A tree-based method for selection of variables in models for ordinal data

In this paper we propose a strategy aimed at variable selection in models for ordinal data. The procedure core exploits the primary splitters, as they arise from a tree-based procedure. Indeed, the latter is based on a recursive partitioning approach that increasingly reduces the so called impurity of the response variable based on the covariate values. Since the reduction in impurity provides a natural importance ranking of the covariates, this allows to identify those relevant among the available ones. The proposed strategy is, then, used for detecting relevant covariates in a case study concerning the consumers’ preferences towards different types of smoked salmons.