Files
Abstract
The discrete choice model is a powerful method that can quantify consumer preferences for both non-market and market goods. In either stated preference surveys with choice experiments or revealed preference data like the retail-level transaction records, discrete outcomes are in the form of multivariate correlated variables. From the view of multivariate probability distributions, this dissertation investigates several methods that improve the estimation and interpretation of discrete choice models. First, for the mixed logit model, Gauss-Hermite integration is found to be a powerful alternative to maximum simulated likelihood when the number of random parameters is moderate($leq$6). It avoids simulation bias and simulation noise and only incurs controllable approximation error. Further, the Bayesian approach and the block delete jackknife are tested to outperform the Delta method in describing the distribution of mean Willingness to Pay in the mixed logit model. The virtues of a normally distributed cost coefficient is also validated with both empirical and synthetic data set. Finally, in the sense of aggregating discrete choices outcomes in a real market, a utility-consistent count system is developed as a tool to analyze consumer brand choices.