Artificial intelligence-enabled causal estimate of Medicare drug plan integration in cancer care: A doubly robust
Xiangxiang Jiang1, Jayda Snipes1, Jun Wu2
1Department of Clinical Pharmacy and Outcomes Sciences, College of Pharmacy, University of South Carolina, Columbia.
Background:
Artificial intelligence (AI) methods are increasingly used to strengthen policy evaluation in managed care pharmacy. Among Medicare beneficiaries with cancer, which is one of the most clinically complex and costly populations, prescription drug coverage is obtained through either integrated Medicare Advantage Prescription Drug plans (MA-PDs) or stand-alone Prescription Drug Plans (PDPs). However, causal evidence of plans' impact remains limited because of nonrandom enrollment.
Objective:
To apply an AI-enabled causal inference framework to estimate the causal effect of PDP vs MA-PD enrollment on health care utilization and spending among Medicare beneficiaries with cancer.
Methods:
We conducted a nationally representative analysis among patients with cancer aged 65 years or older using Medicare Current Beneficiary Survey data linked to Medicare claims, supplemented with data from the Area Health Resources Files from 2019 to 2022. Outcomes included annual inpatient, outpatient, and prescription drug events, as well as total, Medicare, and out-of-pocket (OOP) expenditures (inflation-adjusted to 2025 USD). Guided by the National Institute on Aging Health Disparities Framework, 63 multidimensional covariates were incorporated. To address nonrandom plan selection, we implemented conventional regression, two-stage residual inclusion (2SRI) instrumental variables (IVs), and an AI-enabled Doubly Robust Machine Learning IV (DML-IV) approach. The IVs in this study included the county-level PDP penetration rate and the percentage of white-collar workers.
Results:
A total of 3,140 unweighted patients with cancer, corresponding to 22,207,248 weighted patients, were included, with 51.20% enrolled in PDP. For health care use, the naive model showed higher inpatient (incident rate ratio [IRR] = 1.30) and outpatient events (IRR = 1.86) among PDP enrollees; after 2SRI adjustment, only outpatient events remained significant (IRR = 1.56), and no utilization differences were significant in the DML-IV model. For health care costs, the naive model indicated higher total (cost ratio = 1.69), Medicare (cost ratio = 9.94), and OOP spending (cost ratio = 1.63). In the 2SRI model, total (cost ratio = 1.35), Medicare (cost ratio = 5.81), and OOP costs (cost ratio = 1.86) remained elevated. In the DML-IV model, total costs were no longer significant, whereas Medicare (cost ratio = 4.14) and OOP costs (cost ratio = 2.18) remained significantly higher.
Conclusions:
After rigorous AI-enabled causal adjustment, differences in health care utilization and costs between PDP and MA-PD plans largely reflect enrollment selection, whereas financial exposure, particularly beneficiary OOP spending, remains higher under stand-alone PDP coverage. These findings highlight how AI-based causal methods can support managed care and managed care pharmacy leaders in evaluating benefit integration and designing strategies to improve financial protection for high-cost populations.
