Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

The Oregon Medicaid Lottery and Insurance Outcomes

Abstract. The Oregon Medicaid lottery of 2008 created an unusually informative experiment in the consequences of public health insurance for low-income uninsured adults. Random selection from an oversubscribed reservation list produced a substantial difference in the probability of obtaining Medicaid, permitting researchers to estimate both the effect of receiving an opportunity to apply and, under instrumental-variable assumptions, the effect of coverage among people whose enrollment was induced by that opportunity. This research dossier reconstructs the institutional allocation process, the statistical estimands, and the evidence concerning care utilization, clinical measurements, mental health, financial exposure, labor-market activity, household spillovers, and the economic incidence of public financing. It places the principal findings of Finkelstein and colleagues (2012), Baicker and colleagues (2013), Taubman and colleagues (2014), and subsequent analyses into a common causal and insurance-economic framework while preserving the distinct data sources, follow-up periods, populations, units, and uncertainties. The most secure combined finding is that Medicaid substantially increased access to and use of medical services and reduced financial strain. The experiment found improved depression-screening outcomes and increased diabetes detection and treatment, while it did not detect statistically significant changes in several measured physical-health indicators over approximately two years. Emergency-department records showed increased visits over approximately eighteen months, thereby challenging a simple account in which expanded office-based care rapidly substitutes for emergency use. Later welfare analysis emphasizes that coverage transfers resources among patients, providers, and other payers as well as protecting households from expenditure risk. The dossier develops original source-based interpretations of the lottery as a sequence of contingent institutional transitions, the distinction between risk transfer and service consumption, and the boundaries of policy transport. These interpretations are analytical syntheses of the published evidence, not new formal theories. The investigation is based on verified scholarly and institutional texts and public data documentation, without an independent patient-level replication or external peer review.

Keywords: Oregon Medicaid lottery; Oregon Health Insurance Experiment; Medicaid; health insurance; random assignment; instrumental variables; intention to treat; local average treatment effect; insurance take-up; low-income adults; OHP Standard; emergency department; medical utilization; financial risk protection; catastrophic medical expenses; depression; diabetes; clinical outcomes; welfare incidence; uncompensated care; causal identification; external validity; public insurance; social protection; enrollment spillovers.

Introduction: Object, Questions, and Evidentiary Method

Health insurance is a contingent financial and institutional arrangement. It affects prices faced by patients, the funds available to providers, the allocation of unpredictable medical bills, and expectations about the affordability of care. Such an arrangement can change behavior before it changes a clinical biomarker; it can produce financial protection even when an individual's measured health remains stable over a short horizon. A rigorous account of its consequences therefore requires separate outcomes and separate causal questions. The Oregon lottery is an exceptionally useful setting for this task because the state created a randomized opportunity to apply for public insurance during a period when the program could serve only a fraction of interested people.

The object of this dossier is the sequence linking the Oregon Health Plan Standard (OHP Standard) enrollment lottery to Medicaid take-up, use of medical services, financial exposure, health outcomes, and the distribution of economic benefits. The intervention occurred in Oregon in 2008, and the principal experiments measured outcomes during roughly the following one to two years. The research distinguishes that historical program and its participant pool from later Affordable Care Act expansion populations. The historical and study-design accounts are anchored in the Oregon experiment background and the randomized first-year Quarterly Journal of Economics study (Finkelstein et al., 2012).

Four questions organize the inquiry. First, what exactly was randomized, and how did random selection become actual coverage? Second, which differences in medical care, clinical endpoints, household finances, and other outcomes can be attributed to this source of variation? Third, what assumptions connect the lottery's intention-to-treat effect to an instrumental-variable estimate of coverage, and which people does that estimate describe? Fourth, which conclusions can inform the design and evaluation of insurance elsewhere, given differences in population, legal entitlement, benefit design, providers, and time?

The analysis treats a scholarly result as a statement containing an estimand, measurement, comparison group, study period, and uncertainty interval whenever those elements are available. It distinguishes the evidence generated by random assignment from descriptive differences among insured and uninsured respondents. It also distinguishes outcomes collected by mailed or in-person survey from administrative emergency-department, enrollment, and employment records. The principal clinical results come from the NEJM clinical-outcomes trial (Baicker et al., 2013); the principal emergency-use estimates come from the Science emergency-department analysis (Taubman et al., 2014). Differences between these analyses are substantively important rather than editorial inconsistencies.

Methodologically, this is a critical synthesis of original trial reports, scholarly extensions, research-data documentation, and identification theory. The analysis adopts the potential-outcomes vocabulary of Imbens and Angrist's local-effect analysis (1994) and Angrist, Imbens, and Rubin's instrumental-variable framework (1996). It interprets numerical findings in the units and observational windows specified by the authors, rather than treating a reported percentage, visit count, or expenditure estimate as universally comparable. Historical program descriptions are sourced to original study teams or named agencies, while later interpretations are separated from the original randomized comparisons.

The scientific novelty of the Oregon experiment lies in the published studies and in their empirical contribution to identification of Medicaid's consequences. This dossier's contribution is a source-critical explanation of how those studies fit together, including points at which apparently competing results concern different outcomes or observation systems. No independent experimental result, invented confidence interval, new causal model, or proprietary doctrine is claimed. The audit trail for the documentary evidence, exact hyperlinks, source limitations, and claim map accompanies the manuscript in EVIDENCE.json.

Foundations and Source-Critical History

Insurance coverage within Oregon's earlier policy architecture

Oregon's 2008 allocation process emerged from the history of its low-income adult Medicaid program. OHP Standard covered a category of adults for whom public coverage had been restricted by financing and enrollment limits. The program had closed to new enrollment in 2004. When additional places became financially possible in 2008, the number of prospective applicants exceeded the number of places. The state opened a reservation list, received registrations from approximately 90,000 people, and held drawings over subsequent months. Selection created an opportunity to make an application; eligibility and administrative processing still determined enrollment. The research team's institutional background account describes this allocation problem as the immediate origin of randomization.

That distinction matters for the institution of insurance. A person can desire coverage, register interest, receive an offer, complete an application, satisfy program criteria, and become insured. Those transitions change both the population observed at each stage and the interpretation of a causal comparison. For OHP Standard, the eligibility design covered low-income, uninsured adults outside other categorical coverage groups. The published first-year study (Finkelstein et al., 2012) documents the lottery mechanism and the interpretation of program access in this specific group. Some alternative presentations report approximately 30,000 names drawn, while other descriptions count roughly 35,000 people offered an opportunity in the relevant drawing period. These figures reflect differences in rounding, reporting window, and counting individuals or people in selected households; their numerical equivalence should never be assumed. The underlying reservation-list count likewise appears as nearly 90,000 and, in later exact counts, 89,824. This dossier uses rounded figures only for institutional orientation and design-specific analysis denominators for causal results.

The state's allocation decision was made for administrative reasons. Researchers subsequently recognized that randomized access permitted more credible inference than a conventional comparison of people who voluntarily enrolled and those who remained uninsured. The trial was therefore a policy-generated randomized opportunity, rather than a clinical experiment in which researchers administered an insurance product with complete compliance. The NBER research overview identifies the state–research collaboration and the study's major domains. The randomized origin of lottery selection is the evidentiary advantage; the imperfect conversion of selection to coverage is the principal causal-design complication.

A precise definition of the intervention

The original intervention is the lottery's offer to apply for Medicaid. Actual insurance coverage is an endogenous, post-lottery variable. Someone offered the opportunity could fail to apply, be found ineligible, lose coverage after initial enrollment, or obtain other insurance. Someone not offered the opportunity could sometimes secure Medicaid through another eligibility route or obtain other coverage. These possibilities mean that the difference in insurance status between groups was substantial but smaller than one. The NEJM experimental methods report a 24.1-percentage-point increase in Medicaid coverage in the two-year clinical-survey sample, with a 95% confidence interval of 22.3 to 25.9 percentage points. The first-year research reports a roughly 25-percentage-point increase in the probability of insurance (Finkelstein et al., 2012).

A claim that every lottery winner received Medicaid would therefore redefine the intervention and distort effect sizes. A comparison restricted to people who ended up enrolled would also discard the lottery's balancing property, because application and eligibility were themselves affected by circumstances related to outcomes. The experiment addresses this problem using assignment as an instrument. The observed first-stage shift is central to every reported effect of coverage. In one sense, imperfect enrollment reduces the immediate practical effect of a lottery invitation; in another, the instrument allows the intervention's clinical and financial consequences to be interpreted among people whose insurance status changed because of the lottery.

What the original sequence of publications established

The 2011 NBER working-paper version and the definitive 2012 journal article documented first-year insurance coverage, medical care use, financial strain, and subjective health. Their sequence is recorded in the NBER first-year paper and the published journal record. The 2013 NEJM study added approximately two-year in-person clinical measurements and extended follow-up on utilization and financial outcomes. The 2014 Science study used emergency-department administrative records over an approximately eighteen-month period, providing a better-measured outcome for ED visits than the earlier self-report measures. The 2016 further emergency-use evidence examined whether increased ED use dissipated with time; the reported pattern persisted over the first two years.

Additional research examined outcomes not exhausted by these primary studies: 2014 administrative labor-market findings, later welfare valuation, changes in enrollment among eligible children in participants' households, rural–urban variation, and machine-learning assessments of subgroup heterogeneity. These analyses help define what the original experiment can legitimately establish and how its findings have evolved. The NBER publications registry is useful as a study-family index, but the claims made here derive from the corresponding research reports rather than from the registry alone.

Evidence hierarchy and relation to insurance economics

A randomized offer identifies the mean causal effect of receiving that offer in the defined list population when assignment and outcomes are measured appropriately. An instrumental-variable interpretation of actual insurance additionally requires assumptions about how the offer changes treatment and outcomes. Later subgroup analysis raises questions about effect heterogeneity without overriding the original design. Comparative studies of other Medicaid programs provide evidence for policy transport only to the extent that institutional differences are explicitly confronted. These are different levels of evidence about different objects, and no one study eliminates the need for the others.

Insurance economics supplies a further distinction. Medical care utilization is a resource-use outcome. Reduced out-of-pocket liability, collection exposure, or catastrophic spending is a financial-risk outcome. Changes in diagnosis, mental health, blood pressure, or glycated hemoglobin are clinical outcomes. All may arise from the same insurance intervention, yet they can differ in timing and direction. A coherent causal evaluation therefore preserves the multiple functions of health insurance rather than reducing the intervention to one aggregate score. The subsequent empirical evidence illustrates why this separation is necessary.

Lottery Design and Take-Up

Assignment, household exposure, and compliance

Selection was randomized from a list, but the effective probability that an individual was affected depended on household membership. A selected registrant could make coverage accessible to listed household members, and multiple listed individuals could increase a household's chance of being selected. Consequently, an unadjusted comparison of selected and unselected people could give disproportionate weight to larger households even when the lottery mechanism was random as implemented. The published methods adjusted for the number of household members on the reservation list and clustered uncertainty at the household level. These design choices are explicitly described in the 2013 NEJM methods and the emergency-use methods (Baicker et al., 2013; Taubman et al., 2014).

The unit of assignment and the unit of clinical measurement thus require separate specification. A measured blood-pressure result is individual. A household-level opportunity to apply can affect multiple people. A hospital encounter is a service episode. A Medicaid record measures enrollment over specified dates. A household's medical debt is a financial relationship. Moving between these units without noting the level of exposure can manufacture apparent precision, conceal spillovers, or conflate service use with financial incidence.

Potential outcomes and the treatment of interest

Let Z denote randomized lottery selection, with Z = 1 for selected and Z = 0 for not selected. Let D(z) be actual Medicaid coverage under selection status z, and let Y(d,z) denote an outcome potentially depending on both coverage and selection. The intention-to-treat effect is E[Y | Z = 1] − E[Y | Z = 0], adjusted as required by the randomization scheme. It measures the effect of the offer, including the real-world frictions that prevent an invitation from becoming coverage. The first-stage effect is E[D | Z = 1] − E[D | Z = 0]. Under suitable assumptions, the ratio of these effects provides the local average treatment effect of coverage for compliers:

LATE=E[Y∣Z=1]−E[Y∣Z=0]E[D∣Z=1]−E[D∣Z=0].\mathrm{LATE}=\frac{E[Y\mid Z=1]-E[Y\mid Z=0]}{E[D\mid Z=1]-E[D\mid Z=0]}.

This identity follows the established local average treatment-effect theory (Imbens & Angrist, 1994). Its denominator is a proportion, not a treatment effect expressed as an arbitrary percentage figure; a first stage of 0.241 implies an amplification of about 4.15 for a compatible outcome's intention-to-treat contrast. This is an arithmetic illustration based on the 2013 reported first stage, not a newly estimated Oregon treatment effect. The exact published instrumental-variable estimates use study-specific regression specifications, weights, and uncertainty calculations.

The model relies on the lottery being as good as random conditional on the appropriate assignment design; the offer affecting the outcome through coverage rather than through an independent channel (exclusion restriction); selection not making anyone who would otherwise have Medicaid lose coverage (monotonicity); and a relevant, consistently defined treatment. The instrumental-variable identification framework (Angrist, Imbens, & Rubin, 1996) explains why the causal population is the group whose enrollment responds to the lottery. Stable treatment specification and interference assumptions also matter when household members or providers respond to another person's selection.

Compliers and the meaning of a coverage effect

A complier is an individual who would obtain Medicaid if selected and would not obtain it if not selected, under the treatment definition and time window. A person covered regardless of selection is an always-taker; one uncovered regardless of selection is a never-taker. These are potential-compliance types, not directly observable enrollment labels for any particular respondent. An instrumental-variable estimate recovers a weighted local effect among compliers under the required assumptions. The conclusion cannot be read as the average effect for every adult in Oregon or for all Medicaid beneficiaries in the United States.

Compliers may differ in their medical needs, knowledge of the application process, stability of address, ability to navigate documentation, and eligibility. Even if selection itself was independent of measured and unmeasured characteristics, the coverage effect is local to those induced to enroll. When Medicaid is made automatic, mandatory, less administratively costly, or broadly available, it may reach a different population. Estimand clarity is therefore a substantive requirement for future reform comparisons.

Enrollment friction as an insurance outcome

Enrollment is itself evidence about the practical content of a statutory or administrative opportunity. The original studies indicate that lottery selection was far from equivalent to insured status. An offer traversed an application process, means testing, and continued eligibility or recertification. The emergency-department study notes staggered draws, enrollment delays of approximately one to two months, and losses of coverage during the follow-up; people who obtained Medicaid through the lottery accumulated an average of 13.2 months of coverage in an eighteen-month ED observation period. That duration contextualizes the estimated effect and limits any characterization of it as the consequence of eighteen uninterrupted months of treatment.

An enrollment-friction interpretation has direct relevance for insurance design. The promise of access can have little effect if eligible people cannot complete or maintain enrollment, while support for completing administrative procedures can affect the number covered without modifying the formal benefit package. In a later analysis of enrollment spillovers, Sacarny, Baicker, and Finkelstein (2022) found that adult lottery wins temporarily increased Medicaid enrollment among children who were already eligible regardless of the lottery. At three months, approximately one additional child enrolled for every nine adults newly enrolled because of selection. This documented spillover shows that treatment effects are not exhausted by the coverage status of the original adult applicant.

Estimand discipline

Three statements must be kept separate. Winning the lottery changed the probability of obtaining insurance. Medicaid coverage, instrumented by that win, changed specified outcomes among those induced to enroll. Separately, enrollment of previously eligible children responded to an adult application opportunity. Each claim has a different population or outcome; the third potentially violates a naive assumption that the lottery has no consequence outside the treated adult's own coverage. The original adult IV analysis rests on an exclusion restriction for adult outcomes in the study context, while the child result demonstrates a wider household mechanism. It does not by itself quantify how household spillovers affected any of the adult clinical outcomes.

The most important interpretive consequence is that real coverage policy has multiple margins: invitation, application, acceptance, duration of enrollment, and related household action. Random assignment gives a powerful identification foothold but does not collapse these steps into a single binary institutional event. That insight will remain central when evaluating service and financial outcomes.

Health Utilization and Financial Risk

The first-year pattern: access, use, and expenditures

In the first year after random assignment, the treatment group became substantially more likely to hold insurance and to use medical services. The first-year randomized evidence (Finkelstein et al., 2012) identified increases in outpatient care, prescription drugs, preventive services, and hospitalizations. These findings arise from differences between lottery-selected and unselected people, transformed into coverage effects under the instrumental-variable interpretation. They support a causal account in which eliminating a large part of the immediate patient price and enabling payment to providers expands realized demand. They do not independently establish that every additional visit was clinically necessary or that each treatment generated proportionate health improvement.

A critical economic distinction is the extensive margin of care—whether a person uses a service at all—and the intensive margin—how many encounters, medications, or services the user consumes. Insurance can alter each through a different mechanism. Someone who had postponed preventive care may start using a provider; an existing patient may receive more frequent monitoring or prescriptions; a provider may be more willing to schedule follow-up because reimbursement is available. The original Oregon studies do not assign a unique cause to each additional encounter. Their causal contribution is to establish that coverage changes the realized care portfolio, given the prevailing provider environment and benefit design.

The first-year study reported reduced out-of-pocket medical expenses, medical debt, and unpaid bills sent to collection alongside greater use. This conjunction is central to the meaning of insurance. A household can consume more medical care while bearing less direct financial liability, because the cost is transferred to a public financing pool. The 2011 NBER research account and 2012 journal results describe improved self-reported physical and mental health in the same period. Subjective health reports are meaningful welfare indicators, but they should be separated from subsequently measured clinical biomarkers. A respondent's assessment of health can respond to psychological security, easier access, more diagnosis, or changed expectations even when blood pressure and cholesterol are unchanged over twelve months.

Emergency-department use and the importance of measurement

Emergency use became an important interpretive dispute because a simple substitution argument predicted that insurance would allow beneficiaries to replace emergency-department visits with office-based primary care. The first-year and two-year surveys did not show statistically significant changes in reported ED use in their particular survey analyses. The administrative-record study led by Taubman and colleagues (2014), however, examined approximately 25,000 people using Portland-area hospital records over approximately eighteen months and found that Medicaid increased emergency-department visits by 0.41 per person, about a 40% increase relative to the control mean of 1.02 visits over that entire period. The administrative Science study also reported that Medicaid increased the probability of any ED visit by 7.0 percentage points from a control-group rate of 34.5%.

The figure of 0.41 is a cumulative effect over approximately eighteen months. It is not an annual increase of 0.41 visits, and translating it into annual units without preserving the observational construction would change the published finding. The 40% expression is a proportional effect calculated against the study-specific control mean, not a percentage-point increase in the probability of visiting. The 7.0-percentage-point figure concerns the binary probability of any visit, not visit frequency. These three numerical representations can coexist because they describe related but different estimands.

The administrative study categorized encounters and found increases across visits considered non-emergent, primary-care-treatable, and emergent but potentially preventable, without a statistically significant increase in one category of emergent, nonpreventable visits. Categorization is a research construct applied to observed encounters; a retrospective classification cannot by itself establish what an individual patient knew before deciding to seek emergency care. Nor does a utilization increase prove that the emergency department was always an efficient treatment venue. The experiment team's results synthesis reports that greater ED use accompanied rather than replaced expanded office-based use.

The later NEJM follow-up on emergency care (Finkelstein et al., 2016) found that increased ED use persisted over the first two years, weakening the claim that a short initial surge necessarily disappears once the newly insured establish ordinary care relationships. This finding still describes the early post-coverage horizon; it does not establish a lifetime trend, the equilibrium response to a large national expansion, or the quality of individual ED decisions. It also clarifies why survey and hospital-record results need not be treated as contradictory estimates of the same precisely measured outcome. The source, population, recall process, and window differ.

Clinical outcomes: diagnosis, treatment, and measured disease indicators

The two-year clinical experiment (Baicker et al., 2013) used in-person surveys and objective measurements from 6,387 selected and 5,842 unselected adults. It identified higher rates of diabetes diagnosis and medication use following coverage, as well as several increases in preventive services. The reported increase in a post-lottery diabetes diagnosis was 3.83 percentage points (95% confidence interval 1.93 to 5.73); use of diabetes medication increased by 5.43 percentage points (95% confidence interval 1.39 to 9.48). These are instrumental-variable estimates of coverage in the specified survey population, not general statewide changes in disease prevalence.

At the same time, the study did not detect statistically significant effects on mean glycated hemoglobin, the share exceeding a specified diabetes threshold, mean blood pressure, or mean cholesterol. For example, the estimated systolic blood-pressure difference was −0.52 mm Hg (95% confidence interval −2.97 to 1.93). This interval includes improvements and deteriorations of modest size; a statistical failure to reject zero cannot be interpreted as proof of a precisely zero effect. The study provided better evidence about the short-term average of measured risk factors than about rare clinical events or longer-term consequences of treatment. Increased diagnosis and medication use can occur before measurable population-level risk-factor changes, particularly when therapy is recent or adherence is incomplete.

A further interpretive question concerns what a diagnosis rate represents. Insurance may increase detection of conditions that were already present but previously undiscovered. That generates a positive effect on recorded diagnoses without necessarily increasing the underlying incidence of disease. For diabetes, improved diagnostic access and increased treatment were observed together. The proper conclusion is that Medicaid increased identified and treated disease in the measured period; the underlying health benefit of additional detection requires further clinical follow-up beyond the observation of diagnosis itself.

Depression, perceived well-being, and financial security

The clearest positive clinical-screening result concerned depressive symptoms. Baicker and colleagues reported that Medicaid reduced the probability of screening positive on the PHQ-8 depression measure by 9.15 percentage points (95% confidence interval −16.70 to −1.60; P = 0.02). Relative to the control-group prevalence near 30%, that represented about a 30% decline. This was an effect on a validated screening threshold, not an independently adjudicated psychiatric diagnosis for every respondent and not necessarily a uniform change in depression severity. The same analysis recorded an increase in post-lottery diagnosis of depression, a combination consistent with expanded recognition of illness and improved symptoms for some beneficiaries. The clinical study's full results specify the PHQ-8 measurement and the study's limitations.

The psychological consequences of health insurance may follow several overlapping channels. A person may feel better because medical care becomes affordable, anxiety about debt declines, treatment becomes possible, or familiar care pathways become available. The trial established a change in depression screening and self-reported well-being; it did not uniquely decompose the contribution of financial security, clinical treatment, or perceived access. Attribution to one channel would require a separate identifying design or strong assumptions about mediation. The absence of a definitive mechanism does not diminish the causal evidence for the measured screening outcome.

Financial protection was substantial. The 2013 study reported that Medicaid nearly eliminated catastrophic out-of-pocket medical expenditures under its specified threshold, with an estimated reduction of 4.48 percentage points (95% confidence interval −8.26 to −0.69) from a low control-group rate. It also found less financial strain and reduced exposure to burdensome bills. These measures do not imply that public financing eliminated the medical services' resource costs, removed all household liabilities, or covered every category of care. Rather, they identify how the allocation of payment risk changed for people induced to obtain Medicaid during the study. The mechanism is consistent with the first-year evidence on debt and collections (Finkelstein et al., 2012).

The catastrophic-spending result is consequential even for households that would not have suffered an unusually high bill. Insurance is purchased or assigned before the realized health state is known. A reduction in the probability of a severe expenditure may have ex ante value beyond the observed average change in expenditures. Conversely, eliminating out-of-pocket charges can increase the volume of care consumed. These are two connected functions of coverage: pooling risk and financing services. Their relative welfare contributions cannot be inferred merely from the increase in utilization.

Labor-market activity and the indirect insurance budget

Some theoretical accounts predict that extending publicly funded insurance may reduce labor supply if eligibility depends on income or if workers no longer need employment to obtain insurance. Others predict improved job search, financial stability, or health sufficient to support employment. The Oregon lottery offers a way to test these mechanisms empirically in the early follow-up period. The randomized administrative labor-market analysis (Baicker, Finkelstein, Song, & Taubman, 2014) found no statistically significant effects of Medicaid on employment or earnings in the measured period, while detecting an increase in food-stamp participation. The associated NBER working-paper record specifies that its confidence interval ruled out an employment decrease greater than 4.4 percentage points and an increase greater than 1.2 percentage points.

The labor finding should be given its exact evidentiary weight. It narrows the set of plausible short-term effects in the observed group; it does not prove that all labor effects are zero in every population or under other eligibility regimes. Nor does it show that coverage was fiscally costless. Public insurance affects public outlays, uncompensated care, providers' receipts, household spending, and potentially other transfer programs. The observed SNAP participation increase is a reminder that program information and administrative contact can alter engagement with other support systems, an effect relevant to economic incidence even when earnings remain statistically stable.

Medical spending and economic incidence

The experiment's first-year spending calculations estimated resources used, not just amounts paid directly by enrolled households. A later Journal of Political Economy welfare analysis (Finkelstein, Hendren, & Luttmer, 2019) explicitly separated gross Medicaid expenditure from the incremental resource cost of care and transfers to parties that would otherwise finance uncompensated treatment. The authors estimated that the resource cost of covering an additional recipient was about 40% of Medicaid's gross spending, while the remaining approximately 60% represented a transfer to providers of implicit uncompensated-care insurance for the uninsured. They also estimated recipient willingness to pay in a range of $0.50–$1.20 per dollar of the resource cost under their alternative welfare frameworks.

These numbers have a specific interpretation. They are estimated within the authors' welfare methods and Oregon study context, not an accounting identity for every Medicaid program. The authors' willingness-to-pay range varies because the value of insurance in unobserved health states requires assumptions beyond randomized differences in realized spending. In their analysis, a substantial portion of payments replaced financing that had previously occurred through uncompensated care. Hence gross Medicaid spending cannot simply be equated to new net social resources consumed, and the amount paid to providers cannot simply be equated to a transfer valued equally by the insured. This distinction is essential when comparing public insurance with other redistributive instruments.

The original 2015 working-paper version of the welfare analysis reported its ratios against government spending and employed an earlier framework. Its headline ratios should not be spliced into the final 2019 publication's resource-cost denominators. The final published welfare study supplies the appropriately identified later estimates and explains where the uncertainty enters. Such source chronology is part of scientific editorial control: a value can change because a model is revised or a denominator is redefined, even when the underlying experiment has not changed.

A unified reading of observed insurance outcomes

Taken together, the observations establish three coexisting effects over the short follow-up. Coverage increased the use of medical services; it reduced the household's exposure to unmanageable medical payments; and it produced a mixed set of clinical outcomes, with evidence of lower depression screening prevalence and more disease detection but imprecisely estimated or statistically undetected changes in several physical markers. A higher ED visit count and a lower probability of catastrophic household spending are both causal consequences of the intervention under the reported assumptions. Treating one as sufficient to negate the other imposes an external definition of insurance success that the experiment itself did not test.

The interpretive task is to characterize the insured benefit relative to its cost, incidence, and timing. Some additional care may have large future benefits, some may have limited clinical value, and some may substitute for care that would otherwise be financed elsewhere. The Oregon evidence informs this analysis while leaving room for clinical effectiveness research, price measurement, provider-response studies, and longer-run welfare evaluation. It establishes that risk protection, access, and clinical health occupy distinct empirical dimensions, each of which must be reported with its own unit and uncertainty.

Causal and Generalizability Limits

Internal validity of the offer effect

The lottery's central strength is that selection was random within the state’s implemented drawing scheme. Where the design and sample are properly handled, selected and unselected people are comparable in expectation before the intervention. Thus an intention-to-treat comparison can attribute a subsequent average difference to the opportunity created by selection, subject to appropriate handling of follow-up and measurement. This does not require researchers to explain why every applicant took a particular action. It requires that assignment itself be implemented and analyzed as randomized and that outcome measurement retain a defensible relationship to that assignment.

Survey participation complicates otherwise clean random assignment. A biomarker cannot be measured in someone who did not participate in the examination, and survey response may correlate with health or insurance experience. The 2013 team employed sampling and field-collection weights, controlled household size, and assessed alternative specifications, as detailed in the NEJM methods section. These procedures support inference within the observed design but are not a magical recovery of every unobserved outcome. The validity of conclusions involving the examined subsample rests on the sampling and response assumptions that justify its analysis.

Administrative data can reduce some problems associated with respondent recall and nonresponse. Yet administrative coverage is also bounded by participating institutions, geography, dates, and coding practices. The ED analysis's hospital-record system provided greater precision for the specific emergency-use outcome than the survey estimates; it did not automatically make that study an unbiased national census of all emergency visits or all health care. A sound assessment chooses the data source appropriate to the outcome instead of ranking every study solely by its sample size.

The exclusion restriction in practice

An instrumental-variable estimate of coverage normally assumes that winning the lottery affects the adult's outcome through Medicaid coverage. Lottery selection might also provide information about public programs, stimulate application behavior, alter expectations, or affect household circumstances before final coverage. The published designs interpret the instrument through the assumption that these alternative channels do not materially affect the target adult outcome, or they are part of the coverage mechanism as specified. The assumption is not directly testable from the lottery comparison alone. The formal instrumental-variable account (Angrist, Imbens, & Rubin, 1996) makes clear that identifying a complier average treatment effect requires more than an instrument correlated with treatment.

Household spillovers provide a concrete reason to be precise. The 2022 child-enrollment study found that an adult's selection affected the enrollment of children who were already eligible independently of that selection. This is a documented institutional pathway operating beyond the original adult insurance indicator. It does not automatically invalidate all adult instrumental-variable results: its effect on each adult outcome must be established, rather than presumed. Nevertheless, the child finding reveals that the instrument can activate administrative and family mechanisms wider than a single treatment variable. A later replication should specify whether the desired causal quantity is own coverage, household access, or the combined consequences of selection.

Monotonicity, compliance, and treatment definition

Monotonicity asserts that being offered a Medicaid opportunity does not make an otherwise-covered individual forgo Medicaid. In the specific setting, the assumption is plausible as an account of program enrollment, but treatment definition and temporal measurement matter. A person can cycle in and out of coverage, switch eligibility categories, or enroll in private insurance. The original first stage was calculated for the study's chosen coverage indicators and time windows. One cannot extend its numerical size to a different follow-up horizon without new data.

The population of compliers is neither the whole reservation list nor the entire low-income uninsured population. It consists of people who would be covered when selected and uncovered when not selected. The conditioning is cross-world and cannot be directly observed on a case-by-case basis. Statistical methods identify its average effect under assumptions, while descriptions of applicant characteristics can only approximate how it relates to external populations. Consequently, an average effect among Oregon compliers is a result about a defined response group within an existing institutional environment.

Statistical power and interpretation of null findings

A nonsignificant estimate is evidence about an interval of effects compatible with the data and statistical method; it is not a demonstration that the true effect equals zero. The difference becomes especially important when physical-health endpoints are heterogeneous, rare, measured with error, or responsive over more than two years. For example, the reported 95% interval for systolic blood pressure, −2.97 to 1.93 mm Hg, excludes some large changes but allows both modest improvements and modest worsening. The absence of a significant average cholesterol or glycated-hemoglobin effect is likewise bounded by outcome variation and study power.

The trial tested numerous outcomes: utilization, diagnosis, objective biomarkers, self-reported health, depression, preventive services, expenditure risk, and more. When many comparisons are made, some nominally significant results can occur by chance even under null hypotheses. Researchers and later readers must inspect prespecification, multiple-testing practices, and the coherence of effect patterns rather than selecting one P value as a complete policy verdict. The clinical study and its accompanying supplementary methods constitute the appropriate source for specific uncertainty intervals. Any subsequent secondary analysis proposing differentiated cardiovascular effects requires its own multiple-testing and sample-selection assessment.

Moreover, a physiological change is not the only reason that a patient or payer may value insurance. Protection against a financially catastrophic state can be valuable in expectation even when the event does not occur in the observed period, and an earlier diagnosis may be an intermediate condition for later treatment benefit. The inability to demonstrate a two-year mean biomarker change therefore cannot logically refute the independently measured financial and utilization effects.

Heterogeneity and exploratory subgroup analysis

Treatment effects may depend on prior health, age, household composition, distance to providers, baseline service use, and the conditions of access. The question is whether observed differences reflect real effect modification or noisy estimation across many subgroups. The causal-forest reanalysis (Hattab, Doherty, Ryan, & O'Neill, 2024) found variations in lottery-related insurance take-up across demographic subgroups but relatively weak evidence of heterogeneous treatment effects across the examined health and financial outcomes. Its machine-learning methods can generate hypotheses; they do not turn an observational subgroup classification into a new randomization or overcome the loss of precision created by instrumenting coverage.

Rurality is another example. The rural–urban Oregon analysis (Allen, Wright, & Broffman, 2018; first online 2017) combined comparative survey outcomes with interviews of 86 rural participants. It found that rural participants appeared to benefit from Medicaid on studied dimensions at least comparably to urban participants, while the qualitative evidence highlighted persistent practical barriers. This result is more informative than an assumption that a rural address universally blocks insurance gains. At the same time, provider availability, transport, and time costs can mediate the effect of a coverage entitlement. The Oregon experiment did not randomize entire local provider networks or regional medical capacity.

Temporal and institutional external validity

The original intervention occurred in 2008 in a geographically and administratively specific program. Later national Medicaid expansions reached different states, beneficiaries, provider systems, and enrollment rules. In 2014, the Affordable Care Act's expansion changed eligibility in many jurisdictions on a far larger scale. The comparative ACA expansion study (Wherry & Miller, 2016) used quasi-experimental differences across states and periods rather than a comparable lottery; its findings support separate statements about coverage, access, and selected utilization under those policy conditions. It does not retroactively change the Oregon randomized effect or establish that every national outcome shares Oregon's magnitude.

A large expansion can affect providers' workload, appointment availability, wages, and negotiated or administered prices. A small program's marginal effect while most provider capacity remains unchanged need not equal the general-equilibrium effect of broad coverage. Likewise, an insurance design with high patient cost-sharing may produce different behavior from a low-cost-sharing Medicaid benefit. A later program with different eligibility verification or continuous enrollment rules may produce a different effective duration of coverage. Translating Oregon's results requires an institutional crosswalk: target population, financing method, benefit scope, enrollee price, provider payment, access constraints, and follow-up duration.

Ethical and policy interpretation

The experiment's evidentiary credibility arises from an administrative lottery used to allocate scarce places. The ethical assessment of allocating medical coverage through scarcity is a separate subject from the validity of the inference. A policy maker could find the randomized evidence highly informative while also believing that a waiting-list lottery is an undesirable permanent insurance institution. Empirical accuracy does not require endorsement of the historical allocation mechanism, and a normative proposal does not alter the historical causal estimate.

An evaluation also requires an explicit welfare criterion. A budgetary analysis may prioritize fiscal cost; a health-outcome analysis may prioritize measurable clinical changes; a financial-protection analysis may prioritize the reduction of catastrophic household risk; an equity analysis may emphasize who receives access. These are legitimate but different questions. The Oregon study provides inputs to them; it does not settle their relative normative weights. The economic-incidence work of Finkelstein, Hendren, and Luttmer (2019) formalizes one way of evaluating recipient value and provider transfers, with explicit assumptions. Any further policy ranking must explain the criterion used and its uncertainty.

Data, Measurement, and Validation

Source provenance and the reproducibility boundary

The Oregon experiment generated several overlapping datasets rather than a single universal matrix of outcomes. Lottery administrative records establish offers and enrollment possibilities. Medicaid program files measure insurance participation. Mail and in-person surveys collect reported care, perceived access, expenditure exposure, and selected diagnoses. Direct clinical examinations generate blood pressure, lipid, and glycated-hemoglobin measures. Hospital records capture specified emergency encounters, while employment and transfer-program records provide administrative measures of labor and public-benefit outcomes. The original studies combine these materials according to different research questions, consent procedures, linkages, geographic samples, and observation intervals.

The Oregon experiment public-use repository, maintained through NBER, lists public-use data and replication materials for the 2012 QJE, 2013 NEJM, 2014 Science, and 2016 NEJM analyses. Its stated data-use conditions prohibit reidentification, contacting subjects, and commercial use. These conditions are part of the scientific and ethical provenance of any new empirical project. The present manuscript analyzes published results and documentary methods; it has not downloaded the restricted-by-terms dataset, reproduced a regression, or inspected individual participant records. Source-based calculations below are checks of interpretation, not claims of statistical replication.

The NBER project portal makes the distinction between publications and data/protocols explicit. Replicability is a chain: discover the documented source, comply with data terms, recover the appropriate analytic sample and version, execute the original code or a properly disclosed independent specification, compare outputs with the publication, and document differences. Merely finding a public-use dataset establishes availability of materials, not the successful reproduction of estimates. This dossier does not mislabel literature synthesis as a completed code replication.

The unit and denominator audit

A transparent research record attaches a denominator to each numerical statement. The number of people on the reservation list describes interested applicants, not an outcome sample. The approximately 12,229 two-year in-person survey respondents describe a measured clinical-survey population, composed of 6,387 selected and 5,842 unselected adults. The approximately 25,000 ED participants reflect an administrative hospital-record study. The control mean of 1.02 emergency visits refers to all recorded visits per person over the ED observation period in that sample. A 34.5% control rate measures having any observed emergency visit. These denominators cannot be interchanged.

The primary studies' use of randomized assignment does not authorize unweighted comparison when the selection mechanism makes individual treatment probabilities depend on household size. Their reported methods include household-size indicators and household-clustered uncertainty. The first-year journal publication also relies on its particular survey and administrative samples. A replication should preserve study-specific weighting and collection design, because an apparent disagreement between two estimates may be caused by a changed sample or outcome definition rather than a failure of causal identification.

For money, an equally important unit audit separates out-of-pocket expenses, total medical spending, government payments, uncompensated care, and household debt. A decline in unpaid bills is not necessarily equal to the aggregate cost of treatment avoided. A transfer to providers may change who finances care without changing the amount of medical service delivered. In the published Medicaid value analysis, gross Medicaid spending, incremental resource cost, and transfers are separately defined within a welfare exercise. Using the authors' resource-cost denominator for one ratio and a public-budget denominator for another would create a numerically plausible but analytically invalid result.

A compact verification matrix

Empirical object Original source Measurement and follow-up Claim supported Important restriction
Insurance first stage Finkelstein et al. (2012); Baicker et al. (2013) Assignment versus Medicaid coverage, approximately one to two years Lottery opportunity raises coverage roughly 25 percentage points Offer is not enrollment; precise first stage varies by sample
Outpatient care, prescriptions, hospitalizations Finkelstein et al. (2012) Survey and linked administrative measures in first year Utilization rises on several margins Not a separate clinical-effect estimate for each encounter
Depression screening Baicker et al. (2013) PHQ-8 screening at approximately two years Positive-screening probability declines 9.15 percentage points Screening threshold, not a universal clinical diagnosis
Blood pressure, cholesterol, glycemic control Baicker et al. (2013) Objective examination at approximately two years No statistically detected mean improvement in selected indicators Confidence intervals allow some nonzero effects
Emergency visits Taubman et al. (2014) Portland-area hospital records, approximately eighteen months 0.41 additional visits per covered person Visits over the period, not visits per year
Employment and earnings Baicker et al. (2014) Linked 2009 administrative data No statistically significant short-run effect Limited time window and specified confidence bounds
Household child enrollment Sacarny et al. (2022) Medicaid enrollment following adult lottery selection Temporary additional enrollment among already-eligible children Distinct spillover outcome and household unit
Economic incidence Finkelstein et al. (2019) Modeled welfare decomposition using experimental estimates Resource cost and gross payments differ; provider transfers are material Sensitive to welfare assumptions and counterfactual financing

The matrix is an interpretive index, and each reported estimate remains tied to its original published paper in References. It does not combine effect sizes into a pooled estimate, because the outcomes and study populations are not homogeneous observations of one dependent variable. The first row makes especially clear that the same approximate 25-percentage-point first stage should not be imposed mechanically on a later report if its own estimated first stage differs.

Quantitative sanity checks

The first numerical consistency check concerns emergency use: 0.41 added visits divided by the 1.02 control mean equals approximately 0.402, or 40.2%. This verifies the relationship between the reported absolute and relative ED visit effects, allowing for rounded estimates. A separate calculation shows that a 7.0-percentage-point increase from a 34.5% rate corresponds to an approximate 20.3% increase in the probability of any ED visit. Neither calculation is an independent regression, and neither converts visits into days or service costs. The published ED study remains the source for confidence intervals and regression specifications.

The second check concerns the first stage. An increase of 24.1 percentage points equals a probability difference of 0.241. Its reciprocal is approximately 4.149. For a simplified Wald estimator, a compatible intention-to-treat outcome difference is therefore multiplied by about 4.15 to express the complier effect of insurance. The formula must be applied to a treatment definition and instrument from the same analytic dataset. It should not be used to derive a clinical effect from the ED first stage or vice versa. The NEJM trial report supplies the 24.1-percentage-point figure and its confidence limits.

The third check concerns the change in a depression-screening rate. A 9.15-percentage-point decline against a control-group prevalence of roughly 30% is a relative reduction of approximately 30.5%, consistent with the paper's statement of about 30%. It is an effect on an outcome prevalence, not a 30.5-point fall and not a 30.5% change in continuous PHQ-8 score. Because source values are rounded, the derived percentage is approximate.

The fourth check concerns welfare incidence. If an author's decomposed gross expenditure is normalized to $1, and the estimated resource cost of incremental coverage is 40 cents, the other 60 cents of the financing flow takes the form identified as transfers in that author's model. These figures do not mean that a Medicaid recipient necessarily receives 40 cents in benefits, or that every hospital receives 60 cents in additional profit. They are the estimated allocation of gross costs within a model of counterfactual uncompensated care. The 2019 original economic analysis supplies the relevant model and welfare assumptions.

Link integrity and source identity

A scientific link must identify the work or institutional record that supports the adjacent assertion. The manuscript uses contextual link text such as the published study title or the exact result being discussed. Sources repeated in several chapters retain the same canonical external address. Research-reference identifiers are internal to EVIDENCE.json; the reader sees author–year attribution and a direct link rather than cryptic codes. The final References section provides full bibliographic identities for the linked primary and methodological works.

The web-visible verification performed for this dossier checked publisher records, PubMed/PMC scholarly records, NBER research and data pages, and source-specific statistical descriptions. Some publisher-hosted full texts or PDF supplements are inaccessible in the available browsing environment; those sources are used only to the extent supported by their accessible abstract, journal metadata, or separately available author manuscript. Individual numerical estimates were checked against accessible primary research text or authoritative project summaries. A direct HTTP status test from the local authoring environment is a separate technical operation and cannot be inferred from search indexing alone. The evidence ledger records this boundary rather than claiming exhaustive full-text access or full internet-link crawling.

Documentary revision controls

Primary papers can appear as working papers, online-first articles, print-issue articles, public author manuscripts, and corrected versions. The same study can therefore have multiple dates and display formats. For Finkelstein et al., the 2011 working paper precedes the definitive 2012 QJE publication. For Allen, Wright, and Broffman, first online publication occurred in 2017 and print issue placement in 2018. The 2015 working-paper version of the Medicaid welfare paper and the 2019 journal publication use different framing and denominators. Research traceability requires naming the version whose results are used.

The corpus can be checked against the original authors' Oregon publication index, and the broader physical-health interpretations should be read against actual estimates instead of compressed media summaries. Where the underlying policy changed after the study, the historical 2008 intervention retains its own date and provisions. A retrospective claim about this experiment is therefore validated against what existed and was measured at the relevant time, not against a current Medicaid application page.

International Synthesis and Scientific Conclusions

What Oregon contributes to insurance science

The Oregon lottery provides a relatively clean empirical answer to an otherwise difficult question: what happens to a subset of uninsured, low-income adults when public insurance becomes available to them through an exogenous administrative opportunity? Its answer is multidimensional. Coverage increased observed consumption of health services, reduced medical financial strain, improved some self-reported and mental-health measures, and did not generate statistically detectable short-run improvements in several measured physical-risk factors. These findings were generated under a distinctive policy architecture and causal design, but the theoretical distinctions they demand are relevant across insurance systems.

The most important general scientific implication is that insurance operates simultaneously as financial risk transfer, payment authorization, and access infrastructure. In a low-income population, ex ante insurance protection may be valuable even when short-run biomarkers are stable. In the same population, reduced patient price can increase both primary-care and emergency-care utilization. More diagnoses can indicate detection rather than a rise in disease incidence. Funding can replace informal and uncompensated care rather than produce an equivalent increment of net medical resources. Each of these statements belongs to established insurance economics and causal evaluation; the present dossier integrates their meaning in a bounded, source-based synthesis.

The translation from eligibility to effective protection

An insurance reform's legal or administrative eligibility is only one point in its operational chain. Oregon's lottery induced coverage but left substantial non-take-up among those offered a chance to apply. Its household spillovers show how the same intervention can improve knowledge of or connection to programs serving other eligible family members. The child-enrollment spillover evidence (Sacarny et al., 2022) indicates that existing eligibility can remain unused until another event makes application salient. International comparisons should therefore separate the eligibility population, the enrolled population, and the population with uninterrupted effective benefits.

The distinction is especially important where coverage is conditioned on documentation, frequent recertification, income tests, or a complex provider network. Two countries with identical statutory coverage percentages can differ in realized financial protection if one has more enrollment churn or greater point-of-care charges. The Oregon evidence does not quantify those differences internationally; it provides a rigorous warning against treating formal entitlement as measured insurance receipt.

The relationship between provider capacity and care use

A meaningful insurance benefit requires clinical capacity through which it can be exercised. Coverage can increase demand for physician visits while appointment supply remains constrained. It can also improve the ability of insured patients to obtain reimbursement for emergency encounters, leading to increased visits in a facility that operates under different access conditions. The reported absence of rapid substitution from ED to office-based care is therefore understandable without attributing irrationality to beneficiaries. A medical-access explanation requires attention to location, continuity, hours, scheduling, and provider acceptance, not just benefit eligibility.

The rural–urban extension of the Oregon survey and interviews suggests that considerable coverage benefits can arise under rural constraints, even as those constraints persist. This finding permits a constructive comparative question: which parts of insurance protection can be achieved immediately through payment coverage, and which require complementary investment in clinical supply? The Oregon design estimates outcomes under the observed capacity conditions; the answer for a different jurisdiction requires new evidence about its providers and population.

Health outcomes and longer horizons

The clinical evidence also carries a general lesson about temporal inference. Access and expense changes can appear within months. Detection and initiation of treatment may follow. Biological changes, complication reductions, and mortality can require substantially longer observation or different sample sizes. These are plausible sequences, not a causal guarantee that all additional use will eventually improve health. The correct research strategy is to identify intermediate measures, prespecify clinically meaningful endpoints, and maintain sufficient follow-up to estimate them.

The clinical-outcomes article supplies unusually concrete limitations for interpreting physical results. It shows that an insurance intervention can improve specific mental-health outcomes while leaving several population averages inconclusive. The absence of a significant two-year average effect should neither be erased nor generalized into a lifetime assertion. Future studies need sufficient power for clinically important effect sizes, complete follow-up or defensible missing-data assumptions, and explicit control of multiple comparisons where many endpoints are examined.

Economics of public and implicit insurance

The financial evidence connects directly to the allocation of social costs. Uninsured people do not necessarily receive zero care: emergency providers and hospitals may deliver services without receiving full payment from patients. Public insurance can therefore change the payer and the timing and certainty of receipts in addition to changing utilization. The JPE valuation study (Finkelstein, Hendren, & Luttmer, 2019) formalized the relevance of uncompensated care in interpreting Oregon's fiscal spending. Its estimates require a counterfactual account of who would pay for care without Medicaid. That counterfactual matters wherever hospitals, charities, governments, and households share the burden of uninsured treatment.

The choice of social welfare metric must therefore be explicit. Gross budget costs, incremental real resources, patients' willingness to pay, protection from tail risk, provider cash flows, and effects on tax-funded or employer-funded systems are related quantities but cannot be collapsed into one reported dollar. A reform judged solely by increases in utilization could look expensive while omitting financial risk transferred from households. A reform judged solely by reduced household debt could ignore changes in medical resources or the incidence of public financing. Both would be incomplete economic evaluations.

Comparability with later Medicaid expansions

The Oregon experiment's estimand is localized to its compliers under its 2008 institutional rules. Later reforms—including ACA-era expansions—operate with larger and different populations, more sustained program exposure, and sometimes changed provider incentives. The quasi-experimental expansion evidence (Wherry & Miller, 2016) contributes evidence about a different treatment contrast, but its comparison across expansion and non-expansion areas depends on assumptions distinct from the Oregon lottery's random assignment. A direct numerical comparison between studies must preserve the targeted population, conditional trends or assignment mechanism, outcome type, and follow-up window.

Oregon's results can serve as a methodological reference for international public insurance research: identify a credible source of assignment variation, measure actual enrollment rather than entitlement alone, observe both financial and service outcomes, and specify the conditions needed to transport an effect. The experiment cannot supply an international coefficient for the health benefit of insurance or an automatic answer for schemes with different coverage, delivery, or cost-sharing structures. Those require evidence generated in the relevant country and period.

Residual research questions

Several questions remain scientifically open within the published Oregon study family and its extensions. How do longer periods of continuous coverage alter chronic-disease control, mental-health trajectories, and avoidable complications? What proportion of additional utilization is preventive, beneficially diagnostic, substitutive, or low-value? How do changes in provider supply modify the use of emergency departments and ambulatory care? How much of uncompensated care's cost is borne by providers, paying patients, local public budgets, or other actors? How robust are welfare estimates when willingness to pay and counterfactual payment burdens are defined differently? Which mechanisms of enrollment friction or household assistance explain the observed response of already-eligible children?

These questions specify research needs rather than promising predetermined answers. Some require longer administrative records; others require linked clinical data, natural experiments in benefit design or provider capacity, formal welfare comparisons, or ethically implemented field studies. The public experimental data catalogue provides a documented path for authorized replication of certain original results, subject to its use conditions. Further research should disclose the treatment definition, sample scope, assumptions, and estimation uncertainty before drawing generalized insurance policy conclusions.

Conclusion

The Oregon Medicaid lottery is a landmark source of causal evidence on health insurance because an oversubscribed administrative program generated randomized access to an opportunity to apply for coverage. Its identification strength begins with lottery assignment, while the meaningful effect of actual Medicaid coverage requires attention to incomplete take-up, household-level assignment features, instrumental-variable assumptions, and the complier population. The first-stage increase of roughly one-quarter in the probability of coverage is essential to understanding every subsequent treatment-effect estimate.

The combined empirical findings are substantial and internally coherent. Medicaid increased outpatient care, prescriptions, and selected preventive services; the emergency-department study demonstrated about 0.41 additional visits per covered person over approximately eighteen months. Financial outcomes showed reduced out-of-pocket burden, medical debt, and catastrophic spending risk. Measured clinical outcomes included lower depression-screening prevalence and greater detection and medication treatment of diabetes, together with no statistically significant short-run changes in average blood pressure, cholesterol, or glycated hemoglobin. Administrative research found no statistically significant early employment or earnings effect. Further analysis documented household enrollment spillovers and a substantial transfer to providers implicit in Medicaid financing. The specific units, intervals, populations, and limits of these results are part of the findings themselves.

The experiment establishes no universal verdict that insurance can be assessed through a single endpoint. Its findings describe a public insurance institution operating across medical access, household financial security, provider payments, and health outcomes whose response times and measurement systems differ. The proper scientific treatment preserves those dimensions while requiring a common causal grammar. The question of whether a program's social benefits exceed its costs then calls for a clearly specified welfare criterion and transparent assumptions about who ultimately pays and who gains.

For international insurance science, Oregon's enduring contribution is methodological as well as substantive. A policy-generated randomized offer, properly analyzed, can illuminate the effects of coverage that observational comparisons often confound. The strength of that evidence does not automatically extend beyond its compliers, historical rules, or observation windows. A rigorous comparative application combines credible identification with exact documentation of enrollment, benefit design, provider supply, financial incidence, and outcome timing. This dossier's original synthesis establishes that analytical correspondence; it does not claim a new theory or independent patient-level empirical discovery.

Appendix: Estimands, Numerical Checks, and Evidence Boundaries

A. Definitions and reproducible arithmetic

  1. Lottery offer (Z): randomized permission to submit an application for OHP Standard, not a guaranteed insurance contract. Medicaid receipt (D): the study-specific actual coverage indicator, potentially observed over dates and subject to enrollment, renewal, and loss of eligibility. Outcome (Y): a prespecified clinical, utilization, financial, or economic variable with its own measurement window.
  2. Intention to treat: difference in mean outcome by Z under the implemented randomization design. First stage: difference in D by Z. Local average treatment effect: their ratio under relevance, random assignment, monotonicity, and exclusion assumptions, with household clustering and other original design adjustments handled in the actual regression.
  3. NEJM 2013 insurance first stage: 0.241, 95% confidence interval 0.223 to 0.259. The reciprocal 1 / 0.241 = 4.1494 is only an illustrative Wald scaling factor.
  4. ED administrative example: 0.41 / 1.02 = 0.40196, approximately 40.2% increase in visits over the study period; 0.07 / 0.345 = 0.20290, approximately 20.3% increase in probability of any ED visit relative to its control-group level.
  5. Depression screening: 9.15 / 30 = 0.305, approximately 30.5% relative reduction using a rounded control prevalence of 30%. The exact effect is −9.15 percentage points and has the confidence interval reported in the clinical study.
  6. Welfare incidence in the 2019 model: approximately 40% incremental resource cost plus approximately 60% transfers to providers of uncompensated care, each stated relative to gross spending within that study's cost definitions. This is an analytical decomposition attributable to its authors, not a directly observed cash-flow identity for every provider.

B. Interpretation safeguards

An absence of statistical significance permits effects within the reported interval. Administrative ED visit counts and survey self-reports have different ascertainment properties. A physician diagnosis is an observed diagnostic event, not a new disease case. A positive depression screen is not interchangeable with clinical psychiatric diagnosis. Household spillovers and effective insurance duration require their own definitions. Exact enrollment and allocation figures depend on whether one counts reservation-list individuals, selected registrants, or members of affected households. The physical-health study's 12,229 respondents and ED analysis's approximately 25,000 individuals are distinct outcome samples. One should not annualize the 0.41 ED effect by assertion or attach it to the first-year survey results.

C. Research integrity and access statement

The cited source family includes peer-reviewed original papers, authoritative bibliographic records, the original study's public data and institutional documentation, and established causal methodology. Publisher pages, PubMed records, NBER materials, and accessible full-text mirrors were used for documentary verification as recorded in EVIDENCE.json. No original microdata regression was run. Some complete journal texts and online supplementary appendices were not accessible in the authoring environment; no full-text assessment is claimed for those restricted items. No external peer review, prospective clinical intervention, national legal audit, independent systematic literature search, or public release is represented as completed. All causal and numerical claims are restricted to their identified publication and study period.

References