Valuing river characteristics using combined site choice and participation travel cost models

Journal of Environmental Management 80 (2006) 237–247 www.elsevier.com/locate/jenvman Valuing river characteristics using combined site choice and participation travel cost models C Johnstone a,*, A Markandya a,b a Department of Economics and International Development, University of Bath, Bath, UK b Fondazione Eni Enrico Mattei (FEEM), Milan, Italy Received September 2004; received in revised form 19 July 2005; accepted 29 August 2005 Available online 27 December 2005 Abstract This paper presents new welfare measures for marginal changes in river quality in selected English rivers The river quality indicators used include chemical, biological and habitat-level attributes Economic values for recreational use of three types of river—upland, lowland and chalk—are presented A survey of anglers was carried out and using these data, two travel cost models were estimated, one to predict the numbers of trips and the other to predict angling site choice These models were then linked to estimate the welfare associated with marginal changes in river quality using the participation levels as estimated in the trip prediction model The model results showed that higher flow rates, biological quality and nutrient pollution levels affect site choice and influence the likelihood of a fishing trip Consumer surplus values per trip for a 10% change in river attributes range from £0.04 to £3.93 (£2001) depending on the attribute q 2005 Elsevier Ltd All rights reserved Keywords: Valuation; River quality; Angling; RUM Introduction The aim of this study is to provide new welfare estimates of use value for changes in river quality in the UK Recent cost-benefit analyses of large-scale environmental improvement projects such as the 4th Periodic Review of the Water Industry (PR04) Environment Programme have highlighted the need for more specific and up-to-date values for marginal changes in river quality The extensive benefits transfer carried out in previous cost-benefit analyses showed that, currently, values for angling are only available for broad-scale changes in quality of fishery, e.g ‘coarse-poor’ to ‘coarse good’, or ‘coarse-good’ to ‘gamemoderate’ However, where environmental improvements will result in specific outcomes such as reductions in phosphorous concentrations or increases in biodiversity, what is needed are values for marginal changes in these specific river attributes In addition, such specific marginal values will be useful in meeting * Corresponding author Address: Environment Agency, Economics, Rio House, Aztec West, Bristol BS32 4UD, UK Tel.: C44 1454 205580; fax: C44 1454 205566 E-mail address: claire.johnstone@environment-agency.gov.uk (C Johnstone) 0301-4797/$ - see front matter q 2005 Elsevier Ltd All rights reserved doi:10.1016/j.jenvman.2005.08.027 the new demands in the field of policy and environmental management in implementing the water framework directive (WFD) The main gaps in the literature that this study seeks to fill can therefore be summarised as follows: † lack of UK use values for marginal changes in a range of river quality indicators, e.g flow, species richness, nutrient pollution levels; † lack of UK use values for habitat-level physical river characteristics, such as extent of river modification; † lack of values for different types of rivers, e.g lowland, upland etc The aim of the study is therefore to generate new economic use values that would meet the needs of policy and project appraisal in valuing specific and marginal changes in environmental quality, for different river types The rest of this paper is structured as follows Section briefly outlines the background to the econometric models used Section describes the study area and ecological data used to measure river quality In Section the angling data gathering and some descriptive data for the study sample are provided In Section 5, the models and results are set out, and the welfare measures shown The results are discussed in Section 6, and in Section some conclusions are offered 238 C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 Methodology The study combined two types of revealed preference travel cost models commonly used in calculating welfare measures for changes in river quality These were a random utility site choice model (RUM) and a trip prediction or participation model Because RUM site choice models cannot predict total recreational trips taken in a season, researchers have proposed various methods for linking participation and site choice decisions in a single model Parsons et al (1999) compare four models for doing this The first, developed by Morey et al (1993) is a repeated nested logit model, where the participation decision is the first level, and site choice a second level The second approach uses the inclusive value index from the site choice model as an explanatory variable in the trip prediction model The last two are variations on this in that they split the inclusive value term into two separate price and quality terms, but differ from each other in the specification of the quality term1 In this study, the trip and site choice models are linked by substituting the actual number of trips in the site choice model with the predicted number of trips from a change in river quality as estimated in the participation model This expands the site choice model by embedding the participation decision inside it, and as such allows the researcher to estimate welfare gains from both site characteristics and trip behaviour A similar approach was originally proposed by Bockstael et al (1987), later modified by Hausman et al (1995) In Bockstael’s approach, the per trip welfare measure from the site choice model is multiplied by the total number of trips per season estimated in the participation model This study is the first application of such an approach to recreational use of rivers in the UK, and as such is expected to generate useful empirical results, which will help inform policy decisions for future environmental legislation such as the water framework directive Study area and ecological data 3.1 Study area The study area comprised a range of ecologically varied regions around England, and the spatial unit of analysis for the study was the river reach, as defined for water quality monitoring purposes by the Environment Agency In order to get a broad range of river types, the rivers for the study were selected from the natural areas/countryside character initiative characterisation system devised by English Nature and the Countryside Agency Fig below shows the geographical location of the study areas (natural area shadings randomly assigned) The study area was split into upland and lowland areas2 These broad categories were created to test whether significant differences in angling participation and choice existed for different types of river and in different parts of the country, and to produce more specific and policy relevant welfare estimates Table below shows the principal rivers and total number of river stretches for each study area Table Study area dataset of rivers and stretches Study area Principal rivers Berkshire and Marlborough downs Exmoor and quantocks Midland clay pastures Midland plateau South chalk South Devon Southern magnesian limestone Southern pennines The fens Kennet, Lambourne 10 Exe, Bray, Mole Avon, Nene Severn, Stour, Tame Itchen, Test, Cuckmere Avon, Tavy, Plym, Dart Nidd, Ure, Wharfe, Don 29 30 51 21 23 24 Aire, Calder, Wharfe Ouse, Cam, Witham, Lark Wharfe, Ure, Ribble 25 56 Yorkshire dales Total 34 303 3.2 Ecological and environmental data The river quality variables included in the study encompass physical/structural data; chemical water quality data; indicators of the river’s biological quality, and also indicators of angling quality, in terms of fish population data The river quality indicators thus cover a wide range of river attributes, so welfare estimates for a range of environmental outcomes could be produced Fig below shows the environmental/ecological variables used to describe river quality 3.2.1 Chemical data The most well-established means of measuring freshwater quality is through reporting on the chemical composition of river water Three determinants are commonly used— biological oxygen demand (BOD), ammonia and dissolved oxygen (DO) Organic wastes are generally considered to be the most widespread pressure on river systems Nutrient data, namely orthophosphates and total oxidised nitrogen (nitrates) levels were also included in the dataset, as a recent report estimating the costs of eutrophication in freshwaters (Pretty et al., 2002) suggests that nutrient pollution is a pervasive and significant pressure on river quality 3.2.2 Habitat data The physical structure or habitat, along with hydrological factors, determine the biodiversity and wildlife potential of Both use a vector of quality indices, but one includes the estimated coefficients of the quality indices, and the other does not No of river stretches A third sub-sample ‘Chalk’ was also defined for the welfare estimation Chalk rivers are one of the priority habitats identified under the UK Biodiversity Action Plan Consequently estimates of the recreational value of changes in the environmental quality of chalk streams would be expected to provide useful input into policy making C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 239 Fig Geographical location of the study areas a river system Habitat quality was measured with an indicator that describes the extent of physical modification of the river channel, the habitat modification score (HMS)3 3.2.3 Biological data A biological assessment of rivers and aquatic life gives a more complete picture of the ecological health of a river system, as the chemical and physical measures cannot account for other types of environmental stresses, for example, heavy metals and pesticides This biological assessment of river quality is based on the diversity and pollution tolerance of families of macroinvertebrates—animals such as snails, shrimps, mayflies/dragonflies etc Two variables were used—NTaxa which is a measure of species diversity; and ASPT, which is a measure of organic pollution instream Raven et al (1998) 3.2.4 Fish population data As this research Johnstone (2004) focuses on angling, it was felt to be important to have at least one variable (fish populations), which could measure both recreational, i.e angling, and ecological quality The data used were: the number of fish species present in a river, which is a measure of species richness; an estimate of the number of fish per 100 m2, which is a measure of density, and thirdly status, a dummy variable that takes the value of one when game (salmon and trout) fish species are present, and when only coarse fish are to be found Survey design and sample 4.1 Survey design The questionnaire was designed to be as concise and simple to answer as possible and fits onto one sheet of A4 paper The first part elicited the travel cost information, and the second part the 240 C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 Fig Environmental and ecological variables used to measure river quality questions on motivations for choice of river site (not reported in this paper) The first five questions ask the angler to state their age, gender, occupation and home postcode, and the names of any angling clubs they belong to This information is used to calculate the respondent’s travel cost, in terms of the distance travelled to the fishing site, and their wage rate, in order to estimate the value of their leisure time4 Respondents were not asked directly for their income level, as it was thought that this might be perceived as intrusive and reduce the response rate In question six the respondents were asked to give three pieces of information for the five main rivers fished in the last year5: the name of the river fished, the site on that river, and the approximate number of visits made to that site per year their survey of anglers The final total number of responses obtained in the survey is within the accepted sample size range for a study of this scale, i.e between 300 and 500 useable records (Ward and Beal, 2000) Although such a sample frame is not ideal and may be slightly biased towards anglers who buy Angling Times or look at angling websites, half of the responses received were from angling clubs Also, the data were scrutinised to ensure duplicate entries were removed Whilst it is possible that the relatively low response rate means that the sample may be biased, comparing the sample to some recently collected statistics shows it is broadly representative of the angling population as a whole On their website the Environment Agency (EA) give some recent statistics on the angling population from a survey of the general public carried out in 20018, which can be roughly compared to the study sample: 1% of the study sample is female compared to their estimates of between and 20%; 72% are 40 years old or over compared to 70% over 35 years old The EA survey found that anglers were most frequently in social class C2; approximately 45% of the study sample were in social class C2 Analysis of the study data showed that 37% of the sample were retired, where weekly income was estimated to be £172 (ONS, 2000), 30% had weekly incomes of between £240 and £440, 32% between £440 and £640 and 1% earned more than £700 p/w Thus in terms of age and income the study sample is fairly representative of the angling population as a whole, but is slightly biased in terms of gender balance, with male anglers over represented Table below shows the descriptive statistics for the explanatory variables Table Descriptive statistics of independent variables Variable 4.2 Data gathering and sample The data were gathered in four ways Initially around 1300– 1500 questionnaires were sent out to angling clubs and as a regional insert into Angling Times magazine, which resulted in approximately 300 responses, about two-thirds of which were returned from angling clubs, giving a 20–23% response rate An online version of the questionnaire was also created and linked to national angling websites6; this generated about 100 responses As originally a large number of questionnaires had been produced, the remainder were sent to fishing tackle shops7 In total, 421 responses to the questionnaire were received The response rate from the primary data collection method is similar to that achieved by Davies and O’Neill (1992)—22%—in It is worth noting again here however that whilst the majority of valuation studies have applied, and continue to apply, this standard procedure, it has recently been argued (e.g Feather and Shaw, 1999) that the process relies on assumptions regarding the labour market that are unlikely to hold in a number of cases It is acknowledged however that ‘past year’ may not precisely equal the last 12 months, and so the sample-period may not be exactly the same for each angler Each electronic questionnaire submitted via the angling websites generated a file containing the data that was sent to the University of Bath internet server This permitted us to keep track of responses gathered in this way However this was found to be too indirect a method of data collection and did not generate any significant number of responses a Age Incomeb Travel costc No fish species No fish 100 m2 Status (dummy) Orthophosphates Nitrates ASPT Ntaxa BOD Ammonia DO HMS Flow Mean SD Min Max 4.44 365.09 19.84 5.96 34.70 0.38 0.60 6.99 0.89 0.84 2.23 0.17 93.56 18.14 2.89 1.47 179.83 15.33 2.88 47.41 0.48 0.68 3.63 0.12 0.22 1.05 0.22 8.82 15.64 2.03 1.00 41.40 0.00 1.00 0.22 0.00 0.01 0.32 0.47 0.20 0.77 0.01 44.92 0.00 1.00 8.00 713.80 103.69 14.00 294.20 1.00 3.72 16.76 1.15 1.37 7.63 1.94 112.95 60.00 10.00 a The variable age is a categorical variable reflecting the age group of the respondent, where 1Zteenager, 2Ztwenties, 3Zthirties, etc; there are seven categories b Income is the mean weekly wage for the respondent’s occupation and age, derived from the Office for National Statistics (2000) c Travel cost is calculated as the marginal cost of motoring (10 pence per mile) plus the travel time cost at 40% of the respondent wage rate (A sensitivity analysis was conducted for two different percentages of the wage rate—20– 60%) Travel cost data were derived from travel distances and times for each fishing trip, obtained from Multimap.com Survey of Rod Licence Holders, Simpson, D and Mawle, G, Environment Agency 2001 C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 Models and results 5.1 The trip prediction model The first type of travel cost model estimated was a count data model, which provided estimates of proportional changes in trips following marginal changes in the river quality attributes The model is specified as X X Tij Z bCij C gk Xkj C hi Ii C 3ij (1) i k Where TijZnumber of trips made by individual i to site j; Cij is the travel cost for individual i to go to site j, Xk are site characteristics (e.g quality of fishing), and Ii are individual characteristics (e.g occupation) The coefficients b, g and h determine the impact of the explanatory variables on the number of trips and is an error term9 Count data such as numbers of trips often follow a Poisson distribution, so a Poisson regression may be an appropriate analysis for the ecological-economic travel cost model The Poisson distribution describes the occurrence of sparse events, for example in this case on how often a river stretch j will be fished or not (including the possibility that it will not be fished at all), and can be written: À Á eKlij ðlij ÞTij Pr Tij jCij ; Xj ; Ii Z Tij ! (2) Where Tij is the number of trips made by individual i to site j and the Poisson parameter, lij, which is the expected (mean) value of Tij, is constrained such that X X gk Xkj C hi Ii lnðlij Þ Z bCij C k i whereas the Poisson distribution has only one parameter (the mean), the Negative Binomial distribution has two separate parameters, the mean and the variance which gives it more flexibility The Negative Binomial distribution is given by   ðxij C rij K1Þ! rij PrðTij Z xij Þ Z qij ð1Kqij Þxij ðrij K1Þ!xij ! (3) 241 Here xij is the number of trips made by individual i to site j In this case theP mean number P of trips rij/qij is constrained so that rij Z bC C g X C ij k j kj i h i Ii qij Both the Poisson and negative binomial models were estimated, plus the zero-inflated versions of each—‘ZIP’ and ‘ZINB’ The zero-inflated versions are designed for datasets with excess numbers of zeros created by two distinct stages, firstly where a binomial probability distribution (logit or probit) is used in a ‘transition’ or ‘hurdle’ stage, where the observation either moves from to or not—in this case whether a person decides to visit a river stretch or not (the participation decision) The second or ‘event’ stage is then modelled with a symmetrical Poisson or negative binomial distribution in which the ‘event’ (trip) could have a zero or a positive value Table below shows the results The highly significant (at a probability level of !0.001%) likelihood ratio (LR) test of alpha10 in the bottom row of the table shows that the negative binomial distribution provides a better fit of the data than the Poisson distribution and the positive significant (O1.96) Vuong statistics in the second-to-last row show that the zero-inflated models are a better fit than the standard normal versions Thus overall, these measures of fit suggest that the zero-inflated negative binomial (ZINB) model is preferred This model gives generally the same results as the others, in terms of which variables are significant and have the expected sign, except for the fish species richness variable number of fish species, which has changed to be a significant negative predictor of trips The relatively low pseudo R2 values suggest that the models not explain much (less than 10% in the preferred model) of the variance in trips: the explanatory power of the trip prediction models is relatively low The preferred model (ZINB) finds the river quality variables that significantly decrease the likelihood of a fishing trip to be higher levels of orthophosphates and nitrates and a higher number of fish species instream11 The significance of status meant that the likelihood of a fishing trip is increased in rivers supporting coarse fish species Significant positive predictive variables were NTaxa, DO and Flow, suggesting that anglers make more trips to rivers with greater macroinvertebrate species diversity, higher levels of dissolved oxygen and higher flow rates xij Z 0; 1; 2; with the properties rij Mean Z qij Variance Z rij ð1Kqij Þ q2ij The error term captures specific individual departures from the population model that are not captured in the set of explanatory variables C, X and Y; in our case these could include family traditions of fishing, family status, health of the individual etc The error terms could be correlated across individuals (e.g individuals belonging to a club may all go to one site a fixed number of times), and across sites (two sites may be visited alternately by the same group of individuals) 5.1.1 Welfare measures from the trip prediction model The model coefficients can also be used to predict the total numbers of trips with the current levels of river quality and 10 The negative binomial model command in Stata includes an ancillary parameter alpha a which is an estimate of the degree of overdispersion—when a is zero, negative binomial has the same distribution as Poisson The larger a is the greater the amount of overdispersion in the data, and the worse fit a Poisson distribution 11 This rather counter-intuitive result may be reflecting anglers preferences for upland rivers The counter-intuitive signs on the variables in the count-data models may also be due to multicollinearity among the variables—this was explored by dropping variables known to be collinear and seeing if this changed the results Whilst no variables changed sign, dropping certain variables made others less, or non-significant; this analysis showed that the variables that were consistently significant and correctly signed were travel cost, orthophosphates, dissolved oxygen and flow rate 242 C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 Table Results of the trip prediction models Model Variables Poisson Travel cost Age Wage No fish species No fish 100 m2 Status (dummy) Orthophosphates Nitrates ASPT Ntaxa BOD Ammonia Dissolved oxygen HMS Flow _Cons LR c2 (15) Log likelihood Pseudo Adj R2 Vuong LR of alpha (NB over pois) a b c Neg Bin a K0.1262 0.0096 0.0002b K0.0023 0.0030a K0.7911a 0.0387 K0.1688a K1.6139a 1.6630a K0.1788a K0.0207 0.0123a K0.0055a 0.4938a K0.4212 15555.28a K10955.6 0.41 K K ZIP a K0.0459 K0.0299 0.0004 0.0676 0.0006 K0.1196 K0.7218c K0.1213c 1.6361 1.8071 K0.0866 2.3459a 0.0527a 0.0021 0.7453a K10.5181 450.38a K2605.58 0.07 – 17000.00 ZINB a K0.0283 0.0049 K0.0002c 0.0072 0.0022a K0.1485a 0.0407 K0.0443a K0.2086 K0.0660 0.0421 0.9061a 0.0244a 0.0017 0.1585a K0.1317 991.97a K4644.67 0.11 9.96 – K0.0353a 0.0528 0.0001 K0.2187a 0.0141c K1.6363a K0.6165b K0.1215a K2.8324 2.4006c K0.1664 0.1289 0.0603a 0.0117 0.8025a K4.2645 360.11a K2523.56 0.09 6.82 4226.93 Significant at the 001% level or lower Significant at the 05% level Significant at the 01% level with respect to changes in the river quality levels The expected numbers of trips with current levels of river quality based on the zero-inflated negative binomial count data model as described above is written as EfNTripsij jXj ; Yi ; Cij g ( ) X X gk Xkj C hi Ii KbCij h T ij Z exp k (4) i where Xj, Ii and Cij are the explanatory variables as defined in Eq (1); g, h and b are corresponding coefficients and T ij is the predicted number of trips Using this formula, the expected number of trips as predicted by the count model is 5937, and the actual number of trips in the dataset is 4853, thus actual number of trips is 82% of predicted trips This is similar to the results achieved by Hanley et al (2003), who estimated that actual trips were approximately 70% of their predicted trips The expected number of trips with an increase in river quality can be written EfNTripsij jXj ; Yi ; Cij g ( ) X X à gk Xkj C hi Ii KbCij h T Ãij Z exp k (5) i where X* is the river attribute vector with the quality changes The consumer surplus is calculated by dividing the expected number of trips by the coefficient on the travel cost variable, b ÈP É P exp T ij k gkj Xkj C i hi Ii KbCij CSij Z (6) Z b b Thus the consumer surplus per trip can be calculated by dividing the total consumer surplus by the total number of trips, which gives a value of £25 per trip12 This is based on a linear form of the demand function Although this particular model could not be tested for sensitivity to the assumption of linearity, similar calculations of consumer surplus for alternative forms of the demand function did not generate significant differences when applied to the kinds of changes in quality being considered here The consumer surplus from a change in river quality is therefore: É P gj Xkjà C i hi Ii KbCij DCSij Z b ÈP É P exp g X T Ãij KT ij k k kj C j hi Ii KbCij K Z b b exp ÈP k (7) In other words, the change in consumer surplus resulting from a change in river quality is the estimated number of trips in the changed condition minus the estimated number of trips in the original condition divided by the coefficient on travel cost This procedure was used to estimate the percentage reduction in trips and the associated per trip consumer surplus values for a 10% increase in each significant river attribute, which are shown in Table 12 This value is for all rivers as there were insufficient observations to estimate values for the three river types with the trip prediction model C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 Table Predicted changes in trips from a 10% increase in the significant river attributes with the expected signs from the ZINB count model River attribute Percentage change in trips Change in consumer surplus per trip (£2001) No fish 100 m2 (C) Orthophosphates (K) Nitrates (K) NTaxa (C) Dissolved oxygen (C) Flow (C)a K2 K7 26 79 123 0.51 K0.63 K1.79 5.21 11.06 13.72 243 visiting a fishing site, and is composed of at least one trip to a site in a set of n number of locally available fishing sites This n is the set of river stretches in the study Area where the respondent fished (see Table 1), and therefore will vary across individuals The log-likelihood of observing the pattern of fishing trips using the conditional logit (CL) model is therefore lnðLðbÞÞ Z 319 X n X iZ1 jZ1 ( À Á ) P P exp bCij C gk Xkj C hi Ii À Á P P dij ln Pn gk Xkj C hi Ii jZ1 exp bCij C (10) a A category increase, e.g from 4–5; a category increase is a doubling of flow rate Table Results of the conditional logit model for three sub-samples of the dataset 5.2 The site choice model The second type of travel cost model estimated is a site choice random utility model (RUM) The RUM theoretical framework is commonly used in both stated and revealed preference studies where the focus is on valuing changes in specific attributes The theory is based on a framework for modelling individual choice developed by McFadden in the 1970s, and states that an individual’s choice is informed by an evaluation of measurable alternatives, plus a random component, which the researcher cannot measure The basic RUM model is specified as: X X Uij Z bCij C gk Xkj C hi Ii C 3ij (8) k where dij takes the value one if individual I visits site j and the value zero if she does not The results are given in Table i Where UijZthe utility of individual i at site j; Cij is the travel cost of travelling to site j; Xk is a vector of the characteristics of site j, I is the vector of individual characteristics and 3ij is the random error term An individual will choose the site that maximises her utility U, thus an individual will choose site if U1OUj for all jZlocally available substitute sites The regression model estimates the parameters: b,g and h so as to maximise the likelihood of the observed pattern of fishing trips A regression model that is commonly used in RUM site choice models is the multinomial logit or conditional logit (CL) model13 The probability of an individual choosing site j* out of the set of n alternatives is formally written as: À Á P P exp bCjà C gk Xkjà C hi Ii P P P prðjÃÞ Z n (9) gk Xkj C hi Ii ÞÞ jZ1 ðexpðbCj C This is the exponential of the utility of site j divided by the sum of all of the exponentiated utilities The probability thus depends on the attributes of all the river sites in the individual’s choice set, as well as the chosen site, in other words, the model takes account of the locally available substitute sites In this study, the dataset contains 319 respondent observations—an observation is created for each respondent 13 These two terms are used interchangeably in the literature, although there are subtle differences, in that the multinomial logit can incorporate individual specific variables such as age and income, i.e variables that are constant within groups, whereas the conditional logit cannot River type Explanatory variables Upland Lowland Chalk Travel Cost No fish species No fish 100 m2 Status (dummy) Orthophosphates Nitrates ASPT NTaxa BOD Ammonia DO HMS Flow No Obs LR c2 (13) Log likelihood Pseudo R2 K0.376a 0.110b 0.024a K2.464a 0.687 0.119 K5.431a 3.129a K2.266a 21.588a 0.152c 0.003 1.173a 4243.00 2482.83 K868.65 0.59 K0.375a 0.198a K0.002b K37.892 K1.733a 0.128c 16.917a K4.741a 0.527a K2.184a K0.0302a 0.030a 0.602a 8042.00 3681.20a K1383.28 0.57 K0.22a K0.50 K0.01 K9.34a 3.44 0.60 29.84c K3.49 K3.62 15.40 0.23a K0.19a 0.71a 1626.00 1118.00a K238.46 0.70 a b c Significant at the 001% level or lower Significant at the 05% level Significant at the 01% level On first inspection, the upland and lowland site choice models give a rather mixed message As would be expected, travel cost is always negative and highly significant, most of the river quality variables are highly significant and the models explain a relatively high proportion of the variance in site choice However, many of the river quality variables have unexpected signs The only river quality variable that is consistently signed as expected and significant is Flow14, which is actually a measure of quantity, although it is often used as a proxy indicator of river quality As shown in Table above, the sub-sample models upland lowland and chalk produce markedly different results, and with the exception of flow and status, the river quality variables vary in terms of their signs and significance The chalk model has the fewest significant variables— status, ASPT, DO, HMS and flow—but the variables that are 14 Status is consistently negatively signed, suggesting anglers predominantly chose rivers supporting mixed and coarse fish species 244 C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 significant have the intuitively expected signs, in that they confirm what would be expected: that rivers with lower levels of pollutants/higher biological and chemical quality would be chosen over those with higher levels of pollutants This model also has the highest Pseudo R2 measure, explaining over twothirds of the variance in site choice A possible reason for the conflicting signs on the river quality variables in the upland and lowland models is that the quality variables are highly collinear Analysis showed that some of the river quality variables were fairly highly correlated, particularly the biological quality indicators ASPT and NTaxa In order to explore whether the unexpected signs are the result of multicolinearity, some of the variables that are known to be collinear were dropped from the upland and lowland models and the models re-run This analysis showed that variables that were significant and had the expected signs were status, NTaxa and flow in the upland model, and number of species, orthophosphates, ASPT and flow in the lowland model Another possible reason for the unintuitive results for some of the river quality variables is temporal mismatch between river quality and angling data, where the river quality data does not reflect the current river conditions More up-to-date and temporally commensurate data, in particular fish population data would undoubtedly improve the models and resulting welfare estimates One of the shortcomings of the conditional logit model is the independence of irrelevant alternatives (IIA) assumption This states for example that the probability of choosing between two fishing sites is not affected by the presence of other sites in the choice set It is reasonably likely that the alternative sites within each anglers choice set will in fact affect the probability of choosing site x over site y, for example if they supported similar fish species It is possible therefore the unexpected signs on the variables could also be a result of the restrictiveness of the IIA assumption One way to deal with this is to use a slightly different modeling approach, a nested or mixed logit, for example as used in Parsons and Massey (2003) Nested or mixed logit models partitions the choice set of sites into n number of groups, and thereby allows for correlation among sites by specifying separate price and attribute aspects of the error term This is a potentially useful way the research dataset could be extended in the future 5.2.1 Welfare measures from the RUM conditional logit site choice model The ‘log sum’ approach is used to calculate the consumer surplus associated with the changes in the river quality variables from the RUM site choice model In this process, the total welfare each individual gains from each site under a hypothetical improved condition is compared to the total welfare from the original, unimproved condition Dividing the difference by the marginal utility of money gives an estimate of the change in consumer surplus CSimpr This can be written as: CSimpr Z fWimpr KWorig g l (11) where CSimpr is the change in consumer surplus with the improved river quality, Wimpr is the welfare level with the improved river quality, Worig is the welfare level in the original river condition, and l is the individual’s marginal utility of income, which is the travel cost coefficient from the site choice model, and translates the utility into monetary terms As the conditional logit was used, this is expressed for each individual i and each site j as: Wimpr Z jKn iZ319 XX n  o X X dij ln exp bCij C gx Xkà C hi I i iZ1 jZ1 (12) Worig Z jZn iK 319 X X n  o X dij ln exp bCij C g k X k C hi I i iK1 jZ1 where, as before, Cij is the travel cost of individual I to site j and Xk are the river quality variables In this example, one or more of these change from the original to the improved valuations and the new values are marked with a star, while the original values are marked with an ‘0’ dij takes the value one for each site if the individual visits that site and zero if he does not These welfare estimate calculations were carried out for each of the sub-sample models, upland, lowland and chalk Table presents the results; the values are in grey where the characteristic has the expected sign but is not significant Table Consumer surplus per trip for a 10% increase in significant river attributes with the expected sign (£2001) Upland No fish species No fish 100 m2 Orthophosphates Nitrates ASPT NTaxa BOD Ammonia DO HMS Flow Lowland Chalk 0.53 0.04 K0.36 3.93 0.32 0.32 K0.40 K0.13 1.50 1.17 1.59 0.23 K0.01 0.07 A one category change for the variable flow To calculate the welfare gain from the total trips the welfare in the original unimproved condition is divided by the coefficient on travel cost: Worig/l The welfare gain per trip is then simply {Worig/l}/Stij, i.e the total welfare gain divided by the total number of trips Applying these values results in an estimated welfare gain per trip of £47.31 for lowland sites, £19.27 for upland sites and £5.78 for chalk sites C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 5.3 Linking the count and site choice models The count and site choice models can be linked by using the predicted number of trips from a 10% change in the river quality variables as estimated in the trip prediction model as shown above in place of the actual numbers of trips in Eq (12) above In other words dij is replaced with the predicted number of trips under the original and improved conditions Worig and Wimpr, are then as shown below: o XX N n  X X T ij ln exp bCij C Wimpr Z gx Xkà C hi I i i j (13) Worig Z XX i n  o X X T 0ij ln exp bCij C gk Xk0 C hi I i j The predicted number of trips under the original and improved conditions are represented by T 0ij and T Nij respectively Table below shows the welfare estimates with respect to a 10% increase in the river quality variables using the number of trips estimated by the trip prediction model In general using the estimated number of trips in the improved condition results in slightly larger welfare estimates, except for ASPT and ammonia, which reduce the welfare per trip slightly; this is due to the opposite signs for these coefficients in the trip prediction model Table Consumer surplus per trip for a 10% increase in significant river attributes with the expected sign using the predicted number of trips from the count model (£2001) Upland No fish species No fish 100 m2 Orthophosphates Nitrates ASPT NTaxa BOD Ammonia DO HMS Flow Lowland Chalk 2.49 0.05 K0.49 3.27 0.29 0.57 K0.43 K0.13 2.09 1.97 3.70 0.29 K0.01 0.15 Discussion 6.1 Empirical results The importance of river quality in angling participation and site choice confirms the results of previous studies and the associated consumer surplus values estimated are broadly of a similar magnitude to previous studies For example in £2002, ECOTEC Research and Consulting (1993) estimated £26–£40 per trip, Radford et al (1991) estimated £22 per trip, and an earlier travel cost study by Radford et al (1984) estimated £33–£50 per trip Overall, most of the empirical results were reasonable in that more explanatory variables had the right signs and the 245 travel cost variable was always negative, which confirms economic expectations The importance of river flow rates on recreational use supports a number of previous studies, for example Willis and Garrod (1999), who found (using a different method of elicitation—the CVM) that anglers preferred, and were willing to pay for, more natural flow levels in rivers The statistically significant and positive relationship between dissolved oxygen (DO) and angling site choice in the upland and chalk models supports an early study by Smith et al (1986), who also found this river quality variable was a significant predictor of the number of recreational trips to rivers The site choice models have also shown that, as would be expected, different river characteristics are important in predicting fishing site choice in the three sub-samples In upland areas, the variables shown to be significantly (p! 0.01) related to site choice were NTaxa, (reflecting macroinvertebrate species richness and thus indirectly, habitat quality) and flow (representing rivers with higher volumes of water); BOD, DO and number of fish 100 m2 were also less significant predictors The importance of aquatic invertebrate species richness (NTaxa) in choice of fishing site in upland areas confirms anglers’ preferences towards rivers with higher numbers of species in regions that are not predisposed geologically or physically to high species richness In lowland areas, significant variables were ASPT, which is a measure of organic/nutrient pollution levels; number of fish species, a measure fish species diversity; orthophosphates, which are a particular type of nutrient pollution and flow That nutrient enrichment should be more of an influential factor in recreational use of rivers in lowland areas, where agriculture and human inputs of nitrates and orthophosphates are greater, is consistent with ecological theory Upland aquatic areas are less prone to nutrient enrichment problems such as eutrophication as they have pre-existing lower levels of nutrients before anthropogenic inputs are taken into account (Petts and Foster, 1985) 6.2 Methodological results The RUM site choice model used in this study has the advantage that it allows the researcher to include the effects of the substitute sites available to each respondent within the study areas in which they were observed to have made fishing trips The models were able to evaluate the influence of the cost and quality attributes of the substitute sites on site choice, and the significance and coefficients of these variables reflect this Overall, the model design can be considered to have been a success: the travel cost variable was always negative and significant and there were more consistently statistically significant river quality variables with the expected sign than unexpectedly signed variables It could be argued that one should compare the standard conditional logit (CL) site choice model with a mixed version that allowed for collinearity between sites, to see how this affected the results However, in their paper, Parsons et al 246 C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 (1999) found that the models that split the inclusive value term into price and quality terms resulted in significantly smaller welfare measures, which were inconsistent with the site choice model per trip estimates The results of the count data ‘trip prediction’ model generally corresponded with the results of the site choice models, in that biological quality, as measured by number of fish 100 m2 (fish population density) and NTaxa, was important in both participation and site choice This was also the case for nutrient pollution (although demonstrated through the variable nitrates as opposed to orthophosphates), and the most significant variable was flow Interestingly, the welfare estimates from the trip prediction model were higher for number of fish per 100 m2, NTaxa, dissolved oxygen and flow by a factor of 10 or more than those from the RUM site choice model (see Table 8) Only for orthophosphates, were the estimates at similar levels in the three methods One explanation for the higher welfare estimates in the trip prediction model is that count data models not explicitly incorporate substitution effects, while RUM models allow for them This can be seen in the markedly larger influence of travel cost in the RUM models Also the welfare value per trip was very similar for the Count and RUM models Note, however, that the relatively low explanatory power of the count model meant that the trip predictions and associated consumer surplus should be used with some caution Linking the trip prediction and site choice models was also successful in that the welfare estimates for improvements in river quality with the estimated number of trips from the trip model are slightly larger, as would be expected As such the travel cost models used in this study provide some initial estimates of the likely impact of river quality on recreational use of rivers, and have usefully identified a number of ways this analysis could be extended with respect to rivers and angling in the UK Conclusions This study has provided consumer surplus values for marginal changes in a number of river quality indicators for three different types of river, ‘Upland’, ‘Lowland’ and ‘Chalk’ The study also produced per trip angling welfare values for these three river types As noted above, it is anticipated that these per trip and quality change values will be useful as inputs into various environmental and resource management policy decisions, for example where it is necessary to gauge the value of increases in river species richness, diffuse pollution or abstraction Methodologically, the study has successfully demonstrated an integrated approach to modelling participation and site choice, which allows the researcher to produce welfare estimates that take both decisions into account Table below summarises the welfare estimates produced by the trip prediction and RUM travel cost models Table Per trip consumer surplus values for a 10% increase in the river quality variables that were significant in both trip prediction and RUM travel cost models No fish 100 m2 Orthophosphates Ntaxa Dissolved oxygen Flow Value per trip Upland Lowland Chalk Count model RUM model Combineda 0.51 K0.63 5.21 11.06 13.72 25 – – – 0.04b K0.36c 0.32b 0.87d 0.94e 24e 47.3 19.3 5.8 0.05b K0.49c 0.57b 1.19d 1.94e 24 47.3 19.3 5.8 A one category (e.g 4–5) increase for flow, which is a doubling of flow rate (£2001) a The combined model calculates the change in consumer surplus by substituting the predicted number of trips from the count data model in place of the actual number of trips in the RUM model Wimpr and Worig are then calculated as shown in Equation (14), and the consumer surplus change is calculated as shown in Eq (11) b Upland rivers only c Lowland rivers only d Average of chalk and upland rivers values e Average of all three river type values Given the fairly low response rate in the angling survey, these welfare estimates should be used with caution to give broad estimates of welfare impacts There are a number of ways future studies could develop this area of research For example, whilst there were insufficient observations in this study, both the trip prediction values and the linked welfare estimates could be improved by splitting the trip prediction model into upland lowland and chalk, as was done in the site choice models Using GIS to select the set of locally available substitute sites, or to select the relevant river quality data would also improve the data gathering process Acknowledgements This paper is part of Dr Johnstone’s PhD thesis carried out at the University of Bath, which was joint funded by ESRC and NERC The authors would like to thank Nick Hanley for his input and advice, Pam Mason, David Howard and Mike Hornung for their help on the wider research project, and Ron Thomas at the Environment Agency for assistance with GIS software and providing much of the environmental data They are also indebted to two anonymous referees who made a number of important comments that have improved the paper None of these people are responsible for any errors that remain References Bockstael, N., et al., 1987 Estimating the value of water quality improvements in a recreational oemand framework Water Resources Research 23 (5), 951–960 Davies, J., O’Neill, C., 1992 Discrete-choice valuation of recreational angling Journal of Agricultural Economics 43 (3), 452–457 ECOTEC Research and Consulting, 1993 A Cost Benefit Analysis of Reduced Acid Deposition: UK Natural and Semi-Natural Aquatic Ecosystems: A C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 Contingent Valuation Study of Aquatic Ecosystems ECOTEC Research and Consulting, Birmingham, and the Department of Environment Food and Rural Affairs (Defra), London Feather, P., Shaw, D., 1999 Estimating the cost of leisure time for recreation demand models Journal of Environmental Economics and Management 38, 49–65 Hanley, N., et al., 2003 Valuing the benefits of coastal water quality improvements using contingent and real behaviour Environmental and Resource Economics 24, 273–285 Hausman, J.A., Leonard, G.K., McFadden, D., 1995 A utility-consistent, combined discrete choice and count-data model Journal of Public Economics 56, 1–30 Johnstone, C., 2004 An Ecological and Economic Approach to Valuing River Quality PhD Thesis Dept Economics and International Development, University of Bath, Bath, UK Morey, E.R., et al., 1993 A repeated nested-logit model of Atlantic Salmon fishing American Journal of Agricultural Economics 75, 578–592 Office of National Statistics (ONS), 2000 Labour Market: New Earnings Survey Parsons, G.R., Massey, D.M., 2003 A random utility model of beach recreation In: Hanley, N (Ed.), The New Economics of Outdoor Recreation Edward Elgar, Cheltenham Parsons, G.R., et al., 1999 A comparison of welfare estimates from four models for linking seasonal recreational trips to multinomial logit models of site choice Journal of Environmental Economics and Management 38, 143–157 247 Petts, G.E., Foster, I., 1985 Rivers and Landscape Edward Arnold, London Pretty, J., et al., 2002 A Preliminary Assessment of the Environmental Costs of the Eutrophication of Fresh Waters in England and Wales Centre for Environment and Society and Department of Biological Sciences, University of Essex, Colchester, UK Radford, A.F., et al., 1984 The Economics and Value of Recreational Salmon Fisheries in England and Wales: An Analysis of the Rivers Wye, Mawddach, Tamar and Lune R8, Centre for the Economics and Management of Aquatic Resources, University of Portsmouth, Portsmouth Radford, A.F., Hatcher, A., Whitmarsh, D., 1991 An Economic Evaluation of Salmon Fisheries in Great Britain A Research Report to the Ministry of Agriculture Fisheries and Food Research Report 16, Centre for the Economics and Management of Aquatic Resources, University of Portsmouth Raven, P.J., et al., 1998 River Habitat Quality: the Physical Characteristics of Rivers and Streams in the UK and Isle of Man River Habitat Survey, Report No 2, May 1998 Smith, K., et al., 1986 A comparison of direct and indirect methods for estimating environmental benefits American Journal of Agricultural Economics, 280–290 Ward, F.A., Beal, D., 2000 Valuing Nature with Travel Cost Models — A Manual Edward Elgar, Cheltenham, UK Willis, K., Garrod, G., 1999 Angling and recreation values for low-flow alleviation in rivers Journal of Environmental Management 57, 71–83  ... for lowland sites, £19.27 for upland sites and £5.78 for chalk sites C Johnstone, A Markandya / Journal of Environmental Management 80 (2006) 237–247 5.3 Linking the count and site choice models. .. inspection, the upland and lowland site choice models give a rather mixed message As would be expected, travel cost is always negative and highly significant, most of the river quality variables... welfare measures for changes in river quality These were a random utility site choice model (RUM) and a trip prediction or participation model Because RUM site choice models cannot predict total