) (6.8-12.2) Good Very good Very good Very good 95% confidence intervals between brackets. In Austria, the CS welfare measures for the policy scenarios increase gradually, reflecting sensitivity to scope. Keeping water quality constant, a reduction in the flood return period yields a higher CS. As expected based on the results shown in Table 4, the same applies when keeping flood risk constant: the CS is significantly higher for a higher water quality level. The results for Hungary and Romania only differ due to water quality changes (higher quality yields, as expected, a higher value) in view of the fact that flood risk is not a significant 24 determinant of choice behavior. The estimated CS therefore are the same when keeping water quality levels constant and varying flood risk. The CS in Austria are significantly higher than the CS for the same policy scenario in Hungary, while the CS in Hungary are significantly higher than the CS in Romania for each policy scenario. Confidence intervals are also largest for Austria. Based on the standard error (σ), the variation coefficient underlying the estimated CS (σ/CS) is, on average, twice as high in Austria compared to Hungary and Romania. Generally, a higher inaccuracy of estimation results is expected to result in more problematic transfers. Based on the results in Table 5, transfer errors vary between 62 and 263 percent for transfers between Austria and Hungary (average 138%), 89 and 1442 percent for transfers between Austria and Romania (average 580%), and 70 and 325 percent for transfers between Hungary and Romania (average 172%) (see Table 6). Table 6: Performance of the two main value transfer methods for different policy scenarios (transfer errors in %) Policy AU to HU HU to AU AU to RO RO to AU HU to RO RO to HU scenario UAT AVT UAT AVT UAT AVT UAT AVT UAT AVT UAT AVT 1 244 125 71 55 1362 482 93 66 325 52 76 52 2 263 136 72 58 1442 519 94 67 325 52 76 52 3 165 76 62 39 792 276 89 60 237 39 70 49 4 177 83 64 42 832 294 89 62 237 39 70 49 5 200 97 67 47 912 331 90 65 237 39 70 49 210 103 67 48 1068 380 91 64 272 44 73 50 Average AU: Austria; HU: Hungary; RO: Romania; UAT: Unadjusted Value Transfer; AVT: Adjusted Value Transfer 25 Transfer errors are reduced substantially when transferring the entire choice model from one river basin country to another instead of the predicted unadjusted average values in Table 5. The average respondent characteristics in the river basin country to which the estimated model is transferred are now used to predict the CS for each policy scenario: CS 1 = βˆ 2 X 1 (where β̂ is a vector with coefficient estimates, X the vector with relevant design attributes and respondent characteristics; 1 is the country to which the model is transferred, 2 the country in which the model was originally estimated). The errors associated with this ‘value function approach’ (from country 1 to country 2 and the other way around) are also presented in Table 6 (labeled as adjusted value transfer or AVT). On average, errors are reduced by 45 percent based on AVT as compared to unadjusted value transfer (UAT) for transfers between Austria and Hungary, 62 percent for transfers between Austria and Romania and 73 percent for transfers between Hungary and Romania. However, the predicted values remain nontransferable based on the outcome of the two one-side t-test if an error tolerance level of 20 percent is used (test results are available from the authors). Errors are lowest when transferring the choice models between Hungary and Romania (39-52%), and highest when transferring the models between Austria and Romania (60-519%). To what extent these errors are acceptable in a cost-benefit policy evaluation will depend on policy-maker demand for accurate estimates and the stage of the policy and decision-making cycle for which the estimated CS are used. 5.6. Welfare aggregation using GIS A final step in the welfare estimation procedure is the aggregation of the estimated CS across the population benefiting from the welfare gains associated with the river restoration policy scenarios. This step is often critical to arrive at a valid and reliable (and hence credible) estimation of total economic value (TEV). Most studies simply use the number of people 26 living in an administrative unit or geographic jurisdiction (e.g. county, province, state or country), and average values are transferred unconditionally (i.e. uncorrected) across the population living within the boundaries of this geographical unit. Depending on area size and population density, aggregated TEV can differ enormously (Bateman et al., 2006). Strict guidelines for welfare aggregation do not exist, making the estimation procedure vulnerable to manipulation. In view of the fact that the population from which the samples in this study were drawn and their characteristics are unevenly distributed over space, the aggregation procedure is carried out using GIS. To this end, GIS data about Europe’s major rivers including the Danube from the 2008 ESRI database were combined with (1) the JRC 100 by100 meter population density grid (Gallego, 2008) and (2) NUTS-3 level information about per capita income disaggregated from NUTS3 regions to 100 x 100 m grid cells5. Euclidian distances were calculated per 100 by 100m grid cell to the Danube river in meters. A TEV is calculated for two policy scenarios: improvement of water quality in the Danube river to (1) good and (2) very good conditions (keeping flood conditions in both cases constant). The importance of accounting for preference heterogeneity in welfare aggregation procedures, in this case income and distancedecay for which secondary GIS data were available and additional calculations could be made, will be illustrated by comparing the outcomes of two different TEV approaches: TEV calculated by aggregating the unadjusted average values across the whole population living within the boundaries of the administrative units surveyed in this case study (TEV1)6 and TEV 5 The Nomenclature of Territorial Units for Statistics (NUTS) is a breakdown of territorial units to harmonize regional European statistics. NUTS-3 is the lowest aggregation level, and usually follows a European member state’s own regional administrative structure. 6 The administrative units surveyed in this case study were the states of Lower Austria and Vienna in Austria, Komárom-Esztergom County in Hungary and the counties Braila, Constanta, Ialomita and Tulcea in Romania. 27 adjusted for the estimated income and distance-decay effects (TEV2). In the former case, the sample population and the estimated CS per capita are assumed to represent the population at large in the administrative units. The steps in the latter aggregation procedure are summarized below. First, information about population density was converted to number of inhabitants per 100x100m grid cell. Second, the average CS for reaching good and very good water quality with the help of river restoration were converted from per household to per capita values based on each sample’s average household size, and multiplied with the number of people in each 100 by 100m grid cell. For each grid cell also average per capita income was determined based on the specific geographical NUTS-3 area to which a grid cell belonged. This average per capita income was subtracted from mean per capita income in the sample and multiplied on a cell by cell basis with the estimated income coefficient for each country and the number of people living in each cell7. In this way, the economic value per grid cell was modified upwards or downwards depending on the income difference. The distance of each grid cell to the Danube river was used to correct the economic value per capita per grid cell for the distance-decay effects detected in each country8. The estimated distance-decay factor was multiplied by the calculated distance of each grid cell from the river in kilometres and multiplied by the number of people (illustrated for good water quality in Austria in Figure 4). In a final step, the income and distance adjusted values (in the case of Hungary only distance adjusted) are added up to estimate the TEV of river restoration to good 7 Except in Hungary. The area surveyed in Hungary falls within the boundaries of a single NUTS-3 region, so no income variation can be found within this area based on the available GIS data. 8 Modified for the percentage of respondents who visited the case study area in Romania and the percentage of flood affected households in Hungary (see Table 4). 28 and very good water quality (illustrated for very good water quality in Romania in Figure 5). The results of the two aggregation procedures are presented in Table 7. The values are adjusted for differences in purchasing power between the three countries. Figure 4: Illustration of distance-decay effects for good water quality in Austria The most striking observation from Table 7 is the considerable difference between TEV1, calculated by simply multiplying the number of inhabitants living inside the administrative units, and TEV2 where the CS is adjusted for (a) distance-decay only and (b) distance-decay and income differences. On average, accounting for distance-decay yields a 30 percent lower TEV than the unadjusted TEV, while the additional income correction reduces TEV by another 10 percent. These differences are most pronounced for Romania. Accounting for distance-decay results in Romania in a 35 to 50 percent lower TEV (for very good and good water quality respectively). To put the statement the 29 other way around, not accounting for distance-decay overestimates the TEV by a factor of 1.5 to 2. Based on the available NUTS-3 information about average income levels in the study area (which were lower than the sample average possibly due to the fact that the sample consisted of a relatively high share of more higher educated respondents), TEV is reduced by an additional 9 to 12 percent (for very good and good water quality respectively). Figure 5: Illustration of TEV for very good water quality in Romania adjusted for distance-decay In Hungary, unadjusted TEV is 42 percent higher than TEV adjusted for distance-decay in the case of good water quality, and 17 percent in the case of very good water quality. No income correction 30 could be carried out because the survey area falls completely within one NUTS-3 region, and so there was no variation in income levels within that area. Table 7: TEV for good and very good water quality based on two aggregation procedures (€106/year) TEV1 TEV2 unadjusted aggregation GIS adjusted aggregation Distance-decay Distance-decay and correction income correction Austria Good water quality 54.0 46.3 41.9 Very good water quality 73.8 na 69.4a Good water quality 2.0 1.4 na Very good water quality 3.9 3.3 na Good water quality 10.2 5.0 4.4 Very good water quality 15.3 9.9 9.0 Hungary Romania a Income correction only. na: not available. In Austria, accounting for distance-decay produces a TEV for good water quality that is 14 percent lower than the unadjusted TEV. Because the distance-decay effect was only significant when included as an interaction term with good water quality (see Table 4), no adjusted TEV for very good water quality is presented for Austria in Table 7. Adjusting TEV also for the spatial variation in income levels across the survey area lowers TEV by another 10 percent. The sample income average was somewhat higher than the average in the whole survey area, possibly due to the overrepresentation of respondents from Vienna where average income levels are relatively high 31 compared to the rest of the country. The income adjusted TEV for very good water quality is 6 percent lower than unadjusted TEV. 6. Discussion and conclusions The main objective of this paper was to estimate the welfare impacts of the joint benefits generated by ecological restoration of heavily modified river stretches (i.e. flood risk reduction and water quality improvement) and test their transferability across the second largest river basin in Europe. The welfare impacts were measured through a stated choice experiment, carried out in three different countries, using an identical research format. The common experimental design was developed to enable direct comparison of the non-market benefits of WFD implementation. Recommendations to develop and adopt a common approach for water quality valuation go back to the 1992 special issue on value transfer in Water Resources Research (Desvousges et al., 1992), but their use is rare and in stated preference research mainly limited to contingent valuation (e.g. Brouwer and Bateman, 2005). One of the most important challenges faced during the development of the common valuation design was to keep the design generally applicable to all three countries. This meant compromising on the inclusion of case study specific detail while keeping the design at the same time sufficiently meaningful to policy makers and lay public in each country. This may also have been a reason why flood risk, often a poorly understood concept in valuation research (e.g. Loomis and duVair, 1993), was only a significant determinant of public choice in one of the three river basin countries. The estimated choice models in the three countries are significantly different, both in terms of model parameters and variances, indicating that their potential for transferability is limited. Observed and unobserved preference heterogeneity plays a significant role in the three 32 countries, and is an important reason for the non-transferability of the results. The PPP adjusted marginal values for water quality improvements were the same in Hungary and Romania, but not the PPP adjusted CS estimates for different floodplain restoration policy scenarios. None of the estimated CS measures were transferable at the 20 percent error tolerance level imposed. When controlling for preference heterogeneity, most of the transfers also exceeded the 50 percent tolerance level (transfer errors were lower than 50% in only 30% of the cases), even through transfer errors were reduced by up to almost 75 percent. Hence, the ‘function approach’ to value transfer clearly outperformed the ‘unadjusted value approach’, but transfer errors remained relatively high, varying on average between 48 and 380 percent for different floodplain restoration scenarios. Accounting for preference heterogeneity due to distance-decay and income effects in the welfare aggregation procedure of the CS estimates across the population of beneficiaries also consistently resulted in lower TEV across the three countries. Compared to unconditional aggregation of average CS estimates over the population living within the boundaries of the administrative units from which the survey samples were drawn (common practice in many value transfer studies), the calculated TEV turned out to be 10 to 50 percent lower in the more sophisticated GIS-based aggregation procedure presented in this paper. In conclusion, the welfare impacts of floodplain restoration in the context of the WFD differ significantly throughout the Danube river basin, i.e. within and between the countries through which the river flows. Errors are substantial when using country-specific values and models to predict values elsewhere, even if one takes preference heterogeneity and different socioeconomic conditions in the three countries that were part of this study, such as purchasing power, into account. Similarly, substantial error may be created if average values are aggregated in the same country without controlling for preference heterogeneity and socio- 33 economic conditions that are unevenly distributed in space. As such, both sources of error will affect the outcome of the policy analysis for which the values are to be used. Acknowledgements This study was carried out as part of the EU DG Research funded project AquaMoney (SSPI022723) (www.aquamoney.org). The useful comments of the project group on the study design and results presented at several project meetings are gratefully acknowledged. 34 References Bateman, I.J., Day, B.H., Georgiou, S. and Lake, I. (2006). The aggregation of environmental benefit values: Welfare measures, distance decay and total WTP. Ecological Economics, 60(2): 450-460. Ben-Akiva, M. and Lerman, S.R. (1985). Discrete choice analysis: Theory and application to travel demand. The MIT Press, Cambridge Massachusetts. Bennett, J. and Blamey, R. (2001). The choice modelling approach to environmental valuation. Edward Elgar, Cheltenham, UK. Bhat, C.R. (2001). Quasi-random maximum simulated likelihood estimation of the mixed multinomial logit model. Transportation research Part B, 35: 677-693. Brander, L.M., Florax, R.J.G.M. and Vermaat, J.E. (2006). The empirics of wetland valuation. A comprehensive summary and meta-analysis of the literature. Environmental and Resource Economics, 33(2): 223-250. Brouwer, R., Langford, I.H., Bateman, I.J. & Turner, R.K. (1999). A Meta-Analysis of Wetland Contingent Valuation Studies. Regional Environmental Change, 1(1): 47-57. Brouwer, R. and Bateman, I.J. (2005). Benefits Transfer of Willingness to Pay Estimates and Functions for Health-Risk Reductions: A Cross-Country Study. Journal of Health Economics 24: 591-611. 35 Brouwer, R. (2008). The potential role of stated preference methods in the Water Framework Directive to assess disproportionate costs. Journal of Environmental Planning and Management, 51(5): 597-614. Carson, R. T. and Mitchell, R.C. (1993). The value of clean water: the public’s willingness to pay for boatable, fishable and swimmable quality water. Water Resources Research, 29(7): 2445-2454. Desvousges, W.H., Naughton, M.C. and Parsons, G.R. (1992). Benefit transfer: conceptual problems in estimating water quality benefits using existing studies. Water Resources Research, 28(3): 675-683. Gallego, F.J. (2008). Downscaling population density in the European Union with a land cover map and a point survey. Unpublished document. IPSC, JRC, ISPRA. Downloaded 12 Oct. 08 from http://dataservice.eea.europa.eu/download.asp?id=18555&filetype=.pdf Greene, W.H. (2003). Econometric Analysis. Upper Saddle River (NJ): Prentice-Hall International. Hanley, N., Wright, R. E.and Alvarez-Farizo, B. (2006). Estimating the economic value of improvements in river ecology using choice experiments: an application to the water framework directive. Journal of Environmental Management, 78(2): 183-93. Hensher, D.A. (1986). Sequential and Full Information Maximum Likelihood Estimation of a Nested Logit Model. The Review of Economics and Statistics, 68(4): 657-667. 36 Hensher, D.A., Rose, J.M. and Greene, W.H. (2005). Applied choice analysis. A primer. Cambridge University Press, Cambridge. Hohensinner, S., Habersack, H., Jungwirth, M., and Zauner, G. (2004). Reconstruction of the characteristics of a natural alluvial river-floodplain system and hydromorphological changes following human modifications: the Danube River (1812-1991). River Research and Applications, 20(1): 25-41. ICPDR (2007). Significant water management issues in the Danube river basin district. Document number IC 132, International Commission for the Protection of the Danube River, Vienna, Austria. Institute for Water Quality (2008). Wassergüte der Donau. Federal Agency for Water Management, Vienna, Austria. Kristofersson, D. and Navrud, S. (2005). Validity tests of benefits transfer — are we performing the wrong tests? Environmental and Resource Economics 30, 279–286. Loomis, J.B. and duVair, P. (1993). Evaluating the effect of alternative risk communication devices on willingness to pay: Results from a dichotomous choice contingent valuation experiment. Land Economics 69: 287-298. Louviere, J.J., Hensher, D.A. and Swait, J.D. (2000). Stated choice methods: analysis and application. Cambridge University Press, Cambridge. 37 McFadden, D. (1974). Conditional logit analysis of qualitative choice behaviour. In: Zarembka, P. (ed.). Frontiers in econometrics. Academic Press, New York, pp. 105-142. McFadden, D. and Train, K, (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics, 15: 447-470. Morrison, M. and Bergland, O. (2006). Prospects for the use of choice modelling for benefits transfer. Ecological Economics, 60: 420-428. Porras, I., Grieg-Gran, M. and Neves, N. (2008). All that glitters. A review of payments for watershed services in developing countries. Natural Resource Issues No. 11. International Institute for Environment and Development. London, UK. Ready, R. and Navrud, S. (2006). International benefits transfer: Methods and validity tests. Ecological Economics, 60: 429-434. Schiemer, F., Baumgartner, C. and Tockner, K. (1999). Restoration of floodplain rivers: The Danube restoration project. Regulated Rivers: Research & Management, 15(1-3): 231-244. Swait, J. and Louviere, J. (1993). The role of the scale parameter in the estimation and comparison of multinomial logit models. Journal of Marketing Research, 30 (3): 305-314. Train, K. (2003). Discrete Choice Methods with Simulation Cambridge University Press, Cambridge. 38 Troy, A. and Wilson, M.A. (2006). Mapping ecosystem services: Practical challenges and opportunities in linking GIS and value transfer. Ecological Economics, 60: 435-449. 39 View publication stats