Essays on choice modeling

The increased interest in modeling preference heterogeneity in Choice Experiments (CE) data has promoted the use of choice models within the family of Logit Mixture models. However, researchers have scarce guidance on how to select the appropriate model. A Monte Carlo study is set up to analyze the performance of different information criteria and tests used to discriminate between models, either nested or non-nested. In particular, within non-nested models, four criteria and three tests are assessed, all of them based on the Kullback-Leibler Information Criterion (KLIC): the AIC, AIC3, CAIC and BIC information criteria, and the tests for non-nested models proposed by Vuong, Horowitz and Ben-Akiva and Swait. Our results indicate that some criteria (especially CAIC) work better than others; and that, when feasible, information criteria should be complemented by the Vuong test, which has a low power, but it virtually never selects the wrong model, while both the Horowitz and the Ben-Akiva and Swait tests too often provide wrong indications. The paper concludes with a CE application dealing with public acceptance of wind farms, where the indications drawn from the Monte Carlo analysis are used to inform model selection.