Variable selection in general multinomial logit models

The use of the multinomial logit model is typically restricted to applications with few predictors, because in high-dimensional settings maximum likelihood estimates tend to deteriorate. A sparsity-inducing penalty is proposed that accounts for the special structure of multinomial models by penalizing the parameters that are linked to one variable in a grouped way. It is devised to handle general multinomial logit models with a combination of global predictors and those that are specific to the response categories. A proximal gradient algorithm is used that efficiently computes stable estimates. Adaptive weights and a refitting procedure are incorporated to improve variable selection and predictive performance. The effectiveness of the proposed method is demonstrated by simulation studies and an application to the modeling of party choice of voters in Germany.