ENHANCING RELIABILITY ESTIMATION IN PERSONALITY ASSESSMENTS USING MACHINE LEARNING APPROACHES

Abstract

Reliability estimation is central to psychometrics because it ensures that personality assessments produce consistent and dependable results. Traditional methods, such as Cronbach’s alpha and McDonald’s omega, are widely used but depend on strong assumptions like unidimensionality and tau-equivalence. These assumptions are often violated in real personality data, which can lead to under or overestimation of reliability. In this study, we examine how machine learning (ML) methods specifically Random Forests, Gradient Boosting, and Neural Networks can be applied to improve reliability estimation. We used both simulated data and a secondary dataset of 250 participants who completed the Big Five Inventory (BFI-44), with a four-week retest interval. Results show that ML-based approaches provide more stable and accurate estimates of internal

consistency and test–retest reliability than traditional indices, especially in small samples and multidimensional scales. These findings suggest that combining psychometric theory with computational modeling offers a promising way to advance reliability assessment in applied psychology.