Full-Information Selection Bias Correction for Discrete Choice Models with Observation-Conditional Regressors

Abstract

We examine self-selection in polychotomous choice models that construct attribute values for each alternative conditioned on observed choices. Using observations made only when the alternative was chosen ignores private information which was a basis for the decision, biasing resulting estimates. We suggest a full-information maximum likelihood procedure that performs well at the extremes of the choice set in our sample, and use an “identification at infinity” weighting to identify levels. We apply the model to understanding fishing location choice in the economically significant Bering Sea pollock fishery, where expected catches at each location are constructed from harvests observed when that location is chosen.