Assessing Opt-Out Options for Discrete-Choice Stated Preferences: Results From a Saltwater Angling Survey by Melissa C. Ruby F. Reed Johnson Kristy E. Mathews Triangle Economic Research 1000 Park Forty Plaza, Suite 200 Durham, NC 27713 (919) 544-2244 (Voice) (919) 544-3935 (Fax) Paper Presented at the 1998 AAEA Meeting, Salt Lake City, UT Draft: 15 May, 1998 Do Not Cite or Quote Without Permission Copyright 1998 by: Triangle Economic Research All Rights Reserved
Assessing Opt-Out Options for Discrete-Choice Stated Preferences: Results From a Saltwater Angling Survey Abstract An important methodological issue in stated-preference (SP) experiments is the inclusion and format of an opt-out option. A split-sample design is used in a saltwater angling SP survey to test how including a no-trip option influences attribute salience relative to an option that allows respondents to choose their customary site.
Introduction Assessing Opt-Out Options For Discrete-Choice Stated Preferences: Results From A Saltwater Angling Survey Stated preference (SP) surveys are increasingly applied to a variety of environmental economic contexts. SP methods have been used as an alternative to contingent valuation (CV) and advanced travel cost models to value recreational trips and other public goods (Gan and Luzar, 1993; MacKenzie, 1993; Roe, Boyle, and Teisl, 1996; Opaluch et al., 1993; Johnson and Desvousges, 1997). In addition, SP methods are increasingly combined with revealed preference (RP) methods to obtain a richer view of environmental preferences (Adamowicz, Louviere, and Williams, 1994; Adamowicz et al., 1997). Finally, the National Oceanic and Atmospheric Organization has recently endorsed the use of SP to determine the scale of in-kind compensation for lost naturalresource services. Thus, SP techniques are playing a more prominent role in many areas of environmental economics. Many methodological issues associated with SP techniques remain unresolved in environmental economics. One of these issues is the decision to include an opt-out alternative in discrete-choice SP questions. Employing an opt-out option avoids a forced choice by allowing respondents to select another alternative if they do not prefer any of the hypothetical goods in the choice set (Carson et al., 1994; Olsen and Swait, 1997). There are two main forms of opt-out options used in market research: the nopurchase option and the my current brand option. The equivalent formats in a recreation context are no trip and my current site, respectively. Little attention has 1
been given to determining under which circumstances each format should be used. Yet, the type of opt-out option may have a substantial impact on results because the format may induce respondents to evaluate the choice sets in different ways. The purpose of our SP study is to identify the most important features of a saltwater fishing experience along the central Gulf Coast of Texas. The results of this analysis will be used to inform the selection of restoration alternatives in a natural resource damage assessment (NRDA). Respondents to the survey were prescreened saltwater anglers. In addition, this central coast region of the Texas supports significant amounts of recreational fishing and offers numerous fishing sites. Thus, the argument for including a my current site opt-out option appears to apply in this context. In this paper, we evaluate the effect of opt-out options on the choices of recreational saltwater fishing sites. We compare the results from an opt-out split-sample treatment for differences in attribute salience and potential policy applications. Because of the panel nature of the data, we estimate the data using random-parameters logit techniques in addition to standard conditional logit models. Study Design The experimental design consists of the specific attributes, attribute levels, and fishing site pair profiles that appear in the SP questions. Table 1 contains the attributes and levels used in our SP survey. The experimental design consists of main-effects, nearly orthogonal arrays of 30 choice sets using the attribute levels in Table 1. We blocked the SP experimental design into two sets of 15 SP questions each. We randomly assigned a 15-question choice set to each respondent. 2
In addition to the two different design blocks, we included an experimental treatment on the opt-out alternative in each choice set. The first treatment asks respondents to assume that Site A and Site B are the only two sites available for their next fishing trip. Respondents who do not like either of the two hypothetical fishing sites offered can select the third alternative of Neither Site A nor Site B, I will not go saltwater fishing. In the second treatment, respondents who do not like either of the two hypothetical fishing sites may opt-out by selecting the third choice of Neither Site A nor Site B, I will fish at another site, and providing the name of that fishing site. The questionnaire was administered to 3,488 anglers from three counties on the central coast of Texas from December 1996 to March 1997. The study used a combination mail-telephone mode for survey administration. The response rate varies from a low of 68 percent to a high of 83 percent, depending upon method of calculation. The final sample size for the SP survey is 1,345 respondents. Conceptual Framework The SP survey includes a series of choice judgments with three alternatives in each choice set. The linear specification of utility for the three alternatives is: i i i (1) U = V + ε X β+ ε where U i, j = 1, 2, is the utility of each of the two hypothetical sites and U i 0 is the respondent's utility for the opt-out option in all t choice sets. X is a vector of site attributes, β is a vector of marginal-utility parameters, and ε designates error terms. i 3
Assuming ε follows a type-one extreme-value error structure, the probability that alternative j will be selected from choice set t is a standard conditional-logit expression. Conditional logit models are known to be subject to violations of the irrelevance of independent alternatives assumption (IIA) and do not account for correlations within each respondent's series of choices. Revelt and Train (1998) recently have proposed using random-parameter logit (RPL) for SP data similar to ours. RPL is not subject to the IIA assumption, accommodates correlations among panel observations, and accounts for uncontrolled heterogeneity in tastes across respondents. We modify Equation (1) to introduce respondent-specific stochastic components, η i, for each β: i i i i i (2) U = V + ε X ( β+ η) + ε The standard conditional logit expression now becomes: 3 j1 j2 j15 8 i i i i (3) Prob C = C, C,... C, = 15 t= 1! i i i exp µ V ( β+ η ) 2 k= 0 i i i exp µ V ( β+ η ) We compare both the standard logit and the RPL estimates of the SP model in this study. Analysis Results The first model shown in Table 2 is a standard conditional logit model for the Not Go Fishing treatment. Many of the variables are attributes described earlier. The results of this model indicate that almost all of the site characteristics are highly significant in kt " $ # 4
influencing site choice. In addition, all of the site attributes, except for RESTROOMS and LNMILES, are positive, indicating that the presence of these attributes increases the probability of site selection, other things equal. The negative coefficient on the distance variable suggests that anglers derive less utility from a fishing trip when they have to drive farther, which is consistent with expectations. The basic model for the Prefer Another Site treatment, also shown in Table 2, reveals similar patterns. For instance, the NOADVISORY coefficient is the largest in magnitude and by approximately the same magnitude as in the Not Go Fishing treatment. In addition, all of the site characteristic coefficients are positive except for RESTROOMS and LNMILES. However, the site amenities of restrooms and bait shops are not significant, suggesting that these amenities do not enhance site attractiveness relative to limited parking in this treatment model. Also, the aesthetic variable NOVIEW is not significant, indicating that aesthetics did not influence site selection. As discussed in the previous section, the RPL technique avoids potential sources of bias. Table 2 reports RPL results corresponding to the basic models. However, in the RPL models, each parameter includes both a systematic and a random component, and thus, the model estimates a mean and a standard deviation for each distribution. Treating the site characteristics as random parameters allows us to test for the degree of heterogeneity in preferences across respondents. The only site characteristic that is fixed is LNMILES in order to provide a basis for normalizing utilities. Allowing preferences for site attributes to vary across respondents shows that there is quite a bit of unexplained heterogeneity in respondent preferences in the Not Go 5
Fishing treatment. Most of the EST_STD variables are highly significant, indicating statistically different preferences for these characteristics across respondents. We can assess differences in the amount of heterogeneity across attributes by comparing the relative magnitude of the standard deviation coefficient with the mean parameters. For instance, there are four variables for which the standard deviation coefficient is larger than the mean estimate, suggesting very large variation in preferences across respondents. These variables include BOAT, GOODPARK, NOVIEW, and LOWCONGEST. Comparing the RPL model to the basic model in the Prefer Another Site treatment yields somewhat different disparities between the two models. For the RPL model, no variables lose significance and several gain significance. Thus, the basic model in this treatment shows more bias than did the basic model in the Not Go Fishing treatment and in the direction of suppressing attribute salience for most characteristics. One reason for this difference could be that the Prefer Another Site model incorporates site characteristics from alternative sites that respondents specified. We would expect respondents favorite sites to differ greatly from other respondents favorite sites based on differences in location and preferences. Thus, the Prefer Another Site models involve greater preference heterogeneity and much of this variation is not controlled for in the basic model. For both conceptual as well as empirical reasons, we use only the RPL models in subsequent analysis because they are more accurate and unbiased measures of angler preferences. Because the models may reflect different underlying scales, we cannot compare the coefficients directly across models. Consequently, to understand how 6
treatment of the opt-out option affects attribute salience, each model must be rescaled by a common coefficient. In this case, we rescale all of the attribute level mean parameters by the negative of the coefficient on LNMILES. Table 3 contains the rescaled parameters for each treatment, as well as the t- statistic comparing the attribute levels across the two treatments. First, the variables NOADVISORY, REDDRUM, FLOUNDER, TROUT, BOAT, and LOWCONGEST have significantly higher coefficients in the Prefer Another Site treatment than they do in the Not Go Fishing treatment. These results suggest that when respondents are given the choice of specifying an alternative site, they are more likely to specify sites with these characteristics, thus increasing the salience of these attributes. Another interesting result of this comparison is the number of site attributes that are not significantly different between the two treatments. Thus, treatment of the opt-out option does not seem to affect variables that are not as important to anglers. Implications For Policy Applications Although treatment has some effect on individual attribute results, it is important to understand how the choice of opt-out treatment may affect potential policy applications. To this end, we have created two policy simulations to determine how changes in welfare estimated by the two treatments differ for a given policy improvement. In the first scenario, fifty saltwater fishing sites exist in the study area. For simplicity, all fifty of these sites have average site characteristics, except for the fish consumption advisory attribute. For this attribute, 49 sites have no advisory and one site 7
has an advisory. The policy implemented is to improve water quality at the one polluted site and thus remove the advisory. We estimate welfare before and after the policy change, using log-sums, and interpret the difference between the two estimates as the change in welfare resulting from the policy. The results are almost identical across treatments, 0.0195 for the Not Go Fishing treatment compared with 0.0199 for the Prefer Another Site treatment. In the second scenario, only ten saltwater fishing sites exist in the study area. Of these ten sites, all have average site characteristics and all have fish consumption advisories. In this case, the policy implemented is to improve water quality at three of the ten sites and add access to two more sites with no advisories. As a result, there would still exist seven sites with fish consumption advisories but there also would be five sites without fish consumption advisories. For this scenario, the policy would yield a change in welfare of 2.7223 under the Not Go Fishing treatment but a change in welfare of 3.5794 under the Prefer Another Site treatment, a difference of approximately 30 percent. These two scenarios yield different results concerning the opt-out treatment because the differences in individual attribute levels produced by the two treatments will matter in policy analysis when they have the biggest opportunity of influencing the probability of site selection. Thus, changing the attribute for the fiftieth site in the first scenario does not substantially change the probability of visiting that site under either treatment. However, in the second scenario, there are both fewer sites and the sites are of poorer quality. Improving several existing sites and adding two new good sites 8
substantially increases the probability that those sites would be visited over the existing ones. Conclusions The results of this study show that the opt-out treatment can influence results under certain circumstances. The effect of the opt-out treatment appears to be most influential on site characteristics that are the most salient to respondents. We therefore expect to find differences between the two treatments in policy situations in which changes in these characteristics are likely to be important determinants of welfare. These results suggest that is important to consider in which situations each treatment would be applicable. In this context, our sample consisted of anglers in a study area that contains many substitutes. Thus, it seems that applying a Prefer Another Site treatment is more appropriate than a Not Go Fishing treatment because in reality anglers are frequent purchasers of saltwater fishing sites. In addition, by including the site characteristics of these anglers preferred sites in estimation, we obtain a more accurate picture of the salience of site attributes. Although the Prefer Another Site treatment seems to be more appropriate in this situation, other policy contexts may govern the use of the Not Go Fishing, or similar non-participation, opt-out treatments. For example, in situations when the sample is not likely to consist of usual recreators or when there are no real substitutes present in the study area, applying a no-purchase opt-out treatment may be more appealing conceptually. Regardless of the context, this aspect of the discrete-choice SP questions should be given careful thought in designing an SP survey. 9
Table 1. Attributes and Levels Used in SP Survey Fishing mode Attribute Additional distance to fishing or launch site Species and catch rate Surroundings Congestion Amenities Fish consumption advisory Pier Boat Levels 5 additional miles 15 additional miles 30 additional miles 1 Red Drum 3 Red Drum 2 Flounder 10 Flounder 2 Speckled Trout 10 Speckled Trout No view of industrial plants View of industrial plants Many people or boats in sight Some people or boats in sight Limited parking Good parking Good parking and restrooms Good parking, restrooms, and bait shop No advisory (Fish can be eaten) Fish should not be eaten 10
Table 2. Comparing Model Results Between Basic Logit Models and Random-Parameters Logit Models for the Opt-Out Option Treatments Not Go Fishing Treatment Prefer Another Site Treatment Variables Basic Coefficient T-ratio Random-Parameters Logit Coefficient T-ratio Random-Parameters Logit Coefficient T-ratio Random-Parameters Logit Coefficient T-ratio RED DRUM 0.1992 *** 9.714 0.2600 *** 8.869 0.1799 *** 9.604 0.3682 *** 13.457 EST_STD 0.3033 *** 9.361 0.2931 *** 8.017 FLOUNDER 0.0803 *** 14.644 0.1121 *** 14.026 0.0703 *** 13.123 0.1353 *** 16.360 EST_STD 0.0888 *** 9.286 0.1118 *** 11.080 TROUT 0.0883 *** 15.711 0.1418 *** 19.505 0.0850 *** 15.523 0.1491 *** 20.084 EST_STD 0.0640 *** 7.144 0.0936 *** 8.446 BOAT 0.2762 *** 8.550 0.2506 *** 3.803 0.2555 *** 8.340 0.5332 *** 7.165 EST_STD 1.5941 *** 30.193 1.8785 *** 27.521 GOODPARK 0.2415 *** 4.979 0.2449 *** 2.612 0.1635 *** 4.055 0.1965 *** 2.998 EST_STD 1.0879 *** 18.071 0.7805 *** 8.683 RESTROOMS -0.1489 *** -3.055-0.1394 * -1.907-0.0439-0.985-0.1042 * -1.679 EST_STD 0.0124 0.078 0.4212 *** 5.474 BAITSHOP 0.0884 * 1.731 0.0983 1.292 0.0157 0.348 0.1581 ** 2.278 EST_STD 0.0616 0.312 0.6645 *** 7.975 NOVIEW 0.1360 *** 4.179 0.0501 0.754 0.0411 1.419 0.0735 1.447 EST_STD 0.9016 *** 14.814 0.7933 *** 13.656 LOWCONGEST 0.1127 *** 3.245 0.1452 *** 2.629 0.1540 *** 4.406 0.2917 *** 3.417 EST_STD 0.5035 *** 6.500 1.1740 *** 16.972 NOADVISORY 2.0916 *** 60.782 3.3684 *** 36.519 1.9018 *** 53.527 4.2528 *** 32.159 EST_STD 2.2590 *** 25.162 2.8301 *** 25.842 LNMILES -0.1954 *** -6.982-0.5212 *** -16.864-0.2193 *** -8.598-0.3754 *** -14.992 ASD_AB -1.2307 *** -10.654-0.9158 *** -7.462 0.5829 *** 15.778 1.0117 *** 21.838 ASD_CMISS 0.6033 *** 4.753-1.4226 *** -28.858 No. of observations 10410 10410 9765 9765 Madalla's pseudo R-square 0.398 0.556 0.410 0.556 McFadden's pseudo R-square 0.237 0.378 0.242 0.372 *** Significant at the 1 percent level ** Significant at the 5 percent level * Significant at the 10 percent level 11
Table 3. Comparing Rescaled Coefficients Across The Opt-Out Option Treatments And Testing For Significant Differences Parameters Not Go Fishing Rescaled Coefficient Prefer Another Site Rescaled Coefficient T-ratio of Differences RED DRUM 0.4988 0.9808-5.235 *** FLOUNDER 0.2151 0.3604-5.415 *** TROUT 0.2721 0.3972-5.170 *** BOAT 0.4808 1.4204-3.996 *** GOODPARK 0.4699 0.5234-0.214 RESTROOMS -0.2675-0.2776 0.047 BAITSHOP 0.1886 0.4212-0.987 NOVIEW 0.0961 0.1958-0.536 LOWCONGEST 0.2786 0.7770-1.987 ** NOADVISORY 6.4628 11.3287-12.343 *** LNMILES -1.0000-1.0000 0.000 ASD_AB -1.7571 2.6950-16.747 *** ASD_CMISS -3.7896 *** Significant at the 1 percent level ** Significant at the 5 percent level * Significant at the 10 percent level 12
References Adamowicz, W., J. Louviere and M. Williams. Combining Revealed and Stated Preference Methods for Valuing Environmental Amenities. J. Environ. Econ. and Manage. 26(1994)271 92. Adamowicz, W., J. Swait, P. Boxall, J. Louviere and M. Williams. Perceptions versus Objective Measures of Environmental Quality in Combined Revealed and Stated Preference Models of Environmental Valuation. J. Environ. Econ. and Manage. 32(1997)65 84. Carson, R.T., J.J. Louviere, D.A. Anderson, P. Arabie, D.S. Bunch, D.A. Hensher, R.M. Johnson, W.F. Kuhfeld, D. Steinberg, J. Swait, H. Timmermans, and J.B. Wiley. Experimental Analysis of Choice. Mktg. Letters 5:4(1994)351 368. Gan, C. and E. J. Luzar. A Conjoint Analysis of Waterfowl Hunting in Louisiana. J. Agr. and Appl. Econ. 25(1993)36 45. Johnson, F. R. and W. H. Desvousges. Estimating Stated Preferences with Rated-Pair Data: Environmental, Health, and Employment Effects of Energy Programs. J. Environ. Econ. and Manage. 34(1997)79 99. Mackenzie, J. A Comparison of Contingent Preference Models. A. J. Agr. Econ. 75(1993)593 603. Olsen, D.G., and J. Swait. Nothing is Important. Working Paper, Alberta Canada: Faculty of Management, University of Calgary. 1997. R-1
Opaluch, J.J., S. K. Swallow, T. Weaver, C.W. Wessells and D. Wichelns. Evaluating Impacts from Noxious Facilities: Including Public Preferences in Current Siting Mechanisms. J. Environ. Econ. and Manage. 24(1993)41 59. Revelt, D., and K. Train. Mixed Logit With Repeated Choices: Households Choices of Appliance Efficiency Level. Rev. of Econ. And Stats. Forthcoming. 1998. Roe, B., K.J. Boyle, and M.F. Teisl. Using Conjoint Analysis to Derive Estimates of Compensating Variation J. Environ. Econ. and Manage. 31(1996):145 159. R-2