Employee Mental Health and Firm Performance Machine-Learning Assessment: Empirical HR analytics Study

Abstract

Employee mental health is increasingly discussed as a determinant of productivity and sustainable firm performance, yet evidence that integrates management theory with machine-learning (ML) based HR analytics remains limited. Using cross-sectional survey and HRIS data from 612 employees in 27 knowledge-intensive firms, we examine whether mental health is a robust predictor of a multi-source performance composite (supervisor-rated task performance and discretionary effort, combined with unit financial and quality KPIs). Guided by JD-R and AMO, we specify a process view in which job demands and resources shape mental health and engagement, and engagement is a key pathway linking mental health with performance. We compare a baseline linear model with random forests and gradient boosting machines and use permutation importance and SHAP to interpret model predictions. The best performing ML model explains substantially more out-of-sample variance than the linear baseline, and mental health and engagement consistently rank among the strongest predictors. Importantly, we distinguish predictive association from causal explanation: SHAP explains the fitted model’s predictions, not mechanisms. We discuss how non-linear prediction patterns can motivate refined theorizing about boundary conditions in JD-R/AMO and outline implications for responsible, ethics-aware HR analytics.