Combining RP and SP data while accounting for large choice sets and travel mode – an application to forest recreation

Abstract

This paper analyses the use of forests for recreational purposes in Lorraine, France, a region with many forests and easy access for recreational users. This implies that residents in Lorraine can choose between a large set of forests if they decide to visit a forest. The abundance of forests in Lorraine makes identification of the visited forests difficult. To facilitate identification of forests actually visited, we have incorporated an interactive map in a Web-based survey intended to include both revealed and stated preference data. We compare different sampling schemes to define the choice set used for site selection modelling when the actual choice set considered is potentially large and unknown to the analyst. Easy access to forests also implies that around half of the visitors walk or bike to the forest. We apply an error-component mixed-logit model to simultaneously model the travel mode decision and the site selection decision and to combine revealed and stated preference data. Finally, the effect on the willingness-to-pay of changes in forest quality and access is evaluated based on alternative choice set specifications, model specifications and data sources (revealed and stated preference data).

Keywords:

• travel cost method
• choice experiment
• site selection
• choice set
• error-component mixed logit
• travel mode choice

Acknowledgements

The authors would like to thank Christian Piedallu, Jean-Marc Rouselle, Max Bruciamacchie, and Vincent Perez for their contribution to the design of the questionnaire and the data collection and Serge Garcia and two anonymous reviewers for their useful comments.

Notes

1. An overview of French studies on the recreational value of forests can be found in Montagne et al. (2008).

2. Kuriyama, Michael Hanemann, and Hilger (2010) considered 52 beaches in a study with trip information on 4367 trips taken by 617 respondents. However, in general, only a few alternative sites are considered.

3. Endogenous estimation of a choice set was proposed by Haab and Hicks (1997). However, this approach focuses on identifying the ‘actual’ consideration set (Campbell, Hensher, and Scarpa 2014) and is less relevant in coping with the issue of large choice sets since the computational requirements increase exponentially with the size of the potential choice set.

4. Note in the analysis of travel mode that only the RP data are relevant since we did not consider the travel mode in the choice experiment.

5. We have excluded income here since it has the same value in all of the alternatives and therefore cancels itself out.

6. Digital maps have previously been used in online questionnaires, for example, by (Horni, Charypar, and Axhausen 2010) in the empirical formation of choice sets for analysis of shopping behaviour.

7. Budget of the French driver, June 2011 (www.automobile-club.org).

8. The alternative cost of time is generally a source of discussion in the literature. Regardless of the assumption, it is true that individuals will consider a limited time budget and that the transport time will influence the choice of the forest to be visited. Even though we find this discussion to be relevant, it is beyond the scope of this paper.

9. The results presented are based on MXL models. We also estimated conditional logit models. These results are available on request from the authors.

10. The log-likelihood function is estimated using Halton draws (150 draws). Estimation of the model was carried out using Nlogit 4.0 software, whereas the strategic sampling scheme was carried out in MATLAB.

11. As in Lemp and Kockelman (2012), the results are an average of 10 replications of the simple and of the strategic sampling scheme.

12. Here, we use the choice set for the RP data based on the second iteration of the strategic sampling strategy.

13. Estimates in parentheses are based on strategic sampling and MXL estimation.

14. With simple random sampling, the log-likelihood function is often flat at its optimum value.

15. Pooling of RP and SP data is also rejected in Huang, Haab, and Whitehead (1997).

Additional information

Funding

This work was supported by the French Ministry of Agriculture in the framework of the Pôle de Compétences FABELOR and The UMR 357 Economie Forestiére is supported by a grant overseen by the French National Research Agency (ANR) as part of the “Investissements d'Avenir” program (ANR-11-LABX-0002-01, Lab of Excellence ARBRE).