Quantile-based classification trees for ordinal responses

This paper introduces a class of non-parametric classification trees for ordinal responses, hinging on one-way Quantile Anova for the recursive partitioning algorithm. The procedure allows to identify homogeneous groups of observations for the response whose conditional distributions given the selected split are significantly different at possibly many quantiles. Two applications are discussed to show how the proposed method allows to learn response drivers for specific locations of the response range. The first case study illustrates how to determine separately drivers of satisfaction and drivers of dissatisfaction for professional placement of Italian research doctors. The second example, instead, focuses on the propensity to vote for one of the main German political parties, and it aims at identifying the most dissimilar voters' profiles on the basis of differences throughout the whole distribution. Throughout the presentation, conditional inference trees are assumed as the benchmark in a comparative perspective for both interpretation and prediction.