Nested Logit Models Which Are You Using?

Abstract

The adoption of disaggregate analysis in transportation and other fields has led to widespread use of choice models to describe the influence of the characteristics of decision makers and the attributes of alternatives and choices. The multinomial logit model (MNL) is the most used because of its relative simplicity, the potential to add new alternatives, its ease of estimation, and the wide availability of estimation software. However, concerns about the restrictive assumptions of the MNL (independent and identical distribution of error terms) and its properties have led to a search for more robust model structures. The nested logit model has become widely used in a variety of contexts because of its ability to represent similarities (covariance of error terms) among groups of alternatives. It is not widely recognized that there are two forms of the nested logit model. These different models, the utility maximization nested logit (UMNL) and the nonnormalized nested logit (NNNL), have very different properties, which produce different estimation results, behavioral interpretations, and forecasts. The use of nested logit estimation requires a thoughtful choice from these model structures. Model structure and properties of the UMNL and NNNL models are described and compared, and the differences between them are illustrated analytically and empirically. Although the selection of a structure is a matter of judgment, the UMNL model is preferred because it is based on utility maximization and it has intuitively reasonable elasticity relationships and interpretation of utility parameters across alternatives.