Estimating individual preferences with flexible discrete-choice-models

Discrete-choice-models have come a long way since the field erupted in the seventies with the seminal work of McFadden and colleagues at Berkeley. For more than 25 years the multinomial logit (MNL) model and its close relative, the hierarchical or nested logit (NL) model (a generalization that allowed treating correlated alternatives grouped in nests, the simultaneous brain child of Williams (1977), and Daly & Zachary (1978)) reigned supreme, using mainly revealed preference data. At the end of the 90s the modeller’s tool kit expanded considerably with the installation of both stated preference data, as a strong ally to help treating the problems posed by new alternatives and latent or secondary variables, and the most flexible member of the family, the mixed logit (ML) model, that allows considering two remaining deficiencies of the old guard, heteroskedasticity and taste variations. This paper provides a glimpse of the field and speculates on the potential use of these methods to complement sensory work in the effort to better understand customer preferences.