Predicting return-to-work outcomes: A comparative analysis of logistic regression and machine learning in occupational mental health

This investigation compares the analytical performance of logistic regression (LR) and targeted maximum likelihood estimation (TMLE) methodologies for occupational health research using simulated workforce data. The study employs a simulation framework based on empirical observations from 738 Brazilian public university employees, generating a reference population of 10,000 cases with predetermined distributions of 2 critical exposure variables and some covariables. Through systematic sampling across multiple sample sizes (150-5000 observations), the research evaluates the comparative efficacy of traditional regression techniques versus TMLE framework. Key findings revealed important differences between the methods. For analyzing changes in diagnosis chapters, LR consistently produced more accurate estimates, with odds ratios (ORs) closer to the true population value (0.28) compared with TMLE (0.44-0.46). The precision of LR estimates, as measured by root mean squared error, was significantly better, especially in larger samples (0.02 vs 0.16 at n = 5000). The analysis of mental health conditions showed more complex results. While LR estimates were initially unstable in small samples, they converged to the true OR (27.35) as sample size increased. TMLE produced more stable but consistently conservative estimates across all sample sizes (ORs = 15-17). Notably, TMLE confidence intervals only included the true population value in the smallest sample (n = 150), suggesting potential limitations in its application for this type of analysis. These findings carry significant implications for occupational health research methodology. The results challenge prevailing assumptions about TMLE framework automatic superiority, demonstrating that conventional regression techniques can outperform more complex approaches in specific epidemiological contexts. The study highlights the continued relevance of LR for analyzing occupational health outcomes, particularly when investigating well-defined exposure-outcome relationships with adequate sample sizes. The research contributes to methodological literature by providing empirical evidence for analytical decision-making in public health research contexts.