A generalized additive model for discrete-choice data

The usual assumption of a linear-in-parameters utility function in a multin omial legit model is relaxed by a sum of one-dimensional nonparametric func tions of the explanatory variables. The model generalizes the logistic regr ession of the generalized additive model for a binary response to a qualita tive variable that can assume more than two values. Simulation studies show that the proposed method can recover underlying nonlinearity in utility of various shapes. The model is applied to consumer panel data collected by b ar-code scanners from two product categories, and the marketing implication s are sought.