The Effects of the Training Sample Size, Ground Truth Reliability, and NLP Method on Language-Based Automatic Interview Scores’ Psychometric Properties

Abstract

While machine learning (ML) can validly score psychological constructs from behavior, several conditions often change across studies, making it difficult to understand why the psychometric properties of ML models differ across studies. We address this gap in the context of automatically scored interviews. Across multiple datasets, for interview- or question-level scoring of self-reported, tested, and interviewer-rated constructs, we manipulate the training sample size and natural language processing (NLP) method while observing differences in ground truth reliability. We examine how these factors influence the ML model scores’ test–retest reliability and convergence, and we develop multilevel models for estimating the convergent-related validity of ML model scores in similar interviews. When the ground truth is interviewer ratings, hundreds of observations are adequate for research purposes, while larger samples are recommended for practitioners to support generalizability across populations and time. However, self-reports and tested constructs require larger training samples. Particularly when the ground truth is interviewer ratings, NLP embedding methods improve upon count-based methods. Given mixed findings regarding ground truth reliability, we discuss future research possibilities on factors that affect supervised ML models’ psychometric properties.