The use of personality measures to predict work-related outcomes has been of great interest over the past several decades. The present study used machine learning (ML) to examine the optimal level in the personality hierarchy to use in developing predictive algorithms. This issue was examined in a sample of incumbent police officers (N = 1,043) who completed a multifaceted personality measure and were rated on their job performance. Criterion-related validity was investigated as a function of level of operationalization in the personality hierarchy (dimensions, facets, items), scoring method (unit weighting, ordinary leastsquares regression, elastic net regression), content relevance (all items vs. job-related items), and sample size (100, 200, 300, 500, 800). Results showed that empirically derived scores outperformed unit weighting across all levels of the personality hierarchy. The highest validity estimates were consistently obtained using elastic net scoring (with hyperparameter tuning resulting in solutions closer to ridge regression) at the item level, with minimal differences between ordinary least squares and elastic net for dimensions or facets with at least moderate sample sizes (N ≥ 200). An exploratory modeling approach where all item content was used did not outperform scoring when the item pool was relegated to only job-relevant personality traits. Taken together, findings suggest that personality scoring should occur at narrow operationalizations down to at least the facet level. In addition, this study demonstrated how ML can be used to not only maximize criterion-related validity but also to test long-standing theoretical problems in the organizational sciences.
- Bandwidth–fidelity trade-offs
- Employee selection
- Machine learning