Relation Between EBA and Nested Logit Models

We show that elimination by aspects (EBA) generalizes nested logit and cross‐nested logit models. The latter two models are equivalent to a special case of EBA called preference trees. The transformations between preference trees and nested logit models become more complex when the utilities of alternatives are functions of covariates. In this case, a simple model in one domain corresponds to a complex model in the other. An extended EBA model, in which the utilities of alternatives are functions of covariates, represents a two‐stage choice process. Alternatives are first screened using a probabilistic lexicographic rule and then compared in terms of their compensatory utilities. We provide a typology of the relations between EBA and other logit models, and we discuss issues concerning estimation, statistical testing, and data collection. We describe an application illustrating (1) the process of constructing a preference tree with covariates and (2) the different implications obtained from a preference tree and a comparable nested logit model.