Questions: Causal Inference in Health Economics

Question 1

The Oregon Health Insurance Experiment (2008) used a lottery to assign Medicaid coverage to low-income adults. Two years later, Medicaid coverage significantly increased emergency room usage, contrary to the prediction that insurance would shift care from ERs to primary care. This finding was controversial. Why is the lottery design so valuable despite producing an unexpected result?

Model answer: The lottery created true random assignment to Medicaid eligibility, eliminating selection bias — the fundamental problem that people who choose to enroll in Medicaid differ systematically from those who do not (sicker, lower income, more health-aware). Without randomization, observational comparisons between Medicaid enrollees and non-enrollees confound the effect of insurance with the characteristics of people who seek insurance. The lottery also had high internal validity: the comparison group (lottery losers) was identical in expectation to the treatment group (lottery winners) on all observed and unobserved characteristics. The unexpected ER finding is precisely why the design is valuable — it overturned a widely held but empirically untested assumption. Prior quasi-experimental studies could not credibly distinguish whether insurance reduced ER use (substitution to primary care) or increased it (newly insured people use more of everything, including ERs).

The Oregon experiment also found that Medicaid significantly reduced financial strain and depression, increased preventive care use, and increased diabetes detection — but did not produce statistically significant improvements in measured physical health outcomes (blood pressure, cholesterol, HbA1c) after two years. This null finding on physical health was heavily debated: the study may have been underpowered for health outcomes, two years may be too short to detect health improvements, or the effect of insurance on health may genuinely be small in the short run. The study illustrates both the power of randomization (credible causal estimates) and its limitations (finite sample size, limited follow-up, local average treatment effect for compliers only).

Question 2

A researcher wants to estimate the effect of hospital competition on quality. She uses a difference-in-differences design, comparing changes in mortality at hospitals that experienced a nearby hospital closure (treatment group) with hospitals that did not (control group), before and after the closure. The key identifying assumption of this design is:

• A. Hospital closures are randomly assigned
• B. The treatment and control hospitals had identical mortality levels before the closure
• C. In the absence of the closure, mortality trends at treatment hospitals would have been parallel to those at control hospitals (the parallel trends assumption)
• D. All patients at treatment hospitals are identical to patients at control hospitals

Answer: (C)

Difference-in-differences does not require equal levels — it allows treatment and control groups to differ permanently, as long as their time trends would have been the same absent the treatment. The parallel trends assumption states: whatever factors were driving mortality changes over time affected treatment and control hospitals equally. The method removes time-invariant differences between groups (the 'first difference' in a cross-sectional comparison) and common time trends (the 'second difference' in a before-after comparison), isolating the treatment effect. This assumption is not directly testable (we cannot observe the counterfactual), but researchers assess its plausibility by checking whether pre-treatment trends were parallel over multiple prior periods. If treatment hospitals already showed diverging trends before the closure, the design is not credible.

Question 3

Instrumental variables (IV) methods in health economics require an instrument that is correlated with the endogenous treatment variable but affects the outcome only through the treatment — the exclusion restriction. In the context of estimating the effect of insurance on health, geographic distance to a hospital has been proposed as an instrument for insurance coverage.

• T. True
• F. False

Answer: False

Distance to a hospital is a poor instrument for insurance coverage because it likely violates the exclusion restriction — distance affects health outcomes through channels other than insurance. People living far from hospitals may have less access to care regardless of insurance status, may differ in socioeconomic characteristics (rural vs. urban), and may face different environmental health risks. A valid instrument must affect the outcome ONLY through the endogenous variable. Better instruments for insurance include lottery-based assignment (Oregon experiment), age-based eligibility cutoffs (Medicare at 65, used in RD designs), and employer mandate thresholds (firms above vs. below a size cutoff, used in DiD designs). The exclusion restriction is the most important and most commonly violated assumption in applied IV work.