Testing Predicted Choices Against Observations in Probabilistic Discrete-Choice Models

Probabilistic discrete-choice models, such as multinomial logit models, are widely used to predict changes in market shares or total demand resulting from changes in policy variables under management control. These models often are evaluated in terms of their ability to predict choices in a holdout sample. This paper presents a new test for comparing predicted and observed choices. The results of a Monte Carlo experiment indicate that the new test has good finite-sample properties and high power in several circumstances likely to arise frequently in applications.