Examining response aberrance as a cause of outliers in statistical analysis

Abstract

This study examined to what extent participants who produce aberrant responses were in fact outliers in statistical analysis. Participants of this study were high school students (N=2983) who filled out three personality questionnaires. Response aberrance for these instruments was detected using infit, outfit, and person-fit statistics under Rasch modeling, all of which reflect the degree to which response patterns conform to the model. According to the person-fit cutoff, participants were divided into three categories: overfit, fit, and underfit. Mahalanobis Distance (MD) was used to identify participants classified as outliers, based on a simple regression analysis. Analysis of variance highlighted significant differences between these three categories. The study found that underfit persons were more likely exhibited higher MD values than overfit or fit persons, meaning that they tended to perform as outliers. The correlation coefficients between two variables considerably increased after underfit persons were excluded in subsequent analyses. Another result showed that participants tended to consistently produce aberrant responses across the questionnaires, but that they did not consistently perform as outliers.