Empirical Comparison of Causal Machine Learning and Post-Hoc AI Interpretability Models for Risk Factor Analysis: An...

Area of Science:

• Medical Education
• Health Services Research
• Computational Statistics

Background:

• Specialty choice among medical students is critical for equitable healthcare access.
• Previous research primarily described specialty choice patterns.
• Emerging predictive models can identify influential factors but may conflate association with causation.

Purpose of the Study:

• To investigate the causal relationship between medical student motivations and personality traits with their eventual specialty career choice.
• To compare causal inference methods with predictive model explanations (SHAP) for understanding specialty selection drivers.

Main Methods:

• Utilized Double/debiased machine learning (DoubleML) for causal effect estimation.
• Analyzed data from 399 medical students, assessing Year 4 motivations and Big Five personality traits.
• Correlated these factors with Year 6 specialty career choices (person-oriented vs. technically oriented).

Main Results:

• High motivation for surgery (level 6) decreased person-oriented choices by 0.37 (p < .001).
• High motivation for general practice increased person-oriented choices by 0.265 (p < .001).
• Psychological traits did not show significant effects (p > 0.05); SHAP explanations diverged from causal effects for weaker motivations and personality traits.

Conclusions:

• Specific motivations, particularly for surgery and general practice, are strong predictors of medical specialty choice.
• SHAP values from predictive models should be interpreted cautiously as they may not reflect true causal influence, especially for weaker or correlated factors.