High Leverage Points and Outliers in Generalized Linear Models for Ordinal Data

Abstract

A generalized hat matrix based on Φ-divergences is proposed to determine how much influence or leverage each data value can have on each fitted value to a generalized linear model for ordinal data. After studying for evidence of points where the data value has high leverage on the fitted value, if such influential points are present, we must still determine whether they have had any adverse effects on the fit. To evaluate it we propose a new family of residuals based on Φ-divergences. All the diagnostic measures are illustrated through the analysis of real data.