Diagnostics for categorical response models based on quantile residuals and distance measures

TL;DR

This paper introduces a new diagnostic approach for polytomous categorical data models using randomized quantile residuals and distance measures, improving visualization and interpretation.

Contribution

It proposes the use of randomized quantile residuals combined with Euclidean and Mahalanobis distances for better diagnostics in categorical response models.

Findings

Quantile residuals perform well in simulations.

Distance measures enhance graphical interpretation.

Method applied successfully to real data examples.

Abstract

Polytomous categorical data are frequent in studies, that can be obtained with an individual or grouped structure. In both structures, the generalized logit model is commonly used to relate the covariates on the response variable. After fitting a model, one of the challenges is the definition of an appropriate residual and choosing diagnostic techniques. Since the polytomous variable is multivariate, raw, Pearson, or deviance residuals are vectors and their asymptotic distribution is generally unknown, which leads to difficulties in graphical visualization and interpretation. Therefore, the definition of appropriate residuals and the choice of the correct analysis in diagnostic tools is important, especially for nominal data, where a restriction of methods is observed. This paper proposes the use of randomized quantile residuals associated with individual and grouped nominal data, as well as Euclidean and Mahalanobis distance measures, as an alternative to reduce the dimension of the residuals. We developed simulation studies with both data structures associated. The half-normal plots with simulation envelopes were used to assess model performance. These studies demonstrated a good performance of the quantile residuals, and the distance measurements allowed a better interpretation of the graphical techniques. We illustrate the proposed procedures with two applications to real data.