Sampling issues are of greater importance in the estimation of statistical models in general, and of discrete choice models in particular. Althoughsimple random sampling strategies are theoretically convenient, they are practically never feasible in a transportation context. However, it is well known that maximum likelihood estimation of the multinomial logit (MNL) model on stratified samples yields to consistent estimates for all parameters except the constants, which must be corrected afterward. This propertydoes not generalize to non-MNL generalized extreme value (GEV) models, sothat classical estimation procedures do not yield to maximum likelihood estimators from choice-based samples. In this paper, we propose a simple estimator providing consistent estimates of all parameters, including the constants, for non-MNL GEV models, which does not require an a priori knowledge of the sampling probabilities. The new estimator require minor modifications of existing estimation codes. We illustrate the concept on synthetic and real data for nested and cross-nested logit models. The results showthat the classical "exogenous sample maximum likelihood" estimator produces biased estimates, even for parameters other than the constants. The newestimator is able to identify all the parameters, including the constantsif the GEV model is non-trivial. For the covering abstract see ITRD E135582.
Abstract