Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Reference Dependence and Loss Aversion in Insurance

Abstract. Reference dependence changes the economic interpretation of an insurance contract because the premium, the insured loss, the deductible, the expected claim payment, and the insurer's promise may each be evaluated against a different benchmark. Loss aversion introduces an asymmetric response to deviations from these benchmarks; probability weighting, expectations, salience, limited attention, and the institution of insurance interact with that response. This monograph provides a source-critical account of the foundations of reference-dependent valuation, reconstructs cumulative prospect theory and expectations-based reference points, develops explicit comparisons of premium and deductible alternatives, and evaluates experimental and revealed-preference evidence from insurance markets. It distinguishes price sensitivity from premium-as-loss framing; expected indemnification from an entitlement to payment; a higher deductible from an anticipated out-of-pocket loss; and probability weighting from subjective risk beliefs. The empirical synthesis covers foundational insurance-framing experiments, automobile and homeowners deductible choices, probabilistic insurance, health plan design and inertia, and low-probability disaster protection, with cautious international interpretation. Its original contribution is an integrated interpretation of the contract's multiple reference points, transparent algebraic demonstrations of ranking reversals under specified assumptions, and a hierarchy of identification requirements. Neither a new behavioral theory nor an unverified universal loss-aversion coefficient is proposed. Reference dependence is a meaningful component of insurance choice when the reference point can be specified, measured, and distinguished from alternative mechanisms; sound insurance analysis requires the contractual, temporal, and institutional context of that choice.

Keywords: Reference dependence; loss aversion; insurance; prospect theory; cumulative prospect theory; expectation-based reference points; premiums; deductibles; insurance claims; indemnification; probability weighting; mental accounting; endowment effect; status quo bias; insurer default; adverse selection; revealed preferences; health insurance; automobile insurance; homeowners insurance; catastrophic risk; experimental identification; welfare economics; consumer protection.

Introduction

Insurance is an exchange of a certain premium for a contingent promise. Its consequences are distributed across time and states of the world: the premium is paid now, the insured event may never occur, and the contractual payment, if due, follows a claims process. Each element of this sequence provides a possible reference point. A buyer may compare the premium with a current spending budget, the deductible with last year's deductible, a claim payment with the face amount of the policy, or actual consumption with anticipated consumption. The same contractual cash flow can therefore enter more than one psychological account.

The foundational contribution of Kahneman and Tversky's prospect theory (1979) was to analyze risky outcomes as changes relative to a reference point and to separate the value of outcomes from the decision weights attached to their probabilities. Cumulative prospect theory (Tversky and Kahneman 1992) subsequently provided a cumulative representation applicable to more complex distributions of gains and losses. In insurance, Johnson and colleagues' framing experiments (1993) demonstrated that descriptions of risks and coverage can influence hypothetical purchasing decisions. Later studies of actual insurance decisions showed strong preferences for low deductibles but also supplied reasons for resisting any immediate inference from that pattern to a single behavioral parameter.

A theoretical account must distinguish three questions. What reference point does the decision maker use? What is the shape of utility around it? How are probabilities, beliefs, and contractual outcomes represented? There is no direct empirical equivalence between a consumer buying comprehensive coverage and that consumer having a particular coefficient of loss aversion. Expected utility with standard concavity, distorted probabilities, misperceived loss frequencies, precautionary motives, default effects, transaction costs, and illiquidity can generate overlapping observations. Barseghyan and colleagues' insurance estimates (2013) provide an especially important warning against treating these distinct mechanisms as interchangeable.

The scope here is insurance-specific reference-point formation and insurance-specific behavioral valuation. General prospect theory, generic consumer endowment, actuarial risk classification, and the full economics of insurance demand are indispensable contexts, while the object of this study remains the way references change evaluation of premiums, deductibles, insured events, and claim outcomes. The authorial work consists of reconstructing the scientific arguments, drawing conditional implications, testing them against primary evidence, and specifying where identification succeeds or remains open.

The Research Object, Sources, and Method

The unit of analysis is the insurance decision, represented by a set of available contracts, a loss distribution, a payment schedule, a set of information states, and one or more perceived benchmarks. A contract specifies the premium and coverage schedule, exclusions, limits, and conditions of settlement. A subject's reference point may refer to wealth, expenditure, a status quo contract, expected medical spending, previous annual premiums, or an insurer's anticipated payment. These are different variables even when expressed in the same currency.

The primary theoretical sources are the 1979 and 1992 prospect-theory papers, Tversky and Kahneman's reference-dependent choice analysis (1991), Kőszegi and Rabin's expectations-based model (2006), and their analysis of reference-dependent risk attitudes (2007). The empirical sources include peer-reviewed insurance experiments, structural analyses of contract choices, and health insurance market studies. The evidentiary method links each attributed finding to the original publisher, primary institutional research page, or a directly accessible original document. Bibliographic identity is separate from full-text access; a publisher's abstract verifies a published claim or abstract-level finding, while technical derivations are identified as reconstruction rather than quotations from inaccessible appendices.

Comparison proceeds across four levels: formal prediction, experimental manipulation, revealed choices, and institutional description. An experiment that randomly assigns a premium frame may identify an effect of that frame within its population and task; it does not, by itself, identify the psychological reference point. A revealed deductible choice may constrain a structural model, but the estimates remain conditional on the loss-frequency specification, choice set, unobserved heterogeneity, and assumed preference function. A regulator's survey can document reported attitudes and obstacles, while causal behavioral mechanisms require a separate design.

An insurance claim is also a contractual and administrative event. Coverage exists only under defined terms; a policy limit, exclusion, deductible, waiting period, or lapse can alter the legally due payment. Analyses of perceived disappointment should compare what was due with what was anticipated rather than assume that every perceived shortfall is a contractual breach. The same distinction prevents consumer frustration from being mislabeled as a preference parameter.

Throughout, factual findings are attributed to specific studies and their settings. Formal examples use hypothetical monetary amounts and explicitly stated probabilities. They are illustrations, not estimated market quantities. International sources are used to investigate transferability, and absence of representative multi-country causal evidence is treated as a limit of what can be inferred.

Reference-Dependent Valuation: Historical and Conceptual Foundations

Classical expected utility evaluates a lottery of final wealth values through a utility function and objective or subjective probabilities. In a two-state insurance problem, an individual with initial wealth W, loss L, probability p, premium P, and indemnity I obtains final wealth W−P in the no-loss state and W−P−L+I in the loss state. Expected utility is (1−p)u(W−P)+p u(W−P−L+I). Concavity of u provides a conventional account of willingness to pay a loading for risk transfer; reference dependence is not a prerequisite for insurance to have value.

The conceptual change introduced by prospect theory's original formulation (Kahneman and Tversky 1979) is that a decision maker may encode a deviation x from a benchmark r, assigning v(x−r) rather than evaluating only the final asset position. Its common descriptive representation is steeper on the loss side than the gain side, and its curvature differs by domain. This is an empirically motivated functional representation; its parameters and reference level require particular evidence in any application.

Tversky and Kahneman's loss-aversion account (1991) relates asymmetry to changes in initial endowments and to valuation of the status quo. Kahneman, Knetsch, and Thaler's experimental market study (1990) found differences between willingness to accept and willingness to pay for consumer objects under conditions that constrained simple transaction-cost explanations. The subsequent endowment and status quo synthesis (1991) connected these phenomena. The direct experiments were primarily about goods and endowments, not indemnity contracts. Their insurance relevance is therefore inferential: inherited coverage and anticipated protections can become psychologically salient holdings, but the analogy needs direct testing.

A status quo reference point must not be mistaken for one universal benchmark. Holding an existing policy may generate resistance to reducing coverage, while an uninsured individual may regard premium payment as a departure from normal expenditure. The two people face the same actuarial prices but different contract histories. Samuelson and Zeckhauser's status quo study (1988) analyzed disproportionate retention of existing choices and included health-plan selections. It provides institutional relevance while leaving open how much of any particular decision is driven by endowment, information costs, familiar-provider networks, or active preference.

Reference dependence is therefore a family of empirical propositions involving identifiable benchmarks and valuation around them. It cannot be inferred merely from the verbal existence of a loss in an insurance contract. An insured event is economically a loss, a premium is a payment, a deductible is retained risk, and an unpaid expectation is a discrepancy; their psychological status requires observation.

Prospect Theory and the Insurance Domain

For a simplified prospect, let outcomes measured against a reference level be x_i and let decision weights be π_i. A standard illustrative value function is v(x)=x^α for x≥0 and v(x)=−λ(−x)^β for x<0, with α,β>0 and λ>1 representing steeper valuation of losses for matched magnitudes under that parameterization. The value assigned to a prospect is a suitably weighted sum of the v(x_i), not an expected monetary amount. The original 1979 theory used decision weights; the 1992 cumulative representation supports rank-dependent weighting of ordered outcomes and accommodates both gains and losses in multistate problems.

This structure contains two distinct behavioral components: outcome evaluation and probability transformation. In a binary insurance example, one might overweight a small loss probability and purchase extensive coverage even with weak or absent loss aversion. Conversely, a steep value function on the loss side may induce a preference for avoiding a premium if the premium is a certain loss and the uninsured event is represented against a different benchmark. The sign of an insurance-demand prediction depends on where the reference point lies, how the premium enters the accounts, and whether the risk was expected.

The well-known fourfold pattern in Tversky and Kahneman (1992) concerns attitudes over high- and low-probability gains and losses. Applied carefully, it helps explain why small-probability losses can be salient and why risk seeking may occur in some loss-domain comparisons. An insurer cannot infer that customers universally demand complete coverage from the mere presence of loss aversion: a certainty-valued premium may itself be experienced as a loss, a deductible may be treated as an accepted status quo, or a contingent claim may be insufficiently trusted.

A useful distinction is between the economic event and its frame. A $500 deductible represents a monetary obligation only if an insured event produces otherwise-covered costs above the relevant threshold. Advertising it as "saving $500" relative to another deductible introduces an alternative comparison. Calling a claim reimbursement a "gain" may describe the arrival of cash but neglect the simultaneous physical or medical loss and the previously purchased entitlement. A scientifically defensible model specifies the accounting and temporal boundaries of both the events and the frame.

The original insurance framing experiments (Johnson et al. 1993) found decisions that changed under framing manipulations and probability descriptions. Those results support the practical relevance of representation, while the experiments' hypothetical choices and sample/task specificity qualify their extrapolation to premiums paid from real household budgets. Their importance lies in demonstrating a mechanism to investigate, rather than fixing its magnitude across products and countries.

Expectations-Based Reference Points and the Timing of Coverage

A reference point need not be a fixed current asset level. Kőszegi and Rabin (2006) developed a model in which recent rational expectations about outcomes enter reference-dependent preferences. Their framework makes expectations part of the choice environment. In an insurance setting, anticipated coverage may constitute a benchmark before any loss occurs; an unanticipated loss and an anticipated risk with an anticipated coverage decision can then have different utility consequences.

The reference-dependent risk-attitudes analysis (Kőszegi and Rabin 2007) makes this distinction explicit. Its abstract-level predictions differ sharply depending on whether exposure and the possibility of insurance were anticipated: an unexpected risk may reproduce the classical prospect-theory distaste for insuring losses, while planned exposure can generate first-order aversion over small risks. The timing of information is thus analytically important, not a minor presentation feature.

Consider four moments: the customer anticipates future exposure, chooses a policy, experiences either the loss or no loss, and learns the claim result. The premium can be evaluated relative to a spending plan formed before the quote; insurance protection can be evaluated relative to a coverage expectation formed after purchase; claim settlement can be evaluated relative to an estimate formed during adjustment. The reference object can change across moments without implying any inconsistency in the financial contract. Likewise, auto renewal after a long insured history may differ psychologically from first purchase even if expected accident risk is identical.

Expectation-based reference points supply a discipline for making such claims testable. Record elicited forecasts before revealing the choice, vary contract terms while holding physical risks fixed, and observe whether changes in expectations predict valuation. Mere retrospective reports of being "surprised" by a deductible can reflect unfamiliarity with terms, omission of an important exclusion, genuine processing costs, or post-event recall. A study must distinguish these channels.

Temporal sequencing is especially relevant where losses are discovered gradually, as in chronic healthcare expenses, property restoration, and liability disputes. Long claims delays create additional expectations about time and settlement amount. Their costs include discounting, liquidity, and administrative effort; loss aversion may operate on the gap between expected and received payment but must be estimated separately from these objective burdens.

Premiums as Expenditure and Perceived Loss

A premium has three simultaneous economic functions. It is the price paid for contingent claims on the insurer; it finances expected losses, expenses, and capital costs; and it reduces the consumer's present liquid resources. Only the last is directly observed by the consumer at purchase. The consumer may consequently experience premium payment as a certain loss against a preexisting cash budget even when the policy increases expected utility by reducing future financial risk.

Suppose two equivalent quotes differ only in presentation: one describes a $600 annual premium as "pay $600 now," while another describes the same amount as "protect up to $20,000 in covered losses for $50 a month." The second frame emphasizes a contingent benefit and periodic cash flow. These descriptions have identical annual nominal totals but may differ in financing timing, fees, renewability, and coverage. Only after holding such contractual differences constant is a causal attribution to framing warranted. The premise of Johnson et al.'s insurance experiments (1993) was that risk and benefit presentation can change perceived values and purchase choices; it does not justify assuming a universal promotional treatment effect.

An informative analysis separates premium-as-loss coding from ordinary budget constraints. An individual with insufficient cash cannot make a large immediate payment regardless of utility curvature. A household facing alternative expenses can rationally reject coverage when premiums exceed expected risk-transfer benefits. Taxes, premium subsidies, payroll deduction, employer contributions, installment arrangements, auto-renewal, and payment frequency also shift observed demand. Some arrangements change transaction costs and liquidity, not merely perception. The mechanism is therefore identified only if the economic terms are equalized or measured.

The reference level may be last year's renewal premium. A 10% increase from $500 to $550 appears as an unfavorable change to an incumbent, whereas the same $550 initial quote may be evaluated against an entirely different expected budget. This difference is compatible with reference-dependent value, but it is also compatible with learning about inflation, search costs, anticipated future increases, and asymmetric information about underwriting. Price-change experiments require transparent causal separation.

Premiums can additionally be coded in different mental accounts: transport, property maintenance, healthcare, or household protection. The meaning of the account affects attention to competing expenditures. Such categorization remains a hypothesis unless independently measured. The strong conclusion from the research is methodological: the reference point must be established before the premium response can be interpreted as loss aversion.

The relation between willingness to pay and welfare deserves care. A person may accept a small premium loading because of the protective value of insurance under conventional utility, because of probability weighting, or because of relief from worry. Conversely, strong risk aversion need not produce a purchase where coverage is unaffordable, excluded, untrusted, or offered at a prohibitive price. The insured/uninsured comparison is structurally incomplete without the feasible set of offers.

Deductibles as Contractual Thresholds and Psychological Anchors

A deductible defines the first part of a covered loss borne by the policyholder. Let D be a per-loss deductible and L a covered loss before application of exclusions or limits. A standard stylized indemnity is I_D(L)=max(L−D,0). If the insurer offers deductible D_1 at premium P_1 and deductible D_2>D_1 at a lower premium P_2, the premium saving ΔP=P_1−P_2 is certain while the extra retained loss under the high deductible is contingent. For a loss L, that extra cost is min(L,D_2)−min(L,D_1), bounded above by D_2−D_1. A comparison based solely on the deductible gap assumes that the loss is large enough to activate the entire gap.

This economic structure permits multiple benchmarks. A customer with a $250 deductible may regard a move to $1,000 as surrendering established protection; a first-time purchaser may focus on the lower premium of the $1,000 plan; a customer accustomed to a $2,500 deductible may regard $1,000 as added security. The historical contract and the chosen frame generate different perceived changes although the marginal risk transfer between the two offers is the same.

Consider a hypothetical annual premium saving of $180 for moving from a $250 to a $1,000 deductible. If every covered claim exceeded $1,000 and the annual probability of exactly one claim were 10%, the expected additional retained loss would be $75. The actuarial expected saving would be $105 before other costs. A person may nevertheless prefer the lower deductible because of utility curvature, perceived frequencies, probability weighting, liquidity, or a reference-dependent aversion to the additional $750 at claim time. The observed purchase cannot identify a unique mechanism. If claims can occur below $1,000, the distribution of their severities matters; if there can be multiple claims, annual expected amounts depend on frequencies and per-claim terms.

A deductible can also become a mental anchor. A policy advertisement may foreground the "maximum out-of-pocket" difference, even though standard deductible products may also include coinsurance, limits, sublimits, and exclusions. Health insurance especially requires separation of deductible, copayment, coinsurance, and out-of-pocket maximum. Misreading these features can generate apparent loss aversion that is actually a misunderstanding of the payment schedule. Loewenstein and colleagues' health-insurance surveys (2013) documented significant misunderstandings of traditional coverage terms and explored responses to simplified plans.

The deductible should therefore be analyzed as a threshold in the claim-payment function and as a possible benchmark in the decision maker's representation. Conflating the two erases the central scientific question. A reliable experiment fixes the distribution of covered losses, explicitly explains contract terms, and manipulates the psychological comparison independently of actual payouts.

Claim Payments, Entitlements, and the Experience of Nonpayment

Insurance promises contingent compensation conditional on covered events and policy terms. A policyholder who incurs a covered loss may anticipate full indemnity, partial reimbursement, repair services, or a defined benefit. The amount anticipated at reporting time can become a reference point. If the settlement falls below this forecast, the policyholder may experience an additional psychological loss even when the payment is contractually correct. A legal disagreement over exclusions or proof of loss is a separate issue requiring contractual and jurisdiction-specific analysis.

Three quantities must be distinguished: gross physical or economic loss L, legally due indemnity I(L), and anticipated indemnity R. The discrepancy I(L)−R is a candidate gain–loss evaluation, while the net economic loss L−I(L) remains a state-dependent financial outcome. The two can move in opposite directions. For example, a customer may receive $4,000 against a valid $4,000 claim and feel satisfied after anticipating only $3,000; another may receive the same $4,000 but feel a shortfall after expecting $5,000. The change in sentiment need not change the amount due. If the policy terms legally required $5,000, the second case also involves an actual shortpayment and calls for a different assessment.

Settlement timing is equally important. An approved payment delayed for months imposes discounting and liquidity costs. An advance payment can alter the reference level for the remaining settlement, potentially making a later adjustment feel like a loss. But losses following a disaster often also involve fear, relocation, damage to valued objects, and disruption of family life. A monetary loss-aversion coefficient alone cannot capture those dimensions.

Wakker, Thaler, and Tversky's probabilistic insurance study (1997) tested another dimension of promise reliability: a small chance that a nominally insured loss would not be reimbursed. Their survey evidence found strong reluctance to accept probabilistic indemnification; the published abstract reports that many respondents required more than a 20% premium reduction for a 1% nonpayment risk. This is a result from its specified experimental scenario, not an estimate of how all consumers price insurer insolvency or every form of coverage exclusion.

The insurer's promise creates an expectations channel alongside actuarial risk. Credibility of the insurer, transparent claims procedures, prior experiences, and the difference between contractual limits and perceived protection all influence the expected quality of indemnification. Measured claim dissatisfaction may reflect administrative failures, fairness concerns, or reference-dependent evaluation. Their causal separation requires independent records of policy wording, communications, legally due benefits, timing, and ex ante expectations.

A sound scientific study of claims therefore begins with the contract, not the label "loss aversion." The behavioral hypothesis has content when an independently elicited reference point predicts reactions to the same legally valid payment and when the analysis controls the direct welfare effects of delay, disagreement, and uncovered damages.

Insurance Framing, Mental Accounts, and Endowment

The same physical risk may be presented as a chance of incurring an uninsured loss, a chance of avoiding that loss, a premium required to preserve coverage, or a deductible reduction described as a benefit. Each presentation foregrounds different states. Tversky and Kahneman's framing analysis (1981) demonstrated preference reversals with altered descriptions of decision problems. Insurance-specific evidence from Johnson et al. (1993) showed framing effects and distortions in evaluations of premiums and benefits under hypothetical choices.

A framing experiment is compelling when descriptions are semantically equivalent and incentives are controlled. A comparison between a policy promising "90% coverage" and one emphasizing "10% out-of-pocket" must also establish that the benefit calculation really is equivalent over the entire relevant loss distribution. Cost-sharing rules involving thresholds, network restrictions, exclusions, and multiple claim types often defeat superficial equivalence. Insurance language can modify perceived products when its meaning is misunderstood, even if the researcher intended a pure frame manipulation.

Endowment enters through existing policies, guaranteed renewability, employer plans, accumulated no-claim discounts, and familiarity with claim processes. The experimental endowment evidence (Kahneman, Knetsch, and Thaler 1990) concerns consumer goods and cannot automatically be transposed to insurance. An incumbent may rationally value a long-held policy because switching introduces underwriting risk, waiting periods, lost continuity, benefit uncertainty, or administrative costs. Empirical status quo effects must be net of these material differences.

Samuelson and Zeckhauser (1988) documented status quo bias in experiments and decisions involving health programs. Later, Handel's health insurance analysis (2013) measured substantial inertia in an employer setting and emphasized that reducing inertia can change the market composition of risk pools, potentially increasing adverse selection. The direction of the market-wide welfare effect need not match the direction of any individual's improved active choice.

The perception of holding coverage can create a potential loss from cancellation, whereas the perception of paying for an unused service can create regret or dissatisfaction with retention. Either response depends on the account against which the decision is assessed. Long-term observed renewal is accordingly a mixture of preference, default settings, switching costs, information, and risk dynamics. A model should allow these to vary rather than assign every persistence effect to loss aversion.

For editorial presentation, two comparisons are useful: the full cash-flow schedule and a frame-neutral description of protected and retained losses. Replacing vague marketing language with accurate structured examples can help measure comprehension. It is an empirical question whether such disclosure changes participation, contract selection, or welfare; evidence of one outcome does not establish the others.

Probability Weighting, Risk Beliefs, and Contract Reliability

Loss aversion is a feature of outcome valuation. Probability weighting concerns how stated or perceived probabilities influence a choice. Subjective beliefs concern which loss probabilities a person expects to face. These concepts should be estimated separately where possible. A consumer who overweights a known 1% loss probability may prefer a low deductible; another who believes the true probability to be 5% can make the same choice under linear weights. In a simple binary design the two explanations may be observationally equivalent.

Cumulative prospect theory (Tversky and Kahneman 1992) supplies an explicit rank-dependent weighting mechanism. Barseghyan et al.'s methodology paper (2013) studies the identification problem between probability weighting and risk misperceptions. Its methodological discussion emphasizes environments with more than two outcomes and changing ranks across lotteries as useful for separating mechanisms; its insurance illustration uses simulated data rather than purporting to supply direct new field identification for all markets.

In an insurance contract, objective claim frequency is rarely known perfectly to an individual. Frequency may vary with age, geography, occupation, property conditions, driving exposure, health, or climate. Consumers receive signals from personal claims, friends, media reports, and insurer disclosures. Claim probabilities estimated by a researcher from administrative data also contain selection, reporting, and censoring assumptions. Even a perfectly elicited perceived probability need not match a modeler's actuarial benchmark.

Contract reliability is a further independent dimension. Probabilistic insurance (Wakker, Thaler, and Tversky 1997) demonstrates a marked penalty for a small possibility of nonpayment in its study setting. That penalty can involve probability weights, the salience of failure, perceived betrayal of the insurer's promise, or distrust. It should not be confused with the ordinary probability of an insured loss. One probability concerns the hazard; the other concerns performance conditional on the hazard and contract.

To identify loss aversion, a data set should ideally contain elicited beliefs, several lotteries or contract schedules with varying outcome rankings, comparable premium increments, and a reference-point manipulation or expectation measure. Without these, rich models can still fit deductible choices, but parameter interpretation is conditional. A rigorous source-critical conclusion retains that condition visibly in prose and does not collapse every departure from expected utility into a single behavioral diagnosis.

Automobile Deductibles: Structural Evidence and Its Limits

Automobile coverage offers repeated contract menus in which premium differences and deductible levels are visible, and ex post claim histories can help estimate risk. Cohen and Einav (2007) estimated preferences from auto deductible choices while allowing unobserved heterogeneity in claim risk and risk aversion. Their estimates exhibited substantial and skewed heterogeneity. This is evidence about the sample and structural model; it does not establish an invariant population-wide psychological parameter independent of deductible framing and loss-frequency assumptions.

The next major distinction comes from Barseghyan, Molinari, O'Donoghue, and Teitelbaum (2013), who modeled deductible choices allowing standard risk aversion and probability distortions. Their findings indicated a substantial role for overweighting small probabilities and showed that loss aversion alone, as represented in the examined Kőszegi–Rabin specification, or disappointment aversion alone could not reproduce their estimated distortions. The result weakens single-mechanism interpretations and strengthens the empirical importance of probability weighting within their data and model comparisons.

A methodological tension remains. Ex post claims improve measurement of underlying risks, yet claims are not always identical to accidents: policyholders may avoid reporting small losses below the deductible, change behavior when coverage changes, or encounter reporting requirements. A deductible alters both retained losses and sometimes incentives. Claims frequencies observed under different policies can therefore be endogenous to contract choice. Structural models address these issues through assumptions and additional information, not by making them vanish.

Choice sets may also be nonuniform. Different customers encounter different quotes because underwriting and risk classification vary. A model of risky choice must condition on the actual alternatives offered. A customer selecting the lowest deductible from a narrow menu is not equivalent to one offered a much higher deductible discount. Renewal histories, bundles, brokers, and automatic payments further complicate preference inference.

The analytic role of automobile evidence is unusually clear: actual monetary decisions are consistent with strong aversion to certain deductible risks, but their explanatory decomposition remains model-dependent. The empirical program should seek quasi-experimental premium shifts, randomized frame changes, elicited probability beliefs, and within-person data across products. That combination can distinguish price and risk responses, while separate reference-point measurement is still needed to establish loss aversion.

Homeowners Insurance and Modest-Risk Overinsurance

Homeowners insurance permits observation of voluntary deductible protection for losses that may be modest relative to lifetime wealth. Sydnor's homeowners study (2010) documented widespread low-deductible choices at premiums well above the associated expected value, and reported that fitting them through a conventional risk-aversion model implied implausibly high risk parameters. This is an important descriptive and structural result, since the observed choices expose the limits of a purely smooth-wealth-utility interpretation for modest-stakes risks.

The inference needs precision. The study directly concerns real homeowners deductible menus and observed decisions. It does not individually identify loss aversion as the cause. Alternative interpretations include small-probability overweighting, perceived claim frequencies, desire to avoid administrative trouble, liquidity needs, heuristic comparison, and contract comprehension. Some of these may be present simultaneously. Standard expected utility can explain loaded insurance in principle; the issue is the very large curvature required by particular modest-risk selections under its maintained assumptions.

The apparent tension between low-deductible overpurchase and catastrophe nonpurchase is analytically instructive. The same individual may purchase extensive coverage for familiar low-stakes claims while declining flood or earthquake protection, because the menu, reference status, lender requirement, trust, affordability, subjective beliefs, and attention differ. It would be unsound to label either pattern as the single "true" degree of loss aversion. A quantitative model must describe which risks the household perceives and which contracts are available and enforceable.

Homeowner coverage also varies significantly by jurisdiction and peril. Standard policies may exclude certain flood or earthquake risks; special endorsements, public programs, or separate policies can be required. Cross-country comparisons must normalize definitions of insured perils and government assistance. A nominal deductible in one market may have a different exposure meaning in another. The psychological reference point follows the locally available contract and institutional history, not the English word used to describe it.

A robust inference from the homeowners evidence is that field choices provide stringent tests for parsimonious utility models. A comprehensive explanation should jointly consider reference dependence, probability processing, beliefs, and economic constraints, with sensitivity analysis rather than post hoc selection of the favored mechanism.

Health Insurance: Complexity, Defaults, and Choice Architecture

Health coverage links monetary risk to access, treatment decisions, provider relationships, and the timing of care. Its premium can be deducted from wages; the insured can incur a deductible, coinsurance, and copayments; and the coverage network may constrain the feasible treatment set. Any account of perceived financial losses must respect those institutional components. Expected medical out-of-pocket expenditure is a distribution over treatments and health states, not merely the deductible stated on the first page of a policy.

Loewenstein and colleagues (2013) used surveys to study how people understand conventional health plans and evaluate a simpler all-copay design. Their findings support the importance of limited comprehension, yet the authors reported weaker evidence for large corresponding changes in plan appeal or actual healthcare choices. This distinction matters: better comprehension is an outcome in its own right, but it does not entail a particular sign for demand, risk preference, or welfare.

The status quo and choice architecture receive stronger empirical attention in Handel's employer insurance study (2013). The paper used a change in employer coverage provision to estimate substantial inertia and to model the effect of reducing it. The studied setting illustrates a market equilibrium issue: encouraging active choice can improve a consumer's individual match while changing the risk composition of plans, possibly worsening adverse selection. The reference-dependent interpretation of a familiar plan must be compared with switching costs and market-selection externalities.

Ericson and Starc (2016) examined product standardization on the Massachusetts Health Insurance Exchange. Their study linked a reform that standardized cost-sharing parameters and changed presentation to shifts toward more generous plan choices and different brands; decomposition distinguished valuation from availability effects. The evidence demonstrates that changing the architecture of available plans can affect choices without requiring the analyst to assume a change in stable tastes. It is not, by itself, a direct estimate of a loss-aversion coefficient.

A deductible may become a reference for expected expenditure during the insurance year. A household anticipating steady medication costs can understand that it will cross the deductible early; an individual facing uncertain emergency needs may focus on worst-case exposure. A provider network may also be an endowed relationship whose disruption entails real quality and continuity costs. The relevant welfare measure includes health outcomes, treatment access, and financial risk, not just the number of plans purchased or the average premium.

Insurance communication can be evaluated using identical total-cost examples at several utilization levels. Comparison tools should show premiums, typical and high-cost out-of-pocket payments, network differences, and coverage limits in a form the purchaser can understand. Such a design makes the benchmark more explicit. Whether it reduces harmful reference dependence is an empirical question requiring comprehension measures, actual choice behavior, and later outcomes.

Catastrophic Risk, Worry, and the Problem of Nonpurchase

Catastrophic risks are often rare, spatially correlated, difficult to assess, and financially severe. Insurance demand for them can be low despite large possible losses. A simplistic application of loss aversion might predict large protective expenditures, yet real households may have limited liquidity, underestimate exposure, rely on disaster assistance, or not understand exclusions. The reference point may be a recent period without losses, so a catastrophic event is poorly represented as a probable financial contingency.

Schade, Kunreuther, and Koellinger (2012) conducted a high-stakes insurance experiment with rare loss probabilities and ambiguity. Their evidence linked worry to willingness to consider protection and reported a distribution containing both nonbuyers and participants willing to pay substantial amounts. This heterogeneity is significant: emotional concern and probabilistic judgments can change the initial decision to consider insurance as well as the willingness to pay conditional on consideration. The research setting is experimental; the distribution of household purchasing in a particular national catastrophe market requires separate field evidence.

Kunreuther and Michel-Kerjan's analysis of homeowners' disaster decisions (2009) discussed behavioral influences and institutional constraints surrounding protection against natural hazards. Such risks provide an important counterweight to universal explanations. A policyholder may strongly dislike paying a recurring premium while assigning little perceived probability to an exceptional event. Another may pay high premiums to avoid the anticipated emotional burden of a loss. Both can be rationalized by various behavioral specifications; the precise mechanism cannot be identified from purchase status alone.

For public policy, the reference-dependence question interacts with risk mapping, warnings, lender requirements, subsidies, and post-disaster support. Mandatory insurance may solve one participation problem but creates questions of affordability, risk classification, distribution, and enforcement. Premium subsidies can promote access while shifting incentives and fiscal burden. Policy design should evaluate material protection and long-run financial exposure rather than maximizing take-up as an isolated goal.

A severe hazard also creates uncertainty about the insurer's ability to pay promptly, especially when many claims occur at once. Perceived nonpayment can diminish the value of promised protection and can interact with the uncertainty about the loss itself. This mechanism is conceptually related to probabilistic insurance but cannot be estimated simply by applying the numerical results of Wakker, Thaler, and Tversky (1997) to modern catastrophe markets.

Measurement and Identification: What Would Establish Loss Aversion?

Behavioral inference must be based on observed contrasts that identify a particular latent mechanism. A preference for comprehensive coverage is compatible with numerous models. An estimate of loss aversion requires a value function, a specified reference point, independent information about beliefs and probability weighting, and an empirical design with sufficient variation to separate these variables.

First, measure the economic choice environment. Record the actual quotes offered, coverage and exclusions, premium payment schedule, deductibles, limits, underwriting restrictions, expected loss distributions, wealth proxies, and liquidity. Record renewal history and whether a default option was assigned. If risk type and offered prices are correlated, the sample of chosen contracts alone cannot identify willingness to pay or behavioral primitives. The general analysis of such selection and welfare issues in Einav, Finkelstein, and Levin (2010) provides a methodological foundation.

Second, elicit the benchmark before the decision or intervention. Questions can measure anticipated premiums, expected claim reimbursement, current deductible, expected household expenditures, and beliefs about loss frequencies. Pre-treatment measurement limits retrospective rationalization. It does not eliminate elicitation effects; the act of asking about a loss may itself make it more salient. Researchers should randomize elicitation order or compare elicited and unobtrusive conditions where practical.

Third, randomize framing separately from contractual content. A study could present identical expected net payments with the premium as an annual payment, a monthly equivalent, and an opportunity cost, while holding installments, fees, and eligibility fixed. A second randomization could change whether the low deductible is the default. A third could present identical claim payments against different experimentally induced expectations. Estimation of framing effects and status quo effects is possible under randomized assignment; identification of loss aversion remains dependent on the validity of the link from treatment to reference-point changes.

Fourth, elicit beliefs under varied outcome rankings. The identification discussion in Barseghyan et al. (2013) makes clear why simple binary choices can confound risk misperception and probability weighting. A factorial design including multiple contract states, varied loss probabilities, and more than two relevant outcomes can test which mechanisms fit across tasks. In a full field implementation, elicited beliefs should be compared with claims records, while accounting for the fact that claims experience is affected by deductibles and reporting thresholds.

Fifth, observe actual behavior and later outcomes. Hypothetical willingness to pay can illuminate cognitive mechanisms, but real renewal decisions, contract lapses, filing behavior, claim satisfaction, healthcare use, and financial protection are essential for external validity. The original Johnson et al. framing studies provide an early example of experimental representation effects, while Cohen and Einav's deductible analysis illustrates what can be learned from observed contracts under structural assumptions. Neither design alone resolves all the other's limitations.

Sixth, evaluate estimation stability. Report likelihood or fit metrics, parameter confidence regions, sensitivity to utility curvature and belief assumptions, alternative reference-point definitions, exclusion restrictions, sample selection, and the extent to which preference heterogeneity is estimable. A single fitted λ value can look precise while depending materially on assumed probability weights or reference points. Sensitivity analysis should be a substantive part of scientific interpretation rather than a postscript.

Finally, design replications across insurance lines and countries. Definitions of coverage, social insurance, medical cost sharing, claims obligations, distribution channels, and disaster financing differ. Pooling populations without harmonizing these institutions risks estimating a composite that lacks a common behavioral interpretation. The proper generalization is conditional on equivalent risks, offers, references, and institutional structures.

Formal Illustrations, Comparative Statics, and Robustness

This chapter uses stylized algebra to test internal consistency and to show why apparent paradoxes arise. The equations are reconstructions from established expected-utility and prospect-theoretic representations, not a newly introduced proprietary model. Each example is conditional on its specified assumptions.

A two-contract deductible comparison

Assume exactly one period and either no loss or a fixed covered loss L larger than both deductible levels. Contract A has low deductible d_A and premium P_A; contract B has high deductible d_B>d_A and premium P_B<P_A. Define the annual premium saving s=P_A−P_B>0 and incremental deductible exposure h=d_B−d_A>0. Relative to A, the monetary payoff from choosing B is +s if there is no claim, and s−h if the claim occurs. If probability of claim is p, the expected monetary advantage is s−ph. Thus choosing B raises expected wealth when s>ph. This equality is an accounting result under the stated assumptions; it does not say which contract maximizes expected utility.

For a purely illustrative example let s=$180, h=$750, and p=0.10. The expected advantage of B is $105. A risk-neutral chooser with no transaction effects would choose B. A risk-averse chooser can prefer A; a decision maker who places unusually high decision weight on the $570 claim-state disadvantage can also prefer A. A model that encodes the premium as a current loss and the extra claim expense as a prospective loss may generate still another ranking.

Under a simple illustrative two-outcome reference valuation that compares B against A and takes the A outcome as a constant benchmark, with linear branches v(x)=x for x≥0 and v(x)=λx for x<0, the signed score is (1−p)s−pλ(h−s) when h>s. For the example, this becomes 162−57λ and changes sign at λ=162/57≈2.8421. This threshold has meaning only for this reference choice, these linear branches, actual probability weighting, and a decision rule selecting the positive score. It is not a population estimate and is not an invariant property of the contracts. Changing the benchmark or decision weights changes the threshold; introducing conventional consumption utility changes it again.

Changing the reference point

The last calculation treats A as the reference outcome. Suppose instead the household regards B as normal, so choosing A is the deviation. Relative to B, A gives −s without a claim and h−s with a claim. A simple linear loss-averse score is −(1−p)λs+p(h−s). In the example that is −162λ+57, which is negative for λ>1. Thus a fixed λ and the same two contracts can predict retention of either option when the current option is made the reference, depending on how the comparison is specified. This is a conditional illustration of status quo sensitivity; it establishes that reference-point selection is essential to model predictions.

A complete expectations-based model requires a consistent reference distribution rather than arbitrarily assigning the desired contract after the fact. The Kőszegi–Rabin model offers a theoretical approach to endogenous references, and their risk-attitudes extension emphasizes the role of anticipated choices. A valid empirical specification must implement that logic rather than freely switch benchmarks to explain each outcome.

Probabilities and beliefs

Replace p with a perceived probability q in the preceding score, or, in an appropriate cumulative weighting representation, replace the probability terms by model-consistent decision weights. The same observed choice can be produced by a larger loss multiplier λ, a larger perceived probability q, a stronger weighting of the claim state, or combinations of these. Consider the loss-averse comparison with p fixed at 0.10: increasing q strengthens the disadvantage of B without changing λ. This algebra makes explicit the observational equivalence motivating Barseghyan et al.'s field-identification discussion.

The effect of adding states is not a cosmetic complication. If loss severity varies and may fall below one deductible, h is replaced by the state-contingent difference min(L,d_B)−min(L,d_A). Under expected monetary value, the premium advantage remains s while the expected additional retained loss is E[min(L,d_B)−min(L,d_A)]. Under cumulative prospect theory, each ordered payoff can receive rank-dependent weights. Researchers must check that their payoff ordering, probabilities, and reference differences match the specification being estimated.

Dynamic contract history

If a policy renews, the current deductible can become an expectation for next year. A premium increase may be encoded as a loss relative to last year's bill, while the policy's continued protection is experienced as a retained entitlement. Let the absolute premium increase be ΔP. The objective payment rises by ΔP; reference-dependent displeasure additionally depends on the expectation the consumer formed before the renewal quote. A retention decision confounds such displeasure with updated claims experience, switching costs, underwriting differences, and future premium beliefs. Longitudinal field records improve measurement, but identifying the psychological mechanism still requires pre-quote expectations or experimental variation.

The actuarial interpretation

Expected retained loss, premium, and insurer obligation remain economic quantities that should be reconstructed from the actual contract. A high willingness to pay for coverage can coexist with positive expected net cost because the value of reduced financial risk is distinct from expected indemnity. If a covered loss is correlated with poor health, disrupted earnings, or temporary illiquidity, state-dependent utility further alters the demand for protection. These ordinary welfare channels remain relevant whether or not preferences are reference-dependent. The formal exercise demonstrates a central evidentiary requirement: comparisons of model fit should include appropriately specified economic alternatives rather than a deliberately weak expected-value baseline.

Insurance Welfare, Regulation, and Consumer Communication

Welfare analysis asks whether insurance choices improve ex ante well-being given risks, prices, information, contracts, and institutions. Reference-dependent preferences complicate the question because the reference point may itself depend on choice architecture, expectations, and previous exposure. A policy that increases a measured take-up rate can still create excessive premium burdens, distort pooling, or leave important hazards uncovered. Likewise, a low-deductible choice can be expensive and yet provide genuine liquidity or peace-of-mind benefits not captured by a narrow expected-claim calculation.

One practical implication is that consumer communication should make the full distribution of contract payments visible. A comparable table of premiums, out-of-pocket payments at representative loss sizes, coverage exclusions, claims limits, and the likelihood assumptions can reduce the gap between an advertised anchor and the actual contract. The recommendation follows from the evidence of misunderstanding and frame sensitivity; its effectiveness should be tested prospectively rather than asserted as a universal remedy.

The G20/OECD financial consumer protection principles, amended in 2022, recognize the role of behavioral insights, product testing, fair treatment, and suitable consumer outcomes in financial markets. Their relevance to insurance is institutional. They provide a governance rationale for asking whether consumers understand products and whether choice environments produce appropriate outcomes, while not prescribing a single psychological theory of insurance demand.

The EIOPA Consumer Trends Report 2024, released in January 2025, addresses perceived value for money, digital distribution, and potential exclusion. EIOPA's 2025 Eurobarometer reports EU-wide survey results on insurance ownership, distribution, and trust. Such indicators illuminate consumer experiences and institutional heterogeneity; they are not randomized measures of loss aversion. Their value lies in describing contexts in which perceived costs, comprehension, and expectations may matter.

Market-wide effects deserve additional care. Handel (2013) shows how interventions that reduce inertia can affect selection and welfare in an employer health insurance market. Market equilibrium analysis must therefore distinguish improved individual information from changes in premiums and risk pooling. Einav, Finkelstein, and Levin (2010) formalize why demand, insurer costs, and selection are essential for welfare evaluation. Behavioral evidence should be integrated into those economic relationships rather than used as a substitute.

A consumer-oriented communication regime can report total annual costs under multiple clinically or physically relevant scenarios, explain claims examples, make exclusions comparable, and disclose how automatic renewals and defaults operate. Supervisors can evaluate comprehension and subsequent household outcomes using pre-registered randomized designs. A welfare claim requires specifying the reference standard: realized consumption, informed ex ante preference, protected essential expenditure, or another explicit criterion. The psychological frame alone does not resolve the normative choice.

International Scope and Evidence Boundaries

The canonical scientific horizon is international, but most strongly identified field estimates cited here arise from particular U.S. insurance markets, and the foundational theory experiments were conducted in narrower research populations. The Danish field risk-attitudes experiment (Harrison, Lau, and Rutström 2007) provides a contrasting empirical setting for elicited risk attitudes; its experiment is not itself an estimate of insurance-specific reference dependence. The European insurance consumer survey provides broad descriptive geographical coverage yet does not offer individual-level randomized loss-aversion identification. Combining these sources creates contextual triangulation, not a pooled causal coefficient.

Several mechanisms can vary by institutional environment. Social insurance reduces the range of private financial shocks; compulsory automobile coverage changes participation margins; catastrophe assistance changes the ex ante cost of going uninsured; high medical cost-sharing interacts with provider prices and network rules; insurance supervision and guarantees affect credibility. Cultural and linguistic differences may affect how terms such as "deductible," "excess," "co-payment," and "indemnity" are understood. These factors can move both objective contract values and psychological reference points.

Cross-country studies should predefine the intended estimand. Is the question whether a premium frame changes immediate take-up within a country, whether the same mechanism operates in two regulatory systems, or whether a structural loss-aversion parameter is transportable across countries? Each requires different harmonization and different assumptions. International comparison also requires common timing: nominal premiums, annualized installment costs, inflation, purchasing power, public subsidies, and claim horizons must be made comparable.

The available literature supports the relevance of reference dependence and framing as candidates for understanding insurance decisions. It does not establish a single global prevalence, a common loss-aversion parameter across insurance lines, or an invariant response to deductible changes. A comprehensive international research program would combine contract microdata with harmonized randomized framing protocols and prospective reference elicitation. Until then, external validity remains conditional and should be reported as such.

Integrated Scientific Synthesis

The literature yields five interconnected findings. First, insurance is a sequence of temporally separated cash flows and entitlements; premium, deductible, insured loss, and indemnity can serve as distinct reference objects. Second, reference-dependent valuation and probability weighting are analytically different. Third, the same observed deductible purchase can arise from several mechanisms. Fourth, contract history and default settings influence the environment in which preferences are expressed. Fifth, welfare cannot be recovered from choice frequencies without measuring contracts, prices, risks, and the effects on other policyholders.

The clearest directly insurance-specific causal evidence is that frames can affect hypothetical insurance choices in Johnson et al. (1993). The probabilistic insurance experiment (Wakker, Thaler, and Tversky 1997) establishes notable sensitivity to a small specified chance of nonpayment in the survey context. The homeowners field analysis (Sydnor 2010) finds modest-stakes coverage choices difficult to reconcile with ordinary smooth utility curvature alone, while automobile deductible structural evidence (Barseghyan et al. 2013) places significant explanatory weight on probability distortion and rejects some single-mechanism accounts within its tested models.

Those findings should be read together. Experimental frames establish that presentation can matter; field data establish that the phenomena are economically relevant; structural restrictions illustrate which explanatory channels fit a particular set of observations. None licenses the inference that every insurance purchase manifests a known degree of loss aversion. The appropriate causal hierarchy is: establish actual contract economics; measure beliefs and expectations; identify frame and reference variations; estimate alternative preference mechanisms; test model robustness; and evaluate welfare with market responses.

The formal illustrations also identify a fundamental issue in interpretation. Taking the current insurance contract as the benchmark can reverse the sign of a choice comparison without changing the actuarial terms or the assumed steepness of loss valuation. That logical fact follows from the specified reference-dependent representation. It is not offered as a new theory; its contribution is to make explicit the assumptions that must be carried into empirical studies and policy evaluation.

This synthesis supports a research agenda focused on experimental manipulations of ex ante reference points, multi-outcome contract menus, real premium expenditures, verified claim outcomes, and longitudinal measurements. It also supports a consumer-protection agenda centered on accurate contract comparisons and demonstrated comprehension. Both agendas require distinguishing psychological causes from genuine economic disadvantages created by bad coverage, inadequate affordability, or weak contractual performance.

Conclusion

Reference dependence in insurance is the evaluation of contract outcomes against salient or expected benchmarks. Loss aversion is an asymmetric valuation of unfavorable and favorable deviations from those benchmarks under a specified representation. Insurance presents multiple benchmarks over time: premiums against anticipated expenditure, deductibles against retained-loss expectations, claim payments against anticipated indemnity, and policy changes against existing protection. A meaningful explanation must identify which benchmark is operative and how it interacts with probability beliefs, weighting, liquidity, utility curvature, contracts, and market institutions.

The foundational theories provide coherent descriptions and testable conditions. Insurance experiments support framing and nonpayment sensitivity; field studies document substantial deductible-choice heterogeneity and the importance of probability distortion. Health-plan research reveals the separate influence of comprehension, defaults, and product design. Catastrophe experiments demonstrate the coexistence of high willingness to pay and complete nonpurchase, consistent with heterogeneous beliefs and concern. The collective evidence favors careful mechanism separation and discourages reduction of complex insurance behavior to one preference parameter.

The original contribution of this Research Dossier is the integrated, insurance-specific analysis of reference points across the full contract lifecycle, the explicit worked comparisons of deductible choices under alternative benchmarks, and a structured account of what experimental and revealed-choice data can actually identify. These are analytical interpretations and conditional deductions from established science. No new named theory, conceptual system, or unverified scientific priority is asserted.

Future high-quality research should record expectations before choice, compare objectively equivalent contracts, vary framing and default status independently, measure beliefs and actual payouts, and test how results travel across institutional settings. Welfare evaluation should include the full distribution of financial outcomes, the quality of contractual performance, and consequences for risk pooling. These requirements provide a precise scientific basis for interpreting reference dependence and loss aversion in insurance.

References