You are on page 1of 9
me 59 Numbers 11-12 December 2016 Journal of Environmental Planning and Management, 2016 J Routledge Vol. 59, No. 11, 2064~2078, http://dx.doi.org/10.1080/09640568.2015.1119105 ae The effect of experience on choosing where to go: an application to a choice experiment on forest recreation Gengyang Tu*>* and Jens Abildtrup* “Laboratoire d'Economie Forestiére, AgroParisTech, INRA, 54000, Nancy, France; ’Université de Lorraine, France. (Received 17 April 2015; final version received 9 November 2015) The main objective of this paper is to investigate the impact of experience on the choice of visits to forests in a stated discrete choice experiment. Recent literature has indicated that experiences with the environmental services valuated may increase the respondents’ certainty in their choice of hypothetical altematives. We apply two indicators of experiences: the number of visits and the number of different forests visited during the last year. Applying the generalized multinomial logit model, we find that an increase in the number of visits to forests makes respondents’ choices more predictable. However, the number of different forests visited reduces the predictability of choices. Furthermore, we investigate the relationship between respondents’ experience of forest recreation and the self-reported choice certainty, controlling for respondents’ social- demographics and other design characteristics. Finally, we show that self-reported choice certainty is positive correlated with the scale factor, as expected. Keywords: forest recreation; choice modeling; scale; uncertainty; experiences 1, Introduction A decision-maker’s uncertainty in the choice of alternative may be related to her personal experience with the choice of good or service. In choice studies, it is a common assumption that all attributes are traded in a fully compensatory decision-making process and respondents are able to assess the exact utility they derive from the good presented to them. Respondents are, therefore, assumed to answer any valuation question with absolute certainty (Hanemann 1984). However, stating exact utilities and willingness to pay (WTP) for a non-market good is a difficult task for most people. Also, uncertainty may be derived from the misunderstanding of the words, the ambiguity of the survey instrument, and the unfamiliarity of relating to the goods in monetary terms. Preference uncertainty has been recognized when conducting stated preference studies (Li and Mattson 1995; Wang 1997). Several contingent valuation studies have investigated the cause of respondent uncertainty, and the ways to handle it (e.g., Ready, Whitehead, and Blomquist 1995; Loomis and Ekstrand 1998; Alberini, Boyle, and Welsh 2003). In recent years, this issue has been given an increasing attention in the literature of choice experiment (CE) studies (e.g., Lundhede et al. 2009, Hensher and Collins 2011; Olsen et al, 2011; Hensher, Rose, and Beck 2012; Rose, Beck, and Hensher 2015). In the case of ‘experience goods’ consumption,' a consumer’s uncertainty over his type is reduced as the experience with the good increases. Therefore, direct experience with a good may have an impact on respondents’ valuation for that good (Nelson 1970, “Corresponding author. Email: gengyang.tu@nancy.inra.fr © 2016 University of Newcastle upon Tyne Journal of Environmental Planning and Management 2065 1974; Ackerberg 2003). In the environmental literature, studies have found that experience with a public good has an effect both on measures of the predictability of consumers’ preferences and as a determinant of willingness to pay for public goods (Boyle, Welsh, and Bishop 1993; Adamowicz, 1994; Whitehead ef al. 1995; Cameron and Englin 1997; Breffle and Morey 2000; Ferrini and Scarpa 2007; Hanley, Kristrom, and Shogren 2009). According to Fiebig ef al. (2010), the scale factor of a random utility model (RUM) may be derived from the inconsistency of respondents’ choices due to differences in experiences, attitudes, perceptions, or some other unidentified construct held by the population for the various alternatives under consideration. It represents the amount of idiosyncratic error in the responses provided by each respondent. The scale factor may therefore be linked with choice certainty, since the increase of scale factor makes respondents’ choices more deterministic (Swait and Adamowicz 2001; Hensher 2006; Rose et al. 2009, Beck, Rose, and Hensher 2013). A recent study found evidence that a subject’s scale in a RUM applied to a discrete choice model increases with experience, i.e., the subjects’ decisions are more predictable from an econometrician’s point of view (Czajkowski, Hanley, and LaRiviere 2015). This study analyzed the link between number of visits to beaches and the hypothetical choices of beaches in a discrete choice experiment. So far, the relationship between experiences and choice uncertainty in forest recreation has been paid little attention in forestry literature. The first objective of this study is to investigate the impact of respondents’ personal experiences on the personal- specific scale factor in the RUM of choices of forest for recreational purposes. Two measures of experiences of forest recreation are included in our model: the number of visits during the last 12 months and the number of different forests visited during the same period. As shown theoretically in Czajkowski, Hanley, and LaRiviere (2015), one can expect that the number of visits to forests will help respondents to learn about their preferences and therefore reduce uncertainty. In addition, it may be argued that visiting a large number of different forests will, in particular, be important for learning, as this allows respondents to experience a large diversity of forests. However, visiting many different forests may also indicate that the respondent has no strong preference for forest attributes and her choices are, therefore, guided by unobservable factors. This study applies, as in Czajkowski, Hanley, and LaRiviere (2015), the generalized multinomial logit (GMNL) model to test whether the indicators of experience have an impact on the scale factor. Therefore, this study serves to assess whether the results of this previous study on choice of beach can be generalized to other settings. The second objective of this paper is to investigate the impact of experience with forest recreation on respondents’ self-reported certainty in a choice. A significant impact of experience on self-reported choice certainty will support our hypothesis that experience for the modeling of discrete choices in a RUM. Moreover, our model explaining self-reported certainty controls for the toughness of choice, which is described by utility difference (UD) between the chosen alternative and best alternative among the other alternatives left in a choice set. According to Wang (1997), one would expect uncertainty to be high for indicated payment close to the individual's true WTP, and low for payments distinctly smaller or larger than the true WTP. A significant relationship has been found between UD and self-reported uncertainty (Brouwer ef al. 2010; Olsen ef al. 2011). Finally, we combine the two modeling approaches, the GMNL model and the self-reported certainty model, by including the predicted self-reported choice certainty as an explanatory variable of the scale factor in the GMNL model. EEE EE EES<~— 2066 G. Tu and J. Abildtrup 2. Method 2.1, Scale effect and experience In this part of study, we apply the GMNL model to account for the scale heterogeneity and the effect of experience on personal specific scale factor. Studies have found the potential to explain respondent heterogeneity by differences in error scale, since the last decade (Louviere ef al. 1999, 2002). Since then, models that allow for Continuous error scale heterogeneity have been developed (Sonnier, Ainslie, and Otter 2007, Fiebig ef al. 2010). The GMNL model appears to outperform other models as it allows for heterogeneity both in error scale and all attribute preferences (Fiebig et al. 2010). In the GMNL model, for individual i, his utility for choosing alternative j in the choice situation n is given by U Gin) [oiB + yn; + (1 = youn) Xn + Bin @ where a; is the scale of the error term for person i, which should be positive. It is assumed to be log-normal distributed with mean | and standard deviation t. ; is the vector of individual specific standard deviation from the mean of parameters B. The parameter y € [0,1] describes how the variance of residual taste heterogeneity varies with the scale factor. If y equals to 1, from Equation (7), we obtain GMNL model type! (GMNL-1): U (jin) = (o;8 + nj) Xin + Gin 2 GMNL-I assumes that the standard deviation is independent of the scale factor influencing the mean values of the parameters. If gamma equals to 0, from Equation (7), we obtain GMNL model type 2 (GMNL-II): U (jin) =o; (B+ n,)Xjin + 81m (G3) GMNL-II assumes that the means and standard deviations vary proportionally with the scale factor. In this case, the scale effect has an impact on both taste and taste heterogeneity. oe With both of the models, we could write the individual parameter as: B= 01 B+nj 4 where the random variable o; captures the scale heterogeneity, and the individual’s taste ity is captured by 7}. cates eee parameters associated with all forest attributes are random parameters which are assumed to follow a normal distribution. Assuming that all the random taste parameters 8; are normally distributed and the individual scale parameter o; is log-normally distributed leads to the GMNL-II which assumes the individual parameters are a mixture of normal with proportionally different means and standard deviations (Fiebig e¢ al. 2010). Furthermore, if all the random taste parameters 8; are specified as fixed parameter, then the GMNL model became a scaled multinomial logit model. Journal of Environmental Planning and Management 2067 Suppose there are J available alternatives at choice situation 7, the probability that individual i choosing alternative j can be presented by: exp((oiB+ yn + (1 — y)oin)) jin) Pr (yjpme p= — eM tym t Loma) Sean eae) (3) The probability is simulated over D draws from the assumed distribution. To explain the scale heterogeneity across individuals, the variable ; can be presented by the function: 0; =exp(—o + 6zin + Te) (6) The objective of this step is to investigate the impact of experience with recreation on the scale heterogeneity of forest choice decision. The self-reported choice certainty level for each choice situation may also be correlated with the scale heterogeneity. Therefore, in our model, we use zj,, which is a vector of experience variables and self-reported certainty as covariates of the scale parameter. @ is a vector of parameters of zj,, representing the marginal effect of z;, on the scale factor, Compared to Czajkowski, Hanley, and LaRiviere (2015), we also include interaction terms between attributes and experience variables representing the correlation between experience variables and preferences to account for deterministic taste heterogeneity of preferences caused by respondents’ personal experiences. 2.2, Experience and self-reported choice certainty In the next step, controlling for the toughness of choice situation, we investigate the relationship between respondents’ experience and their stated certainty level concerning their choice. In the survey, respondents were requested to indicate their certainty level after each choice set, using a Likert scale. This stated certainty level may, of course, differ from true certainty, which limits how far one can make conclusions on true certainty from the observed stated certainty. However, we believe that the stated choice certainty is a relevant indicator for the true certainty. The method applied in this step follows a recent study of Olsen et al. (2011). With the estimated GMNL-II model, we calculate the individual-specific expected aggregate utility of each alternative in each choice set and then calculate the expected UD between the alternative chosen, j, and the best alternative of the two alternatives left (either w or J, i.e., for each choice set: UD = E(uy(xy,¢y)) — max{E (u(r, €1)); E(w xiv 6iv))} (o) =A 2y—maxl Bn: Bie} As described by Equation (7), the expected aggregate utility of each alternative is in function of the estimated utility weights and the attribute levels. The UD is finally the product of estimated individual-specific utility weights and the attribute levels of each alternative. Assuming respondents can be more certain about their choice if the UD between alternatives is bigger, the UD can be used to explain the toughness of each choice set, Then, the UD is used as an explanatory variable in a random effects ordered 2068 G. Tu and J. Abildtrup probit model, where the dependent variable is self-reported level of certainty in each choice situation. We estimate the latent certainty variable assuming the equation: CERT) =; ky + oy +61 (8) where CERT} is the latent variable of choice certainty of response i, ky, & € [1,..., K] is a vector of explanatory variables for respondent / in choice situation j, which may affect respondents’ certainty level in each choice. The term é;x is distributed as N (0, 1]. The term e; is N [0, 07] distributed, and is the same for every stated certainty made by the same respondent. The observed variable CERT), which is a set of stated certainty scores obtained from ordinal scale and polychotomous choice method, is an ordinal variable. The ordered probit model is appropriate in this case (McKelvey and Zavoina 1373): The respondents’ stated certainty level varies from ‘very uncertain’ to ‘very certain without presumed cardinality. Assuming the level of certainty h € (1, H] which is defined through its unknown lower bound jz,_ and upper bound j1,,, CERTy is determined from the model as follows: CERT; = h if Uy) < CERT; €48,750 21 6 Source: Age and gender: INSEE — Population estimations; income: taxable income 2008; www2.impots.gouv.fr/documentation/statistiques/ircom2007/ tegion/region.htm In Table 2, the main demographic and socioeconomic characteristics of the effective sample (1,054 respondents) used to estimate the choice model are presented and compared with the total population in Lorraine. The share of female respondents is lower in the sample than in the population and the 40-60 year old respondents are overrepresented in the sample. The sample exhibits an overrepresentation of people in high income classes. The relatively high rates of middle-aged people and high-income groups in the sample are not unusual for Internet and mail surveys (Olsen 2009). Thus, even though the response rate might raise some concerns regarding the representativeness of the sample, the skewness of the sample for central socio-demographic characteristics does not seem to be much worse than similar surveys with much higher response rates. Table 3 gives the descriptive statistics of the variables used to model self-reported choice certainty and scale heterogeneity. Among the 1,144 completed questionnaires only 1,054 respondents had visited a forest at least once during the last 12 months. Only people who had visited a forest where asked to participate in the choice experiment, as the respondent was asked to consider if she would have gone to one of the hypothetical forests, if they had existed, instead of the forest chosen (most often) during the last 12 months. 4, Results 4.1, Scale heterogeneity and experience i 1 on the left side is a Table 4 presents the results of GMNL regressions. The first model SMNL model. The model in the middle and the model on the right side are all GMNL-II models. The only difference between these two GMNL models is that the scale factor of Journal of Environmental Planning and Management 2071 Table 3. Descriptive statistics Variable Description Number obs, Mean sd. Min. Max. Broadleaves Is 1 if attribute Dominant tree 6,324 ei 43 6 1 species = Broadleaves; otherwise 0; Mix species Is 1 if attribute Dominant tree 6,324 04 #OS 0 1 species = Mixed tree species; otherwise 0; One_path Js 1 ifattribute Hiking paths = One 6,324 03 03 0 1 marked hiking path; otherwise 0; Several_path Is 1 if attribute Hiking paths = More 6,324 os oF 6 1 than one marked hiking path; otherwise 0; One_facility Is 1 if attribute Facilities = Parking 6,324 a Os 6 1 or picnic places indicates presence of pienic or parking place; otherwise 0; Both_facilities Is 1 if attribute Facilities = Parking 6,324 03 0S oO 4 and picnic places indicates presence of picnic or parking place; otherwise 0; Water Is | ifattribute Access to water (lake 6,324 os 05 0 1 or river) = 1; otherwise 0; dist Distance to forest from home of 6,324 148 196 0 260 each alternative nbvisit Number of visits during last 12 1,054 278 469 1 365 months nbforest Number of different forests visited 1,054 45 63 0 90 cert Self-reported choice certainty 6,324 45 13...1 6 Choicenumber Number of choice sets in experiment 6,324 35 wi 1 6 hunter 1 if respondent is a hunter 1,054 Ot os 6 1 the one in the middle does not have ‘certain’ as covariate. Comparing the two GMNL models, it appears that the parameters of the experience covariate do not change much with the presence of the covariate ‘certain’. The GMNL-II full model which is presented in column 4 has lower Akaike information criterion (AIC)/Bayesian information criterion (BIC) values, indicating that this model is preferred in terms of model fit, Following Scarpa, Ferrini, and Willis (2005), an Alternative Specific Constant (ASC) is specified for the status quo alternative in order to capture the systematic component of a potential status quo effect. The positive Parameter estimate for the ASC captures a systematic status quo effect. All other things being equal, respondents prefer the status quo alternative, i.e., the forest they have visited most often in the past 12 months, In other words, respondents show an affinity for this alternative beyond what the specific attribute levels for this alternative, relative to the two other alternatives, would predict, All Parameters have the expected sign and are significantly different from zero. The utility of Visiting a broadleaf forest or a mixed species forest is higher than visiting a forest dominated by conifers. Visitors generally prefer a forest with one marked hiking path to a

You might also like