In this application, participation and recreational site selection decisions for Delaware trout fishing are jointly modeled in the context of both a repeated standard multinomial logit and random-parameter logit (mixed logit) using a 1998 participation data set gathered by state biologists for 316 randomly drawn anglers who took at least one trout fishing trip at one of the eight Delaware designated trout fishing sites. The data set has been augmented with a data file of site characteristics collected separately and including trip costs to and from each site, physical dimensions and available amenities.

The mixed logit generalizes the standard logit and captures unobservable heterogeneity by allowing coefficients to randomly vary over anglers rather than being fixed. The model does not exhibit the restrictive "independence from irrelevant alternatives" (IIA) property of logit and can represent any substitution pattern. Also, in situations with repeated choices over time, accounting for preference heterogeneity implies that the unobserved utility associated with any alternative is necessarily correlated over time for each angler.

A series of hypothetical and actual policy scenarios of interest for evaluation are simulated. The hypothetical welfare scenarios include (1) closures of individual and multiple trout fishing sites and (2) changes in advisory scenarios. The actual policy scenarios consist of actual decisions made by the State Division of Fish and Wildlife and include (1) site change scenarios and (2) changes in the quantity of trout stocked. The unbounded normal and bounded uniform and triangular distributions are considered to investigate and assess the impact of the choice of the distribution of anglers' tastes on estimated coefficients and predicted welfare changes.

The results of the heterogeneous models are compared to each other and the standard multinomial logit. Overall, I find evidence of significant heterogeneous preferences throughout the population for all trout fishing site attributes. I also find the bounded distribution specifications to be a significant improvement in fit and plausibility. Finally, the bounded-support distributions generally produce welfare estimates that are bounded below by the standard logit and above by the normally distributed welfare estimates.