Causal Inference in Health Economics

Core Idea

Health economics relies on causal claims — insurance reduces financial risk, cost-sharing reduces utilization, hospital competition lowers prices — but healthcare markets make causal inference exceptionally difficult because people select into insurance, treatments, and providers based on unobservable characteristics correlated with outcomes. The field has developed a distinctive toolkit of research designs to address this endogeneity. Randomized experiments (the RAND Health Insurance Experiment, the Oregon Medicaid lottery) provide the cleanest evidence but are rare, expensive, and ethically constrained. Quasi-experimental methods exploit natural experiments: difference-in-differences (comparing changes in outcomes before and after a policy change between affected and unaffected groups), regression discontinuity (exploiting eligibility cutoffs where assignment is as-if random), and instrumental variables (using an exogenous source of variation in the endogenous variable). Every major empirical finding in health economics — from the price elasticity of healthcare demand to the effects of insurance expansion — rests on the credibility of a specific causal identification strategy.

Explainer

The central problem of empirical health economics is that you cannot simply compare people with insurance to people without insurance and attribute any health or utilization difference to the effect of insurance. People who have insurance differ from people who do not in ways that independently affect health outcomes — they tend to be employed, higher income, more health-conscious, and less chronically ill. This selection bias contaminates naive observational comparisons, and it pervades every important question in the field: the effect of insurance on health, the effect of competition on hospital quality, the effect of pharmaceutical patents on innovation, the effect of physician supply on costs.

The RAND Health Insurance Experiment (1974-1982) addressed this problem definitively for one key question — the effect of cost-sharing on utilization — by randomly assigning 2,000 families to insurance plans with different coinsurance rates. Random assignment guaranteed that the groups were identical in expectation on all characteristics, observed and unobserved. The result — a price elasticity of demand for healthcare around -0.2, meaning a 10% increase in out-of-pocket price reduces utilization by about 2% — remains the benchmark estimate forty years later. But the RAND experiment cost over $300 million in current dollars and took a decade. Health economists cannot run randomized experiments for most policy questions.

Quasi-experimental methods exploit naturally occurring variation that mimics randomization. Difference-in-differences (DiD) compares the change in outcomes over time between a group affected by a policy and a group not affected. The Medicaid expansion studies exemplify this: states that expanded Medicaid under the ACA (treatment) vs. states that did not (control), comparing outcomes before and after 2014. The identifying assumption is parallel trends — absent the expansion, outcomes would have evolved similarly in both groups. Regression discontinuity (RD) exploits sharp eligibility cutoffs: Medicare eligibility at age 65 creates a discontinuity where 64-year-olds and 65-year-olds are nearly identical in all respects except insurance coverage, allowing credible estimation of the effect of Medicare on utilization, spending, and health outcomes. Instrumental variables (IV) use an exogenous source of variation in the treatment variable — for example, a state-level policy change that affected insurance coverage but plausibly had no direct effect on health.

Each design has its strengths and weaknesses. RCTs provide the highest internal validity but are expensive, often ethically infeasible, and measure effects only for the specific population and setting studied (limited external validity). DiD is flexible and widely applicable but relies on the untestable parallel trends assumption. RD provides highly credible local estimates but only at the cutoff — the effect of Medicare at age 65 may not generalize to the effect of insurance at age 40. IV estimates are only as good as the instrument's validity, and the exclusion restriction (the instrument affects the outcome only through the treatment) is never provable. The credibility of any empirical finding in health economics rests on the credibility of its identification strategy — and the field's methodological sophistication has advanced precisely because the stakes of getting causal claims wrong in health policy are so high.