Are simple regression models sufficient for predicting mental health from working conditions? A case study using machine learning approaches

This study explores whether complex machine learning models offer a substantial advantage over simple regression models in predicting employees’ mental health scores based on workplace conditions. Data were collected in an applied setting from N = 23,764 participants across 186 companies in Germany and Switzerland. Mental health was measured using a composite score that included emotional exhaustion, concentration problems, social impairments, and psychosomatic issues. Four predictive models—linear regression, elastic-net regression, decision tree, and random forest—were compared based on out-of-sample predictive accuracy. Results identified no significant performance differences between linear regression and the more complex models, suggesting that working conditions, as measured in this study, primarily predict mental health in an additive manner, i.e., without complex interactions between the working conditions. Furthermore, a linear regression model using only the eight most important predictors, identified through feature importance analysis, achieved predictive accuracy similar to that of the full model (i.e., a linear regression with all predictors), highlighting the potential for the creation of efficient and practical screening tools.