Background

The United States has long debated the effects of public health insurance programmes. The central empirical challenge is that Medicaid enrollment is not random: people who enroll differ systematically from those who do not in ways that are correlated with health outcomes. A naive comparison of insured and uninsured individuals conflates the effect of insurance with the selection of sicker or healthier people into coverage.

Prior to the OHIE, the literature relied on instrumental variables strategies — exploiting legislative changes in Medicaid eligibility thresholds — or regression discontinuity designs at age cutoffs like the 65-year Medicare eligibility threshold. These studies found suggestive evidence that insurance improves health and reduces financial stress, but each faced objections about the validity of the identifying variation.

The Natural Experiment

In early 2008, Oregon had sufficient funds to expand its Oregon Health Plan Standard (OHP Standard) programme — a Medicaid programme for low-income adults without dependent children — but not enough to cover all eligible individuals. The state decided to use a lottery to determine who would be allowed to apply for the programme. Approximately 90,000 individuals signed up for the lottery; 30,000 were selected.

This lottery created random variation in the probability of Medicaid coverage. Not all lottery winners ultimately enrolled — some were found ineligible or did not complete the application — so the lottery is best understood as an instrument for coverage rather than as direct random assignment to coverage. The lottery winners are the intent-to-treat group; the effect of actual coverage is estimated by instrumental variables using lottery selection as the instrument (Finkelstein et al.(2012)).

The exclusion restriction requires that the lottery affected health outcomes only through its effect on insurance coverage. This is plausible but not guaranteed: winning the lottery could affect outcomes through other channels, such as stress reduction from financial security, independent of coverage.

Data and Design

Finkelstein et al.(2012) linked lottery records to administrative data on emergency department visits, hospital admissions, and credit reports, and also conducted a mail survey and in-person interviews with a subsample of lottery participants. The study team collected measures of:

• Healthcare utilisation (doctor visits, prescriptions, hospital admissions, emergency department use)
• Financial outcomes (medical debt, collections, bankruptcy)
• Self-reported health status, mental health (depression screening), and happiness
• Physical health measures: blood pressure, cholesterol, glycated haemoglobin (HbA1c)

The clinical measurements were taken approximately two years after the lottery. The sample for clinical measurements was a probability subsample of approximately 12,000 individuals, of whom roughly 6,400 were lottery winners.

The LATE Framework

Let \(Z_i \in \{0,1\}\) denote lottery selection, \(D_i \in \{0,1\}\) denote Medicaid coverage, and \(Y_i\) denote an outcome. The ITT effect is: \[ \text{ITT} = \mathbb{E}[Y_i \mid Z_i = 1] - \mathbb{E}[Y_i \mid Z_i = 0] \] The IV (LATE) estimate of the effect of coverage is: \[ \text{LATE} = \frac{\text{ITT}}{\mathbb{E}[D_i \mid Z_i = 1] - \mathbb{E}[D_i \mid Z_i = 0]} \] The denominator is the "first stage" — the fraction of lottery winners who actually enrolled in Medicaid minus the fraction of non-winners who enrolled through other means (a small number). In the OHIE, the first stage was approximately 0.25, meaning that winning the lottery raised the probability of Medicaid coverage by about 25 percentage points (Finkelstein et al.(2012)). The LATE therefore scales up the ITT by a factor of roughly 4 to recover the effect of coverage on compliers.

Main Findings

Healthcare Utilisation

Medicaid coverage substantially increased healthcare utilisation. Lottery winners were significantly more likely to have a primary care visit, more likely to receive a prescription drug, and more likely to be hospitalised (the last result being positive from a welfare standpoint, suggesting previously unmet need). Emergency department use also increased, contrary to predictions from models in which ED and primary care are substitutes.

Financial Protection

One of the clearest results was on financial outcomes. Medicaid coverage dramatically reduced the probability of having medical bills sent to collections and reduced out-of-pocket medical expenditures (Finkelstein et al.(2012)). The reduction in catastrophic medical spending and the associated reduction in financial stress were substantial — results that have been replicated in subsequent studies of Medicaid expansion under the ACA (Sommers et al.(2017)).

Mental Health

Medicaid coverage led to a significant reduction in the probability of screening positive for depression. This finding, which is large and robust, is an important result: it demonstrates that public health insurance has substantial mental health benefits, even over a two-year horizon.

Physical Health: The Null Results

The most controversial finding was the absence of statistically significant effects on physical health measures. After two years, lottery winners did not have statistically significantly lower blood pressure, lower cholesterol, or lower HbA1c (a measure of blood sugar control relevant to diabetes management) compared to non-winners (Baicker et al.(2013)).

These null results were widely interpreted as evidence that Medicaid does not improve health. But this interpretation requires care. Several considerations counsel against reading the null results as evidence of zero effect:

1. Statistical power. The clinical subsample was designed to detect an effect of the size observed in prior observational studies. If the true effect is smaller than these studies suggested (which might reflect positive selection bias in observational work), the study would be underpowered.
2. Time horizon. Two years may be too short for changes in chronic disease management to manifest in biomarker levels, particularly for conditions like hypertension and diabetes where the physiological damage accumulates over decades.
3. Compliance and appropriate care. Not all Medicaid enrollees received appropriate care for their conditions. The study measures the average effect of coverage, not the effect of high-quality, protocol-consistent care.
4. The LATE qualifier. The effect is identified for compliers: those who enrolled because they won the lottery. Compliers may differ systematically from the full population of Medicaid-eligible individuals.

Subsequent Work and the ACA Expansion

The Oregon experiment prompted a large literature on the effects of Medicaid and the ACA Medicaid expansion. Sommers et al.(2017) review this literature and find consistent evidence of improvements in self-reported health, access to care, and financial protection following ACA Medicaid expansion. Several studies using difference-in-differences designs exploit the 2014 ACA expansion, which extended Medicaid to adults with incomes up to 138% of the federal poverty line in states that opted in.

The quasi-experimental evidence on ACA Medicaid expansion has found effects on mortality that are difficult to detect in a two-year experiment. Miller et al.(2019) use a regression discontinuity design at the 138% income threshold and find significant reductions in all-cause mortality among newly eligible adults. This result suggests that the OHIE null results on physical health may partly reflect the two-year time horizon.

Lessons for Causal Inference

The OHIE illustrates several general lessons for causal inference practice.

The value of randomisation. The lottery provided the cleanest possible identification strategy for the effects of Medicaid. It eliminated confounding by selection. The debate about the results — particularly the null physical health findings — took place on methodological ground much firmer than any observational study could provide.

The ITT/LATE distinction matters. The fact that the first stage was only 25 percentage points means that the LATE estimates are scaled up by a factor of four relative to the ITT. This scaling increases standard errors correspondingly. A null result in the LATE should not be interpreted without reference to the statistical power of the study.

Null results are informative, not dispositive. The absence of a statistically significant effect is not the same as evidence of absence. The appropriate interpretation of the OHIE null results is that the effects of Medicaid on biomarkers over two years are either zero or too small for this study to detect. These are different statements with different policy implications.

Conclusion

The Oregon Health Insurance Experiment is a landmark study not because it resolved the debate about Medicaid's effects, but because it reframed it. By providing randomised evidence, it raised the evidentiary standard for the debate. Its findings on financial protection and mental health are clear and large. Its null results on physical health are honest and important, even as their interpretation remains contested. The study remains one of the best examples of how a well-designed natural experiment can simultaneously illuminate and complicate a policy question.