Estimating Discrete Choice Models with Incomplete Data

In this paper a new approach is developed for estimating discrete choice modeling parameters for data sets in which one of the alternatives is never observed to have been chosen. This estimation approach, inspired by methods used in revenue management, is applied to a multinomial logit model that has a full set of identifiable alternative specific constants. Parameter estimates are found by combining disaggregate data (obtained from a random or exogenous sample) with aggregate observations of market shares for the observed choices. The method provides an estimate of the market share for the alternative that was not observed to be chosen in the estimation data set. This estimate is significant because it suggests that the transportation community can forecast demand for a new alternative or demand for alternatives that are expensive to collect with intercept surveys without having observations for individuals who have chosen these alternatives. The methodology can be extended to other generalized extreme value models.