A Comparison of Model Selection Procedures for Latent Class Discrete Choice Models

The use of the latent class model (LCM) is a popular way of modelling discrete choices. It is a good alternative to the traditional multinomial logit model (MNL), but the larger number of parameters comes with a risk of overfitting the data. A good model selection procedure is critical when the LCM is applied in an empirical setting. In this paper, three existing model selection methods are researched and placed in the context of the LCM. The methods are then applied in an empirical setting. Panel data on the purchases of saltine crackers in the Rome (Georgia) market is used to illustrate the methods. This research finds that (i) existing selection methods other than the most commonly used Bayesian Information Criterion (BIC) can provide valuable information during the model selection process and (ii) none of the three methods considered in this paper provide enough information on their own to adequately select a single model.