Treatment-Effect-Based Versus Risk-Based Targeting of Care Management Outreach in Medicaid: A Retrospective Cohort Study with Machine Learning

Medicaid care-management programs typically allocate scarce outreach capacity to beneficiaries with the highest predicted risk of an acute event, assuming that risk and responsiveness are aligned and stable across short intervals. The authors tested whether targeting outreach by predicted individualized treatment effect-the conditional average treatment effect (CATE) recomputed each calendar month-outperforms risk-based targeting. The authors analyzed 164,063 adult Medicaid beneficiaries (2,670,806 person-months) enrolled in community-based care management in Washington and Virginia between January 2023 and December 2025. The exposure was a completed care-manager telephone contact within a calendar month; the primary outcome was an emergency department visit or hospital admission within 30 days. CATEs were estimated using a causal forest with cross-fitted propensity and outcome-model nuisance functions, augmented by within-person fixed effects. Two monthly allocation rules were compared at 10% population capacity: a risk-based rule (top decile predicted event probability) and an effect-based rule (top decile predicted CATE). Policy values were estimated via doubly-robust off-policy evaluation. Findings were validated by cross-state replication, a marginal structural model targeting for time-varying confounding, and a staggered-rollout instrumental variable with pretrend, exclusion-restriction, and monotonicity diagnostics. Effect-based targeting prevented 13.3 (95% CI, 12.8-13.9) acute events per 2000 members per month versus 2.5 (95% CI, 2.3-2.7) under risk-based targeting-a 5.3-fold (95% CI, 4.9-5.7) improvement. Within-person variance accounted for 63.6% of total CATE variance. Gains were similar across racial/ethnic groups and states. Monthly allocation of Medicaid care-management outreach by predicted treatment effect substantially increased the number of acute events prevented at fixed capacity without widening between-group disparities.