Boosting conditional logit model

A componentwise smoothing spline-based boosting procedure is developed for the conditional logit model to estimate the covariate effects nonparametrically. The proposed method can be applied to discrete choice modeling to predict the choice outcomes. Our boosting procedure possesses the properties of slow over-fitting behaviour, automatical variable selection, consistent approximation to the utility function, and the ability to capture the potential nonlinear covariate effects. We show in the simulation studies that the method can provide accurate estimates of the true functional forms of the covariate effects and can select the predictors that are most related to the choice utility. The proposed boosting conditional logit procedure is also applied to two real datasets and its prediction accuracy is demonstrated to be superior to that of the conventional conditional logit regression.