Graphical Structure Learning Identifies Hypothesized Mechanisms for Heterogeneous Treatment Effects in Medicaid...

Area of Science:

• Health Services Research
• Causal Inference
• Population Health Management

Background:

• Heterogeneous treatment effect models identify high-benefit Medicaid beneficiaries but lack mechanistic detail.
• Population health programs require understanding intervention-specific mechanisms for optimization.

Purpose of the Study:

• To combine graphical structure learning and regression-based effect estimation for mechanistic hypothesis generation.
• To investigate intervention-specific mechanisms in a Medicaid managed care program.

Main Methods:

• Employed Peter-Clark algorithm and Greedy Equivalence Search for conditional dependency identification.
• Analyzed associations between baseline characteristics, interventions (behavioral health therapy, clinical pharmacy, community health workers, care coordination), and acute care outcomes.
• Utilized false coverage rate correction for post-selection inference.

Main Results:

• Behavioral health therapy linked to reduced psychiatric admissions (RR, 0.27).
• Clinical pharmacy demonstrated dose-dependent cost reductions.
• Community health workers associated with emergency department visit reduction (RR, 0.62).
• Care coordination linked to emergency department reduction in women (RR, 0.38).

Conclusions:

• Graphical methods can generate mechanistic hypotheses complementing heterogeneous treatment effect models.
• Identified specific intervention-pathway associations warranting further investigation.
• Findings support targeted population health strategies for Medicaid enrollees.