Graphical structure learning identifies hypothesized mechanisms for heterogeneous treatment effects in Medicaid population health programs.

Heterogeneous treatment effect models identify Medicaid beneficiaries who benefit most from population health programs but provide limited mechanistic insight. We combined graphical structure learning with regression-based effect estimation to generate hypotheses about intervention-specific mechanisms in a program serving 6396 Medicaid managed care enrollees (2023-2025). The Peter-Clark algorithm and Greedy Equivalence Search identified conditional dependencies between baseline characteristics, intervention exposures (behavioral health therapy, n = 434; clinical pharmacy, n = 632; community health workers (CHWs), n = 2169; care coordination, n = 1042), and acute care outcomes. Four associations warranted investigation: behavioral health therapy was associated with reduced psychiatric admissions (risk ratio [RR], 0.27; 95% confidence interval [CI], 0.19-0.37; E-value, 7.6); clinical pharmacy showed a dose-dependent association with reduced costs; CHWs were associated with emergency department (ED) reduction (RR, 0.62; 95% CI, 0.51-0.74; E-value, 4.5); and care coordination was associated with ED reduction among women (RR, 0.38; 95% CI, 0.27-0.52; E-value, 9.1). False coverage rate correction for post-selection inference (Benjamini and Yekutieli, J Am Stat Assoc. 2005;100(469):71-81) yielded 98.44% CIs that continued to exclude the null for all reported associations. Graphical methods can complement heterogeneous treatment effect models by generating mechanistic hypotheses for confirmation through future trials. This article is part of a Special Collection on Latino Health.