Using Machine Learning to Estimate the Heterogeneous Impacts of Medicaid Managed Care

Almost all states currently use managed care to deliver health care services to some of their Medicaid population. There is mixed evidence regarding the extent to which managed care affects health expenditures, utilization, and health outcomes. Apart from the problem of non-random selection into managed care plans, the considerable variation across managed care plans regarding key plan features hinders their evaluation. Moreover, the literature has not systematically explored heterogeneity in the impacts of managed care on health care spending or health outcomes. Such heterogeneous impacts could have important implications, especially in low-income populations
where the role of insurance coverage on health disparities is not well understood.

This paper examines the effects of Medicaid Managed Care (MMC) plans in the State of New York using machine learning approaches to uncover treatment effect heterogeneities along observed attributes that do not need to be pre-specified. The fact that managed care plans differ in plan features (e.g., reimbursement incentives) logically motivates studying their heterogeneous impacts. Such heterogeneous responses to treatment can arise from differences in observed and unobserved characteristics. If the effects of managed care plans are different along observed dimensions such as race/ethnicity, gender, age, socioeconomic status, or other complex combinations of attributes, then such differences in response to coverage could have significant implications for the design of health insurance plans and policies targeting health disparities. Also, understanding the heterogeneous impacts of MMC can shed light on the extent to which treatment effects can be generalized to other target populations.

We address these issues using the universe of administrative claims data from New York City. The data contains a subset of Medicaid recipients who were auto-assigned into an MMC plan and a subsample of recipients who made an active plan choice. First, this unique feature of our data allows us to overcome selection issues by focusing on the auto-assigned group. Second, it also permits us to extrapolate our findings to the self-selected group of recipients and investigate the empirical relevance of performing a principled search for treatment effect heterogeneity. We first use recent generalizations of random forest algorithms to estimate the heterogeneous impacts of MMC plans on
health care spending, utilization, and health outcomes in the subsample of auto-assigned enrollees. Using the auto-assigned sample estimates as the gold standard, we then estimate treatment effects for the self-selected sample (conditional on covariates) with the aim of understanding the practical importance of selection and sorting behavior in plan choice.