Joint estimation of angler revealed preference site selection and stated preference choice experiment recreation data considering attribute non-attendance

ABSTRACT

We estimate angler demand models with revealed preference (RP) and stated preference (SP) site selection marine recreational fishing data. We combine RP data from the Marine Recreational Information Program (MRIP) creel survey with SP discrete choice experiment survey data from 2003/2004. There are eight SP trip decisions and one RP trip decision for each of 1928 anglers who provided enough information to be analysed. Joint RP-SP generalized multinomial logit models are estimated. We find that the SP travel cost coefficient is much lower than the RP travel cost coefficient in absolute value, suggesting hypothetical bias in the SP data. This difference is reflected in the willingness to pay estimates, where the SP estimates for improved catch are much higher than the RP estimates. We use inferred attribute non-attendance (ANA) methods to identify respondents who may be ignoring the SP cost variable. The SP cost coefficient accounting for ANA is much higher in absolute value than the SP coefficient from the model that does not account for ANA. The ANA model indicates much more consistency between the RP and SP data. The smaller difference in the travel cost coefficients is also reflected in the willingness to pay estimates.

KEYWORDS:

• Revealed preference
• stated preference
• joint estimation
• recreational fishing

JEL:

• Q51
• Q22
• Q26

Acknowledgements

A previous version of this paper was presented at the AERE ‘Recreation’ session at the 2017 Southern Economic Association meetings and the LEEPout seminar at the University of Exeter in 2021. The authors thank a journal referee for comments on the paper and David Carter for a number of helpful comments on the technical report upon which this paper is based.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 One can also appeal to substitution and income effects stemming from price changes to explain why the demand for quota is downward sloping.

2 There is a large literature that combines RP and SP data, but these studies use only continuous/count data (Hynes and Greene 2013), a combination of discrete and continuous/count data (Huang et al. 2016), or combine data without joint estimation (Anciaes, Metcalfe, and Sen 2020).

3 For simplicity, throughout this section, we assume that catch is synonymous with harvest (i.e. recreational landings) unless otherwise noted.

4 The ability of the GMNL to separately estimate the individual-level scale and preference heterogeneity has been challenged by Hess and Rose (2012) and Hess and Train (2017), who argue that the utility specifications in GMNL models simply allow for more flexible distributions of the preference parameters.

5 Note that this is the GMNL-I model in Fiebig et al. (2009) where $\gamma =1$.

6 Distances were calculated between individual site (INTSITE) and anglers’ home zipcodes (ZIP) with ArcGIS by NMFS economists.

7 The 2003/04 Stated Preference Survey can be found at https://www.st.nmfs.noaa.gov/Assets/econ-human/pdf/SE_SP_2003.pdf.

8 Our trip cost variable differs from current suggestions for state-of-the-art recreation demand models (Lupi, Phaneuf, and von Haefen 2020). The state-of-the-art in recreation demand modelling is to use the variable cost per mile and an opportunity cost of time of between 33% and 75%, if not higher, of the wage rate. The willingness to pay estimates shown below are different than those that would be produced from the trip cost variable measured by on the current literature. Travel costs that include only operating costs (fuel, maintenance and tires), excluding operating and ownership costs (insurance, registration and taxes, depreciation and finance charges) would produce a lower willingness to pay estimates. Travel costs that include the opportunity cost of time will produce a higher willingness to pay estimates. Our results are constrained by the choice experiment that was developed in the early 2000s and should be considered exploratory and not necessarily suitable for policy analysis.

9 We do not weight these models for onsite sampling. Previous research suggests that the MRIP weights had little effect on WTP estimation (Lovell and Carter 2014). Endogenous stratification is less likely a problem with our dataset, since only those RP observations that can be matched with the SP data are analysed.

10 Separate conditional logit models are estimated and presented in the Appendix. In the revealed preference model only two of the four catch and keep coefficients are statistically significant and all of the catch and release coefficients have negative signs. Comparing statistically significant catch coefficients, the RP dolphin and red snapper catch-and-keep coefficients are much larger than the SP coefficients. This may be a result of the unequal size of the catch variables in the underlying data. The marginal value of additional catch when catch is low, as in the RP data, is higher than when catch is relatively high, as in the SP data. Otherwise, the pattern of results is similar to that from more complex econometric models.

11 When these constraints are not imposed the $\tau$ and $\gamma$ estimates become very large. For example, in an unconstrained $\tau$ and $\gamma$ model with a constrained cost coefficient the estimated parameters are $\tau =54$ and $\gamma =19$. Most of the coefficients are similar across these models with all differences less than 10%. The exceptions are coefficients on dolphin release (31% increase), grouper release (37% increase) and king mackerel release (18% increase). In the unconstrained model the standard deviations are 88%–96% smaller and none of these are statistically different from zero. Since the constrained τ and γ models conform to expectations in terms of the random parameters and there will be little difference in the willingness to pay for the catch and keep values, we proceed with the constrained τ and γ models.

12 The pattern of regression results for scaled and mixed logit models is similar to what is presented there. The willingness to pay estimates are 39% and 15% lower from the scaled, mixed and generalized multinomial logit models with scale differences relative to the GMNL models with scale differences. These results are available upon request.

13 The willingness to pay estimates from the ECLC model are significantly lower than the corresponding willingness to pay estimates from the SP conditional logit model presented in the Appendix. The interpretation of the ECLC model as a hypothetical bias correction, as in Koetse (2017), suggests that the willingness to pay estimates that do not account for hypothetical bias are about 300% higher than those that do account for the bias (i.e., attribute non-attendance).

14 The pattern of WTP results for catch-and-release is similar to catch-and-keep.

15 Note that the data are almost 20 years old at the time of this writing. Stated preference willingness to pay estimates have been found to be temporally reliable up to five years but not as much as 20 years between surveys (Skourtos, Kontogianni, and Harrison 2010). Welfare estimates from revealed preference recreation demand models also may not exhibit temporal reliability (Ji, Keiser, and Kling 2020). Finally, the policy context has changed for several species in the Gulf of Mexico and Atlantic Ocean since the survey was conducted (Ropicki, Willard, and Larkin 2018). We caution the reader that this is only an illustration and these estimates should not be used for policy purposes. Also, recent research has highlighted the importance of an efficient property rights structure for the equi-marginal principal to apply and the need to focus on other important economic factors when determining the appropriate allocation (Abbott 2015; Holzger and McConnell 2014). Given that recreational anglers do not purchase quota, whereas commercial fishermen in the Gulf of Mexico do, however, this would compromise our analysis.

Additional information

Funding

This work was supported by National Marine Fisheries Service [grant number MARFIN Award NA14NMF4330220].