Empirical Comparison of Causal Machine Learning and Post-Hoc AI Interpretability Models for Risk Factor Analysis: An Application to Medical Specialty Choice
How medical students choose specialties shapes access to care. Prior work mostly describes patterns; newer prediction tools can rank influential factors but may blur association with true drivers. Using a curated cohort of 399 students, we examined Year 4 motivations for a given specialty (six items, six levels) and personality traits (Big Five) in relation to Year 6 specialty career choice (person vs technically oriented). We estimated effects with Double/debiased machine learning (DoubleML) and contrasted them with SHAP explanations from an earlier predictive model. Strong motivation for surgery at level 6 lowered the probability of a person-oriented choice by 0.37 (p < .001); high motivation for general practice raised it by 0.265 (p < .001). Other motivation signals were smaller. Psychological traits showed no clear effects (p > 0.05). SHAP broadly matched directions for the strongest items but diverged for weaker ones (e.g., anesthesiology, radiology). Comparing causal and predictive explanations, SHAP directions generally matched DoubleML for strong, well-separated motivations (e.g., surgery level 6, general practice) but diverged for weaker or correlated signals (radiology, anesthesiology, emergency medicine, mid-level psychiatry) and for psychological traits. These discrepancies caution that SHAP values reflect model-conditional associations rather than causal effects, so predictive importance should not be interpreted as causal influence.