Discovering Careless Response Behavior in Psychometric Data

Abstract

In psychological healthcare considering bipolar Likert scales data as compositional data can enhance statistical validity. Applying an isometric log-ratio transformation yields interval scaled real-valued data. It increases the normal approximation of item response means, reduces statistical biases and enhances the statistical power of the Pearson correlation test and two-sample t-tests (paired and unpaired) affecting linear regression, partial least squares path modeling and moderator analysis. Mental overload, missing attention, faking or social desirability can corrupt a test person's answers in a psychometric survey. As a result, the corresponding questionnaire data are useless affecting subsequent analyses and interpretations. Aiming to detect careless response behavior as statistical outliers we compare the well-known Mahalanobis-distance to a multivariate projection pursuit method. Performing outlier detections with traditional and with isometric log-ratio transformed data we point out the superiority of the compositional data interpretation of psychometric bipolar scales data.