Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Natural Experiments in Insurance Economics

Abstract. Natural experiments are a central means of evaluating causal questions in insurance economics when insurance purchase, risk exposure, loss reporting, and insurer behavior are jointly determined. Coverage contrasts constructed from voluntary enrollment are typically confounded by health, risk aversion, income, information, and institutional access; even a precisely measured association between insurance and an outcome may fail to identify a policy-relevant causal effect. This monograph examines the potential-outcomes basis of identification and the institutional circumstances under which lotteries, eligibility thresholds, policy changes, geographically differentiated reforms, and natural hazard shocks provide defensible research designs. It derives the distinction between offer effects and coverage effects, explains the local interpretation of instrumental-variable and discontinuity estimates, and critically assesses difference-in-differences designs under staggered treatment adoption, anticipation, endogenous selection, and interference. Evidence from the Oregon Medicaid lottery, Medicare eligibility and introduction, American Medicaid policy expansions, rainfall and agricultural insurance experiments in India and Ghana, China's agricultural insurance rollout, and United States flood-insurance responses shows that insurance changes financial protection, care utilization, household investment, purchasing behavior, and, in some designs, health outcomes. The magnitude and direction of effects depend on treatment definition, time horizon, population, provider and market responses, and the outcome used to represent welfare. The original contribution is a source-grounded integration of the economic estimand, identification assumptions, observable insurance mechanism, and limits of transportability across contracts and jurisdictions. The paper presents methodological derivations and analytical interpretations, without claiming a new theory, original causal estimate, unpublished dataset, or systematic meta-analysis.

Keywords: Natural experiments; insurance economics; causal identification; potential outcomes; randomized eligibility; encouragement design; intention to treat; instrumental variables; local average treatment effect; regression discontinuity; difference-in-differences; staggered adoption; policy evaluation; Medicaid; Medicare; agricultural insurance; rainfall index insurance; flood insurance; financial protection; risk sharing; adverse selection; spillovers; external validity; welfare economics; causal inference; research reproducibility.

Introduction

Insurance changes the allocation of risk across states of the world, households, firms, and institutions. Its economic consequences also arise before a covered loss materializes: insured people may alter consumption, treatment, borrowing, precaution, work, or investment. The fundamental empirical difficulty is that the decision to obtain coverage already reflects determinants of these outcomes. A household expecting high medical expenditure may be unusually likely to insure; a farmer with access to credit and reliable distribution networks may be more able to purchase an index product; an insurer may expand supply in areas with improving conditions. Comparing insured and uninsured units therefore combines the influence of insurance with the conditions that led units to insure.

Natural-experimental research seeks variation in an insurance-relevant treatment that can be interpreted independently of the unobserved potential outcomes under an explicit research design. State lotteries, administrative age thresholds, geographically differentiated legislation, and plausibly external changes in coverage opportunities can make such variation available. The quality of identification depends on institutional details, not the rhetorical label “natural experiment.” The researcher must establish what changed, why it changed for some units and not others, which outcome is observed, and which counterfactual is credible.

The central claim of this monograph is that causal evidence in insurance economics is meaningful only when four objects are kept aligned: the policy or contractual intervention, the population whose exposure changes, the outcome and time horizon, and the parameter actually identified by the design. This statement organizes an analytical reading of existing research rather than establishing a novel formal theory. A randomized Medicaid offer identifies a different effect from the expansion of universal social insurance; a discontinuity at a statutory age threshold identifies a local effect; and a flood shock that raises policy demand does not by itself measure the welfare value of the policy purchased.

The conceptual basis is the modern potential-outcomes literature, including Imbens and Angrist's identification of local treatment effects (1994) and the wider causal inference framework of Hernán and Robins (2020). The empirical corpus is international but intentionally differentiated: strong identification for one population and product does not automatically transfer to another. Source quality, timing, institutional jurisdiction, assignment integrity, measurement, and reproducibility are treated as part of the economic argument.

The Economic Object of Causal Insurance Research

An insurance intervention can be defined as an offer to enroll, a reduction in premium, a change in statutory eligibility, an indemnity entitlement, a change in the insurer's creditworthiness, or a complete shift in market organization. These interventions are economically distinct. An offer may not lead to enrollment; enrollment may provide different contractual benefits; reimbursement rules may change provider incentives; a public subsidy may change both coverage and the price paid by people who were already insured. An empirical estimate that moves among these objects without redefinition becomes ambiguous.

Let the outcome Y be measured at a specified horizon and let D indicate actual coverage. The effect of insurance on an individual is conceptually Y(1) - Y(0), where the two potential outcomes describe the same individual with and without the relevant coverage arrangement. Only one of those states is observed at a given time. The average treatment effect is E[Y(1) - Y(0)] for an explicitly defined population. Other parameters, including the average effect among those who enrolled or among those induced to enroll by a specific offer, need separate identification.

For insurance economics, Y may be medical spending, unpaid debt, clinical health, agricultural investment, harvested output, insurance purchases, avoided distress sales, subjective confidence, or expected utility. A reduction in uninsured financial risk can improve welfare even where a short-run average clinical indicator does not change. Conversely, higher use of insured services does not itself establish a welfare gain when services vary in marginal benefit and social cost. Financial protection and health production are related through a mechanism and remain separate measurable outcomes.

Treatment versions demand particular care. Medicaid in a United States state and year represents a defined benefit package, provider payment scheme, eligibility and renewal rules, and administrative environment. Private rainfall index insurance in rural India represents a contingent payment based on an index that may differ from a farmer's realized crop loss. Public old-age health insurance alters both individual price exposure and the structure of a provider market. In every case, the estimand concerns an implemented institution, not an abstract homogeneous binary product.

A causal study therefore requires an explicit causal question in the form: What is the effect of an implementable insurance change, relative to an available counterfactual, for which population and time interval? The practical consequence is to record offers, actual coverage, contract terms, prices, claim performance, and institutional supply independently rather than calling all of them “insurance exposure.”

Potential Outcomes and Identification

For unit i, define Y_i(d) as the potential outcome under coverage d, where d belongs to {0,1}. The observed outcome equals D_i Y_i(1) + (1-D_i)Y_i(0), assuming the treatment description captures the relevant regime. The unobserved counterfactual creates the fundamental problem of causal inference. To identify E[Y(1)-Y(0)] by a simple difference in observed means, assignment must be independent of both potential outcomes, or a defensible conditional independence condition must hold given measured covariates.

Voluntary insurance purchase rarely satisfies unconditional independence. Expected loss, beliefs, risk preferences, medical needs, liquidity, ability to navigate enrollment, network access, and prior exposure may influence coverage and Y. Conditioning on observed income or prior utilization does not resolve unmeasured need or behavior by assumption. Insurance markets also create simultaneous causation: insurer pricing responds to predicted risk while purchase responds to price and risk. Selection and moral hazard may operate together.

A research design identifies a parameter through an assignment mechanism. If Z is randomized, E[Y|Z=1]-E[Y|Z=0] identifies the effect of being assigned the intervention, subject to consistent outcome measurement, follow-up and interference conditions. If a legal cutoff determines eligibility, the comparison must establish continuity of potential outcomes near the threshold. If one jurisdiction changes policy while another does not, the unobserved outcome trend in the treated jurisdiction must be represented credibly by the chosen control. These are different justifications for a causal interpretation, and their plausibility must be audited separately.

The potential-outcomes approach to causal inference developed by Hernán and Robins (2020) foregrounds carefully specified interventions, exchangeability, positivity, and consistency. In insurance, positivity may fail where coverage is legally compulsory or unavailable to a population; consistency may fail when the policy differs in claims adjudication or provider networks across settings. A formally correct regression cannot repair an intervention that was never defined precisely enough to correspond to the observed data.

Causal identification is also narrower than economic evaluation. Identifying a change in claims or spending provides a causal component of the welfare calculation, while valuing risk reduction, marginal utility, redistribution, fiscal cost, and insurer adaptation requires further assumptions. Research quality improves when these empirical and normative steps are presented openly rather than collapsed into a single “impact.”

Randomized Coverage Offers and Imperfect Compliance

Many insurance interventions randomize an encouragement, subsidy, enrollment assistance, or eligibility opportunity. Randomization gives credible exogenous assignment of Z, but the actual coverage D can differ between those offered and not offered. Some offer recipients decline; some nonrecipients obtain coverage through other channels. This creates the central distinction between intention-to-treat effects and effects of realized coverage.

The intention-to-treat parameter is ITT_Y = E[Y|Z=1]-E[Y|Z=0]. It describes the effect of assigning the offer or eligibility opportunity. Its interpretation follows the policy actually administered: one randomly selected mailing may increase information and trust as well as enrollment. In a program evaluated for its operational performance, ITT is often the more directly reproducible policy effect because it does not presume universal response.

The first-stage effect is ITT_D = E[D|Z=1]-E[D|Z=0]. An offer can increase coverage materially without moving enrollment to 100 percent. In the Oregon first-year experiment, Finkelstein et al. (2012) report that lottery selection raised insurance coverage by approximately 25 percentage points in the first year. That number is a treatment-compliance result rather than the population's baseline level of coverage. It helps explain why an effect per covered person, obtained by scaling the intention-to-treat contrast, can be much larger than the reduced-form effect of being offered the opportunity.

Randomization does not eliminate the need for design checks. The assignment unit may be a household or individual, outcome data may be linked with errors, and survey respondents may differ from nonrespondents. Researchers must account for the level at which offers were randomized and the sample on which outcomes were measured. A final reported treatment effect can also change with the follow-up horizon because insurance enrollment, utilization, provider behavior, and long-run health accumulation evolve on different schedules.

Randomized offers identify the effect of the offer directly. Effects of actual coverage require additional assumptions addressed below. The distinction matters across international insurance contexts: providing a rainfall policy for free is not the same treatment as offering one at actuarially fair or subsidized price; each policy induces different participation and possibly distinct expectations about claims.

Instrumental Variables and Local Treatment Effects

An instrument Z can identify a coverage effect when it shifts D, is independent of relevant potential outcomes, affects Y through the specified coverage channel, and satisfies appropriate assumptions about compliance. Under a binary assignment and binary treatment with monotonicity, the Wald estimand is LATE = ITT_Y / ITT_D. In ordinary language, it measures the average effect for compliers whose treatment status changes because of the instrument. It is not generally the average effect for all insured people or the entire eligible population.

Imbens and Angrist (1994) explain the conditions in their local average treatment effect analysis. Relevance is empirical: the eligibility offer must shift coverage sufficiently. Independence relies on the assignment process. Exclusion is substantive: the assignment cannot affect the measured outcome through a pathway outside the treatment as defined. Monotonicity rules out a class of people whose treatment moves systematically in the opposite direction to the instrument.

Insurance instruments often raise difficult exclusion questions. A lottery win may change outreach, family enrollment, information, or beliefs independently of the winner's eventual policy. A premium subsidy may affect liquidity even for individuals who do not purchase. A legal eligibility threshold may trigger retirement or other benefits at roughly the same age. In such cases the effect of the policy package remains meaningful, but attributing the entire reduced-form effect to binary coverage requires a defensible exclusion claim.

A weak first stage creates unstable scaled estimates; instrument strength is therefore part of research substance, not a technical footnote. If treatment take-up is concentrated among those with urgent medical need, the complier effect is especially relevant to the particular offer population and cannot be treated as a universal insurance multiplier. A researcher should report ITT, first stage, coverage-effect estimate when justified, uncertainty, and the scope of the complier population.

The economic meaning of LATE is important for public policy. A subsidy aimed at reluctant purchasers may induce a different group from the group induced by automatic enrollment, and the marginal benefits may differ because of risk, preferences, and access. The same overall enrollment gain need not produce the same utilization or welfare response when achieved through different interventions.

Policy Discontinuities and Regression Discontinuity

An administrative eligibility rule can produce a sharp change in access at a known threshold. Let R be the running variable and c the cutoff. A sharp regression-discontinuity estimand compares the limits of E[Y|R=r] as r approaches c from above and below. It has a causal interpretation when the conditional expectations of untreated potential outcomes are continuous at c and units cannot precisely sort across the boundary in a way that violates comparability. A fuzzy discontinuity uses the jump in treatment probability as an instrument and has an analogous local complier interpretation.

The regression-discontinuity methodological review by Lee and Lemieux (2010) provides the relevant logic, threats, and estimation guidance. In insurance, examples include age eligibility for a public program, income-linked contribution and subsidy thresholds, and administrative price or risk classifications. The statutory rule is only the starting point. Eligibility may change at a birthday while care-seeking, retirement, benefit claiming, and enrollment processing also change. Multiple contemporaneous discontinuities require a composite-policy interpretation or additional tests.

Card, Dobkin, and Maestas study the coverage discontinuity at Medicare age, showing that age 65 produces large increases in health insurance coverage and related use of medical services in the United States. Their analysis also finds distributional differences by race and education and variation in hospital admissions, which warns against explaining all changes as a simple binary uninsured-to-insured transition. The generosity and type of existing coverage can change at the same threshold.

A regression-discontinuity design is local in at least two senses: it applies to units close to the threshold and to the institutional arrangements prevailing around that threshold. Estimated effects for people immediately turning 65 need not describe substantially younger adults. A design based on an income threshold may also fail to transport to households with much lower income or different liquidity constraints.

Good practice includes showing treatment and outcome changes graphically, justifying bandwidth and polynomial choices, inspecting density or manipulation around the cutoff, checking predetermined covariate balance, testing alternative windows, and using inference methods appropriate for local estimation. A discontinuity in a recorded outcome is evidence of a policy effect only after alternative discontinuities and measurement artifacts have been examined.

Difference-in-Differences and Insurance Reforms

Policy reforms frequently change coverage eligibility, premium support, or insurer obligations in one region at one time while leaving other regions unchanged. Difference-in-differences (DiD) compares changes rather than levels. In a simple two-period setting, the estimate is (Y_treated,post - Y_treated,pre) - (Y_control,post - Y_control,pre). Its causal interpretation requires that, absent the policy, the treated and control units would have experienced the same average change in potential outcomes, subject to the design's conditioning assumptions.

The parallel-trends condition concerns unobserved counterfactual outcomes and cannot be proved by a visually attractive pretreatment graph. Multiple preperiods provide useful diagnostics, yet absence of statistically significant pretrends is not proof of the identifying restriction. Concurrent changes in provider capacity, wages, economic conditions, social programs, or insurer rules can violate the comparison. Spillovers to supposedly untreated jurisdictions can weaken the contrast even when state policy borders are clear.

Sommers, Baicker, and Epstein (2012) compare earlier United States state Medicaid expansions with neighboring states in their mortality and access study. The study reports an association between expansions and mortality and access changes under a difference-in-differences style design. Its interpretation depends on the plausibility of the comparator-state trajectories and available controls, and it should not be equated with the randomized Oregon findings. Distinct research designs, populations, observation windows, and outcomes can yield different evidence without logical contradiction.

Frean, Gruber, and Sommers (2016) use a triple-difference strategy with income, geography, and time variation to study Affordable Care Act coverage effects. This design addresses several interlocking policy mechanisms: Medicaid expansion, marketplace subsidies, and the mandate. The authors report that the model accounts for a substantial portion of observed 2014–2015 coverage gains. Such an estimate concerns the modeled bundle of policy provisions and their realized implementation, not the effect of a single uniform insurance contract.

DiD is especially valuable when policy exposure occurs at a collective level, including provinces, insurers, or local markets. The corresponding statistical inference must respect common shocks and dependence at the level of policy assignment. A large number of individuals in a small number of treated jurisdictions does not itself create a large number of independent policy experiments.

Staggered Adoption, Heterogeneous Effects, and Event Studies

Insurance policies are seldom adopted in a single synchronized national experiment. States expand at different dates, municipalities revise mandates, and insurers introduce products in waves. Conventional two-way fixed-effects (TWFE) regressions with unit and time indicators can combine comparisons that have different causal interpretations when treatment effects vary across cohorts or with exposure duration. Already-treated groups may become controls for later-treated groups, even though their own outcomes continue changing because of the policy.

The staggered-adoption decomposition developed by Goodman-Bacon (2021) shows why a TWFE coefficient can mix two-by-two comparisons and becomes sensitive to heterogeneous effects. Callaway and Sant'Anna (2021) derive group-time treatment effects for multiple periods, while Sun and Abraham (2021) analyze dynamic event-study contamination. Their contributions are directly relevant to insurance reforms with staggered Medicaid adoption, phased agricultural coverage, or sequential compulsory insurance mandates.

The proper response begins with the estimand. A researcher should decide whether to report cohort-specific effects, calendar-time effects, years-since-adoption effects, or a well-defined aggregation. A modern estimator does not rescue a reform whose implementation is endogenous to emerging losses or political responses; design credibility remains a substantive question. Conditioning variables should also be chosen for their preintervention status, since adjusting for policy-induced mediators may remove part of the treatment effect.

The applied guidance for staggered designs by Freedman et al. (2023) emphasizes comparisons appropriate to treatment timing and practical diagnostics. In insurance applications, already-insured populations, providers, and firms may experience gradual adaptation. One-year and five-year estimates can differ because the intervention has a dynamic causal pathway, not simply because the statistical estimator is unstable.

Premium levels and subsidies are sometimes continuous rather than binary. Callaway, Goodman-Bacon, and Sant'Anna (2024) describe the continuous-treatment DiD problem: even apparently comparable comparisons across different treatment intensities can be distorted by selection and effect heterogeneity. A one-percentage-point change in premium subsidy is not guaranteed to have a constant effect across all initial subsidy levels or liquidity groups. Continuous exposure demands explicit dose definitions and stronger assumptions where marginal dose effects are interpreted causally.

Natural Hazards, Insurance Choice, and Belief Updating

Natural disasters can provide unexpected variation in perceived risk, losses, liquidity, public attention, and institutional behavior. They are attractive for studying insurance because they occur outside ordinary purchase decisions, but their causal pathways are multiple. A flood can destroy wealth, change expected loss probabilities, alter insurer supply, prompt government relief, change media coverage, and induce future mitigation. Simply attributing higher insurance purchase after a flood to one information mechanism is stronger than the raw comparison warrants.

Gallagher (2014) studies flood-insurance take-up after major floods using a national panel of United States floods and policy purchases. He finds a rise in take-up after flooding followed by a gradual return toward baseline; residents in nonflooded communities sharing a television media market also respond, though less strongly. These patterns support a careful account of learning, salience, and information decay. They measure demand responses; they do not directly identify the contract's effect on indemnity adequacy, future property value, or welfare.

Insurance economists should separate three questions. Did the shock alter beliefs? Did altered beliefs change purchase? Did the resulting contract improve the household's welfare after accounting for premium, basis risk, claims payment, and alternatives? The first two can sometimes be approached through regional panels and media spillovers. The third requires realized protection outcomes and a counterfactual for insurance itself.

Hazard intensity and policy risk classifications can create spatial discontinuities. Yet boundaries in flood maps may coincide with topography, income, construction standards, or mortgage rules, which confound simple comparisons of buildings inside and outside a zone. Prior development and anticipatory relocation may also determine exposure. Credible natural-experiment claims must distinguish external hazard occurrence from endogenous location and contract choice.

This research area illustrates the broader discipline of insurance identification: an exogenous environmental event is a useful candidate source of variation only after the path from that event to the proposed causal treatment has been specified. When several channels are inseparable, the result should be labeled as the total effect of the compound shock rather than a precise estimate of an isolated insurance mechanism.

The Oregon Medicaid Lottery

The Oregon Health Insurance Experiment is unusually informative because access to a limited Medicaid expansion was allocated through a public lottery in 2008. Randomization of the opportunity to apply for coverage generated a credible comparison for low-income uninsured adults who participated in the waiting-list process. It did not randomly assign universal Medicaid coverage to every resident of Oregon. Selection into the applicant pool, enrollment compliance, and the specific benefits and provider environment shape its policy interpretation.

Finkelstein et al. (2012), in first-year evidence from Oregon, find that lottery selection increased insurance coverage, health care use, and financial security. They report lower out-of-pocket spending and medical debt together with improved self-reported health. The first-year coverage first stage of roughly 25 percentage points is central: the direct experimental contrast identifies the offer effect, while their instrumental estimates translate it into the effect of coverage under additional assumptions.

Baicker et al. (2013) provide two-year clinical outcome evidence. The study's measured outcomes included blood pressure, cholesterol, glycated hemoglobin, depression screening, use, and out-of-pocket expenditures. It found greater diabetes detection and treatment, reduced depression, and a strong decline in catastrophic medical spending; it did not find statistically significant improvements in several measured physical-health biomarkers within two years. A nonsignificant mean effect is not a proof of zero effects for every individual, every longer horizon, or every clinical endpoint.

Taubman et al. (2014) use emergency-department records from approximately 25,000 lottery participants in the Portland area in their emergency-care utilization study. They estimate an increase of 0.41 emergency-department visits per person over roughly 18 months, about 40 percent relative to a 1.02-visit comparison mean. The evidence challenges the simple premise that insurance access necessarily reduces emergency-department demand by shifting all care to primary settings. Additional coverage can increase access to both types of care, depending on provider availability and patient needs.

The Oregon case is therefore a multidimensional test of insurance value. Stronger financial protection, greater utilization, more clinical detection, and heterogeneous measured health outcomes can coexist. A welfare evaluation should recognize the benefit of reduced financial uncertainty, the potential value of newly consumed care, the public cost of financing the program, and the horizon over which health effects may emerge. Treating any single one of these indicators as the entire policy result would discard meaningful evidence.

The experiment also illustrates how exposure changes other household members' behavior. Sacarny, Baicker, and Finkelstein (2020) show enrollment spillovers to children who were already eligible for Medicaid. They estimate that at three months, for every nine adults whose enrollment was induced by the lottery, approximately one additional eligible child enrolled. This is evidence for a channel through which assignment can affect broader household participation and a reason to define the unit and treatment pathway carefully when estimating direct adult coverage effects.

Medicare Eligibility and Historical Introduction

Medicare creates two distinct opportunities for quasi-experimental research: the individual eligibility shift around age 65 and the national introduction of the program in 1965. These should never be treated as interchangeable experiments. The former examines local changes for cohorts near a current age threshold, while the latter examines a historical transformation in financing, spending, risk protection, and the provider market.

Card, Dobkin, and Maestas (2008) analyze insurance coverage at age 65 and document sharp coverage changes associated with increased routine care and elective hospital procedures, as well as variation by socioeconomic group. This evidence is consistent with a substantial access effect but also reveals that Medicare changes both the extent and generosity of coverage for many people. A discontinuity in outcomes cannot automatically be credited only to movement from no insurance to some insurance.

In their subsequent study of mortality and Medicare, Card, Dobkin, and Maestas (2009) focus on patients admitted through emergency rooms for conditions treated as nondeferrable. They estimate a local reduction in short-term mortality around the eligibility threshold for this severe patient population, accompanied by changes in treatment intensity. The selection restriction to nondeferrable admissions is part of the identification logic, and the result is not an all-population mortality effect of universal coverage.

Finkelstein and McKnight (2008) examine the initial financial and mortality effects of Medicare's introduction. Their research reports substantial protection from out-of-pocket medical expenditure risk and no discernible first-decade effect on elderly mortality in the aggregate framework they examine. These results concern different periods, populations, outcomes, and methods from the hospital-admission discontinuity study. Contrasting them as simple opposing findings obscures the difference between financial protection, conditional survival after severe admission, and aggregate mortality over a decade.

Finkelstein (2007), analyzing aggregate spending after Medicare's introduction, finds market-wide effects on hospital spending and evidence consistent with changes in medical practice and technology adoption. This provides a crucial equilibrium warning: an insurance expansion can alter provider investment and medical practice in a manner that an individual-level comparison within a largely fixed market cannot capture.

Together, these Medicare studies show why policy evaluation must specify its unit: an individual near an eligibility cutoff, a person hospitalized for a severe condition, a birth cohort, or an entire medical market. The induced causal effects differ because the intervention operates at different levels of aggregation. Reliable research should retain these boundaries as part of the substantive conclusion.

Medicaid Expansions and Policy Bundles

Medicaid expansion studies offer the scale and institutional diversity absent from many lotteries but face the challenge that policy choices and implementation dates are not randomly assigned. Jurisdictions can adopt reforms because of fiscal circumstances, political decisions, health system capacity, or prior coverage trends. Credible designs must show that comparison groups represent treated jurisdictions' untreated trajectories after accounting for such factors.

Sommers, Baicker, and Epstein (2012) use historical expansions in Arizona, Maine, and New York to compare mortality and access before and after reform. The source supports an interpretation of expansion-associated outcome improvements in its studied contexts. It cannot isolate every concurrent contextual influence, and causal claims depend on the common-trend comparison. The appropriate conclusion retains the particular jurisdictions, period 1997–2007, outcomes, and research design.

For the Affordable Care Act, Medicaid eligibility expansion operated alongside marketplace subsidies, individual mandate provisions, outreach, and variation in administrative execution. Frean, Gruber, and Sommers (2016) investigate these mechanisms through income-state-time comparisons. Their coverage findings include spillover enrollment among people previously eligible for Medicaid; thus, measured program participation can rise beyond the newly eligible statutory population. Claims about the “effect of the ACA” must identify whether they concern Medicaid enrollment, total insurance coverage, private crowd-out, or broader health and financial outcomes.

Subsequent evidence extends the time horizon. Aboulafia, Gruber, and Sommers (2025), in a longer ACA policy assessment, examine varying coverage effects from 2013 through 2023 and use time, state, and income variation to distinguish Medicaid and marketplace-related contributions. Their findings concern the specific financing and legislative context of the United States across changing administrations, and the empirical contrasts are not portable as universal coefficients for other insurance systems.

Where multiple reforms coexist, a triple-difference design can improve specificity only if the added comparison is itself credible. Comparing income bands across states and time may reduce some confounding but introduce compositional change, differential economic trends, or indirect exposure. A numerical decomposition of observed coverage gains into policy mechanisms is a model-dependent result and must be reported as such.

The policy lesson is methodological: an insurance expansion is a package of eligibility rules, prices, enrollment infrastructure, benefit design, provider supply, and fiscal transfers. A causal design that captures the whole package is valuable even when separate pathway effects remain unresolved. The name of the package should follow the actual intervention measured.

Agricultural Insurance in India and Ghana

Agricultural insurance provides a different economic environment from medical insurance. Production is seasonal, losses are often spatially correlated, payouts may depend on rainfall or vegetation indices, and informal transfers and credit mediate risk exposure. Insurance may affect farm investment before a loss occurs. A causal study must therefore measure both purchase behavior and the productive decisions that follow protection.

Cole, Giné, Tobacman, Topalova, Townsend, and Vickery (2013) use randomized rainfall-insurance marketing experiments in rural India. Their findings indicate substantial price sensitivity while also identifying trust, liquidity, and limited salience as important constraints on adoption. Randomized price offers support a causal interpretation of specific demand responses, yet consumers' actual claims experience and insurer credibility can differ across regions and product generations.

The purchase experiment is an especially clear case of the distinction between willingness to pay and insurance value. Low take-up may reflect an actuarially unattractive contract, uncertainty about claims, lack of cash at the time of sale, social preferences, or poor information. A demand response to a temporary premium subsidy identifies the effect of that subsidy under the experiment's terms. It is not automatically a full estimate of the welfare effects of coverage, and an increase in purchases need not establish that all buyers benefit equally.

Karlan, Osei, Osei-Akoto, and Udry (2014) conduct cash and rainfall-insurance experiments in northern Ghana. Random assignment to cash grants, insurance grants or offers, and combinations helps separate risk constraints from credit constraints in farmers' investment choices. Their evidence shows why insurance evaluation must ask whether additional borrowing or capital alone can reproduce the same behavior: a product that reduces downside exposure may change investment even where cash access has already improved.

Ghana and India provide complementary, not identical, experiments. The Indian study is especially informative about demand barriers and contract communication; the Ghanaian interventions test the production response to relaxing different constraints. Contract design, basis risk, trust in distribution, and agricultural timing influence both the first stage and outcome. International synthesis must preserve the agricultural system, intervention structure, and randomization unit rather than pooling the findings as one generic agricultural-insurance treatment.

Agricultural Insurance Rollouts in China

Policy-driven introductions of subsidized insurance may create geographically and temporally differentiated exposure even when no randomized offer was used. Such designs can identify larger-scale production effects, but the allocation of rollout timing can be related to local agricultural strategy, institutions, and economic trajectories. The counterfactual therefore rests heavily on the implementation record and the credibility of comparison groups.

Jing Cai (2016) studies agricultural insurance and household decisions using a natural experiment and difference-in-differences and triple-difference strategies in China. The published abstract reports that insurance provision increased production of the insured crop by 16 percent and borrowing by 29 percent. It also describes a change in savings composition rather than an overall change in savings, with different persistence across production, savings, and borrowing outcomes.

These numbers apply to Cai's empirical setting and treatment definition. They should not be converted into unconditional returns to insurance or universal responses for all crops and households. The study's insurance introduction can interact with credit supply, crop-specific policy support, and farmer expectations. A good interpretation records these constraints and, where possible, distinguishes direct indemnity protection from changes in institutional expectations generated by the policy.

China's program evidence complements the randomized studies in India and Ghana. A rollout can reveal effects at greater scale and under routine public administration; a randomized intervention can offer stronger assignment-based identification for the tested offer. Each has different external-validity and mechanism strengths. The economic researcher should compare the interventions and markets before comparing only their estimated percentages.

In agricultural insurance, a particularly important omitted variable is basis risk: a farmer can experience a serious loss without a corresponding index payout, or receive an index payout without an equally serious individual loss. Experimental results in one index calibration and geographical area require direct contract-level examination before being used to forecast welfare under a differently calibrated product.

Interference, Equilibrium, and Spillovers

The stable-unit-treatment-value assumption is often summarized by requiring that one unit's outcome depend on its own treatment and that treatment have a sufficiently consistent definition. Insurance frequently violates the no-interference component. A household's coverage can alter enrollment of eligible dependents, local insurance distribution, informal sharing, hospital utilization, market prices, and political support. Large insurance expansions can affect provider capacity and the fiscal position of the state.

Oregon's child-enrollment spillover evidence demonstrates an observable within-household effect of adult lottery assignment. The study estimates a modest and transitory response among children eligible regardless of the lottery. This observation makes household-linked outcomes substantively relevant to the program effect and shows how the direct treatment contrast may coexist with indirect effects beyond the primary applicant.

In a larger medical market, expanded insured demand can encourage provider entry, adoption of technologies, adjustment of prices, or changes in treatment practice. Finkelstein's aggregate Medicare spending evidence is important because these system-level responses were not simply scaled copies of individual-level coverage effects. Holding supply fixed when evaluating a system-wide reform may understate or mischaracterize the policy's longer-run consequences.

In agricultural risk sharing, formal insurance may alter relatives' transfers, savings, credit transactions, and community insurance networks. A design that randomizes offers across neighboring individuals can experience contamination if they pool payouts or share information. Cluster-level assignment, explicit network measurement, and spillover estimands may be required to characterize the total policy effect. The choice of assignment unit is then both a statistical and an economic design decision.

Interference is an analytical opportunity as well as a threat: it can reveal mechanisms that directly matter for welfare. A well-specified study should distinguish direct effects, indirect effects on untreated units, and general-equilibrium changes. A partial-equilibrium LATE from a small offer experiment should not be multiplied by the national target population without examining whether induced coverage and market responses would remain comparable.

Measurement, Administrative Data, and Inference

Causal inference requires reliable outcomes as much as credible assignment. Insurance data may include enrollment files, claims, electronic medical records, credit bureau balances, surveys, agricultural production records, premium invoices, public expenditure, and mortality registers. Each records a different portion of the underlying economic object. Claims data omit uninsured losses; policy files may record coverage without claim recoverability; surveys can capture informal transfers but suffer from recall and nonresponse.

The Oregon work illustrates the gains from multiple outcomes. The two-year clinical assessment measures biomarkers, depression, use, and catastrophic out-of-pocket expenditure, while the emergency-department analysis uses hospital records. Inferences should preserve the samples and measurement mechanisms of each study rather than assuming that all outcomes share one respondent population or observation window.

Attrition creates a particular problem where insurance changes the likelihood of being observed. Insured people may use providers with better administrative records; newly insured patients may be more likely to receive a diagnosis; treatment can influence questionnaire response or mobility. Researchers need to test differential follow-up, report missing-data strategies, and avoid interpreting greater detection as increased underlying disease prevalence without additional evidence.

Inference should match treatment assignment. A clustered state policy must not be analyzed as thousands of independent individual interventions when there are only a few policy units. A household lottery may require treatment assignment and outcome clustering at the household level. Multiple endpoints, interim analyses, and selected subgroups can produce false discoveries unless families of outcomes and uncertainty are handled transparently.

Insurance-related data also change measurement incentives. Insurers classify claims using contractual categories, health providers code diagnoses for reimbursement, and households may misreport damage or coverage. An insurance reform can alter the recording process itself. Measurement invariance is therefore a design-level hypothesis; administrative completeness cannot be assumed merely because the data are routinely collected.

Reproducibility demands that an empirical claim retain its article version, sample period, outcome definition, assignment method, model, and uncertainty interval. Where replication data are inaccessible, the appropriate record is verification of the published method and result, accompanied by a precise limitation. This monograph does not imply that the original microdata of the cited studies were reanalyzed.

Internal Validity and Transportability

Internal validity asks whether the design identifies the stated parameter in the studied setting. External validity asks whether that parameter helps predict outcomes for another population, contract, scale, or institution. These are connected by economic mechanisms but cannot be substituted for one another. A highly credible estimate for a specific complier group can be much less informative for universal coverage than a lower-precision estimate from a broad institutional rollout, depending on the policy question.

Transportability requires comparing baseline risk, benefit generosity, insurer performance, income and liquidity, treatment uptake, provider supply, and institutional credibility. In health insurance, an emergency department may respond differently where primary care is scarce. In rainfall index insurance, lower basis risk, different crops, or stronger claims trust may change both the first stage and subsequent investment. In public social insurance, taxes and financing can change the distribution of gains and market-wide behavior.

A robust source appraisal distinguishes uncertainty about an estimate from uncertainty about extrapolation. Confidence intervals quantify a specified form of within-study sampling uncertainty; they do not measure unobserved institutional differences between regions. Similarly, an event study with stable pretreatment trends may support the internal comparison while leaving generalization to a different legal system unresolved.

The starting point is to identify the policy-relevant target population and its feasible alternative insurance arrangements. A local age-65 estimate does not answer the same question as a nationwide expansion for younger adults. A short-run subsidized purchase experiment does not answer the same question as permanent, self-financed coverage. These are differences in the economic intervention rather than imperfections in the rhetoric of scientific confidence.

Effective synthesis reports transferable mechanisms and conditional conclusions. It may establish that reliable coverage can reduce exposure to catastrophic spending, that subsidies change demand, and that financing and market structure can alter utilization. It should avoid treating a reported percentage as a law of insurance economics.

Economic Welfare and the Value of Insurance

Insurance welfare encompasses the value of financial protection, changes in resource use, production responses, adverse selection, behavioral incentives, provider and insurer profits, and the incidence of financing. The welfare gain from reduced variance in consumption is conceptually independent of measured changes in average consumption. A person can benefit from security even if no claim occurs during the study and even if average health is initially unchanged.

A conventional expected-utility representation evaluates insurance by comparing expected utility with and without a contract. Let W denote wealth, L a stochastic loss, P the premium, and I(L) the contingent indemnity. Insured wealth is W - L - P + I(L), before considering taxes and broader market responses. The individual gains in the expected-utility benchmark if E[u(W-L-P+I(L))] exceeds E[u(W-L)] under specified preferences and available alternatives. This is standard economic reasoning, not a newly proposed valuation model.

An empirical coverage effect on out-of-pocket spending provides only some inputs to this comparison. The full valuation requires information about risk distribution, marginal utility of consumption, premium, taxes, benefit reliability, and any change in health production. Finkelstein and McKnight's study of early Medicare protection explicitly emphasizes the welfare significance of reducing out-of-pocket financial risk even when aggregate mortality impacts are not discernible. It is a concrete example of the difference between protection and realized physical outcomes.

The Oregon lottery evidence supports the interpretation that financial hardship can fall while care use rises. The Chinese agricultural rollout results indicate that coverage may enable productive changes, including higher insured-crop output and borrowing. Neither finding automatically yields a common monetary welfare unit. To compare policies, the analyst must specify whose gains count, how transfers are weighted, and whether insurer and fiscal costs are treated as real resources, transfers, or both.

Distribution matters especially when a subsidy is financed by general taxes and benefits concentrate among particular households. A policy may generate private benefits for purchasers and external benefits for other people while also reallocating public funds and affecting premiums for nonbeneficiaries. A transparent welfare assessment separates causal behavioral outcomes from these normative and accounting judgments. Causal identification is indispensable to the calculation but is not the whole calculation.

Comparative Evidence Across Insurance Markets

The strongest empirical comparison across the studies in this monograph is not an average effect size. Outcomes, scales, and interventions differ too much for a defensible common numerical meta-estimate. The comparison instead identifies what each design can establish. Oregon randomizes a chance to apply for Medicaid; Medicare designs exploit eligibility and historical introduction; state reforms compare changing policy regimes; agricultural experiments vary price, liquidity and risk protection; China's rollout supplies policy exposure; and flood events create changes in salience and belief.

The Oregon and Medicare results reveal insurance's multiple economic functions. Financial protection can improve without a detectable short-run effect on a given biomarker; care utilization can rise; some hospitalized patient groups can experience mortality changes; and aggregate coverage changes can reshape supply and technology. These are compatible effects of different interventions on different dimensions. A synthesis that removes outcome definitions destroys more information than it creates.

In agriculture, India demonstrates that demand depends on price and nonprice constraints; Ghana links risk protection and investment choices under experimental variation; China estimates rollout-associated production and borrowing changes. These studies together indicate that a credible insurance contract can act through financing constraints and production planning before a loss occurs. The strength of this inference depends on contract detail, rainfall and crop risk, market institutions, and the specific intervention.

Flood-insurance behavior adds a distinct mechanism: beliefs and information about rare risks change with experience. Gallagher's documented flood-demand response is evidence about learning and policy uptake rather than a general causal valuation of indemnity. This distinction should guide the design of subsequent studies measuring actual postloss outcomes, premium payment and claims execution.

Cross-country comparison requires an institutional map. Social insurance funded through taxation, regulated compulsory coverage, private actuarially priced contracts, and subsidized index insurance each allocate costs, risk, and claims discretion differently. Terms such as “treated,” “insured,” and “coverage” must be defined against each legal and administrative system. An estimator's algebra may travel internationally while its substantive parameter does not.

Original Analytical Synthesis

The original analytical contribution of this monograph is an integrated reading of insurance-specific causal evidence through the relationship between assignment, realized protection, economic mechanisms, and welfare. It is useful to organize this evaluation as a sequence of questions rather than as a new conceptual theory.

First, establish the implementable intervention. What exactly was randomized or changed by policy: eligibility, a letter, a subsidy, compulsory enrollment, contractual indemnity, or the supply-side regime? The answer specifies the treatment for which the design can potentially identify a causal effect. An “insurance effect” without this description is incomplete.

Second, follow the path from assignment to realized protection. Record the first stage from offer to coverage, contract characteristics from coverage to enforceable indemnity, and actual claims and services from contract to outcome. A design with credible assignment can still identify a broad policy package rather than the isolated effect of binary enrollment. The proper unit of causal interpretation must follow the observed assignment and compliance structure.

Third, define the econometric parameter before choosing a preferred statistic. An offer ITT, a complier LATE, a discontinuity-local effect, a group-time treatment effect under DiD, and a population average effect all answer different economic questions. Substituting one for another converts a legitimate estimate into an unsupported policy claim. Identification assumptions should appear alongside each inference, with documented alternative pathways and limitations.

Fourth, separate outcome channels. Demand, care use, risk exposure, debt, savings, borrowing, investment, production and health are economically linked but do not have the same welfare meaning. Increased claims can signal successful transfer, greater utilization, worse underlying losses, or changes in reporting. An index-insurance payout may respond to a weather index and fail to match a particular farmer's realized loss. The researcher should specify what the observable measures.

Fifth, assess whether the intervention changes the environment in which the effect was estimated. Provider entry, insurer capital, household spillovers, informal transfers and government fiscal response can make individual-level partial-equilibrium results diverge from a system-wide reform. This consideration is especially important for compulsory or universal insurance, where nonparticipants may also be affected.

Sixth, transport the conclusion only under defensible institutional comparisons. A specific result can support a general mechanism while leaving the magnitude for another country uncertain. The right scientific output is a conditional, policy-relevant statement naming the target population and contractual conditions. The original synthesis is an interpretive discipline for reading existing evidence; it establishes no novel formal model, doctrine, or priority claim.

This framework leads to a substantive conclusion: insurance evaluation is most reliable when empirical identification and institutional economics are joined at the level of the estimand. Econometric credibility alone cannot determine whether a measured coverage effect captures the relevant welfare channel; detailed economic descriptions without a counterfactual cannot establish causality. Source-grounded integration yields stronger, more precise conclusions than either can deliver in isolation.

Research Limitations and Open Questions

This monograph critically integrates published studies and official research records. It is not a preregistered systematic review; it reports neither exhaustive database coverage nor an original meta-analysis. It does not reestimate microdata, audit individual replication code, reconstruct raw administrative records, or independently confirm every published standard error. Publisher abstracts and author or research-center records were used according to their documented accessibility, and estimates are attributed to the underlying authors.

Many insurance-specific identification difficulties remain unresolved in the literature. How should researchers identify coverage effects when claim performance and insurer failure are endogenous to the same shock? How can randomized offers measure the long-run value of catastrophic protection when covered losses are rare? What are the best empirical designs for separating information effects from wealth shocks after disasters? How should staggered reforms be evaluated when premium changes have varying intensity, partial reversals, and cross-border spillovers?

Further research is needed on heterogeneous welfare effects for low-income households, providers, insurers, rural communities, and people with preexisting coverage. Better measurement of contract terms, payout speed, appeal outcomes, policy cancellation, informal insurance, and uncompensated losses would improve the link from assignment to realized protection. Network and equilibrium designs are important where households share risk and providers respond to aggregate insurance expansion.

Causal effects also depend on the time horizon. Financial relief may occur quickly, production choices may change over seasons, and chronic health or technology effects may emerge over years. Short-run nonsignificant findings should be interpreted with their precision and chosen endpoints; longer-run studies need to confront attrition, changing eligibility and treatment versions.

The international evidence remains uneven. Much of the most influential identification literature concerns health coverage in the United States or specific agricultural insurance interventions in lower-income regions. Generalizing across different insurance law, supervision, public finance and distribution systems requires new studies and explicit comparative institutional research. This monograph presents established conclusions where sources warrant them and preserves these open questions as a research agenda rather than inventing answers.

Conclusion

Natural experiments permit insurance economists to move from observed correlations toward identifiable causal effects when the assignment mechanism, counterfactual, treatment definition, and outcome measurement are sufficiently clear. The evidence shows that insurance can alter financial exposure, care use, demand, borrowing, investment and health outcomes through different channels. The economic content of these effects depends on what kind of insurance was provided, whom a policy induced to participate, which institutional actors responded, and how long outcomes were followed.

Randomized eligibility identifies the effect of an offer and, under additional conditions, a local effect of realized coverage. Regression discontinuity identifies changes close to administrative thresholds. Difference-in-differences measures policy contrasts under a defensible counterfactual trend, with staggered adoption demanding careful handling of treatment-effect heterogeneity. External risk events can illuminate beliefs and insurance purchase while requiring separate evidence to evaluate indemnity and welfare.

The reviewed primary studies demonstrate that financial protection is an independent benefit of insurance, that expanded coverage can increase medical utilization, that agricultural insurance can influence productive choices, and that learning from rare events can move insurance demand. Their effect sizes remain tied to the design, population, contractual institution, and time period in which they were estimated.

The resulting scientific conclusion is that causal inference in insurance economics is a joint task of research design, economic mechanism analysis, and institutional comparison. Sound conclusions preserve the identified parameter and its limits, connect outcomes to risk-bearing and welfare, and identify where further evidence is necessary. This monograph provides a completed critical synthesis of that evidence and methodology without establishing a new theory or asserting an original causal estimate.

References

  • Aboulafia, Gabriella, Jonathan Gruber, and Benjamin D. Sommers. (2025). “The Rise and Fall (and Rise) of the Affordable Care Act: Varying Impacts on Coverage Over Time and Place.” National Bureau of Economic Research, Working Paper 33615. NBER research record.

  • 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, with the Oregon Health Study Group. (2013). “The Oregon Experiment—Effects of Medicaid on Clinical Outcomes.” New England Journal of Medicine 368(18): 1713–1722. Published study.

  • Cai, Jing. (2016). “The Impact of Insurance Provision on Household Production and Financial Decisions.” American Economic Journal: Economic Policy 8(2): 44–88. Journal DOI.

  • Callaway, Brantly, Andrew Goodman-Bacon, and Pedro H. C. Sant'Anna. (2024). “Difference-in-Differences with a Continuous Treatment.” National Bureau of Economic Research, Working Paper 32117. NBER research record.

  • Callaway, Brantly, and Pedro H. C. Sant'Anna. (2021). “Difference-in-Differences with Multiple Time Periods.” Journal of Econometrics 225(2): 200–230. Journal DOI.

  • Card, David, Carlos Dobkin, and Nicole Maestas. (2008). “The Impact of Nearly Universal Insurance Coverage on Health Care Utilization: Evidence from Medicare.” American Economic Review 98(5): 2242–2258; earlier NBER Working Paper 10365. NBER research record.

  • Card, David, Carlos Dobkin, and Nicole Maestas. (2009). “Does Medicare Save Lives?” Quarterly Journal of Economics 124(2): 597–636; earlier NBER Working Paper 13668. NBER research record.

  • Cole, Shawn, Xavier Giné, Jeremy Tobacman, Petia Topalova, Robert Townsend, and James Vickery. (2013). “Barriers to Household Risk Management: Evidence from India.” American Economic Journal: Applied Economics 5(1): 104–135. Published study.

  • Finkelstein, Amy. (2007). “The Aggregate Effects of Health Insurance: Evidence from the Introduction of Medicare.” Quarterly Journal of Economics 122(1): 1–37; NBER Working Paper 11619. NBER research record.

  • Finkelstein, Amy, and Robin McKnight. (2008). “What Did Medicare Do? The Initial Impact of Medicare on Mortality and Out of Pocket Medical Spending.” Journal of Public Economics 92(7): 1644–1668; NBER Working Paper 11609. NBER research record.

  • Finkelstein, Amy, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gruber, Joseph P. Newhouse, Heidi Allen, and Katherine Baicker, with the Oregon Health Study Group. (2012). “The Oregon Health Insurance Experiment: Evidence from the First Year.” Quarterly Journal of Economics 127(3): 1057–1106; NBER Working Paper 17190. NBER research record.

  • Frean, Molly, Jonathan Gruber, and Benjamin D. Sommers. (2016). “Premium Subsidies, the Mandate, and Medicaid Expansion: Coverage Effects of the Affordable Care Act.” NBER Working Paper 22213, revised December 2016. NBER research record.

  • Freedman, Seth M., Alex Hollingsworth, Kosali I. Simon, Coady Wing, and Madeline Yozwiak. (2023). “Designing Difference in Difference Studies With Staggered Treatment Adoption: Key Concepts and Practical Guidelines.” NBER Working Paper 31842. NBER research record.

  • Gallagher, Justin. (2014). “Learning about an Infrequent Event: Evidence from Flood Insurance Take-Up in the United States.” American Economic Journal: Applied Economics 6(3): 206–233. Journal DOI.

  • Goodman-Bacon, Andrew. (2021). “Difference-in-Differences with Variation in Treatment Timing.” Journal of Econometrics 225(2): 254–277; earlier NBER Working Paper 25018. NBER research record.

  • Hernán, Miguel A., and James M. Robins. (2020). Causal Inference: What If. Boca Raton: Chapman & Hall/CRC, continuously updated online author edition. Author's book site.

  • Imbens, Guido W., and Joshua D. Angrist. (1994). “Identification and Estimation of Local Average Treatment Effects.” Econometrica 62(2): 467–475. University publication record.

  • Karlan, Dean, Robert Osei, Isaac Osei-Akoto, and Christopher Udry. (2014). “Agricultural Decisions after Relaxing Credit and Risk Constraints.” Quarterly Journal of Economics 129(2): 597–652. Journal study.

  • Lee, David S., and Thomas Lemieux. (2010). “Regression Discontinuity Designs in Economics.” Journal of Economic Literature 48(2): 281–355. Published review.

  • Sacarny, Adam, Katherine Baicker, and Amy Finkelstein. (2020). “Out of the Woodwork: Enrollment Spillovers in the Oregon Health Insurance Experiment.” NBER Working Paper 26871. NBER research record.

  • Sommers, Benjamin D., Katherine Baicker, and Arnold M. Epstein. (2012). “Mortality and Access to Care among Adults after State Medicaid Expansions.” New England Journal of Medicine 367(11): 1025–1034. Published study.

  • Sun, Liyang, and Sarah Abraham. (2021). “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects.” Journal of Econometrics 225(2): 175–199. Journal DOI.

  • 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. NIH author manuscript.