RESEARCH DOSSIER
Empirical Tests of Selection in Insurance Markets
Abstract. Empirical tests of selection in insurance markets ask whether the individuals purchasing more extensive insurance are systematically different in expected cost, private knowledge, preferences, or behavioral responses from those purchasing less coverage. Their central challenge is identification: a conditional association between coverage and subsequent claims is an observable equilibrium pattern, while adverse selection, advantageous selection, and moral hazard refer to distinct mechanisms. This monograph reconstructs the empirical logic of positive-correlation tests, conditional-independence approaches, structural demand-and-cost estimation, dynamic experience-rating methods, and investigations using subjective risk expectations and contract attributes. It compares documented evidence from French and other automobile markets, United States health and Medigap markets, long-term care insurance, life insurance, and United Kingdom annuities. The analysis establishes that the absence of a positive coverage–risk gradient is consistent with multidimensional private information; that a positive gradient may arise from selection, behavioral incentives, or institutional sorting; and that consumer welfare cannot be inferred from the sign of a coverage–claims covariance. An integrated evaluation framework separates the insured event from paid claims, the applicant population from the insured sample, measured from privately perceived risk, and selection tests from welfare measurement. Published studies show heterogeneous findings whose interpretation depends on their populations, contract menus, information sets, identification assumptions, and historical market arrangements. The contribution is a systematic synthesis and methodological audit of established research rather than a newly proposed theory or an original dataset-based estimate.
Keywords: Insurance selection; adverse selection; advantageous selection; asymmetric information; positive-correlation property; insurance coverage; expected loss; moral hazard; insurance demand; conditional independence; empirical identification; health insurance; automobile insurance; long-term care insurance; annuities; welfare analysis; underwriting; contract endogeneity.
Introduction
Selection is central to insurance because the premium and contract terms determine both the risks that enter a pool and the behavior that follows enrollment. A price set according to the expected claims of the current pool changes the composition of that pool when consumers respond differently to the price. When insurers and consumers possess different information about relevant risks, a contract menu also becomes an information mechanism. The importance of these propositions for economic theory is established in Akerlof's analysis of market quality and the Rothschild–Stiglitz insurance equilibrium analysis (Akerlof, 1970; Rothschild and Stiglitz, 1976).
Empirical insurance markets pose a different question: what evidence would distinguish one information mechanism from another? Observed claims can be higher among buyers of generous coverage because they expected higher risk before enrollment, because insurance altered behavior after enrollment, because one plan attracted consumers with particular preferences, because underwriting assigned different terms, or because the available records omit nonpurchasers. The econometric problem is therefore to determine what was known, by whom, at which decision date, and how that information affected contracts, participation, and outcomes.
The early empirical literature proposed and applied tests based on the relationship between contract generosity and realized insured events. Puelz and Snow (1994) reported evidence from collision insurance contracts, while Chiappori and Salanié (2000) reported a null finding for a French automobile sample. These findings differ in institutional context, measurements, and empirical design. They should be read as estimates of particular market mechanisms under particular maintained assumptions, rather than universal verdicts on insurance information.
Cohen and Siegelman (2010), in their comparative review of selection tests, document variation across lines and samples. Einav and Finkelstein (2011) organize the evolution of research from the canonical adverse-selection prediction toward comparative evidence and welfare analysis. This monograph develops the inferential consequences of those findings, uses original logical and statistical synthesis, and makes no claim of independently replicating proprietary insurance data or discovering a new named theoretical system.
The object of study is the empirical test, including its population, hypothesis, variable definitions, identifying assumptions, and interpretive reach. The surrounding equilibrium theory belongs principally to the companion scientific subject Adverse Selection in Insurance Markets. The present research follows the approved Insurance Economics dossier boundary by concentrating on what can be learned from actual insurance data.
Selection Hypotheses and Observables
The decision sequence and the meaning of selection
A minimal insurance timeline consists of an initial risk and information environment, an underwriting and premium offer, a household or firm's participation and contract choice, subsequent insured activity and losses, and finally recording and payment of claims. Empirical analysis becomes misleading when variables from different points in this sequence are treated as interchangeable. In particular, a loss following coverage may reflect a preexisting risk type and a response to the coverage itself.
Let C denote an ordered measure of contract generosity, such as a lower deductible or a higher indemnity limit; X the information legitimately observed and used by the insurer at contracting; U the information or preferences retained privately by the prospective policyholder; Y(0) an outcome under a designated baseline contract; and Y(c) the outcome under coverage c. These definitions are analytic notation, not claims that a general empirical dataset contains all variables. One natural conditional association is
where S indicates appearance in the analytic sample. A positive value of this difference is a fact about observed outcomes in an included population, assuming data and estimation are valid. It does not by itself identify the sign of a private-information effect. It may also differ from the corresponding difference in losses among everyone offered insurance.
Adverse selection, in the conventional demand-side sense, means that those with higher privately anticipated expected costs are more inclined to buy a given contract or a more protective option, conditional on the insurer's relevant information and the offer set. Advantageous selection means that those attracted to coverage can instead have lower expected costs, for example because protective behavior or other characteristics correlate with insurance demand. The possibility is formalized in de Meza and Webb's analysis (2001). Moral hazard concerns the causal change in outcomes attributable to contractual incentives or access, including utilization behavior; the term does not establish misconduct. These mechanisms may coexist.
The empirical object must therefore be specified as either a coverage–risk correlation, a causal coverage effect, an information advantage, a behavioral response, an equilibrium pricing effect, or a welfare consequence. Each object requires different assumptions and sometimes different data. A regression coefficient cannot inherit all these interpretations from the word selection.
Insurance risk versus the claim variable
An accident indicator, an insured loss, total medical spending, a submitted claim, a paid claim, and the insurer's cost are economically different. Deductibles and claim thresholds change whether an event appears in paid-claim data. Limits, exclusions, provider prices, fraud controls, and administrative processing change how much is paid conditional on an event. In health insurance, coverage may make previously forgone treatment affordable, changing utilization and sometimes the diagnosis process. In annuities, mortality determines the financial obligation in the opposite direction: longer survival, rather than earlier death, raises the expected cost of lifetime payments.
This measurement distinction is structural. If researchers use paid claims as a proxy for an exogenous accident probability, they require a bridge from claim filing to underlying incidents. Insurers' internal databases may be excellent for accounting and weaker for observing events that never pass a deductible. Researchers must document whether loss indicators are equally observable across contract types; otherwise contract generosity mechanically influences the measured outcome.
Similarly, observed contract size need not be an ordered index of risk protection. A deferred annuity, a joint-life annuity, an escalating payment, and a guarantee payable to heirs distribute mortality exposure differently. Finkelstein and Poterba (2004) find selection on annuity characteristics even while reporting little substantive mortality variation by annuity size. An empirical test that uses face value alone would omit informative contract dimensions.
The insurer information set
A meaningful test of asymmetric information must condition on the variables the insurer knew or could infer when it set terms, rather than merely those available to the analyst later. Let I denote the insurer's contracting information and U additional consumer knowledge. The hypothesis of a residual consumer informational advantage concerns U after conditioning on I. A test controlling for variables revealed after an accident can unintentionally absorb information or consequences created by the contract. A test omitting material underwriting variables can attribute an ordinary rating effect to private information.
The informational boundary itself changes with experience rating, claims-history reporting, health questionnaires, telematics, mandatory pooling, and regulatory restrictions. Information that was private at contract inception may become shared after a claim or renewal. Cohen (2005) presents automobile evidence involving learning and reports a connection between experienced customers, coverage, and accidents. The design of a selection test therefore requires a temporal definition of the insurer's information and a legal or institutional account of which characteristics enter underwriting.
The population to which results apply
Most insurer datasets contain accepted applicants or active contracts. They generally omit people who declined the price, were rejected, chose a competitor, used self-insurance, or remained uninsured. Conditioning on policy ownership can itself induce dependence between otherwise separate determinants of purchase when selection into the sample depends on both. A null within a restricted insured pool says less about the extensive margin of whether to insure than about choices among contracts inside that pool.
For this reason, coverage must be studied along two margins: participation versus nonparticipation, and contract generosity conditional on participation. Price changes, default assignments, employer-plan eligibility and consumer surveys can reveal different parts of the choice process. The population definition is an identifying assumption and an external-validity boundary, not a routine administrative detail.
Coverage–Risk Correlation Tests
The positive-correlation property
A familiar empirical prediction of conventional adverse selection is that buyers of more generous protection have a higher probability of the insured event, after conditioning on information available to the insurer. Let R denote a well-defined event indicator and H a generous contract. A conditional test might examine whether
Under specific competitive-contract and single-index private-risk assumptions, a positive relationship can be consistent with private risk information influencing contract choice. Chiappori and Salanié (2000) derive and employ methods to examine contract choice and accident outcomes. They combine parametric and nonparametric approaches and find no evidence of asymmetric information in the French automobile population studied. The result is a conditional empirical finding, rather than a demonstration that policyholders everywhere reveal all private information to insurers.
The value of the positive-correlation test lies in its relative parsimony: contracts and incidents are often observed more readily than attitudes, private forecasts or utility parameters. Its cost is dependence on interpretive assumptions. An event correlation may be attenuated by incomplete risk observation, restrictive contract menus, heterogeneous preferences or opposing selection channels. It may be enlarged by behavioral response, unmeasured underwriting and mechanically higher claim reporting.
Conditional independence and null findings
A null hypothesis often employed in such studies is conditional independence between coverage and the occurrence of the insured event after controlling for rating characteristics. In practice researchers may estimate binary-response models, conditional covariance or independence statistics, or compare risk distributions within cells. The interpretation of a failure to reject depends on power, measurement, support, sample construction and the maintained model.
A statistically insignificant estimate does not establish exact independence. Even a precise estimate near zero could reflect cancellation of heterogeneous effects. At the same time, a careful null result can reject economically important forms of the simple positive-correlation prediction if the confidence interval is narrow enough. The relevant statement is the magnitude of effects the data permit, including bounds and uncertainty, rather than a binary label.
Cardon and Hendel (2001) use a structural health-insurance choice model with the National Medical Expenditure Survey and report no evidence of informational asymmetries in their studied sample. Cawley and Philipson (1999), investigating life insurance, report patterns challenging conventional predictions. These studies offer important disciplined counterexamples, but neither makes the coverage–claims test universally decisive.
The covariance decomposition
A useful synthesis follows directly from basic covariance algebra. For a measured outcome Y(C), fix a baseline potential outcome Y(0), and define the coverage response T(C)=Y(C)-Y(0). Then, wherever expectations exist,
The first term concerns baseline risk sorting into coverage within the observed sample. The second concerns contractual response and its correlation with the chosen contract. This identity is not a new causal identification theorem. It states precisely why the observed sign alone cannot separate selection on baseline risk from the behavioral effects of insurance. Moreover, both terms may reflect sample selection and the choice of controls.
Consider, solely as an algebraic illustration, an observed covariance of zero where baseline-risk sorting contributes +2 arbitrary covariance units and insurance response contributes −2. The observational null would coexist with nonzero mechanisms. The example contains no real-world estimate. A positive observed covariance of +1 could similarly arise from baseline sorting of −1 plus behavioral response of +2. No reported insurer can be assigned either decomposition without further design and evidence.
Selection may also concern the magnitude of moral hazard itself. Einav, Finkelstein, Ryan, Schrimpf, and Cullen (2013) study selection on moral-hazard responses and show that anticipated utilization responses can enter coverage choice. Thus the heterogeneity in T is substantively important: the people selecting a plan may be those for whom its incentives have the largest or smallest effect.
Why sign tests are asymmetric in evidentiary strength
A strong positive correlation under well-supported controls constitutes evidence of risk sorting, but a causal label still requires exclusion of behavioral and contractual explanations. A null can contradict a restrictive one-dimensional prediction without disproving all possible private information. A negative gradient may indicate advantageous selection, pricing discrimination, behavioral prevention, or changes in who appears in the sample. Hence the same test statistic has different inferential content under different maintained hypotheses.
The best empirical papers report the direction and uncertainty of the association and enumerate which mechanisms their design actually rules out. Published syntheses by Cohen and Siegelman (2010) and Einav and Finkelstein (2011) explain why selection evidence varies between markets. A coherent comparative science must preserve these differences instead of turning them into a pooled assertion that insurance is always adversely selected or always advantageous.
Confounding and Endogenous Contracts
Contract menus as jointly determined outcomes
Insurance coverage is not a randomized treatment in ordinary market data. Insurers set menus, risk classes, underwriting rules, commissions, and prices while consumers select from the available options. High-deductible and low-deductible contracts may differ in price, service access, network restrictions, exclusions, or marketing. An empirical contrast between their customers combines all these attributes unless the research design isolates the contract dimension of interest.
A simplified choice equation is C=g(P,X,U,Z), where P is the offer schedule and Z describes access, defaults or choice frictions. The observed outcome might be Y=h(C,X,U,V), with V encompassing shocks. Since U enters both equations, a regression of Y on C faces selection bias. If C also changes outcomes, the regression combines sorting with incentives. If firms modify P in response to past portfolio losses, price and coverage may be jointly determined in equilibrium. These possibilities make exogeneity a demonstrated condition, not a default statistical convenience.
In Puelz and Snow's automobile study, the premium–deductible menu and demand system are central to empirical interpretation, and the reported evidence supports adverse selection and signaling in that context. A related investigation using another insurer, underwriting era, or jurisdiction must rebuild the applicable contract-menu interpretation rather than import the original estimated sign.
Underwriting, classification and measurement
Premium rating and underwriting determine the relevant comparison groups. A relationship between claims and coverage conditional on age alone may disappear or reverse once driving history, vehicle use, region, contract tenure or prior insurer information are considered. Conversely, controlling for a variable that was unavailable to insurers at purchase may remove precisely the private information the study seeks to detect. The appropriate set of controls is grounded in actual underwriting practice at the decision date.
Data availability differs internationally. Some markets allow broad underwriting classification, some impose standardized contracts, some prohibit risk-sensitive premiums on specific dimensions, and many combine public and private insurance. These institutional facts determine what private information can influence choice, what prices vary, and whether the same coverage–loss statistic has the same economic meaning across countries.
Claims-based covariates also warrant caution. Prior claims may measure underlying risk, but may also reflect past coverage, insurer reporting, incentives, and filing thresholds. Introducing prior claims as a control can be useful when they were part of the insurer's information set, while simultaneously changing the selection question toward any residual information not revealed by that history. This is why temporally precise datasets and underwriting manuals matter as much as statistical sophistication.
Moral hazard and dynamic information
Repeated contract and claim observations create opportunities beyond a one-period cross-section. Under experience rating, past claims affect later premiums or contractual terms, so present claim behavior can respond to expected future pricing. Abbring, Chiappori, and Pinquet (2003) develop dynamic tests of moral hazard and analyze negative occurrence dependence under their model's experience-rating incentives. Their application reports no evidence of moral hazard in the French car-insurance data they study.
The dynamic approach is powerful precisely because its assumptions are more explicit. Claim timing, renewal decisions, individual heterogeneity, contract switching, and persistent risk shocks all need representation. A decline in claim intensity after an accident can reflect incentives, regression toward the mean, changes in driving exposure, selective exit, or administrative thresholds unless the design separates them. A dynamic empirical result should therefore specify the model and observed histories needed for its inference.
The learning dimension is equally important. Cohen (2005) reports that coverage–accident associations among experienced customers differ from those for less experienced drivers, with evidence relevant to underreported claims histories. Information asymmetry is consequently time-dependent: insurers and policyholders may learn at different rates, and reported experience is itself imperfectly informative. This feature helps explain why a result from one driver cohort need not generalize to another.
Quasi-experimental price and menu variation
Changes in employer contributions or plan prices can help identify demand and cost schedules. Cutler and Reber (1998), studying a Harvard University employer-plan reform, document competition and adverse-selection trade-offs. Their published article reports a welfare loss from adverse selection and contemporaneous premium reductions associated with competition in that setting. The numerical results are specific to that reform and should not be generalized as insurance-sector-wide magnitudes.
For a price or menu change to identify selection credibly, its origin and associated changes must be examined. A subsidy reform may alter both prices and plan marketing. New defaults may alter inertia. Eligibility rules may alter the covered population. Anticipation can affect when consumers switch. Instrumental-variable methods require an exclusion restriction; event studies require credible counterfactual trends; randomized offers identify effects within the randomized offer population. No methodological label removes the need for these conditions.
Handel (2013) uses a firm-level insurance change to study consumer inertia and adverse selection. The results illustrate that reducing switching frictions can worsen risk selection in a specific market even while improving individual choice under a narrower criterion. Selection tests thus connect to behavior, competition and welfare, but their policy implications require a model of the full choice environment.
Missing applicants and endogenous retention
Selection into the data is especially consequential when administrative records observe only contract holders. A loss comparison between high- and low-coverage customers conditional on staying insured can be biased if high-risk people face nonrenewal, lower-risk people exit for cheaper carriers, or customers switch after an unexpected event. The observed distribution at the end of the period may not describe the choice population at the beginning.
A stronger research design follows offers, rejections, lapses, renewals, plan switching, pricing changes and outcomes where feasible. Where such information is unavailable, transparent estimates should be labeled conditional on observed insureds. This is a substantive scope restriction. In empirical insurance science, defining the risk pool is part of defining the estimand.
Multiple Private-Information Dimensions
The failure of a single private-risk index
A one-dimensional adverse-selection story ranks individuals solely by their expected insured loss and assumes that greater risk increases demand for more coverage. Real consumers differ in risk aversion, liquidity, income, trust, planning horizon, attention, perceived risk, health-seeking behavior, and knowledge of insurance arrangements. These dimensions need not vary in the same direction as expected loss.
Finkelstein and McGarry (2006) directly investigate private information in long-term care insurance. They show that some people have private information implying higher subsequent risk, while others have a particularly strong taste for insurance and lower subsequent risk. Aggregate coverage and realized risk are not positively correlated in their study, despite evidence of private information along multiple dimensions. This result is a decisive warning against identifying private information solely with a positive aggregate covariance.
The study's methodological contribution is important because it uses information about people's expectations of future nursing-home use to investigate residual predictive information relative to insurers. Subjective expectations can add evidence about information possessed at the choice date, while their calibration and reporting quality must still be evaluated. Their value lies in testing a concrete informational asymmetry, not assuming every subjective forecast is accurate.
Advantageous selection and its mechanisms
Advantageous selection describes a market where purchasers of more coverage tend to have lower expected insured costs. The interpretation is compatible with heterogeneous preferences: a cautious consumer may both protect against losses and take precautions that reduce them. The de Meza–Webb model makes the theoretical point that the relationship between insurance demand and risk can reverse the standard selection pattern. It provides one theoretical mechanism, not an empirical description of all observed negative gradients.
Fang, Keane, and Silverman (2008) find advantageous selection in Medigap. Conditional on their price controls, people with Medigap had average medical expenditures about $4,000 lower than those without; conditioning on health, the sign of the spending difference reversed to about $2,000 higher. The authors identify income, education, longevity expectations, planning horizons and cognitive ability as relevant sources, while risk preferences did not appear to be the principal source in their analysis. These are sample-specific historical estimates and conditional relationships, not current price-adjusted forecasts.
The Medigap evidence is particularly instructive because the sign changes with conditioning. A superficial comparison would call insured persons healthier; a richer specification reveals both observable selection and spending differences among people of similar health. Whether the residual latter difference reflects treatment, selection on other dimensions, or both depends on the model. The example shows why every statement about advantageous selection must name its conditioning set and outcome definition.
Heterogeneous behavioral response as private information
Individuals can know something about how strongly they themselves would respond to coverage. A person expecting large additional consumption when marginal prices fall may select a generous health plan even if two people have identical baseline expenditure risk. The relevant private dimension is then a prospective treatment response, rather than baseline risk alone.
Einav and colleagues (2013) analyze this possibility in their study of selection on moral hazard. Their panel-data choice and utilization framework finds heterogeneous behavioral responses and evidence that people select in part on those responses. Omitting this heterogeneity can bias projected spending changes from a new deductible or plan design. The empirical lesson extends beyond the health context as a methodological possibility, although its quantitative importance must be established separately for each line.
Aggregation and cancellation
Suppose the population has two private dimensions, baseline expected loss R and a taste for protection A. If R raises demand and A raises demand while being negatively related to R, the coverage–loss relationship can be positive, negative, or zero depending on their joint distribution and contract pricing. Aggregating across age, socioeconomic, geographic or underwriting groups can also reverse within-group gradients. This is a familiar statistical aggregation problem given distinctive economic content by insurance choices.
Consequently, a zero average correlation cannot distinguish complete information from offsetting private signals. A negative correlation cannot certify that all individuals are low-risk purchasers. Even a positive gradient may coexist with substantial advantageous sorting among a subgroup. The appropriate empirical response is to examine pre-specified heterogeneity, alternative risk measures, revealed and subjective expectations, and sensitivity to control sets. Researchers should resist interpreting exploratory subgroup significance as conclusive identification without multiple-testing discipline.
Consumer knowledge and cognitive constraints
Advantageous selection may involve financial literacy or cognitive capacity rather than conventional risk aversion. The Medigap evidence points to cognitive ability as an especially relevant mechanism in its setting. Consumers who understand coverage and public-program interactions may make different choices and exhibit different spending patterns; imperfect comprehension means choices need not reveal stable insurance preferences without additional interpretation.
This finding also qualifies welfare arguments. A contract chosen by an informed household is not mechanically welfare-equivalent to an identical contract chosen under misunderstanding or inertia. When adverse selection and decision frictions coexist, a policy that improves consumer decision quality can also reshape the pool and premiums. Handel's study of insurer choice inertia shows why the two effects must be assessed jointly.
Evidence Across Insurance Markets
Automobile insurance
Automobile insurance offers observable contracts, experience ratings and discrete accidents, yet its empirical history illustrates the diversity of possible outcomes. In the United States collision-insurance application of Puelz and Snow (1994), the premium–deductible system was consistent with adverse selection and signaling. In the French application of Chiappori and Salanié (2000), the conditional association tests found no evidence of asymmetric information. The two papers examined different samples and used different approaches; there is no scientifically valid basis for pooling them into a single accident-risk coefficient.
Cohen (2005) documents asymmetric information and learning in an automobile dataset, particularly among new customers with sufficient driving experience. The reported evidence includes incentives and opportunities to underreport previous claims. Abbring, Chiappori, and Pinquet (2003) provide a distinct dynamic test for moral hazard. Each question is narrower than the proposition that an automobile market has or lacks asymmetric information.
Comparative inference requires knowing the contract menu, time period, legal disclosure obligations, deductible structure, claims observation, exposure measures, driver tenure, experience rating, market switching and treatment of losses below deductible. Differences on these dimensions can rationally produce different estimates. A replicable cross-country dataset must harmonize the economic meaning of its variables before estimating a common parameter.
Health insurance and employer-plan choice
Health insurance involves selection on expected treatment costs as well as behavioral utilization, provider networks, employer contribution rules and imperfect consumer understanding. Cardon and Hendel (2001) report no informational asymmetry in their structural study of single individuals from a specified survey. Cutler and Reber (1998), by contrast, find adverse selection following a pricing reform in Harvard employees' plan choices. The studies have different estimands and populations and should not be treated as direct replications of one another.
Einav, Finkelstein, and Cullen (2010) use employer-plan price variation to estimate insurance demand and how average costs vary with enrollment. Their particular application detects adverse selection while finding relatively small welfare losses from inefficient pricing. That result demonstrates that an empirically important selection slope need not generate a proportionately large welfare distortion. Extrapolation requires knowledge of the demand schedule, insurer costs and the counterfactual regime.
The 2013 selection-on-moral-hazard study introduces heterogeneity in the utilization response, while Handel's insurance inertia analysis integrates behavioral friction into equilibrium consequences. Taken together, the research shows that structural features determining how people choose are as important to the measured association as risk heterogeneity itself.
Medigap and supplemental insurance
The Medigap setting is a prominent demonstration of advantageous selection, documented by Fang, Keane, and Silverman (2008). Their observed negative unconditional spending relationship after price controls changes sign when health is held fixed, and several socioeconomic and cognitive attributes help explain purchase patterns. The result is specific to U.S. supplemental insurance rules, the data and cohorts studied.
A crucial distinction is between the outcome under a baseline program and the outcome under supplemental coverage. Supplemental insurance can change the consumer's effective price of care and thus measured spending. A study must separately address who buys coverage and what coverage does to utilization. Cross-sectional conditioning cannot always establish the direction of either causal component.
Long-term care insurance
Long-term care insurance provides a particularly sharp test of the idea that absence of a coverage–risk gradient implies absence of private knowledge. Finkelstein and McGarry (2006) document offsetting private-information dimensions: perceived institutional-care risk on the one hand and insurance preference on the other. Their findings show that conditional on the insurer's prediction, customers can possess information about future care use even when aggregate coverage groups do not display higher realized risk.
The time horizon makes measured outcomes difficult. Nursing-home entry is only one possible event; policy details can define eligible care differently; death and lapse can truncate exposure; and retrospective samples may miss people with different survival histories. These features reinforce the need to specify the insured event, censoring, observable risk factors and policyholder cohort when generalizing from the study.
Life insurance and annuities
Mortality-related contracts are especially sensitive to the direction of the insurer's liability. In life insurance, earlier death usually raises expected payment costs. In a lifetime annuity, longer survival raises expected payment costs. The same observed characteristic can therefore have opposite interpretations under the two contract forms.
Cawley and Philipson (1999) investigate life-insurance information barriers and report evidence difficult to reconcile with simple conventional predictions. Finkelstein and Poterba (2004), investigating the United Kingdom annuity market, find mortality relationships with annuity features even though annuity size does not show a substantive mortality difference. An empirical test that ignores payment timing, guarantees, survivor options or other contract dimensions may miss where selection manifests.
These studies support a general methodological conclusion: risk type, generosity and expected insurer cost must all be defined for the actual benefit schedule. Comparing mortality gradients without mapping the contractual payoff leads to classification error.
Comparative evidence and study heterogeneity
The body of research assembled by Cohen and Siegelman (2010) and Einav and Finkelstein (2011) does not establish a single universal direction of selection. It establishes an empirical program in which market institutions, insurer information and consumer heterogeneity govern whether a risk gradient is visible and whether it has welfare significance. Apparent contradictions among studies are often partly differences in the tested hypothesis.
A formal meta-analysis would require comparable effect sizes, sampling uncertainties, common outcome definitions and reproducible inclusion criteria. This monograph does not invent a pooled effect across incomparable designs. Its comparison instead classifies the direction of each study's reported conclusion, its empirical unit, and its principal identification caveat. Such classification is an original synthesis of existing evidence, with the boundaries of each underlying finding retained.
Interpretation and Validity
The inferential ladder from data to welfare
Four levels of inference should be kept separate. The first is descriptive: how contracts and observed outcomes covary in the sampled population. The second is mechanistic: whether sorting on private baseline risk, tastes, utilization responses or underwriting explains that covariance. The third is causal and structural: how price or contract changes alter participation, behavior and costs. The fourth is normative or policy-analytic: what allocation and welfare consequences follow under stated institutional and social-welfare criteria.
A study establishing a descriptive positive relationship has achieved the first level. A credible mechanism test can advance to the second. Price variation, sufficiently rich behavioral data or a disciplined structural model may support the third. Welfare analysis additionally requires demand, marginal cost, consumer surplus assumptions, market conduct and counterfactual policy rules. Results must be attributed only at the level they genuinely support.
Einav, Finkelstein, and Levin (2010), in Beyond Testing, explain the transition from simple selection tests to estimated demand and cost relationships. Their review makes clear why different assumptions about insurer competition and consumer behavior matter. The empirical science of selection is therefore broader than the traditional coverage–loss test, but a broader model also entails additional identifying commitments.
Price variation, average costs and marginal costs
A particularly useful empirical strategy measures how enrollment changes with price, then estimates how the claims cost of the insured pool changes as additional consumers enter. Let Q(p) be coverage enrollment at premium p and let AC(Q) denote average expected insurer cost for that enrolled quantity. In a stylized single-contract market, marginal cost can be written as
This is an accounting derivative under the specified cost and enrollment mapping. If marginal entrants have lower expected costs than existing purchasers, increasing enrollment can lower average cost, which is the standard adverse-selection pattern under an ordering in which initially enrolled people have higher expected costs. If marginal entrants are more costly, expanding enrollment can raise average cost, consistent with advantageous selection. The interpretation requires that prices move enrollment in an identifiable way and that observed cost changes are not confounded by plan redesign, medical price shifts or other simultaneous changes.
Einav, Finkelstein, and Cullen (2010) implement demand-and-cost estimation using prices in employer-provided health insurance. This approach creates a bridge from a selection slope to an explicit welfare calculation. It also makes welfare sensitive to the quantitative steepness of demand and costs, rather than to a categorical announcement that selection exists.
For clarity, a positive coverage–claims coefficient and a downward-sloping average-cost curve are not numerically equivalent statistics. The former compares chosen contract types or customers; the latter follows cost composition under marginal enrollment induced by price. Their economic relation is mediated by the choice set and identification design.
External validity and transport across jurisdictions
External validity requires specifying what is invariant across settings. A conditional-independence test may travel across insurance lines as a statistical procedure, while its interpretation does not travel automatically. Auto accidents, nursing-home admissions, medical use, mortality and annuity survival generate distinct timing, exposure, censoring and benefit-payment structures. Likewise, risk rating, community rating, employer subsidies, guaranteed issue, defaults and insurer competition change the economic mechanisms entering observed data.
The relevant unit of comparison is therefore a fully described market-and-contract environment. Claims that a finding generalizes should be supported by shared observable definitions, comparable choice margins, underwriting regimes, and a plausible account of how private information and behavior operate. When these conditions cannot be established, comparative findings are evidence of heterogeneity, and any generalized policy estimate remains unsupported.
Policy evaluation and fairness
A policy intervention may address information problems through underwriting, pooling, subsidies, guaranteed access, standardized benefits, disclosure, default settings or risk adjustment. Its welfare effect depends on who enters, who exits, what is paid, how behavior changes and what distributional objective the institution adopts. A positive selection correlation does not mechanically endorse unrestricted risk rating; a negative correlation does not justify a universal subsidy. Policy assessment is downstream from identified economic quantities and explicit institutional objectives.
Cutler and Reber's competition–selection comparison shows how a pricing reform can produce both gains and losses. Handel's analysis of inertia reduction adds the possibility that improvements in individual decision making can change equilibrium risk sorting. These examples establish why the welfare objective, market horizon and incidence of costs must be specified in applied policy work. Fairness and access also involve distributive choices that cannot be resolved by estimating a covariance.
Minimum standards for interpreting an empirical test
A defensible study should state its insured event and denominator; distinguish paid claims from underlying loss; document the actual insurer information set; define the extensive and intensive coverage margins; identify the population from which observations are selected; describe contract menu and pricing endogeneity; clarify moral-hazard and preference mechanisms; report uncertainty and substantive effect sizes; test sensitivity to alternative outcome and control definitions; and distinguish the observed association from causal and welfare conclusions.
These requirements are synthesized from the methodological limitations encountered across the reviewed studies; they are editorial research criteria, not a new authorial theory. Their purpose is to make competing empirical claims commensurable without concealing the assumptions that give each claim meaning. An accessible abstract can communicate a paper's broad result, while the exact identifying conditions often require inspection of the methods and robustness discussions. Where full publisher text is paywalled and accessible author manuscripts or abstracts are used instead, evidence assessment must disclose that access limitation.
Research synthesis and unresolved empirical questions
Three general observations follow from the comparative record. First, the coverage–loss sign is contingent on preferences, information and the contractual response of the insured. Second, dynamic and price-variation designs can identify dimensions that a static correlation leaves unresolved, but only under their own exclusion and model conditions. Third, the welfare magnitude and direction of an intervention must be inferred from behavior, demand, cost and equilibrium structure; the existence of selection alone supplies too little information.
A continuing research agenda should examine selection when consumers have private forecasts and learning differs across parties; the interaction between insurer classification and behavioral responses; the effect of standardized menus and default renewal; the role of cognitive constraints; and the transportability of selection estimates across regulatory regimes. Comparative data with offers, underwriting decisions, contract choices, retention, risk exposures, outcomes and prices are especially valuable. Future causal work should also examine the durability of effects as insurers reprice and consumers learn. These are research questions motivated by existing evidence, not assertions of new verified empirical findings.
Conclusion
Empirical tests of selection in insurance markets concern the relationship between private information, risk-bearing preferences, contract choice, insurer classification and realized costs. The literature supports neither a universal positive coverage–risk relationship nor a universal absence of asymmetric information. Positive, negative and null correlations occur in documented studies of different insurance products, markets and populations.
The core result of this monograph is an inferential distinction: the observed covariance of insurance coverage and claims is an equilibrium outcome, while the claim that one party has superior private knowledge is a causal and informational interpretation. The distinction becomes decisive when coverage affects behavior, preferences correlate with prevention, claim reporting depends on deductibles, insurers screen risks, or data observe only those who purchased and retained insurance. A null gradient can mask multiple private-information dimensions, as the long-term care evidence demonstrates. Advantageous sorting can be visible in Medigap, while adverse selection can emerge after employer-plan pricing changes. Life-insurance and annuity analyses require an explicit mapping from mortality to expected payment obligations.
The methodological implication is to match the research design to the question. Conditional association tests describe risk sorting under clearly delimited assumptions. Subjective forecasts and risk-preference measurements examine dimensions of private knowledge. Dynamic claims histories can help separate incentive responses from persistent heterogeneity. Price variation and structural demand-and-cost methods make the welfare consequences of selection estimable within specified markets. No single method replaces the need to define the information set, choice process, observed population and counterfactual.
The original contribution here is a consolidated interpretation of these identification boundaries and a cross-line source audit of established scholarship. No new named theory, causal estimate, replication result, or scientific priority is claimed. The defensible conclusion is that selection must be measured as a market-specific relationship among contracts, costs, information and behavior, and that welfare claims require evidence beyond the sign of a coverage–loss association.
References
Abbring, J. H., Chiappori, P.-A., and Pinquet, J. (2003). Moral Hazard and Dynamic Insurance Data. Journal of the European Economic Association, 1(4), 767–820. https://doi.org/10.1162/154247603322493159
Akerlof, G. A. (1970). The Market for “Lemons”: Quality Uncertainty and the Market Mechanism. The Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431
Cardon, J. H., and Hendel, I. (2001). Asymmetric Information in Health Insurance: Evidence from the National Medical Expenditure Survey. The RAND Journal of Economics, 32(3), 408–427. https://doi.org/10.2307/2696362
Cawley, J., and Philipson, T. (1999). An Empirical Examination of Information Barriers to Trade in Insurance. American Economic Review, 89(4), 827–846. https://doi.org/10.1257/aer.89.4.827
Chiappori, P.-A., and Salanié, B. (2000). Testing for Asymmetric Information in Insurance Markets. Journal of Political Economy, 108(1), 56–78. https://doi.org/10.1086/262111
Cohen, A. (2005). Asymmetric Information and Learning: Evidence from the Automobile Insurance Market. The Review of Economics and Statistics, 87(2), 197–207. https://hls.harvard.edu/bibliography/asymmetric-information-and-learning-evidence-from-the-automobile-insurance-market/
Cohen, A., and Siegelman, P. (2010). Testing for Adverse Selection in Insurance Markets. Journal of Risk and Insurance, 77(1), 39–84. https://doi.org/10.1111/j.1539-6975.2009.01337.x
Cutler, D. M., and Reber, S. J. (1998). Paying for Health Insurance: The Trade-Off between Competition and Adverse Selection. The Quarterly Journal of Economics, 113(2), 433–466. https://doi.org/10.1162/003355398555649
de Meza, D., and Webb, D. C. (2001). Advantageous Selection in Insurance Markets. The RAND Journal of Economics, 32(2), 249–262. https://doi.org/10.2307/2696408
Einav, L., and Finkelstein, A. (2011). Selection in Insurance Markets: Theory and Empirics in Pictures. Journal of Economic Perspectives, 25(1), 115–138. https://doi.org/10.1257/jep.25.1.115
Einav, L., Finkelstein, A., and Cullen, M. R. (2010). Estimating Welfare in Insurance Markets Using Variation in Prices. The Quarterly Journal of Economics, 125(3), 877–921. https://doi.org/10.1162/qjec.2010.125.3.877
Einav, L., Finkelstein, A., and Levin, J. (2010). Beyond Testing: Empirical Models of Insurance Markets. Annual Review of Economics, 2, 311–336. https://doi.org/10.1146/annurev.economics.050708.143254
Einav, L., Finkelstein, A., Ryan, S. P., Schrimpf, P., and Cullen, M. R. (2013). Selection on Moral Hazard in Health Insurance. American Economic Review, 103(1), 178–219. https://doi.org/10.1257/aer.103.1.178
Fang, H., Keane, M. P., and Silverman, D. (2008). Sources of Advantageous Selection: Evidence from the Medigap Insurance Market. Journal of Political Economy, 116(2), 303–350. https://doi.org/10.1086/587623
Finkelstein, A., and McGarry, K. (2006). Multiple Dimensions of Private Information: Evidence from the Long-Term Care Insurance Market. American Economic Review, 96(4), 938–958. https://doi.org/10.1257/aer.96.4.938
Finkelstein, A., and Poterba, J. (2004). Adverse Selection in Insurance Markets: Policyholder Evidence from the U.K. Annuity Market. Journal of Political Economy, 112(1), 183–208. https://doi.org/10.1086/379936
Handel, B. R. (2013). Adverse Selection and Inertia in Health Insurance Markets: When Nudging Hurts. American Economic Review, 103(7), 2643–2682. https://doi.org/10.1257/aer.103.7.2643
Puelz, R., and Snow, A. (1994). Evidence on Adverse Selection: Equilibrium Signaling and Cross-Subsidization in the Insurance Market. Journal of Political Economy, 102(2), 236–257. https://doi.org/10.1086/261930
Rothschild, M., and Stiglitz, J. (1976). Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information. The Quarterly Journal of Economics, 90(4), 629–649. https://doi.org/10.2307/1885326