RESEARCH DOSSIER
Stated and Revealed Insurance Preferences
Abstract. Insurance preferences are inferred from statements about risk and protection, hypothetical choices, incentivized experiments, contract purchases, renewals, switching decisions, and claims histories. Each observation is generated by a different decision process. This monograph examines the conditions under which stated and revealed insurance behavior can be compared, the economic and psychological mechanisms capable of producing discrepancies, and the evidential limits of interpreting any single behavior as a stable preference. It integrates research in decision theory, consumer insurance, health-plan selection, index-based agricultural insurance, and flood protection. A formal account distinguishes a valuation of contingent indemnity from beliefs about loss, expected claims performance, liquidity, attention, eligibility, available contracts, and the costs of acting. Methodological analysis addresses construct validity, common-choice-set comparison, selection, structural underidentification, randomized information treatments, administrative-data linkage, longitudinal observation, and cross-jurisdictional transportability. Evidence from German health-insurance switching, United States deductible and health-plan choices, Indian rainfall insurance, and flood-policy uptake demonstrates that survey–behavior gaps may represent different decision environments as well as measurement error. The central analytical conclusion is that credible inference depends on specifying the decision, observing its feasible alternatives, and matching claims about latent preferences to the design and reach of the evidence. The work offers source-grounded synthesis and conditional methodological conclusions; it establishes no new named theory or untested empirical estimate.
Keywords: Stated preference; revealed preference; insurance demand; hypothetical bias; construct validity; risk perception; deductible choice; survey–administrative linkage; insurance literacy; information friction; liquidity constraints; selection; basis risk; claims experience; structural identification; behavioral insurance; causal inference; welfare analysis.
Introduction
The expression insurance preference is commonly used for several distinct objects. A household might say that it prefers comprehensive health coverage, choose a low-premium plan with substantial cost sharing, retain an unsuitable contract because switching requires effort, and later file a claim whose outcome changes its expectations about insurers. Those events are jointly informative, yet they answer different questions. A statement reports a judgment at an elicitation moment; a purchase records an action within an available market; a claim records a contractual and behavioral sequence after an insured event. The intellectual task is to reconstruct which determinants each observation identifies.
Classical revealed-preference reasoning evaluates alternatives through observed choices given constraints. Applied to insurance, the observable action includes risk-bearing, payment today for uncertain future indemnification, and a legally defined claim contingent on a peril, policy wording, exclusions, limits, and insurer performance. An apparent divergence between declared and enacted insurance demand therefore requires a comparison of equivalent feasible decisions. Without this equivalence, the difference might be an artifact of premiums, time, liquidity, information, legal eligibility, or a changing probability of loss. It is premature to assign the residual to inconsistency in preferences.
The experimental-methodology literature already recognizes the difficulty. Jaspersen's (2016) review of insurance-demand experiments surveys how design choices influence what experimental insurance decisions reveal. The substantive market studies then expose additional constraints. Cohen and Einav (2007) estimate preferences from deductibles; Handel and Kolstad (2015) combine surveys with claims data; Cole and colleagues (2013) randomize barriers to insurance uptake. These are complementary research strategies whose implications must remain tied to the populations, contracts, interventions, and assumptions actually studied.
The present object is the relationship among stated intentions, hypothetical demand, real policy decisions, and subsequent insurance experience. Experimental elicitation designs in their own right, individual probability-weighting models, and general welfare estimation each have a wider specialized literature and associated Research Dossiers. They enter here insofar as they establish the meaning and credibility of a stated–revealed comparison.
A fundamental distinction governs the analysis. Agreement across measurements can strengthen a claim that a preference operates consistently under the measured conditions. Disagreement can indicate hypothetical response bias, a changing objective, a misdescribed choice, or constraints on implementing a genuine intention. Neither agreement nor disagreement is decisive without information on how each observation was produced.
Scientific scope and evidence conventions
The monograph synthesizes published journal articles, a documented working paper, and accessible publisher or institutional research records. Where full paper or author-manuscript text was available in the reviewed research record, its research design was considered; where only the publisher abstract and bibliographic metadata could be verified, statements remain within that disclosed scope. Numerical results quoted from the literature retain the originating study's population and definition. Mathematical examples in this work are expressly illustrative, not new observations. There is no original household-level dataset, preregistered trial, or causal estimate generated for this dossier.
The term stated preference denotes an expressed ranking, intention, probability assessment, or hypothetical choice under a specified question. Revealed insurance behavior denotes a recorded action in a feasible insurance opportunity, such as enrollment, deductible selection, renewal, cancellation, or switching. Latent preference denotes the underlying evaluative tendency a researcher aims to infer. Hypothetical bias refers to a systematic gap between responses to hypothetical questions and comparable consequential actions, not to any and every difference between a survey and a policy record.
Hypothetical Intentions and Preferences
What an insurance statement measures
A statement that a respondent is “willing to pay $300 for flood insurance” can express anticipated affordability, expected disaster fear, moral approval of risk protection, an informal guess, or a reservation price. A probability statement about becoming ill is a belief; an answer about preferred deductibles can be a ranking; “I will buy coverage next year” is an intention contingent on future income and opportunity. These variables require different empirical interpretations. Treating all of them as a single continuous index of risk aversion would combine beliefs, prices, values, and intention formation.
The respondent may also imagine a contract different from the researcher's intended contract. An insurance label can mask uncertainty about exclusions, claims documentation, waiting periods, basis risk, or the probability of payment. The reference object must therefore be specified as a conditional indemnity schedule and a set of costs. A meaningful stated preference identifies at minimum the insured peril, covered loss, price, deductible, limit, payout conditions, time horizon, and principal alternatives. In health insurance, provider networks and expected out-of-pocket expense are also central. An elicited answer without these elements may still document a general insurance attitude, but its translation into an actual purchase probability requires assumptions.
An individual might state a positive willingness to buy a policy and still rationally defer purchase when the premium is due during a liquidity shock. An individual might state a preference for low deductibles yet choose a higher deductible because the offered premiums differ from those in the survey. An intention formed before a flood may alter after the event. The data become inconsistent only after the analyst demonstrates stability of the relevant environment and the same decision object.
Preference statements in decision theory
Expected-utility reasoning provides one disciplined benchmark. Let wealth be , loss be , loss probability be , premium be , and indemnity upon a covered loss be . With a deductible and exclusions held within the defined contract, an illustrative expected-utility evaluation is
A choice between no cover and a given indemnity depends on the difference between two such valuations. A survey asking “do you fear a loss?” supplies neither a willingness-to-pay function nor the complete ingredients of this comparison. An observed enrollment decision, conversely, need not reveal the utility function unless the feasible premiums, claims states, beliefs, and alternative contracts are observed or identified.
Behavioral accounts provide additional potential determinants. Kahneman and Tversky's (1979) analysis of risky decisions specifies reference-dependent evaluation and decision weights, giving researchers reasons to scrutinize framing, perceived chances, and what counts as a loss. Barseghyan and colleagues (2013) analyze probability distortions in observed insurance deductible choices. These results make it important to distinguish curvature of utility, probability weighting, subjective probability, and certainty-related features instead of calling every observed preference for coverage “risk aversion.”
A preference statement can therefore be credible and incomplete at the same time. For example, “I want less financial exposure” records a direction of desired protection; the actual contract involves a tradeoff between financial exposure and premium. The act of quantifying the tradeoff can change the reported answer because it reveals costs the respondent had not considered.
Intention as a conditional forecast
Future purchase intention is best interpreted as a conditional forecast of action within an imagined future environment. It includes an implicit probability that a suitable product will be offered, a sales channel will be accessible, credit or cash will be available, and the intention will remain salient. Different respondents may implicitly attach different conditions to the same question.
A transparent survey can improve interpretation by separating four questions: perceived probability of a loss; evaluation of the loss's severity; perceived reliability and terms of insurance; and a choice at a clearly presented premium with a specified enrollment opportunity. A prospective follow-up then tests how each answer predicts a subsequent action. Better question wording narrows ambiguity but does not by itself establish that the forecast is incentive compatible or externally valid.
Werbeck (2024) directly addresses stated versus actual German switching. The working paper combines survey data with detailed health-plan information and reports that 64 percent of stated switchers in its analytic setting exhibit hypothetical bias in stated willingness to switch in response to a price increase. This is a population- and task-specific finding about switching; it is not a universal estimate of the proportion of insurance consumers whose stated preferences are misleading. Its value for the present inquiry is direct matching of hypothetical and actual behavior, with variation across measured personal characteristics.
The status of hypothetical insurance experiments
A hypothetical choice task can hold prices and coverage terms constant across respondents, allowing clean comparisons across assigned information frames or contract menus. Such internal control is analytically valuable. Yet a controlled response does not face every consequence of a paid premium, underwriting decision, denied application, or future claims process. Jaspersen (2016) documents experimental design differences rather than supplying a single conversion factor between stated demand and real uptake.
An incentivized experiment may improve the salience of consequences, particularly where a random trial actually determines a monetary payoff. Nevertheless, laboratory stakes, convenience samples, simplified risks, and one-time decisions can remain distant from multi-year insurance contracts. In the other direction, real market records contain rich consequence but can obscure what the consumer knew and considered. The two modes of observation solve different portions of the identification problem.
The empirical claim appropriate to an experimental response is consequently precise: under the experimental payment and description rules, the participant chose option A rather than B. Extrapolating from this to an annual insurance purchase requires a reasoned bridge covering stakes, recurrence, trust, payment timing, contract enforceability, and opportunity.
Revealed Contract Choices
Observation and the feasible set
A recorded insurance purchase identifies an action under a particular opportunity set. Its value as evidence depends on knowing which policies the consumer was offered, when they were offered, which alternatives were understandable, and whether an application could actually result in coverage. Nonpurchase cannot be treated as refusal of a product that was never eligible, visible, affordable, or available.
Formally, let the feasible and noticed contracts for consumer at time be . Observed choice belongs to that set. A model of the evaluation of a contract may distinguish valuation of financial risk transfer , expectations of claim payment , immediate cash burden , and search or enrollment cost :
This is an accounting representation of factors to identify, not an estimated behavioral theory. Its point is observational equivalence: a low purchase rate can result from modest perceived value, weak payout credibility, an immediate liquidity cost, poor contract visibility, or a high hassle cost. The same outcome can arise under very different preference profiles.
Einav, Finkelstein, and Levin (2010) review empirical insurance-market identification, emphasizing that insurance demand and insurer cost must be modeled together when selection matters. A choice is an outcome of both the demand side and the insurance institution that supplies, prices, and restricts the product.
Deductible choice and what it reveals
Auto-insurance deductible menus offer relatively well-defined financial tradeoffs. Cohen and Einav's (2007) automobile deductible study develops a structural model allowing heterogeneity in claim risk and risk aversion. The observed deductible becomes more informative when coupled with claim-frequency data and premium differences. Yet the estimated preference parameters remain conditional on the structural specification, observed menu, and treatment of unobserved heterogeneity.
Consider an illustrative pair of otherwise identical policies with deductibles of $500 and $1,000. Suppose the lower deductible costs $80 more per year and every relevant claim would exceed $1,000. If the annual claim probability is 10 percent, the expected incremental claim payment from the lower deductible is . Under the stated assumptions a risk-neutral consumer prefers to save the $80, while a risk-averse consumer may rationally pay the $30 expected loading for reduced exposure. A survey asserting “I always prefer the low deductible” is not comparable until the premium difference and claim conditions are supplied. If claims fall below a deductible, if expected probability differs, or if reimbursements are uncertain, the calculation changes.
Sydnor's (2010) homeowners deductible evidence documents low-deductible selections that imply extremely high estimated risk aversion under a conventional model applied to modest stakes. Barseghyan and coauthors (2013) estimate behavioral risk parameters and find probability distortion important in explaining deductible decisions. Taken together, the findings show why a contract choice can be genuinely revealed yet its interpretation as a unique, stable utility parameter remain disputed.
Dominance, understanding, and action
The strongest revealed-preference inference occurs when the comparison is robust to plausible beliefs. Bhargava, Loewenstein, and Sydnor (2017) study dominated health-plan choices among 23,894 employees at a United States firm. The plans varied in premiums and financial cost sharing, and their design allowed a test of financially dominated options. The authors report that the majority selected dominated plans, with excess spending equal to 24 percent of premiums on selected plans. Their combination of administrative observations and supplementary hypothetical tasks supports an interpretation involving misunderstanding of insurance rather than a simple willingness to pay for lower deductibles.
Dominance tests are especially valuable because they can reduce reliance on a detailed estimate of subjective risk attitudes. They require exact definition of “otherwise identical,” full relevant cost states, and restrictions on potential nonfinancial features. Outside the study's plan menu, network or quality differences may defeat the asserted dominance relation. A study of dominated plans is therefore powerful within its defined domain and should not be recast as proof that all unusual insurance choices are mistakes.
Abaluck and Gruber (2011) analyze Medicare Part D choices using prescription claims and plan menus. Their findings about premiums, expected out-of-pocket costs, and financial characteristics indicate systematic choice inconsistencies in that setting. These data document behavior with monetary consequences; interpreting the results as welfare losses additionally depends on the study's welfare benchmark and treatment of unobserved consumer values.
Nonpurchase and the institutional opportunity
There is no revealed refusal if the contract was inaccessible. Agricultural index insurance makes the distinction concrete. Giné, Townsend, and Vickery (2008) analyze rainfall-insurance participation in rural India and report associations of take-up with basis risk, wealth, credit constraints, and social familiarity. A farmer may value protection against drought while rejecting a policy whose rainfall index frequently fails to pay in years of farm loss. The observation reveals evaluation of that particular contingent payoff, not a judgment about insurance against every agricultural hazard.
In randomized field experiments, Cole and colleagues (2013) test price and nonprice barriers to Indian rainfall-insurance adoption. They find price sensitivity alongside evidence on trust, liquidity constraints, and salience. The combination of reported risk exposure and low uptake has an empirically grounded explanation in multiple product and implementation frictions. It is inaccurate to read the gap as a universal rejection of insurance.
Renewal and switching
Existing policies add inertia and timing. A renewal may be passive rather than a new, consciously optimized selection. Failure to switch may reflect search costs, plan-comparison difficulty, anticipated disruption of care, administrative uncertainty, or rational preference for continuity. These forces must be considered before using persistence as evidence of a stable evaluation of the incumbent insurer.
Kling and colleagues' (2012) Medicare plan information experiment randomized personalized cost information. Switching was 28 percent in the intervention group against 17 percent in the comparison group, and predicted costs declined by about $100 annually among letter recipients. A relatively modest information delivery changed consequential action. This establishes the causal effect of the assigned information intervention in the study, while the distinct psychological mechanisms within “comparison friction” remain more difficult to separate.
Measurement Error and Incentive Effects
The comparability problem
Suppose records a survey prediction of purchase and a subsequent observed purchase. A numerical difference is meaningful as a calibration error only when both variables refer to the same choice, time interval, premium and set of conditions. If is a stated probability in and is a realized binary outcome, individual differences partly reflect ordinary uncertainty even under perfect calibration. The appropriate test examines whether observed purchase rates match stated probabilities across comparable groups, with adequate adjustment for sampling and changed circumstances.
An empirical design can estimate
where identifies members offered the specified contract with verified eligibility. The estimand must then explain whether is an intention indicator or a probability forecast; whether means application, paid enrollment, policy in force, or renewal; and whether respondents with missing administrative outcomes were excluded. This descriptive difference alone does not isolate its causes.
A hypothetical example illustrates the issue. In a survey of 500 eligible individuals, 70 percent might state that they would buy a named product, while only 40 percent enroll at the announced date. The observed 30-percentage-point gap would be an empirical fact if actually measured; these numbers are invented solely to show the arithmetic. One would still need price continuity, eligibility verification, the full sales funnel, and information on nonresponse to determine whether it represents hypothetical bias, lost access, different choice opportunities, or a shift in beliefs.
Error in the response, error in the record
Survey responses can be affected by recall, rounding, scale use, social desirability, anchoring, framing, sample selection, and interviewer effects. Administrative records can contain equally consequential errors: plan identifiers may change, premium subsidies may be incomplete, applicant denials may be unrecorded, and claims may be observed only for a subset of insurers or providers. The label “revealed” does not guarantee error-free measurement.
Measurement error also differs in direction. A respondent may overstate willingness to purchase because spending is hypothetical; another may understate willingness because the survey describes a product in unfamiliar technical language. A data system may show no purchase even though a policy was obtained through an unlinked intermediary. Consequently, measurement error can attenuate correlations, bias estimated coefficients, or create subgroup differences that are artifacts of observability.
The record should document whether the respondent understood the policy, whether the question matched the enacted premium, which outcome definition was used, and how nonresponse, cross-insurer movement, and cancellation were handled. A survey–administrative link is a major advance only when its record linkage quality is assessed.
Incentives and strategic behavior
Consequential stakes can discipline some hypothetical responses, although incentives do not eliminate every deviation. A cash-rewarded lottery about a simplified loss is not equivalent to annual protection against a rare, multidimensional disaster. A stated maximum willingness to pay may be affected by strategic response if participants believe the answer will influence future offered prices or public subsidies. Conversely, participants can respond more carefully to a hypothetical task than to a complex real enrollment screen.
The experimental economics literature on hypothetical valuation documents why hypothetical responses should not automatically be equated with consequential valuations. For insurance specifically, Jaspersen's (2016) methodological survey of elicitation distinguishes important design variations; Jaspersen, Ragin, and Sydnor (2022) test predictive risk-attitude models in an experiment with 1,730 participants and multiple insurance choices. Their published study reports that commonly used structural preference models predict insurance choices poorly in their tested tasks despite coherence in the choices themselves. The finding cautions against inferring a universal behavioral model from a measured risk-attitude score. It does not imply that every preference measure lacks predictive value in every insurance market.
Identification from an information treatment
Randomizing an information intervention can establish a causal effect on purchase, switching, or deductible choice under a defined exposure protocol. If denotes assignment, the intent-to-treat effect is
provided assignment is random and outcomes are consistently observed. To identify a narrower mechanism—such as corrected beliefs about claims payouts rather than reduced switching hassle—requires that the intervention act exclusively or separably through that mechanism, or that the design measure multiple mediators with justified assumptions. A personalized comparison letter can change attention and perceived ease at once.
The contrast between randomized treatment and an observed survey–action discrepancy is therefore central. The former can credibly attribute behavioral change to the treatment within the study. The latter describes a mismatch whose sources remain jointly underdetermined until variation or supplemental data distinguish them.
Selection at every stage
A completed policy is the end of a sequence: awareness, consideration, quotation, eligibility, application, acceptance, payment, activation and continued coverage. If any one stage is missing from the record, a model of purchase may confound interest with supply or process loss. Written as a probability identity, not a behavioral doctrine,
The conditioning chain is illustrative: actual data systems may have differently ordered steps. Its substantive purpose is to show that high stated intention can coexist with low final coverage even if intention accurately predicts action at one particular stage. Surveys should specify which stage is predicted, and administrative data should preserve the denominator for that stage.
Insurance Survey Comparisons
Germany: stated switching versus actual behavior
Werbeck's (2024) German health-insurance comparison is particularly close to the central question because it addresses stated responses to price increases against actual plan choices. The paper reports hypothetical bias among 64 percent of respondents stating they would switch and heterogeneity by measured personality-related traits. Its analysis underscores that the most meaningful comparison is between matched decisions rather than between a generalized preference for insurance and aggregate market penetration.
Interpretation nonetheless requires attention to its working-paper status, the sample and observation window, and the operational definition of a price response. Actual switching can be influenced by enrollment rules, transaction costs, inertia, and beliefs about continuity. A survey answer might report the desired response to price under a frictionless switch; observed inaction can reflect the cost of accomplishing the change. Those alternatives are research questions rather than grounds to dismiss the measured discrepancy.
United States: automobile and homeowners insurance
Auto deductible records provide a strong opportunity to infer risk attitudes because options are often similar apart from premiums and deductibles. Cohen and Einav (2007) model auto-insurance selection while accounting for heterogeneous risk and preferences. Barseghyan and colleagues (2013) compare explanations of deductible choices, locating an important role for distorted probability weights in their model. Sydnor (2010) documents small-stakes coverage choices in homeowners insurance.
These studies are often grouped together because they analyze deductibles. Their empirical objects differ, however, in contract line, menus, assumptions, sample composition and structural model. Estimates of preference curvature derived from a particular deductible schedule are not themselves measures of the answers that the same people would give on a hypothetical willingness-to-insure survey. A direct stated–revealed test requires linked answers to precisely matched contract decisions.
United States: health plans and the value of linked information
Health insurance creates a demanding but revealing research environment. The plan can bundle premiums, deductibles, coinsurance, provider networks, formulary provisions, and uncertain service utilization. The utility of a plan depends on knowledge of health needs, anticipated service use, and the economic or nonfinancial value of continuity.
Handel and Kolstad (2015) integrate choice, claims and survey measures, showing that information frictions and hassle costs materially affect choice and estimates of risk preferences in the studied setting. Bhargava, Loewenstein, and Sydnor (2017) identify dominated employee plans and use hypothetical follow-up tasks to investigate comprehension. The combined lesson is that real contracts anchor economic consequences while survey and experimental components explain how consumers understood those consequences.
For Medicare Part D, Abaluck and Gruber (2011) link claims to plan options and report economically consequential patterns of choice inconsistency; Kling and colleagues (2012) experimentally deliver plan comparisons. Together these studies justify interpreting some revealed decisions as outcomes constrained by comparison and information, rather than a clean measurement of the underlying willingness to bear financial risk.
India: rainfall insurance and financial constraints
In rural India, rainfall-index contracts provide a case in which stated concern about risk need not translate into sales. Giné, Townsend, and Vickery's (2008) participation analysis associates uptake with basis risk, wealth, credit constraints and familiarity with the selling institution. Cole and colleagues (2013) randomize insurance-demand interventions and find important contributions of price, liquidity, trust and salience. These findings concern specific marketed products and studied communities, and they do not supply a general conversion rate from questionnaire demand to actual agricultural insurance enrollment.
The substantive challenge is that insurance reliability can be doubly uncertain: a farmer may be uncertain whether drought occurs and whether the rainfall index triggers a claim when a drought harms the farm. Traditional elicitation of “willingness to insure a crop” does not necessarily capture willingness to buy index-based coverage. A valid comparison must state the index station, trigger, spatial correlation, expected mismatch, and timing of payment.
Karlan, Osei, Osei-Akoto, and Udry (2014) examine agricultural risk constraints in Ghana and show why the broader interaction of risk transfer and investment conditions matters. This setting adds an important scope clarification: insurance demand is embedded in production and capital decisions. An answer about desire for coverage cannot be transported mechanically from a household consumption risk to an enterprise income risk with different investment opportunities.
United States: flood insurance and event salience
Gallagher (2014) uses flood-insurance take-up records and a nationwide panel of significant floods. The study reports a spike in take-up after flood exposure followed by gradual decline toward baseline; households in nonflooded communities in the same television market also respond at a smaller rate. These results are consistent with an informational and memory component to disaster insurance behavior. They do not imply that every decline in coverage is irrational, because perceived risk, premiums, availability, and public programs can change over time.
The evidence demonstrates an essential temporal difficulty for survey–behavior matching: a survey taken just after a flood may elicit a credible temporary response to new information, while a policy record two years later reflects a different information state. Describing that difference as “hypothetical bias” would confound temporal learning with elicitation format.
Cross-market interpretation
The comparison across Germany, the United States, India, and Ghana establishes heterogeneous mechanisms rather than a single universal behavioral percentage. Germany supplies an unusually direct switching comparison; U.S. deductible records help estimate risk valuation under explicit assumptions; health-plan studies integrate cognition, claims and administrative consequences; Indian rainfall-insurance studies identify product credibility, basis risk and sales constraints; Ghanaian agricultural evidence locates insurance within investment choices; U.S. flood records expose learning and decay.
The settings also vary in law, subsidy, underwriting, distribution, eligibility, social protection, data linkage and perceived claims reliability. These institutions shape which actions reveal a preference. Cross-country comparability should therefore be built around an aligned decision definition and measurement protocol, not an assumption that a similar product label denotes a similar economic choice.
Triangulation with Administrative Data
Why linking data changes inference
A survey reveals beliefs, intentions, comprehension and stated rankings that policy records rarely measure. Administrative files reveal transactions, premium payments, deductible selections, renewals, cancellations and claims that a survey may misremember. Joint observation can separate several explanations previously confounded in aggregate comparisons. For example, a policyholder reporting that high deductibles are undesirable may have faced a plan menu in which low-deductible policies were unaffordable; the administrative record can verify the premium difference.
Handel and Kolstad's (2015) linked insurance information design offers a methodological example: administrative health-plan choices and claims are analyzed alongside survey measures of information. This improves estimation of decision frictions and clarifies that welfare judgments depend on whether a consumer would choose differently with adequate information or reduced hassle.
The ideal linked observation is indexed by individual, household, market, insurer, product, time and contract version. It contains a documented menu and price quote, the moment of survey, comprehension checks, the response, application outcome, insurer acceptance, actual paid premium and subsequent coverage status. Claims exposure and outcomes should be linked by coverage year and event date, with attention to censoring and privacy controls.
What claims histories can and cannot show
Claims records can clarify realized usage and some components of actuarial risk. They can be used to estimate anticipated out-of-pocket costs conditional on modeling assumptions, and to compare an actual plan against alternative plans in a given choice set. But a claim is observed only after exposure, insurance participation, covered loss, filing, adjudication and often a deductible or threshold. It follows that observed claim frequency among the insured is generally not the same as a population's underlying loss-event probability.
There are several selection problems. Higher-risk households may select richer coverage. Coverage can change preventive behavior or service utilization. Higher deductibles can discourage small claims, altering observed frequency. Rejections and missing provider networks can change whether an event becomes a paid claim. A dataset that excludes uninsured households cannot identify their counterfactual claims solely by extrapolating the insured mean.
A longitudinal design should distinguish experienced loss, reported claim, approved claim, indemnity paid, and subsequent choice. These are not interchangeable indicators of preference. Successful payment can change trust; a denied claim can change beliefs about exclusions; the occurrence of loss can change risk perception independently of the insurer. The most informative research records these components separately.
Data integration and identification steps
A defensible program of source triangulation proceeds through explicit checks:
- Define a common contract and decision period, including premiums, benefits, eligibility, network or trigger conditions, and regulatory constraints.
- Elicit beliefs about loss probability and indemnity reliability separately from valuation and intention.
- Record the actual offer, notice, enrollment channel and feasible alternative set.
- Link survey respondents to policy records under a legally valid and auditable linkage procedure; quantify unmatched records and attrition.
- Check the respondent's knowledge of the offered product before interpreting answer–action disagreements.
- Observe purchasing stages, switching, renewals and claims without treating any one as a direct utility measure.
- Where possible, introduce randomized price, explanation, reminder or comparison interventions that isolate specific frictions.
- Estimate uncertainty, subgroup heterogeneity and sensitivity to alternative assumptions before making claims about mechanisms or welfare.
These are research-design requirements derived from the identified sources and the logic of conditional inference. They are proposed methodological steps, not a newly named research theory.
Model comparison without false certainty
A parsimonious benchmark uses expected utility, stated subjective beliefs and observed budgets. Alternative specifications may introduce probability weighting, loss reference points, limited attention, switching cost, payout distrust or latent consideration sets. If multiple parameter combinations predict the same enrollment decision, the latent preference is not point-identified without additional variation or restrictions.
Three observations can narrow this equivalence. First, a survey measuring subjective beliefs helps separate perceived risk from curvature of utility. Second, independent changes in relative price help distinguish valuation from payment constraints. Third, experimentally assigned, standardized information may distinguish ignorance or inattention from preferences that persist after comprehension improves. Even then, the effect of information on an action is established more directly than the unique psychological process mediating the effect.
The approach to economic welfare must follow the causal interpretation. Einav, Finkelstein, and Levin (2010) develop demand-and-cost perspectives for welfare analysis of insurance markets. In behavioral contexts, the analyst must also specify which expressed or informed preference supplies the evaluative benchmark. An observed choice made under misunderstanding may be a poor measure of a consumer's informed preference, while the consumer's unqualified survey answer may be equally unsuitable. Welfare analysis becomes credible when it states the benchmark explicitly.
Privacy, ethics and inferential integrity
Administrative linkage introduces sensitive financial, health and claims information. A robust study limits data to necessary variables, records permission or lawful authority, maintains secure linkage keys and assesses reidentification risk. Consent and access can themselves select participants; linkage success should therefore be reported by observable group where legally and statistically possible. Data richness can otherwise give a false impression of representativeness.
International datasets differ in the legal basis and practical availability of personal insurance and medical records. A comparison should report the jurisdiction and time of each source and avoid assuming identical permissions or coverage classifications. Where matched microdata cannot be accessed, a transparent review of published results is more credible than constructing invented longitudinal statistics.
Limits of Inference
The boundaries of revealed preference
Observed insurance behavior is evidence of an action under constraints. A contract choice can support an inference about valuation when the relevant feasible alternatives and determinants are identified. It does not automatically yield a unique psychological trait. The deductible studies of Cohen and Einav (2007), Sydnor (2010), and Barseghyan and colleagues (2013) exemplify the role of identifying assumptions: the same apparent willingness to reduce risk can reflect different combinations of utility curvature, probability distortions and product-specific features.
Likewise, expressed intention is evidence of a reported judgment under a described scenario. It predicts behavior best when the scenario and observed opportunity are aligned. A survey can be useful as a forecast without representing a stable, context-free preference.
Contract features and product performance
Insurance is a promise mediated by conditions of indemnification. Two products with the same insured-peril label may differ materially in payout trigger, exclusions, limits and claims process. This is especially important for index insurance, but applies to health plan networks, disaster contracts, and life or disability products. The scientific unit of comparison is the contingent payoff and opportunity, not merely the product name.
Demand estimation that ignores insurer reliability or basis risk may mistake a rational aversion to a weak contract for aversion to risk transfer. The field findings of Giné, Townsend, and Vickery (2008) and Cole and colleagues (2013) underscore this distinction. Trust may represent a belief about performance rather than a taste parameter; liquidity is a constraint rather than a valuation. These differences have direct implications for policy design.
Time, learning and adaptive preferences
Preferences and beliefs can evolve after new information. Gallagher's (2014) longitudinal flood evidence indicates temporal movement in take-up around flood exposure. A longitudinal study should therefore distinguish measurement error from genuine changes in perceived probability, household finances or experience of institutions. A repeat survey with stable wording cannot alone establish stability of the underlying decision context.
An observed change after a claim is similarly ambiguous. It may reflect a revised estimate of peril probability, insurer trust, learning about contract terms, or a change in income after a loss. Tracking all these components supports stronger conditional inference but does not transform a correlation into a causal explanation.
External validity and transportability
A sample of corporate employees confronting an employer health-plan menu, a Medicare population with chronic prescription expenses, and farmers offered rainfall-index insurance inhabit different decision environments. Their measured effects have different immediate estimands. Transport requires evidence that the destination setting shares the mechanism and the relevant constraints, plus a justified adjustment for differences in populations and institutions.
One can transport a research question more readily than a numerical coefficient. “Does limited comprehension alter the relationship between intention and purchase?” is a meaningful comparative question across markets. An estimate of a specific switching response from a randomized personalized letter cannot be transferred automatically to a different insurer, national health system, or online interface.
Welfare and regulation
A gap between stated and revealed behavior is not itself a measure of consumer harm. Welfare analysis must establish what the consumer knew, which choices were feasible, and whether an alternative arrangement would improve outcomes under a specified standard. The dominated-plan evidence of Bhargava and colleagues (2017) makes a stronger efficiency claim within its specific choice environment than could be drawn from mere nonpurchase. The randomized information study of Kling and colleagues (2012) supplies causal evidence that comparative support can alter choices and predicted costs under its experimental conditions.
Effective consumer-protection investigation can therefore examine disclosure, active enrollment, cancellation practices, readability, quote transparency and documented claims performance. A proposed remedy should target the demonstrated friction. Better information may help where information is missing; liquidity relief may matter where premiums cannot be financed; contract redesign may matter where index payouts diverge from losses. None of these remedies follows mechanically from a generic “survey–behavior gap.”
Research priorities and evidential thresholds
The field would benefit from prospective studies with preregistered matching between stated forecasts and real, contract-specific opportunities. Such studies should include a representative sampling frame, repeated elicitation around policy offers, individual linkages to transaction stages and claims, randomized changes in information or premium timing, and transparent attrition reporting. More systematic validation of willingness-to-pay measures against comparable actual purchases would clarify when a survey can be used for market forecasting, inclusion policy, or welfare evaluation.
Empirical generalization requires replication across insurance lines and institutional settings. A meta-analysis should respect heterogeneity in the meaning of “stated,” the definition of “real purchase,” and the sampled population; otherwise pooled gaps could average incompatible estimands. International comparability begins with common definitions, not a pooled percentage.
Integrated analytical findings
The literature supports several coherent conclusions. Stated attitudes are valuable evidence about beliefs and intentions, particularly where markets do not offer a comparable choice. Observed purchases are valuable evidence about behavior under market constraints, particularly where the feasible menu and costs are known. Incentive-compatible tasks can reduce certain features of purely hypothetical responding, while remaining limited by the stakes and environment they reproduce. Administrative claims, plan menus, and survey responses jointly improve identification when linkage and selection are credible. Randomized interventions can identify causal effects of assigned changes, although the internal mechanisms of a complex intervention may remain underdetermined.
The apparent conflict between what a consumer says and what a consumer does often disappears after identifying the missing economic object. A person can simultaneously want protection, distrust a particular insurer, lack the cash needed at enrollment, and rationally decline a contract with large basis risk. In other cases the conflict persists after price and opportunity are aligned, and then misunderstanding, cognitive frictions or measurement bias become plausible explanations. Scientific inference advances by showing which explanation the data can distinguish.
Conclusion
Stated and revealed insurance preferences are complementary observations generated by different methods and opportunities. Their comparison becomes scientifically interpretable when researchers specify the insured contingency, choice set, premium, information state, eligibility, decision horizon, contractual performance and operational definition of action. Survey intentions disclose anticipated evaluations and beliefs; recorded selections document actual decisions under constraints; claims and renewal records add realized experience, institutional performance and time dynamics.
A source-grounded comparison of Germany, the United States, India and Ghana demonstrates the importance of heterogeneous mechanisms. German survey–switching research directly tests a hypothetical–actual gap in one health-insurance market. United States deductible and health-plan studies reveal the limits of treating contract selection as a unique risk-preference parameter. Medicare experiments show that providing personalized comparisons can change enrollment choices. Indian rainfall-insurance studies locate product design, trust and liquidity among the determinants of uptake. Flood insurance records show that temporal learning can affect coverage independently of survey method.
The principal methodological implication is conditional inference: a reported desire to insure, an actual policy, and a claim are evidence about the particular decisions that produced them. Credible claims about latent preference require aligned comparisons and identification of competing explanations. Credible causal claims require appropriately designed variation; credible welfare claims require an explicit and defensible evaluative benchmark. This synthesis consolidates established empirical findings and derives research implications within their evidential reach. It introduces no new named theory, doctrine or unverified novelty claim.
References
Abaluck, Jason, and Jonathan Gruber. 2011. “Choice Inconsistencies among the Elderly: Evidence from Plan Choice in the Medicare Part D Program.” American Economic Review 101 (4): 1180–1210. https://doi.org/10.1257/aer.101.4.1180.
Barseghyan, Levon, Francesca Molinari, Ted O’Donoghue, and Joshua C. Teitelbaum. 2013. “The Nature of Risk Preferences: Evidence from Insurance Choices.” American Economic Review 103 (6): 2499–2529. https://doi.org/10.1257/aer.103.6.2499.
Bhargava, Saurabh, George Loewenstein, and Justin Sydnor. 2017. “Choose to Lose: Health Plan Choices from a Menu with Dominated Option.” Quarterly Journal of Economics 132 (3): 1319–1372. https://doi.org/10.1093/qje/qjx011.
Cohen, Alma, and Liran Einav. 2007. “Estimating Risk Preferences from Deductible Choice.” American Economic Review 97 (3): 745–788. https://doi.org/10.1257/aer.97.3.745.
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. https://doi.org/10.1257/app.5.1.104.
Einav, Liran, Amy Finkelstein, and Jonathan Levin. 2010. “Beyond Testing: Empirical Models of Insurance Markets.” Annual Review of Economics 2: 311–336. https://doi.org/10.1146/annurev.economics.050708.143254.
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. https://doi.org/10.1257/app.6.3.206.
Giné, Xavier, Robert Townsend, and James Vickery. 2008. “Patterns of Rainfall Insurance Participation in Rural India.” World Bank Economic Review 22 (3): 539–566. https://doi.org/10.1093/wber/lhn015.
Handel, Benjamin R., and Jonathan T. Kolstad. 2015. “Health Insurance for ‘Humans’: Information Frictions, Plan Choice, and Consumer Welfare.” American Economic Review 105 (8): 2449–2500. https://doi.org/10.1257/aer.20131126.
Jaspersen, Johannes G. 2016. “Hypothetical Surveys and Experimental Studies of Insurance Demand: A Review.” Journal of Risk and Insurance 83 (1): 217–255. https://doi.org/10.1111/jori.12100.
Jaspersen, Johannes G., Marc A. Ragin, and Justin R. Sydnor. 2022. “Predicting Insurance Demand from Risk Attitudes.” Journal of Risk and Insurance 89 (1): 63–96. https://doi.org/10.1111/jori.12342.
Kahneman, Daniel, and Amos Tversky. 1979. “Prospect Theory: An Analysis of Decision under Risk.” Econometrica 47 (2): 263–292. https://doi.org/10.2307/1914185.
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. https://doi.org/10.1093/qje/qju002.
Kling, Jeffrey R., Sendhil Mullainathan, Eldar Shafir, Lee C. Vermeulen, and Marian V. Wrobel. 2012. “Comparison Friction: Experimental Evidence from Medicare Drug Plans.” Quarterly Journal of Economics 127 (1): 199–235. https://doi.org/10.1093/qje/qjr055.
Sydnor, Justin. 2010. “(Over)insuring Modest Risks.” American Economic Journal: Applied Economics 2 (4): 177–199. https://doi.org/10.1257/app.2.4.177.
Werbeck, Anna. 2024. “Stated Preferences and Actual Choices in German Health Insurance.” Ruhr Economic Papers, no. 1091. RWI – Leibniz-Institut für Wirtschaftsforschung. https://doi.org/10.4419/96973268.