Estimation of Discrete Choice Models in Retrospective Samples

Abstract

The estimation of the parameters of discrete choice models in retrospective samples is often more complex than in prospective studies. In prospective studies estimation usually employs classical maximum likelihood techniques; however, when samples are stratified on the outcome variable, classical maximum likelihood estimators are often biased in large samples. A number of alternative consistent estimators have been proposed by Manski and McFadden (1981). We have found one of their estimators to be applicable to a wide variety of problems and easy to implement. In this article we describe the Manski and McFadden approach, give a variation of the approach when modeling sets of retrospectively sampled outcomes, and present an example of the application of these techniques in a survey research context.