RESEARCH DOSSIER
Health Insurance Coverage Expansion and Causal Evaluation
Abstract. Expansion of health insurance coverage alters legal entitlement, household exposure to medical expenditure, demand for care, provider financing, administrative enrollment, and the distribution of health risks. Its causal consequences require identifying what would have happened to the same populations under an alternative eligibility, subsidy, enrollment, or benefit regime. This monograph examines the identification and interpretation of those consequences through randomized encouragement, coverage lotteries, cluster assignment, age thresholds, difference-in-differences, comparative reform studies, and linked administrative records. Its evidence base combines the RAND Health Insurance Experiment, Oregon's Medicaid lottery, Mexico's Seguro Popular experiment, United States Medicaid and Massachusetts reforms, Thailand's universal coverage expansion, Indonesia's dependent-coverage discontinuity, and studies of medical debt and insurance welfare. It integrates research-design analysis with institutional distinctions between eligibility, offer, take-up, continuous enrollment, effective entitlement, and realized care. Causal estimates are organized by intention-to-treat, complier, population-policy, and treatment-on-the-treated estimands; the study distinguishes statistically undetected clinical effects from demonstrated absence of clinically important effects. The documented literature supports substantial effects on utilization and financial protection in several settings, identifies health improvements in some populations and time horizons, and shows that coverage, access, financial security, and biological health cannot be substituted for one another as outcome measures. An original comparative synthesis establishes practical interpretive requirements for evaluating coverage reforms: specify the policy contrast and population, identify the treatment induced by the policy, examine provider-side and household-side mechanisms, align measurement and follow-up horizons, evaluate heterogeneity, and preserve causal uncertainty. This is a research synthesis of established theories and evidence, not a new authorial causal doctrine. Findings are bounded by the selected studies, heterogeneous programs, available source text, and the absence of new patient-level statistical estimation.
Keywords: Health insurance; coverage expansion; eligibility; enrollment; take-up; randomized controlled trial; randomized encouragement; Medicaid; Oregon Health Insurance Experiment; Seguro Popular; universal health coverage; causal inference; instrumental variables; local average treatment effect; intention-to-treat; compliance; difference-in-differences; regression discontinuity; staggered adoption; event study; parallel trends; external validity; financial protection; medical debt; utilization; mortality; distributional welfare; provider supply; insurance economics; insurance policy evaluation.
Introduction: Research Questions and Method
Health insurance expansion is an intervention into the financial and institutional conditions under which health care is obtained. A change in eligibility may increase the number of legally eligible people without enrolling them. A change in enrollment can produce usable benefits whose value depends on cost sharing, covered services, provider networks, payment systems, and clinical availability. Even a large observed increase in insured persons says little by itself about the causal effects on hospital admissions, medical debt, mortality, or population welfare. Those are separate empirical questions requiring separate outcomes and identification arguments.
The international policy context reinforces the distinction. The WHO–World Bank monitoring report (2025) estimates that 4.6 billion people lacked full essential-service coverage in 2023 and 2.1 billion experienced financial hardship related to out-of-pocket spending in 2022. The indicators measure different conditions and reference years. Similarly, the OECD population-coverage analysis (2025) explicitly defines population coverage as entitlement to a core set of services and states that this measure alone cannot guarantee access. Cross-country comparisons that treat nominal insurance enrollment as identical to universal health coverage confuse policy inputs with realized protection.
Three questions organize this dossier. First, which research designs identify the effects of insurance offers, eligibility, or actual enrollment on outcomes? Second, how do selection, noncompliance, substitution, institutional supply, and spillovers alter the causal interpretation? Third, under what conditions can evidence generated in one jurisdiction, calendar period, or beneficiary group inform other coverage reforms and distributional choices?
The method is documentary, source-critical, comparative, and analytical. Primary peer-reviewed empirical studies are evaluated through their reported design and findings; methodological articles supply identifying assumptions. Institutional reports establish measurement definitions rather than a randomized causal effect. Numerical results are attributed to the specific study, population, period, outcome, and estimand from which they originate. The analysis includes explicit algebraic examples to demonstrate logic; those examples are hypothetical and are not presented as empirical reestimates. Full-population administrative microdata were not downloaded or independently analyzed. For several publisher-restricted papers, verified publisher abstracts, institutional records, and freely available manuscript versions ground the narrower claims recorded here.
Foundations and Source-Critical History
The Object of Coverage Expansion
The unit of the policy intervention can be the individual, household, employer, insurance plan, county, hospital catchment area, province, or state. At each level the assignment mechanism differs. Individual coverage may be offered following a lottery or age threshold; a jurisdiction may expand an insurance program for all residents meeting a revised income rule. A national reform can simultaneously change provider reimbursement, subsidies, benefits, and administrative enrollment. The causal treatment in the first design is an offer to a person; in the last it is a multidimensional institutional change. Their effects answer different policy questions.
Five operational states are useful for measurement: eligibility under prevailing rules; notification or opportunity to apply; completed enrollment; continuous, usable insurance entitlement; and realized financial or clinical protection. These are observable aspects of a policy pathway, not a claim that every country uses the same legal categories. The Medicaid eligibility policy illustrates how mandatory groups and optional expansion groups coexist within one public program. Eligibility thresholds and actual enrollment must be verified at the state and year relevant to a study.
Coverage expansion itself has at least three analytically different margins. The extensive margin moves a person from lacking a recognized entitlement to possessing one. The continuity margin reduces the number or duration of uninsured spells. The intensive margin increases the breadth, generosity, or usability of protection for persons already insured. A study focused on the extensive margin should not automatically be interpreted as identifying all three margins. Changes in public insurance may also displace private contracts, alter uncompensated care, and change claims paid by third parties without a one-for-one change in medical resources consumed.
Entitlement, Enrollment, and Effective Access
Formal coverage can be recorded through survey responses, eligibility files, insurer enrollment, premium payment, or claims authorization. These sources answer different questions. Administrative files document plan membership within the rules of a program; household surveys can capture overlapping forms of insurance and periods without coverage, but are sensitive to recall and questionnaire construction. The AHRQ MEPS documentation for 2024 distinguishes being uninsured for an entire year from having any insurance at some point during the year. Mixing a point-in-time coverage variable with an annual-any-coverage variable creates a definitional change that can be mistaken for a policy effect.
Effective access depends on provider supply and the cost of reaching care. An insurance card may finance consultations only in a network whose providers are distant or overburdened. Deductibles, copayments, formularies, prior authorization, waiting periods, transportation, and waiting time condition realized service use. A reform may increase reported coverage while a binding supply constraint attenuates utilization; it may increase utilization while financial protection remains incomplete because covered treatment requires substantial household payment. Proper outcome construction must therefore specify covered services, actual access, and both medical and nonmedical costs.
Historical Identification Opportunities
The RAND Health Insurance Experiment in the 1970s and early 1980s randomized levels of insurance cost sharing rather than modern uninsurance-to-insurance eligibility. The RAND research brief explains that free care increased service use and that average health consequences were generally limited, with important benefits for particular low-income, sick participants. The intervention primarily identifies the effect of price exposure among insured participants under the assigned benefit structures. It is an essential historical experiment but not interchangeable with a lottery that first offers Medicaid eligibility to previously uninsured adults.
The Oregon Medicaid lottery of 2008 created random variation in the opportunity to apply for public insurance. A different intervention was evaluated in Mexico, where the Seguro Popular trial randomized encouragement and combined demand-side insurance with investments in supply and medication. Massachusetts's 2006 reforms and later United States Medicaid expansions created policy shocks observed through quasi-experimental comparisons. Thailand's 2001 health financing reform altered both financial access and hospital resources. These cases form a sequence of identification opportunities whose differences matter as much as their common objective of protection.
Causal Evidence and the Meaning of Comparison
A correlation between insurance and health does not establish that insurance improved health: prior illness affects insurance demand, employment affects eligibility, wealth affects plan selection and service use, and the prospect of care can change enrollment. Selection can run in opposing directions. Persons with high expected spending may seek insurance, while persons with stronger attachment to work and formal institutions may be more likely to retain employer coverage and also have better observed health. Statistical adjustment is credible only to the extent that relevant confounders are observed and handled correctly.
Randomization, credible thresholds, and well-supported policy comparisons can identify more precise counterfactuals, although they remain subject to design-specific restrictions. The methodological question is always whose outcome under which policy is compared to which counterfactual. The result should be recorded with the assignment unit, follow-up interval, eligibility cohort, adherence rates, uncertainty measure, and identifiable sources of interference. The bibliography is therefore used as an evidence map, with direct links attached to the claim they substantiate.
Randomized and Quasi-Experimental Designs
Potential Outcomes and Policy Estimands
Let Y_i(1) denote the outcome that unit i would have under a specified treatment and Y_i(0) its outcome under an explicitly specified counterfactual. The individual causal effect Y_i(1) - Y_i(0) cannot usually be observed directly because a unit follows only one realized treatment history. Research designs identify averages or other features of these unobserved contrasts under assumptions.
The average treatment effect (ATE) is E[Y(1)-Y(0)] for a defined target population. The average treatment effect on the treated (ATT) conditions the contrast on those who actually took up the treatment. The intention-to-treat effect (ITT) is E[Y|Z=1]-E[Y|Z=0], where Z is randomized assignment or an offer. Under a valid instrumental-variable design, the ratio of the assignment effect on outcomes to the assignment effect on receipt can identify a local effect for those whose receipt was changed by assignment. These are distinct estimands. Their numerical values need not coincide even when calculated using the same experiment.
A policy-maker deciding whether to introduce an offer to an entire eligible population may need ITT at the policy's anticipated offer and take-up rates; a clinical service planner may need the effect among newly insured individuals; a finance ministry may need incremental public cost per additional insured person and downstream savings. Causal interpretation begins with this choice. Treating an ITT estimate as the effect of actual coverage mechanically understates or overstates the latter whenever take-up differs from the offer rate and the populations differ.
Randomized Insurance Offers and Encouragement
A lottery balances observed and unobserved baseline characteristics in expectation across assignment groups. Consequently, the difference in outcomes by original assignment provides a design-based ITT estimate, subject to appropriate weighting, cluster structure, attrition, and preservation of the original randomization. Researchers must preserve the distinction between a lottery winner and an enrolled individual: selection after assignment is a source of bias if winners who enrolled are compared directly with losers who did not.
With partial take-up, randomized encouragement can be an instrument for actual enrollment. The identifying argument generally requires random assignment, a nonzero first stage, an exclusion restriction whereby assignment affects outcomes only through the specified enrollment treatment, monotonicity regarding treatment uptake, and appropriate treatment consistency. The classic Angrist–Imbens local-effect analysis provides the interpretation of instrumental-variable estimates for persons induced to change treatment by the instrument. It does not identify an unconditional population average without additional assumptions.
Insurance encouragement can violate the exclusion restriction if the invitation itself conveys health information, changes perceived security, or facilitates access independently of enrollment. Likewise, a provider investment assigned with insurance encouragement can affect outcomes among people who never enroll. A credible experiment can therefore identify the effect of the bundled offer but fail to isolate the effect of insurance membership alone. Such distinctions should appear in the main conclusion, not only a technical note.
Oregon: A Lottery with Partial Take-Up
The Oregon first-year experiment by Finkelstein and colleagues (2012) used a randomized lottery for access to Medicaid among uninsured low-income adults. Lottery selection raised insurance coverage by approximately 25 percentage points in the first year. In that period the assignment created significant increases in health service use and decreases in out-of-pocket spending and debt; the authors also recorded improvements in self-reported health. The magnitude of this coverage first stage matters because lottery assignment was not identical to program receipt.
The two-year clinical follow-up by Baicker and colleagues (2013) found increases in service use, diabetes diagnosis and medication use, and a 9.15-percentage-point reduction in positive depression screening, with a reported 95% confidence interval from -16.70 to -1.60 percentage points. The study did not find statistically significant effects on measured hypertension, high cholesterol, or average glycated hemoglobin. It reported that catastrophic out-of-pocket expenditures, under its study-specific definition of spending above 30% of income, were nearly eliminated. A lack of statistical significance for a physical outcome does not demonstrate exact zero benefit; the confidence interval and statistical power bound the inference.
One additional finding complicates the widespread assumption that primary care access must immediately displace emergency use. The Oregon emergency-department study by Taubman and colleagues (2014) found that Medicaid coverage increased emergency visits by 0.41 per person over approximately 18 months, about 40% relative to the control mean of 1.02. Expanded coverage can initially raise both primary-care and emergency utilization when price constraints fall and previously unmet needs become expressible demand. These results concern a specific low-income lottery population and observation window, not a universal permanent emergency-use response.
Mexico: Clusters, Encouragement, and Supply Investment
The randomized Seguro Popular evaluation by King and colleagues (2009) assigned encouragement across 74 matched pairs of health facility catchment areas in seven Mexican states. The intervention included insurance encouragement and medical facility or medicine support. Follow-up after approximately ten months in 50 pairs showed a 1.9-percentage-point intention-to-treat reduction in catastrophic expenditure from baseline, reported as a 23% reduction, with a 95% confidence interval of 0.14 to 3.66 percentage points for the magnitude of reduction. The original study found no detected effects on medication spending, health outcomes, or service utilization during this short horizon.
The evaluation is especially important because it is randomized yet bundled. An isolated enrollment effect would require assumptions or variation that separate insurance take-up from associated provider improvements. The estimated complier effect reported in the study refers to compliance with its encouragement design. Researchers should specify its cluster-level assignment, group-level exposure, follow-up retention, outcome definitions, and distinguish the trial's measured result from national claims about long-term universal health coverage.
The Oregon and Mexico trials converge on financial protection while differing in intervention components, populations, benefit architectures, and short-term service-use findings. This difference is substantive evidence about mechanisms and external validity. It should not be erased by averaging their outcomes as if they represented one standardized insurance treatment.
Threshold, Reform, and Comparative Designs
A discontinuity in eligibility around an age or income threshold can reveal local effects if units cannot precisely manipulate the running variable and potential outcomes would otherwise evolve smoothly at the cutoff. In a fuzzy design, a discontinuity in insurance take-up accompanies rather than equals the legal threshold; the ratio of outcome and first-stage discontinuities can identify a local treatment effect under suitable assumptions. The Hahn–Todd–van der Klaauw analysis (2001) formalizes identification for regression-discontinuity designs. Bandwidth, polynomial order, sorting and placebo cutoffs must be examined rather than chosen only for statistical significance.
A health-insurance-specific illustration comes from Indonesia's dependent-coverage cutoff studied by Rizal (2025). The study reports a 14.6- to 20.9-percentage-point drop in coverage at age 21 and a 5.3- to 8.4-percentage-point fall in recent outpatient visits among sick young adults, without a detected change in measured health status. The estimated discontinuity is local to the age rule, the sampled people, and the observed period. It identifies the consequences of aging out of dependent coverage rather than the comprehensive consequences of a new universal national insurance system.
In a difference-in-differences design, policy-affected units are compared over time with units whose relevant outcome trend would have been parallel in the counterfactual absence of reform. The Massachusetts reform mortality study by Sommers, Long, and Baicker (2014) compared changes in county-level mortality among adults 20 to 64 in Massachusetts with matched out-of-state control counties. It found an association consistent with a 2.9% relative reduction in all-cause mortality after reform, but the nonrandomized comparison remains vulnerable to unmeasured differential trends and co-occurring changes. Its identification conditions and transportability differ from those of a randomized lottery.
Staggered Adoption and Modern Difference-in-Differences
Insurance expansions are frequently adopted by different jurisdictions in different calendar years. Simple two-way fixed-effects regression can mix comparisons of newly treated states with already-treated states and can attach undesirable weights when treatment effects vary over time or by cohort. The Goodman-Bacon decomposition (2021) explains the structure of these comparisons; it does not make every two-way fixed-effects estimate invalid, but reveals the restrictions required for its causal interpretation.
The Callaway–Sant'Anna estimators (2021) identify cohort-time treatment effects under explicit assumptions and offer transparent aggregation choices. The Sun–Abraham event-study analysis (2021) demonstrates that standard event-study leads and lags can be contaminated by treatment effects from other periods when adoption is staggered and effects are heterogeneous. This is directly relevant to Medicaid expansion comparisons, where pre-trend evidence, anticipation, choice of not-yet-treated or never-treated controls, and cohort composition materially affect the inferred reform effect.
An event-study plot supplies diagnostic information, not automatic proof of identification. A pre-policy coefficient statistically indistinguishable from zero can reflect low power, changing composition, or imperfect measurement. Researchers should state the pretrend equivalence range they can reasonably exclude, test reasonable alternative control groups, and report cohort-specific effects before summarizing an overall estimate. Across insurance systems with strong intergovernmental financial transfers, an untreated neighbor may also be indirectly affected, challenging the stable-control interpretation.
Selection and Compliance Effects
Selection into Insurance and Care
Enrollment decisions respond to expected illness, risk aversion, prices, application complexity, awareness, trust, employment and household circumstances. Higher-risk individuals may enroll sooner after a reform because expected benefits are high. Conversely, an eligible but administratively excluded individual can remain uninsured despite high need. A post-reform insured-versus-uninsured comparison therefore blends the policy effect with the selection processes that determine who appears in each group. Adjusted regression is useful for describing observed differences, but a causal conclusion additionally requires a justified identifying condition.
Conditioning on downstream variables can itself introduce bias. For example, comparing only those who used a physician after reform restricts the sample based on an outcome that coverage may change. Comparing health conditional on survival may likewise alter the composition of survivors if coverage affects mortality. Insurance studies should locate covariates in time relative to assignment, test for pretreatment imbalance, and avoid adjusting for mediators when estimating the total policy effect. When investigating mediation, the additional assumptions must be stated.
Instrumental Variables and Complier Effects
Let Z denote an insurance offer and D actual insurance receipt. In a simplified binary setting, ITT_Y=E[Y|Z=1]-E[Y|Z=0] and ITT_D=E[D|Z=1]-E[D|Z=0]. Under the instrumental-variable assumptions already described, LATE=ITT_Y/ITT_D. This ratio is an effect for people whose insurance status was changed by the offer. Always-takers and never-takers are excluded from this local margin. The formula is an established identification result; it is not an original theoretical proposal of this monograph.
The instrument's first stage must be estimated with uncertainty. If an offer increases enrollment only slightly, the ratio can become unstable. When coverage is multidimensional, a binary insured flag may conceal switches among public, employer-sponsored, and direct-purchase plans. An offer could cause private-to-public substitution in addition to expansion of total coverage; the resulting treatment effect then combines insurance type and generosity. A study claiming to estimate the effect of becoming insured should show that the instrument actually shifts the uninsured-to-insured margin it names.
The Oregon first-year first stage was approximately 25 percentage points (Finkelstein et al., 2012). This figure is sometimes mistakenly treated as a 25% enrollment rate or as the direct effect of Medicaid on health. It is a difference in coverage probabilities by randomized lottery status under that study's definitions. If an illustrative policy generated a 10-percentage-point difference in the offer-induced probability of enrollment and a 2-percentage-point ITT change in an outcome, the simple IV ratio would equal 0.20 outcome units per induced enrollment; interpreting that number requires the outcome's scale and all identification assumptions. This numerical illustration is hypothetical.
Churn, Substitution, and Spillovers
Administrative churn consists of coverage loss and reentry through renewal, changes in household income, employment transitions, identity verification, residency, or documentation. Two individuals insured at survey date may have experienced very different lengths of protection. A policy might raise annual coverage but leave episodic gaps concentrated among people with unstable employment. The appropriate exposure can be the number of covered months, episode length, or policy receipt at a clinically relevant time, not merely a binary indicator measured at survey interview.
Substitution changes the net effect on insurance counts. Medicaid coverage may crowd out employer or individual plans while lowering households' premium obligations; a subsidized private plan may replace one source of public entitlement. A program can raise government spending more than the increase in the number of insured people would suggest because it changes payer identity. Conversely, it can improve financial protection for already insured individuals by reducing cost sharing. Assessment requires a mutually interpretable insurance hierarchy and, where possible, separate estimates for public coverage, private coverage, any insurance, and continuous coverage.
Spillovers occur when treatment changes provider staffing, appointment queues, uncompensated-care financing, family members' enrollment, household borrowing, or local insurance premiums. The Oregon project account reports related spillovers to children's Medicaid enrollment following household lottery selection. If control households live in markets affected by treated households, their outcomes may also change, making a simple treated-minus-control contrast represent a particular equilibrium experiment rather than a universally separable individual effect.
Provider Capacity, Price, and Insurer Incentives
Insurance expansion reduces the point-of-service price faced by the newly covered beneficiary, but the realized quantity and quality of care depend on the supply curve and institutional purchasing arrangements. If clinicians accept fixed capacity, increased demand can raise waiting time, induce provider substitution, or redirect resources from other patients. Payment methods can influence volume, diagnostic coding, and referral patterns. A large insurance reform may finance formerly uncompensated services, relieve hospital balance sheets, and thereby expand capacity over time.
The Thailand reform evidence by Gruber, Hendren, and Townsend (2014) emphasizes that its "30 Baht" reform simultaneously increased funds available to hospitals serving poorer populations and lowered copayments. The authors find increased utilization and evidence of reduced infant mortality associated with the resulting access expansion. It would be unwarranted to attribute the entire effect exclusively to a change in the legal insurance card when provider resources were also altered. Policy evaluation should document payment reform and supply investments as components of treatment.
Coverage shifts also alter insurer risk pools. An expansion attracting previously uninsured adults with high expected utilization can change short-run claims costs; enrollment of younger or lower-risk individuals can shift the opposite way. Subsidies, risk adjustment, benefit requirements and inter-insurer transfers shape the incidence of these changes. A causal analysis of population welfare that omits insurer finances and provider reimbursement may accurately estimate short-run health effects while missing the fiscal feasibility and cross-subsidy channels of the policy.
Welfare, Public Finance, and Economic Incidence
Insurance's welfare value arises partly from transferring uncertain out-of-pocket expenditure to a pooled arrangement, not only from changing the amount of health services consumed. A household may gain security even if it has no medical encounter during follow-up. Measuring average spending alone can therefore miss risk protection; measuring lower private spending alone can miss public fiscal cost. Evaluation should distinguish insurer or government payments, providers' revenue, households' cash flow, and real resources devoted to medical care.
The Oregon Medicaid valuation by Finkelstein, Hendren, and Luttmer (2019) explicitly translates experiment evidence into welfare-analysis frameworks. Its estimates vary with assumptions about risk protection, valuations, and incidence; the authors emphasize that significant Medicaid expenditure can represent transfers that finance previously uncompensated care. These findings do not provide a single universally transferable fiscal benefit-cost coefficient. They show why the marginal resource cost of coverage, the government's budgetary outlay, and beneficiaries' value should be tracked as separate variables.
Coverage and labor markets interact as well. The Tennessee disenrollment study by Garthwaite, Gross, and Notowidigdo (2014) exploited the loss of public coverage for a large group and found increased labor supply, especially employment associated with private insurance. This is evidence about a reverse-direction coverage shock and employment lock, not a direct estimate of the health benefits of new coverage. It demonstrates that the policy counterfactual must include labor market behavior and alternative insurance acquisition.
External Validity and Distributional Outcomes
Financial Risk Protection
The most immediately observable economic effect of insurance is frequently a reduction in the probability of extreme household medical spending. Financial risk protection must nevertheless be defined before measurement. Studies use catastrophic spending thresholds relative to income, consumption or capacity to pay; medical debt sent to collection; unpaid bills; missed nonmedical payments; and spending that pushes a household into or deeper into poverty. Each outcome isolates a different aspect of exposure. A reported reduction in medical debt collections may reflect a change in billing or debt reporting as well as an economic reduction in unpaid liabilities.
Oregon's two-year experiment nearly eliminated catastrophic out-of-pocket expenditures under its own 30%-of-income definition (Baicker et al., 2013). Mexico's trial found a smaller but statistically discernible reduction in its defined catastrophic expenditure rate within ten months (King et al., 2009). These two results agree on the broad financial protection channel while differing in measures, settings, program components and baseline risk. Their numerical percentages cannot be directly pooled without harmonizing thresholds and sampling frames.
The US medical-debt study by Kluender, Mahoney, Wong, and Yin (2021) used credit-report information for a retrospective description of collections between 2009 and 2020. It reported that 17.8% of individuals had medical debt in collections in June 2020 and documented substantial geographic and socioeconomic concentration. This establishes the scale and distribution of a financial-burden measure; the descriptive national estimate alone is not a randomized causal estimate of Medicaid expansion. Where policy comparisons are used to discuss debt effects, their identification assumptions must be assessed separately.
Service Use and Health Outcomes
Insurance can affect the quantity and composition of care by changing the patient's price, expected burden, and feasible provider choices. Researchers should distinguish preventive visits, primary care, hospital admissions, emergency care, prescriptions, medicines taken as prescribed, screening, disease detection, and continuity of chronic-condition management. More use is not by itself a welfare outcome: additional care can be beneficial, low value, or substitutive depending on need and clinical quality. Conversely, a period with no increase in care can reflect constrained provider supply or improvements in other service margins.
Oregon produced substantial increases in care use and financial protection, while its physical-health measures at two years did not show statistically significant improvements (Finkelstein et al., 2012; Baicker et al., 2013). Its emergency-department analysis demonstrates increased ED use, contrary to a simple displacement hypothesis (Taubman et al., 2014). Mexico's randomized trial detected financial protection without equivalent detected changes in its short-run utilization and health indicators (King et al., 2009). Neither finding permits the generalization that insurance does not affect health. They show that health outcomes may require different intervention components, higher statistical power, and a longer biologically relevant observation window.
A quasi-experimental three-state US coverage study compared low-income adults in Arkansas and Kentucky with Texas after the 2014 coverage expansions. Sommers and colleagues (2016) reported by 2015 a 22.7-percentage-point relative reduction in uninsurance, increased access to primary care, fewer cost-related delayed or skipped services, and improvements in other reported measures. Its survey-based comparative design supports a policy-specific causal interpretation conditional on the comparability and parallel-trends assumptions. It does not inherit the random assignment strength of Oregon.
Mortality and Longer-Term Effects
Mortality is an important endpoint but is difficult to study for relatively short interventions affecting people whose baseline death risk is low. A trial powered to detect changes in preventive visits may have little power to identify small mortality changes. Studies using large administrative data and longer time horizons can be more sensitive to mortality effects, yet causal credibility then depends on the counterfactual and population definitions.
The Massachusetts analysis (Sommers, Long, and Baicker, 2014) reported a 2.9% relative reduction in all-cause mortality after the 2006 reform, equivalent in the study to 8.2 fewer deaths per 100,000 adults compared with a matched control group. The authors explicitly recognized vulnerability to nonrandomized confounding. Their estimate concerns the full package of reform-associated changes, not an experimentally isolated insurance card.
The linked mortality-data analysis by Miller, Johnson, and Wherry (2021, published journal version) compared near-elderly adults likely to benefit from Medicaid expansions in expansion and nonexpansion states, linking survey information to administrative death records. The authors estimated a 0.132-percentage-point annual mortality decrease, approximately 9.4% relative to their sample mean. Their design includes population identification and parallel-trend arguments, with sensitivity analysis, but remains a nonrandomized state-policy comparison. The result is relevant to the potential longer-run and high-risk population consequences of coverage expansion, while its exact numerical magnitude should not be transplanted to other countries.
The Medicare age-65 study by Card, Dobkin, and Maestas (2009) used age-related insurance change among severely ill emergency admissions and estimated a near-one-percentage-point reduction in seven-day mortality around age 65 for its selected hospital population. Admission-based sample selection and the local nature of age eligibility matter for interpretation. It offers evidence about a different margin from low-income adults newly offered Medicaid and should be retained as a separate result rather than pooled into an average "effect of insurance on survival."
International Comparisons
In Mexico, randomized clusters and supply investments jointly shaped the observed financial-protection gain; in Thailand, financing to hospitals serving poorer populations formed a central mechanism; in Indonesia, a sharp dependent-coverage cutoff revealed loss of outpatient use among the ill; and in the United States, distinct Medicare, Medicaid, state and private-purchase institutions generate multiple treatment margins. These systems differ in revenue bases, enrollment procedures, benefit generosity, provider capacity, and baseline service use.
The Ghana NHIS evaluation by Mensah, Oppong, and Schmidt (2010) used propensity score matching and found associations between scheme membership and maternal-care outcomes. Matching observed characteristics can improve comparability but does not remove bias from unobserved differences in health-seeking behavior or program access. The study is useful as evidence about implementation and distribution, with weaker causal identification than random assignment. These differences in evidence strength should be visible in any international synthesis.
The WHO–World Bank UHC report and OECD coverage framework supply essential international context; they are not substitutes for a causal counterfactual. The WHO's service-coverage and financial-hardship time series trace progress and disparities across countries. The OECD reports extensive eligibility for a core benefit set in many high-income countries but identifies remaining access and cost-sharing differences. The proper analytical comparison is between compatible policy interventions and outcomes with transparent institutional descriptions, rather than a ranking that assumes countries with similar insurance enrollment have comparable protection.
Distributional and Equity Analysis
Aggregate effects can conceal radically different experiences. A reduction in mean household expenditure can be concentrated among the previously uninsured with high illness risk; a premium subsidy can transfer resources to eligible low-risk households while a provider shortage leaves rural populations underserved. Distributional assessment should predefine subgroups by baseline income, health risk, age, disability, location, employment, citizenship eligibility where lawful and ethically appropriate, and prior insurance status.
Heterogeneity needs more than a series of unadjusted subgroup regressions. The study should report subgroup-specific first stages, denominators, outcome variability, interaction tests, and intervals. An apparent larger effect among poorer populations can arise from higher baseline uninsured rates even when the causal effect among newly insured compliers is similar. Conversely, the same increase in coverage points can produce different clinical outcomes because the benefit package and provider accessibility vary.
Thailand's reform research places infant mortality and poorer provinces at the center of the empirical contrast (Gruber, Hendren, and Townsend, 2014). The Indonesian threshold study reports larger utilization effects among people who are poorer, less educated, or living in regions with higher care costs (Rizal, 2025). The Oregon trials record financial and mental-health outcomes alongside clinical measures. Together these illustrate why a distributional conclusion must state whether heterogeneity attaches to eligibility, enrollment effects, baseline vulnerability, clinical treatment, or service availability.
Transportability and Treatment-Effect Heterogeneity
Transporting a causal estimate from one policy to another requires more than the source study's internal validity. Target and source populations may differ in baseline disease prevalence, insurance alternatives, provider markets, benefit packages, take-up, unit prices, and administrative complexity. A local average treatment effect for Oregon lottery compliers answers the question of what coverage did for people induced to enroll by that lottery; it does not mechanically equal the effect on all eligible adults in a national mandatory insurance program.
A disciplined extrapolation identifies potential effect modifiers before transport: expected untreated medical use, pre-reform uninsured spells, supply elasticity, eligibility enforcement, health system financing, and duration of follow-up. When underlying individual data and joint covariate distributions are unavailable, qualitative scenario analysis is more defensible than applying a numerical effect as a universal parameter. Heterogeneity is itself a scientific result when it is linked to mechanism rather than treated as inconsistency to be erased.
Data, Measurement, and Validation
Exposure and Denominator Construction
A study must specify whether coverage is measured at a date, during a calendar year, for a continuous spell, or as person-months. It must state whose coverage is included, whether overlapping public and private insurance categories are mutually exclusive, and whether the relevant denominator is all residents, eligible adults, applicants, lottery registrants, surveyed households, claims-eligible members, or patients with a specified diagnosis. Different denominators can produce different percentage-point changes from the same administrative reform.
MEPS 2024 documentation distinguishes full-year uninsured from persons insured during any part of the year. This is an example of how a seemingly minor coding decision changes an outcome definition. A reproducible analysis also records the exact data release, questionnaire and administrative rule version, missing-data processing, weights, survey design, and changes in coding across periods. Annual-person denominators are inadequate for some continuity claims; person-month panels can describe gaps more directly.
Study Data and Record Linkage
Randomized trial records may include randomization lists, offer correspondence, enrollment data, survey responses, clinical examinations, hospital discharges, pharmacy claims, credit reports and mortality records. Linkage improves outcome measurement but can induce missingness if successful matches differ by treatment group or subgroup. The Oregon project combined multiple forms of information across its first-year and later clinical analyses. Each dataset has its own reference period and possible attrition process.
Administrative claims measure billed or reimbursed services conditional on engagement with covered care. They can miss out-of-network payments, untreated illness, nonreimbursed services and shifts between payers. Credit records capture collections reported to particular systems, not all medical liabilities. Mortality files offer a harder endpoint but still require identification, linkage and appropriate population exposure. For cross-country studies, inconsistent registration systems, household definitions and provider payment conventions limit direct pooling.
Internal Validity Diagnostics
For randomized assignments, check treatment allocation, balance in pretreatment covariates, cluster assignment, lottery family-level correlation, survey attrition, protocol deviations, differential outcome ascertainment, and interference. Balancing tests do not substitute for proper randomization, and selectively choosing balanced covariates after seeing treatment outcomes risks model-dependent inference. Report a primary outcome family and distinguish prespecified secondary outcomes from exploratory findings.
For regression discontinuity, inspect the distribution of the running variable around the cutoff, administrative manipulation, covariate continuity, bandwidth sensitivity, polynomial behavior, measurement heaping and other rules that change at the same threshold. The age of 21 in the Indonesian example is an eligibility condition; local age effects could also coincide with schooling or employment transitions. The research design must argue why contemporaneous discontinuities do not account for the outcome jump.
For policy comparisons, examine pretreatment trajectories over an adequately long period, treatment anticipation, co-reforms, comparable control populations, migration, coding changes, and geographic spillovers. State-specific reform dates should correspond to actual implementation and enrollment rather than announcement alone. Matching on observed county characteristics, as in the Massachusetts study, does not guarantee absence of unobserved time-varying confounding.
Sensitivity and Robustness Analysis
Robustness questions are substantive, not cosmetic. Does an inference survive another eligible cohort definition? Another outcome horizon? The exclusion of an anomalous year? A different control group? Different treatment timing? A shift from any coverage to continuous coverage? A placebo reform date? Inference clustered at the policy adoption level rather than at the individual observation level? Each answer can change a policy conclusion.
Modern staggered adoption methods are especially relevant where reforms are phased. A transparent analysis should estimate cohort-time effects, compare alternative control-group definitions, and show dynamic patterns. For a policy with relatively few treated clusters, conventional individual-level standard errors can be misleading because treatment is shared at a much higher level. Confidence intervals should incorporate the actual assignment and sampling structure. Sensitivity checks that do not materially challenge the identifying assumption should not be advertised as proof of causality.
Causal Interpretation, Multiplicity, and Reproducibility
Null-hypothesis tests answer whether data are sufficiently incompatible with a specified zero effect under a model; they do not answer whether clinically important benefits are absent. When a study reports nonsignificant average glycated hemoglobin changes, researchers should inspect intervals, baseline risks, follow-up, adherence, outcome misclassification, and subgroup precision before concluding that insurance has no metabolic benefit. Symmetrically, a single significant outcome among many measures may need adjustment or cautious interpretation in the context of multiplicity.
A reproducible evaluation records a prospectively defined estimand, analytic population, intervention history, comparator, outcome family, observation horizon, model, uncertainty method, and data provenance. Pre-registration, code and de-identified datasets can facilitate reanalysis when privacy law and consent permit. Sensitive insurance and medical information requires careful governance; public release of raw microdata is not always possible or desirable. Reproducibility can instead involve secure data access, documented variable definitions, deterministic code and an independent audit trail.
Evidence Matrix and Claim-Specific Confidence
| Intervention and design | Principal comparison | Supported finding | Central interpretive limitation |
|---|---|---|---|
| RAND cost-sharing randomization | Free versus cost-sharing plans | Higher use under free care; selected health benefits | Benefit price, not universal coverage offer |
| Oregon lottery, first year | Randomized offer, partial take-up | More care and less financial strain | Offer ITT differs from enrollment effect |
| Oregon clinical follow-up | Lottery-based coverage instrument | Less financial strain and depression; no significant measured physical effects at two years | Power and follow-up horizon |
| Mexico Seguro Popular | Cluster encouragement with supply support | Lower catastrophic spending in short follow-up | Bundled intervention and attrition |
| Massachusetts reform | Matched comparative trend | Mortality decline associated with reform | Nonrandomized, co-occurring trends |
| US ACA Medicaid expansions | State-policy comparison | Coverage and access gains; mortality findings in near-elderly cohorts | Staggered policies, assumptions and population selection |
| Thailand 30 Baht | Reform variation with provider funding | Greater use and infant-mortality improvement | Supply and demand change jointly |
| Indonesia dependent coverage | Age-21 discontinuity | Coverage and outpatient-use declines at cutoff | Local age-specific effect |
| Ghana NHIS | Propensity-score matching | Favorable maternal-care associations | Unmeasured selection remains possible |
The matrix is a structured interpretation of cited studies, not a numerical meta-analysis. Its entries differ in treatment, design, population and outcome, so an aggregate standardized effect would require data harmonization and stronger comparability assumptions. The internal confidence hierarchy is strongest for correctly executed random assignment on the offered intervention, conditional for valid thresholds and well-supported quasi-experiments, and more restrictive for observational matching. No design label alone supersedes its actual implementation.
International Synthesis and Scientific Conclusions
What the Evidence Establishes
First, insurance coverage is an institutional means of shifting financial exposure and access conditions. Randomized evidence from Oregon and Mexico demonstrates that expansion or encouragement can improve financial protection in the short term, although specific packages and measures differ. Oregon also demonstrates that increased coverage can increase the use of several types of care, including emergency services. Consequently, emergency-department growth cannot be used by itself as an indicator that the policy failed to improve health or reduce financial risk.
Second, clinical outcomes must be estimated by endpoint, population and time. Oregon's two-year clinical outcomes are not statistically identical to mortality findings from longer or larger observational policy comparisons. Massachusetts and linked Medicaid-mortality research indicate potential survival gains in their specified reform contexts, subject to the nonrandomized assumptions. A review that selectively privileges either the trial's null physical measures or quasi-experimental survival estimates would give an incomplete account of the evidence.
Third, program design and supply capacity influence effects. Mexico's combined insurance encouragement and supply intervention, Thailand's provider financing, and Indonesia's dependent-coverage discontinuity show different causal pathways. A universal proposition about what an insurance card alone does cannot be inferred from interventions that finance hospitals, reform benefit eligibility, alter labor supply, and change local health service capacity at the same time.
Fourth, enrollment is both a response to policy and a mechanism shaping interpretation. Lottery assignment, eligibility and recorded membership should not be conflated. Complier effects answer a narrower question than overall policy offers; instrumental-variable estimates rely on exclusion and monotonicity conditions. Government spending per enrolled member, social resource cost, and household willingness to pay are likewise distinct measures.
Fifth, international generalization is an empirical task. Age-based exits from dependent coverage, voluntary enrollment lotteries, means-tested eligibility, and nationwide compulsory coverage differ in baseline alternatives and implementation. A credible cross-country evaluation should first preserve the full policy architecture, then identify compatible estimands and target populations, and finally report differences that could change the effect. Where these conditions cannot be met, the appropriate output is a comparative explanation and a research agenda, not a falsely precise pooled number.
Evaluation Implications for Insurance Institutions
An insurer, supervisory authority, health ministry, or public financing institution evaluating expansion should state the intended mechanism and select outcomes accordingly. An eligibility reform needs measurements of eligibility, application, take-up, coverage continuity and access. A financial protection program needs household out-of-pocket risk, medical debt, provider payments, and fiscal incidence. A clinical access program needs valid measures of use, appropriateness, clinical process, outcomes and harm. These should be tracked with pre-specified cohorts and observation horizons.
Causal designs should be chosen by the policy's actual assignment rule. Lotteries and controlled phased implementation can identify offers directly when ethics and administrative conditions permit. Thresholds support local designs if exclusion conditions are defensible. State reforms can support policy comparisons if the counterfactual trend can be substantiated. Institutional monitoring without a credible counterfactual can still provide important descriptive accountability, but it should be labeled as monitoring rather than causal attribution.
A useful result report should contain the treatment definition, unit of assignment and analysis, primary and secondary estimands, population exclusions, first stage, outcome scale, absolute and relative effects where interpretable, uncertainty intervals, fiscal implications, data vintage and limitations. This set of reporting requirements follows from the scientific analyses reviewed here. It organizes good evaluation practice without instituting an original theoretical system.
Open Research Problems
Better evidence is needed on long-term insurance continuity, treatment interactions with provider supply, substitution between public and private plans, disability-related needs, and equity across underserved regions. Further questions concern how provider network constraints mediate changes in insurance take-up; how hospital financing alters quality and queues; when reduced financial stress yields measurable health improvements; and whether short-term increases in medical care have long-term benefits offsetting their cost.
Research also needs more standardized definitions of catastrophic household expenditure, medically induced debt, effective coverage and cross-system continuity. The WHO and OECD use explicit frameworks that assist monitoring, while the trials reviewed here show why causal identification still needs intervention-specific data. Distributed policy adoption calls for more robust estimators and transparent event-study diagnostics. Privacy-preserving linkage between enrollment, health, debt, mortality, and provider capacity could improve the evidence base if governance and representativeness are maintained.
Conclusion
Health insurance coverage expansion is a causal intervention into rights, risk sharing, administrative access, patient prices, provider finance and household security. Its scientifically relevant effects are multiple and must be estimated as distinct outcomes of a specified intervention and comparison. Randomized evidence gives strong support for financial protection and demonstrates important utilization consequences in particular settings. Quasi-experimental research finds additional access and mortality effects in defined populations, with causal strength conditional on credible policy counterfactuals. International studies establish that program components, supply constraints, age rules, baseline vulnerability, and time horizons shape what an estimate means.
The central interpretive conclusion is that coverage status, causal coverage effect, realized care, financial protection and health outcomes represent related but nonidentical empirical objects. The valid unit of synthesis is a matched claim with its intervention, population, estimand, identification assumption, measurement horizon and uncertainty. Properly separated, existing experiments and natural experiments support a substantial science of coverage expansion without forcing an artificial universal effect or a new proprietary doctrine. The work constitutes an original source-grounded analytical synthesis. It does not claim a novel general causal theory, conduct primary human-subject research, or substitute for new jurisdiction-specific impact evaluation.
Technical Appendix: Estimands and Worked Examples
Experimental offer and receipt
Suppose a randomized eligibility invitation changes recorded insurance receipt from 35% in the control group to 55% in the assigned group. The first stage is 20 percentage points. If a predefined binary outcome is observed in 18% of controls and 14% of those assigned the invitation, the ITT on the outcome is -4 percentage points. Under exclusion, monotonicity, randomization, no interference and appropriate consistency, the illustrative Wald ratio is -0.04/0.20=-0.20, corresponding to a 20-percentage-point change in this binary outcome among compliers. These invented numbers demonstrate the transformation from assignment effect to local receipt effect; they are not results from any cited experiment.
The policy ITT is -4 percentage points for an offer with that first stage and population. The local complier effect is -20 percentage points only under the stated assumptions. The two should not be described as contradictory or averaged together. If the invitation also changes service availability for non-enrollees, the exclusion restriction fails, and the ratio loses that simple insurance-only interpretation.
Two-period difference-in-differences
Suppose an affected region's uninsured proportion falls from 30% to 18%, while a comparison region falls from 25% to 21% over the same years. The treated change is -12 percentage points; the comparator change is -4 points. The illustrative difference-in-differences is (-12)-(-4)=-8 percentage points. Its causal interpretation requires that, absent reform, the treated region would have experienced a trend comparable to the chosen control and that policy interference, compositional shifts and measurement changes do not undermine the contrast. The arithmetic itself does not establish those assumptions.
Cost and welfare accounting
Let public payments grow by 100 monetary units, household payments decline by 40, and provider uncompensated-care receipts change by 25. These accounting entries do not alone identify net social resource costs, because transfers and real resources have to be reconciled with provider production, taxes, administration, and changes in utilization. A welfare analysis also requires the expected value of risk protection, not only the realized cash transfers. The exercise is schematic and makes no empirical claim about actual public programs.
Transparent outcome interpretation
An outcome that fails a conventional significance threshold may still have an interval including a substantial clinical benefit. Reporting should distinguish a confidence interval that excludes clinically important effects from an interval that includes both benefit and harm. Likewise, a statistically significant reduction in medical debt needs the underlying denominator, debt definition and reporting mechanism before policy significance can be assessed. The statistical threshold never replaces substantive interpretation.
References
Angrist, Joshua D., and Guido W. Imbens. (1994). Identification and estimation of local average treatment effects. Econometrica, 62(2), 467–475. Published-version information and NBER research record. https://www.nber.org/papers/t0118
Agency for Healthcare Research and Quality (AHRQ). (2024 data; documentation accessed 2026). Medical Expenditure Panel Survey: 2024 Full Year Consolidated Data File, HC-256 Documentation. https://meps.ahrq.gov/data_stats/download_data/pufs/h256/h256doc.shtml
Baicker, Katherine, Sarah L. Taubman, Heidi L. Allen, Mira Bernstein, Jonathan H. Gruber, Joseph P. Newhouse, Eric C. Schneider, Bill J. Wright, Alan M. Zaslavsky, and Amy N. Finkelstein, for the Oregon Health Study Group. (2013). The Oregon experiment—effects of Medicaid on clinical outcomes. New England Journal of Medicine, 368(18), 1713–1722. https://pubmed.ncbi.nlm.nih.gov/23635051/
Callaway, Brantly, and Pedro H. C. Sant'Anna. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200–230. https://doi.org/10.1016/j.jeconom.2020.12.001
Card, David, Carlos Dobkin, and Nicole Maestas. (2009). Does Medicare save lives? Quarterly Journal of Economics, 124(2), 597–636. https://academic.oup.com/qje/article-abstract/124/2/597/1905082
Centers for Medicare & Medicaid Services. (2026 access). Medicaid Eligibility Policy. Medicaid.gov. https://www.medicaid.gov/medicaid/eligibility-policy
Finkelstein, Amy, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gruber, Joseph P. Newhouse, Heidi Allen, and Katherine Baicker, for the Oregon Health Study Group. (2012). The Oregon Health Insurance Experiment: Evidence from the first year. Quarterly Journal of Economics, 127(3), 1057–1106. https://pubmed.ncbi.nlm.nih.gov/23293397/
Finkelstein, Amy, Nathaniel Hendren, and Erzo F. P. Luttmer. (2019). The value of Medicaid: Interpreting results from the Oregon Health Insurance Experiment. Journal of Political Economy, 127(6), 2836–2874. https://pmc.ncbi.nlm.nih.gov/articles/PMC8081392/
Garthwaite, Craig, Tal Gross, and Matthew J. Notowidigdo. (2014). Public health insurance, labor supply, and employment lock. Quarterly Journal of Economics, 129(2), 653–696. https://doi.org/10.1093/qje/qju005
Goodman-Bacon, Andrew. (2021). Difference-in-differences with variation in treatment timing. Journal of Econometrics, 225(2), 254–277. https://doi.org/10.1016/j.jeconom.2021.03.014
Gruber, Jonathan, Nathaniel Hendren, and Robert M. Townsend. (2014). The great equalizer: Health care access and infant mortality in Thailand. American Economic Journal: Applied Economics, 6(1), 91–107. https://pubs.aeaweb.org/doi/abs/10.1257/app.6.1.91
Hahn, Jinyong, Petra Todd, and Wilbert van der Klaauw. (2001). Identification and estimation of treatment effects with a regression-discontinuity design. Econometrica, 69(1), 201–209. https://onlinelibrary.wiley.com/doi/full/10.1111/1468-0262.00183
King, Gary, Emmanuela Gakidou, Kosuke Imai, Jason Lakin, Ryan T. Moore, Clayton Nall, Nirmala Ravishankar, Manett Vargas, Martha María Téllez-Rojo, Juan Eugenio Hernández Ávila, Mauricio Hernández Ávila, and Héctor Hernández Llamas. (2009). Public policy for the poor? A randomised assessment of the Mexican universal health insurance programme. The Lancet, 373(9673), 1447–1454. https://pubmed.ncbi.nlm.nih.gov/19359034/
Kluender, Raymond, Neale Mahoney, Francis Wong, and Wesley Yin. (2021). Medical debt in the US, 2009–2020. JAMA, 326(3), 250–256. https://jamanetwork.com/journals/jama/fullarticle/2782187
Mensah, Joseph, Joseph R. Oppong, and Christoph M. Schmidt. (2010). Ghana's National Health Insurance Scheme in the context of the health MDGs: An empirical evaluation using propensity score matching. Health Economics, 19(S1), 95–106. https://pubmed.ncbi.nlm.nih.gov/20730999/
Miller, Sarah. (2012–2013). The effect of the Massachusetts reform on health care utilization. Inquiry, 49(4), 317–326. https://pubmed.ncbi.nlm.nih.gov/23469675/
Miller, Sarah, Norman Johnson, and Laura R. Wherry. (2021). Medicaid and mortality: New evidence from linked survey and administrative data. Quarterly Journal of Economics, 136(3), 1783–1829. https://www.nber.org/papers/w26081
National Bureau of Economic Research. (2026 access). Oregon Health Insurance Experiment: Research Results and Data. https://www.nber.org/oregon
OECD. (2025). Health at a Glance 2025: OECD Indicators, "Population coverage for healthcare." OECD Publishing, Paris. https://www.oecd.org/en/publications/health-at-a-glance-2025_8f9e3f98-en/full-report/population-coverage-for-healthcare_5c968b73.html
RAND Corporation. (2006). The Health Insurance Experiment: A Classic RAND Study Speaks to the Current Health Care Reform Debate. Research Brief RB-9174. https://www.rand.org/content/dam/rand/pubs/research_briefs/2006/RAND_RB9174.pdf
Rizal, Muhammad Fikru. (2025). Public health insurance and healthcare utilisation decisions of young adults. Health Economics, 34(3), 500–517. https://pubmed.ncbi.nlm.nih.gov/39663202/
Sommers, Benjamin D., Sharon K. Long, and Katherine Baicker. (2014). Changes in mortality after Massachusetts health care reform: A quasi-experimental study. Annals of Internal Medicine, 160(9), 585–593. https://pubmed.ncbi.nlm.nih.gov/24798521/
Sommers, Benjamin D., Robert J. Blendon, E. John Orav, and Arnold M. Epstein. (2016). Changes in utilization and health among low-income adults after Medicaid expansion or expanded private insurance. JAMA Internal Medicine, 176(10), 1501–1509. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2542420
Sun, Liyang, and Sarah Abraham. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects. Journal of Econometrics, 225(2), 175–199. https://doi.org/10.1016/j.jeconom.2020.09.006
Taubman, Sarah L., Heidi L. Allen, Bill J. Wright, Katherine Baicker, and Amy N. Finkelstein. (2014). Medicaid increases emergency-department use: Evidence from Oregon's Health Insurance Experiment. Science, 343(6168), 263–268. https://pmc.ncbi.nlm.nih.gov/articles/PMC3955206/
World Health Organization and World Bank. (2025). Tracking Universal Health Coverage: 2025 Global Monitoring Report. World Health Organization. 108 pp. ISBN 978-92-4-011780-8. https://www.who.int/publications/i/item/9789240117808