Bayesian inference in recreation demand models: On linking disparate data sources

In considering changes to the environment, policymakers need information on the value placed in environmental amenities. Unfortunately, information on environmental values is sparse and comes from a variety of disparate data sources, including survey data and behavioral data (such as visitation rates). This makes it all the more important to integrate what information is available, so as to best inform decision makers. Bayesian analysis provides a natural framework in which to integrate different sources of information. In my dissertation I develop Bayesian models to address these integrating problems in context of two key problems of cost benefit analysis: (1) Benefits transfer and (2) the combining of stated preference (SP) and revealed preference (RP) data in valuing the same environmental amenity. My dissertation consists of three essays. In the first essay, a hierarchical linear model (HLM) is developed and used to address the benefits transfer issues. In the context of recreation demand, the HLM assumes that different sites (or studies) share a common linear demand structure, but that each site's coefficients can vary, drawn from a common distribution. The variability of these structural parameters across sites indicates the degree of transferability of estimates among sites. In the second essay, a Bivariate Tobit model is used to link SP and RP data. By construct, this model allows researchers to incorporate prior beliefs about the consistency between RP and SP data. By using Bayes factor we can also compare models with different priors. The Gibbs sampling method and data augmentation method suggested by Chib (1992) are used as the primary tools for Bayesian analyses in these two essays. In the third essay I develop mixed logit models as an alternative approach to combining RP and SP data. In this framework, the discrepancies between SP and RP responses are modeled as having a distribution in the population; a discrepancy whose means and variances also depend on observed attributes of the survey respondents. Understanding the sources of these discrepancies can help to better design SP surveys.