Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Insurance Deductible Choice and Behavioral Risk Preferences

Abstract. An insurance deductible converts part of a contingent loss into a household's own financial obligation. Observed deductible choice consequently appears to reveal willingness to bear risk, yet its interpretation depends on the menu of premiums, the loss process, private information, contract knowledge, liquidity, attention, and the behavioral transformation of probabilities and outcomes. This monograph examines the theoretical foundations, field evidence, experiments, identification strategies, and welfare implications of deductible selection. It distinguishes conventional curvature of utility for wealth from nonlinear probability weighting, subjective beliefs, reference dependence, insurance literacy, and the persistence of prior choices. The central empirical lesson is that low deductibles often command a substantial premium relative to their expected incremental indemnity, while the economic mechanism behind that willingness to pay remains underidentified in ordinary single-choice data. Structural studies of automobile and homeowner insurance document heterogeneous and context-sensitive inferred attitudes; evidence from health insurance shows that financial dominance, confusion, and inertia can defeat simple revealed-preference readings. Evidence from flood policies and European health systems supplies further tests and limits to generalization. An original analytic comparison establishes exact break-even conditions for deductible reductions, explains why claim-size distributions matter, and develops a source-grounded hierarchy of evidence for interpreting observed choices. The conclusions concern identifiable claims and methods rather than a new behavioral theory. No proprietary insurance microdata were obtained or reestimated, and cross-market results are presented within their observed institutional settings.

Keywords: Insurance deductibles; behavioral risk preferences; risk aversion; probability weighting; subjective risk beliefs; revealed preference; automobile insurance; homeowner insurance; health insurance; flood insurance; adverse selection; insurance literacy; liquidity constraints; choice inertia; identification; consumer welfare.

Introduction

A deductible is the portion of a covered claim that remains with the insured, subject to the wording of the contract. Choosing a smaller deductible generally buys a larger indemnity in some claim states at the price of a higher premium. Its economic content is a comparison between money paid with certainty and money paid under specified loss contingencies. Its behavioral content includes what a person believes those contingencies to be, how the contract is understood, and how the decision enters the household's financial and psychological organization.

The significance of deductible choice extends beyond the insurance product. Empirical economics has used actual deductible menus to recover distributions of risk aversion. Cohen and Einav's deductible-choice estimation demonstrates the possibility of joint estimation of claim risk and preference heterogeneity from automobile contracts and subsequent claims. Sydnor's (Over)insuring Modest Risks reveals the implausibly strong curvature of utility that a conventional expected-utility interpretation may require for modest homeowner losses. The later Nature of Risk Preferences study finds that distorted treatment of small probabilities is important in explaining a sample of automobile and home deductible choices. These results make deductible choice a productive research instrument, provided that the interpretation respects its identification conditions.

The object of this dossier is the behavioral interpretation of choices among specified deductible-premium combinations. Contract engineering, optimal indemnity schedules, aggregate pricing, and reserving are important neighboring subjects but have different primary objects. Here the unit of analysis is the person or household facing a feasible insurance menu. Comparison across automobile, property, flood, and health insurance is used to identify what travels across settings and what depends on institutional detail. A deductible attached to annual health expenditure is economically different from an occurrence deductible attached to a single property claim; a percentage deductible or catastrophe threshold introduces further nonlinearity. Such differences enter the analysis before a behavioral parameter is inferred.

The research method is an analytic synthesis of identified primary journal publications, original working papers where relevant, and directly reported empirical results. Formal implications below follow from explicit hypothetical contracts and probability assumptions. They are mathematical illustrations, not reproduced insurer estimates. Published findings are attributed to their own populations and models; the ability to observe an association is not automatically an ability to identify a causal preference. Bibliographic identity and publicly available abstracts were checked at publisher or institutional records. Detailed author-text inspection was possible for the Barseghyan et al. (2013) article and some open research materials; replication packages, restricted insurer microdata, and paywalled proof details were not independently reproduced.

Deductible Menus and Insurance Choices

The contract as a state-contingent transfer

Let a covered loss be a nonnegative random variable L, let d be a fixed per-occurrence deductible, and let P(d) be the corresponding certain premium. For a simple full-indemnity contract without a binding limit, the household's total outlay attributable to the insured event and the policy equals C(d,L) = P(d) + min(L,d). This identity requires a paid, eligible claim and abstracts from exclusions, claim-handling expense, income tax, recovery time, and insurer default. A contract with limits, copayments, coinsurance, or annual accumulation rules has a different outlay function. The premium itself may reflect expected claims, transaction costs, policyholder characteristics, market power, distribution expense, regulation, and other costs. Thus observed menu prices are central evidence, not a dispensable control.

For two options d_L < d_H, denote the extra premium for the lower deductible by ΔP = P(d_L) − P(d_H). The incremental indemnity delivered by the lower deductible in state L is B(L) = min(L,d_H) − min(L,d_L). Its expected amount is E[B(L)]; the additional loading paid for the deductible reduction is ΔP − E[B(L)]. If the loss is always either zero or at least d_H, with probability p of the latter, E[B(L)] = p(d_H − d_L). Otherwise the entire severity distribution between zero and d_H matters. The expected cost advantage of the high deductible is ΔP − E[B(L)]. This comparison is a financial expectation, not a normative judgment about the household's utility, liquidity, contractual risk, or beliefs.

The choice is therefore informative only conditional on the options the person could actually select. If one policyholder receives a $90 premium difference and another receives $180 for the same deductible change, identical choices encode different price tradeoffs. If one person lacks cash for a $1,000 emergency payment, identical loss beliefs can coexist with a higher willingness to pay for cash-flow protection. If one believes a claim is likely and another believes it improbable, choices may differ without a difference in utility curvature. An analysis of risk preferences must reconstruct premium quotes, deductible increments, relevant risk exposures, and information available at the decision date.

Menu geometry and marginal willingness to pay

Multiple deductible levels provide more information than a binary comparison. Imagine a menu with $250, $500, and $1,000 deductibles. The willingness to pay for the first reduction, $1,000 to $500, need not equal willingness to pay for the second reduction, $500 to $250, per dollar of contingent benefit. Under a known loss distribution, the second comparison probes sensitivity to the composition of losses and the marginal utility of consumption in different states. Under a behavioral model, it also probes how the household represents a sequence of prospective out-of-pocket losses.

The 2013 structural insurance study explains why this menu geometry helps separate a coefficient of wealth-based risk aversion from a generic probability-distortion function under its maintained assumptions. The same study reports 4,170 households in a core sample drawn from a larger U.S. insurer dataset, and observes auto collision, auto comprehensive, and home all-perils deductibles, their menus, and claims histories. The smaller core sample is a selected analytical sample; it must not be described as a random cross-section of all American households. The distinction between the whole insurer record and the estimating sample matters for external validity.

A deductible menu is also an instrument of price discrimination and selection. Insurers choose prices under expected claim costs and regulations, and customers sort into them on private information. Different deductible options can attract people with different probability distributions of claims, different anticipated repair behavior, and different preferences. The Cohen–Einav structural analysis models unobserved variation in claim rates and risk aversion together, rather than attributing every deductible selection to one observed demographic characteristic. Subsequent claims are informative about objective risk, although observed claims can themselves change with deductibles through reporting and utilization behavior.

Institutional forms and comparisons

Automobile collision deductibles typically attach to covered incidents and may differ from comprehensive damage in both event rate and severity. Homeowner deductibles are usually embedded in wider property protection; catastrophe endorsements, replacement-cost conditions, and specific peril provisions complicate straightforward comparisons. Annual health insurance deductibles accumulate over care episodes, and the household's optimal expenditure depends on expected use, health status, other cost sharing, provider networks, and a maximum out-of-pocket boundary. In flood insurance, deductible and coverage limit together determine retention at small and large loss levels. Research comparing these markets must place each choice on its own payoff map before estimating psychological preferences.

A useful analytical distinction is the deductible's contract definition, the effective deductible in the relevant claim state, and the deductible as the consumer imagines it. Confusion can arise when a plan's displayed deductible is treated as if it were the entire maximum liability, or when exclusions and annual payment caps are overlooked. The actual out-of-pocket distribution may be easy to compute for researchers yet difficult to understand from a consumer's menu. Evidence from dominated health-plan choices makes this gap empirically consequential: observing the selected nominal deductible alone cannot determine whether the consumer understood the available state-contingent alternatives.

Risk Aversion and Risk Premiums

Expected utility and exact choice conditions

Under expected utility over final wealth W, a decision-maker chooses d to maximize E[u(W − C(d,L))], where u is increasing and concave. With a binary event of probability p and loss at least d_H whenever an event occurs, the lower deductible is preferred when (1−p)u(W−P_L) + p u(W−P_L−d_L) is at least (1−p)u(W−P_H) + p u(W−P_H−d_H). Here P_L and P_H are the respective premiums. Every parameter affects the comparison; there is no context-free inference from the label "low deductible" to a particular level of risk aversion.

For a risk-neutral household, expected utility reduces to minimizing P(d) + E[min(L,d)]. In the binary high-severity setting, the low deductible is preferred precisely when ΔP ≤ p(d_H−d_L). The risk-averse household may rationally pay more than this expected incremental benefit because it trades certain expenditure for a reduction in variable expenditure. The difference between its maximum acceptable ΔP and the expected benefit is a willingness to pay for risk reduction, sometimes called a risk premium in this incremental comparison.

For small wealth risks, a local expansion relates the risk premium approximately to one-half of absolute risk aversion A(W) = −u''(W)/u'(W), multiplied by the variance of the retained random cost. The approximation requires that the stakes be small relative to resources and that the relevant differentiability and higher-order remainder assumptions be credible. It cannot be applied mechanically to large losses, constrained cash flows, or discontinuous out-of-pocket schedules. Furthermore, it estimates curvature at the household's effective wealth position, which is frequently measured poorly in administrative insurance data.

A transparent numerical comparison

Consider a synthetic annual quote with P_L = $250 for d_L = $250, and P_H = $160 for d_H = $1,000. Suppose an eligible event produces a loss of at least $1,000 with probability 0.06 and no loss otherwise. The lower deductible costs $90 more with certainty. Its expected incremental indemnity is 0.06 × $750 = $45. A risk-neutral household with these beliefs spends $45 less on average with the higher deductible. The contract comparison yields a break-even claim probability of $90/$750 = 0.12; it does not yield a recommendation for any actual purchaser.

Now keep the same premiums and event probability but suppose each event loss is $600. The deductible reduction then reimburses only $350 in the event state: min($600,$1,000) − min($600,$250) = $350. Expected incremental reimbursement is $21, and the risk-neutral break-even event probability is 90/350 ≈ 0.2571. A computation based only on the difference in stated deductibles would overstate the economic coverage purchased by more than a factor of two in this example. Adding more loss states reinforces the same point. The correct benchmark is a distribution of incremental benefits, not a single advertised difference.

Both examples are independent arithmetic exercises. Their value is diagnostic: the economic attractiveness of a deductible reduction changes with the severity distribution even when the deductible labels, prices, and claim incidence remain fixed. They also reveal why the degree of risk aversion inferred under a wrong severity assumption can be distorted. A rigorous empirical study requires the joint distribution of event frequency, severity, payment eligibility, and the actual menu.

The calibration problem for modest stakes

The appeal of expected utility is its coherence as a choice framework, yet extreme implied curvature over small stakes raises an internal scale problem. Rabin's risk-aversion calibration theorem establishes that substantial expected-utility aversion to modest gambles, if encoded in a globally concave utility of wealth, can imply implausible aversion to much larger stakes. This is a theoretical cross-scale implication, not proof that individuals are risk-neutral or that small-loss insurance is worthless. It asks whether the same estimated wealth curvature can survive coherent evaluation across stake sizes.

Sydnor's homeowner deductible evidence applies this concern to actual insurance purchases. In his sample many consumers paid premiums for low deductibles that considerably exceeded the expected additional insurance protection; conventional expected-utility fit required very large risk-aversion parameters. This evidence has sometimes been summarized as "overinsurance," but a careful interpretation keeps the term tied to a specified financial benchmark. A household may gain from reduced payment volatility, liquidity assurance, simplicity, or subjective peace of mind, even when the additional premium exceeds expected reimbursement. The challenge is empirical attribution: which component is valued, and can it be measured separately?

Curvature estimates recovered from one small-stake deductible decision should therefore be labeled model-implied local parameters. Calling them stable personal traits requires independent support from other menu positions, risk scales, and coverage domains. The cross-context insurance test explicitly rejects stable preferences under its objective expected-utility and Poisson-claim assumptions in its sample, finding that many households appear more risk averse in homeowner than automobile deductible decisions. A model's rejection does not by itself tell us whether preferences genuinely change, whether beliefs differ, or whether its maintained assumptions are restrictive; the rejection identifies the limits of that interpretation.

Liquidity, precaution, and utility of financial certainty

The ability to pay a deductible after a loss is a household budget fact. When wealth is locked in housing, income is variable, or short-term credit is expensive, a reduction in contingent out-of-pocket costs may be valuable because it preserves consumption at a difficult moment. This is compatible with expected utility, but it requires modeling access to resources and possibly state-dependent marginal utility. A household's current cash balance, liquid saving, credit limits, and co-occurring shocks can matter more than annual income alone.

Health and property losses may covary with lost earnings, displacement, or other expenses. A dollar owed after such a loss can be more consequential than a dollar of premium paid before it. A conventional model that assumes fixed background wealth and independent losses may incorrectly allocate that value to exaggerated curvature of utility. Similarly, repair effort and claim filing costs can make some deductible options effectively more cumbersome. These are empirically testable contract and resource mechanisms, rather than evidence of any new category of risk preference.

The descriptive question must thus distinguish an individual's taste for risk, ability to finance a loss, perceived risk frequency, and the actual contingent consequences of the contract. All can affect the same binary observed decision. The rest of the dossier examines how behavioral economics and research design disentangle them.

Probability Distortions

Probability weighting and subjective event beliefs

An objective annual claim probability is a property of an estimated loss process; the probability actually used by the household may differ. Behavioral decision theory offers at least two distinct accounts. The individual may accurately believe that the probability is p but apply a nonlinear decision weight w(p); alternatively, the person may behave as if the underlying probability were q ≠ p. For a binary covered event and a restricted set of observed prices, these mechanisms can generate the same choice likelihood. Distinguishing them requires information beyond the deductible label and selection outcome.

Kahneman and Tversky's original prospect-theory analysis introduced reference-dependent valuation and decision weights as explanations of systematic departures from expected utility. Tversky and Kahneman's cumulative prospect-theory formulation extended the framework to ranked outcomes and allowed different transformations for gains and losses. Its empirical discussion includes relatively strong aversion to low-probability losses, a pattern relevant to insurance. The implication for deductible choice is a changed valuation of the chance of paying the retained loss; the transformation must still be estimated, and its identity is not established by mere observation of a low deductible.

In a stylized binary model, suppose a household compares a certain premium ΔP with a contingent benefit b of avoiding an additional out-of-pocket loss. A risk-neutral expected-value comparison uses pb. A simple distorted-probability comparison uses w(p)b, subject to the chosen model's complete payoff and normalization rules. If w(p) exceeds p for a small p, the household can rationalize a premium above the actuarial incremental benefit while exhibiting little conventional curvature of wealth utility. The same apparent choice could arise if the household sincerely believes that the event probability is q > p, or if it faces a cash-flow loss that gives b a greater marginal value in the event state.

An empirically estimated "probability distortion" is therefore initially a reduced-form wedge between objective actuarial risk and risk as priced by the decision model. A psychological claim that the person literally misperceives frequency is stronger. Likewise, a claim that the person possesses a stable probability-weighting function across products is stronger. The mapping from estimated wedge to mechanism depends on the experimental variation and exclusion assumptions.

What deductible menus can identify

Barseghyan, Molinari, O'Donoghue, and Teitelbaum's structural estimates of probability distortions jointly allow conventional risk aversion and a generic distortion of event probabilities. Their 2013 analysis uses more than two deductible options and variation in premiums and claim likelihoods. Under their estimated baseline, the paper reports an effective weighting of approximately 0.08 for objective probability 0.02, 0.11 for probability 0.05, and 0.16 for probability 0.10. These are fitted quantities in the authors' sample and model, not transferable population-level beliefs. Their findings are robust to several forms of heterogeneity specified in the article. The authors also acknowledge that their estimated distortion is observationally indistinguishable from systematic misperception of claims probabilities within that data structure.

This last qualification governs how the study should be cited. The empirical model provides evidence that an outcome-weighting or subjective-risk wedge is consequential and that a simple standard expected-utility account fits poorly. It provides further restrictions that do not align with certain loss-aversion-only or disappointment-aversion-only specifications. It does not directly observe a person's private probability belief and thereby settle every psychological explanation. The distinction between reduced-form estimation and structural interpretation is a central scholarly safeguard in behavioral insurance research.

A related identification analysis in field data demonstrates, in theory and simulated deductible examples, how probability weighting and misperception can be distinguished when lotteries have more than two outcomes and their outcome rankings vary across choices. This is a strategy based on informative variation in the ranking of payoffs; it is not an empirical finding that all current deductible datasets contain such variation. Insurance menus frequently have nested coverage benefits with relatively stable orderings, so an identification argument that depends on crossings of ranked outcomes must check the actual menu's rank structure. Simply adding more deductible amounts does not automatically meet the identification conditions.

Framing, reference points, and insurance language

Behavioral valuation changes when identical or closely related insurance consequences are described as losses, savings, liabilities, or protection. The experiments of Johnson, Hershey, Meszaros, and Kunreuther on framing and probability distortions examine hypothetical insurance decisions and report systematic framing effects and deviations from basic probability-consistent choice. The results establish that presentation can affect choices in experimental circumstances. The authors also discuss market illustrations, but hypothetical experiments alone cannot yield causal effect sizes for commercial underwriting populations.

A lower deductible may be portrayed as eliminating a feared payment after a loss. A higher deductible may be portrayed as saving a certain annual premium. Both descriptions refer to the same tradeoff, but one directs attention to the loss state and the other to a certain expense. If decision-makers assign premiums to one mental budget and emergency payments to another, the tradeoff can become less transparent. Reference points may be a previous policy, the quoted default, the expected household budget, or an imagined fully insured state. Distinct representations can induce different marginal willingness to pay with an unchanged objective loss distribution.

These mechanisms must be tested with controlled presentation changes that preserve the underlying financial payoffs. A difference between participants who saw different descriptions is stronger evidence of framing than a correlation between deductible choices and a questionnaire score labeled "loss aversion." Even experimental changes require care: emphasizing a loss may also teach participants relevant facts or increase attention. Manipulations that simultaneously change comprehension, salience, and emotional imagery do not separately identify one pathway.

Distortions, prevention, and experienced claims

The household may update its beliefs after personally experiencing a loss, observing a neighbor's damage, or encountering a memorable news event. Where claims occur rarely, a few events can exert disproportionate influence on subjective frequency estimates. Prior choices may also determine what outcomes the insured sees: a person carrying a high deductible may refrain from filing minor claims, while a low-deductible policy may make reporting more attractive. Observed insurer claim records can therefore be a selected view of the underlying incident process. Any model that equates reported claim counts with exogenous physical loss incidence should state and test that assumption.

Some apparent distortion may reflect the consumer's belief that insurer service, exclusions, or settlement rules vary with the selected option. Such beliefs are about contract reliability, not event likelihood. A subjective probability of obtaining payment conditional on loss combines with the perceived loss probability to produce an effective valuation of coverage. Accurate modeling must separate the event probability from the conditional payment process. This is especially relevant to catastrophe and health coverages in which coverage conditions and claims adjudication can be complex.

The empirical lesson is that observed choices identify preference and belief combinations relative to specified financial consequences. Independent risk elicitation, incentivized probability forecasting, repeated choices over rank-varying lotteries, and controlled information treatments can identify additional components. No single measurement necessarily dominates the others; each removes particular observational equivalences while introducing assumptions about comprehension and stability.

Empirical Deductible Studies

Automobile and homeowner insurance in the United States

Cohen and Einav (2007) estimate risk preferences from a large sample of automobile insurance deductible choices. Their automobile risk-preference model accommodates hidden heterogeneity in both claim risk and preference parameters. They find substantial, skewed heterogeneity in inferred risk attitudes, along with associations with age, sex, and proxies for resources. These correlations concern the authors' insurer population and their maintained model; they are not invariant psychological attributes of demographic groups. The particularly important contribution is methodological: the joint distribution of risk and demand can alter pricing conclusions even if observed choice frequencies appear simple.

Sydnor (2010) turns the research question from average inferred risk aversion to the internal plausibility of the expected-utility interpretation of modest-stakes homeowner protection. In (Over)insuring Modest Risks, homeowners frequently purchase low deductibles despite prices substantially exceeding expected marginal indemnity. Large implied risk coefficients expose a tension between local choice fitting and plausible valuation across the wealth scale. The observed contract decisions are real; the hypothesized psychological explanation is subject to further testing.

Barseghyan, Prince, and Teitelbaum (2011) ask whether insurance risk preference is stable across automobile and home products for the same households. Their cross-line stability test rejects the stable-preference restriction under their structural assumptions and indicates higher apparent homeowner risk aversion in many cases. This does not establish that households randomly change preferences from one purchase to the next. Possible mediators include differences in subjective probabilities, claim severity, product salience, wealth effects, and characteristics of the alternative menus.

The Barseghyan–Molinari–O'Donoghue–Teitelbaum investigation extends the analysis of choice to include effective probability transformations. Its joint utility-and-distortion estimates favor an important contribution from small-probability overweighting under the chosen structural specification. Collectively these U.S. studies tell a richer story than either universal risk aversion or universal irrationality. Insurance purchases supply behavioral data of unusual practical significance, but interpretation changes with the model's admissible explanations and the dimensions of risk observed.

Cross-domain choices and the limits of a single preference parameter

Einav, Finkelstein, Pascu, and Cullen (2012) analyze five employer-provided insurance decisions and a retirement-account investment choice. Their multi-domain risk-preference evidence supports a substantial common component across insurance decisions, with weaker relationships between insurance and retirement asset allocation. That result complements, rather than mechanically contradicts, findings of context dependence: a person can exhibit a detectable shared component while making product-specific choices influenced by beliefs, budget, defaults, or contract understanding. A single global parameter and a mixture of common and domain-specific influences make different predictions at the individual level.

The practical question is whether a coefficient inferred from one deductible menu improves out-of-sample prediction of another. A suitable empirical test holds fixed the time interval, tracks observable changes in risk and resources, reconstructs both menus, and quantifies predictive improvement against a baseline using risk and price alone. High correlation in descriptive deductible rankings is insufficient if the same pricing rule or shared default induces the pattern. Validation should distinguish stability of latent preferences from stability of observed behavior.

Flood insurance: small and large stakes in the same contract

Flood insurance offers a particularly useful comparison because households select both a deductible and a coverage limit. Collier, Schwartz, Kunreuther, and Michel-Kerjan examine decisions by over 100,000 U.S. flood policyholders in their small- and large-stake analysis. They report that more than 94 percent select one of the two lowest deductible options in their studied menu and that 77 percent fully insure by their operational criterion. Standard expected-utility calibrations require very high risk aversion to fit both dimensions in that sample. The comparison is not universal flood-market evidence; it depends on the study's policies, rates, and institutional context.

In a later article, the same authors analyze coverage limits specifically in Insuring Large Stakes. Their normative benchmark, given the premiums and model in the study, indicates lower optimal limits than households typically buy; potential explanations include industry practices emphasizing full insurance as well as probability distortions. Importantly, the normative result is conditional on the calibrated model and source population, not a recommendation that households in general should underinsure dwellings against catastrophic floods. The presence of mortgage covenants, uncertainty about disaster severity, rebuilding costs, and correlations with other losses can alter a specific household's feasible set and utility.

This setting demonstrates a basic identification advantage: deductible and limit choices together interrogate distinct segments of a loss distribution. Yet the two selections can also be influenced by different defaults, narratives, and institutional restrictions. A researcher cannot assume that the same measure of absolute risk aversion translates from the deductible's small retained losses to the limit's catastrophic uninsured tail. Observing both contracts permits a direct test of that assumption.

Health insurance and financial dominance

Health insurance provides especially strong evidence about the danger of treating every selection as informed utility maximization. Bhargava, Loewenstein, and Sydnor (2017) study choices by 23,894 employees of one U.S. firm whose available options differed in financial cost sharing and premiums. In Choose to Lose, most employees selected financially dominated plans in the particular menu; excess spending amounted to 24 percent of chosen premiums by the authors' measure. The analysis shows why some decisions cannot be rationalized merely by the selection of a different risk-aversion coefficient or a different forecast of medical expenses. Subsequent hypothetical choice studies implicated misunderstanding of health insurance in this setting. A dominated contract, when dominance is established across the relevant consumer payoffs, is a stronger diagnostic of choice error than a costly but undominated low-deductible contract.

Handel (2013) studies a major employer plan change and estimates substantial renewal inertia in Adverse Selection and Inertia. In that model, reducing inertia improves certain individuals' choices but aggravates aggregate adverse selection sufficiently to lower welfare in the studied market. This establishes that a seemingly unambiguous improvement in individual choice accuracy can have a different equilibrium implication. It is particularly relevant for the interpretation of deductible persistence: past-plan retention can stabilize observed menus even when latent preferences or medical risk have changed.

Experiments by Kunreuther and Pauly, published in 2022 as low-deductible bias experiment, provide a necessary corrective to a universal low-deductible-bias narrative. Their controlled web experiment, changing defaults and premiums, did not find a general preference for low deductibles. Some participants retained the same choice when defaults changed, while others shifted. The result supports the proposition that a common field pattern can be produced by particular prices, information, institutions, and selection rather than by a fixed behavioral law for all consumers.

Biener and Zou's 2024 Swiss health-insurance maze experiment considers a menu in Switzerland with several deductible options and experimentally varies information and the number of options. Their results indicate that better presentation and menu restriction can improve the likelihood of choosing among economically sensible options in the tested setup, while economically consequential errors remain. This is evidence on an experimental Swiss environment; financial optimality depends on defined medical cost distributions and the regulatory premium-deductible schedule. The study also supports a general research design recommendation: when an intervention makes a menu more transparent, evaluate both movement toward undominated choices and subsequent economic outcomes, not only the average deductible chosen.

The Netherlands: selection, incentives, and persistence

The Netherlands provides longitudinal information on voluntary annual health deductible choices. Croes, Schut, and Varkevisser's Dutch health insurance panel analyzes 2013–2018 deductible behavior and estimates a dynamic logit specification on approximately 266,000 individuals, selected from a much larger insured population. They find that people with higher previous health care costs are less likely to take up or retain the maximum voluntary deductible and that previous deductible choice has a powerful association with the current selection. The authors interpret the pattern as evidence of adverse selection combined with considerable consumer inertia. The study was first published online in 2024 and appears in the 2025 journal volume; the bibliographic record below uses the final volume year.

This setting makes the difference between health risk and choice persistence visible. An individual with high expected utilization may select a low deductible without any probability distortion, while an individual expecting low costs may nevertheless keep a previous low deductible because the renewal environment does not prompt active reconsideration. The researchers' sample restrictions, including exclusion of plan switchers in core analyses, are relevant to the estimand. The results do not directly identify why every particular person persisted; a dynamic logit coefficient on lagged selection can also reflect durable unobserved preferences unless its identification assumptions are satisfied.

The U.S., Swiss, and Dutch findings create a genuine comparative picture without requiring numerical pooling across incompatible menus. The American employer evidence establishes the possibility of systematic financial dominance in a particular benefit context. The Swiss experiment tests presentation and option-count interventions. The Dutch panel links past use and choice persistence over time. Each identifies a different part of the behavioral process. Differences in health system rules, subsidy structures, annual deductible mechanics, and default enrollment prevent a responsible synthesis from treating their reported percentages as estimates of a single international effect.

Identification and Heterogeneity

The inverse problem of observed deductible choice

A researcher observes a chosen deductible d*, a set of quotes M, and possibly characteristics X. The researcher wants to recover a latent object such as a risk-aversion coefficient r, a subjective claim probability q, a weighting function w, or a vector of cognitive and contractual constraints. The observed mapping from these latent objects to d* is generally many-to-one. Even with a complete menu, a choice among two options yields an inequality on willingness to pay, not a directly observed cardinal utility parameter. The distance between those objects explains why structural models need functional restrictions, repeated observations, risk measurements, and assumptions about choice errors.

Suppose a consumer chooses d_L over d_H when the difference in subjective value V_L − V_H is nonnegative. If V includes unknown risk aversion, perceived event probability, subjective severity, default affinity, and disutility from managing a large claim payment, then one observed inequality constrains their sum. The coefficient of one component cannot be recovered unless the others are observed, randomized, normalized, or excluded. A model that omits the other components may still produce a precisely estimated coefficient and a good in-sample fit. Precision conditional on restrictive assumptions and identification of a psychological cause are different achievements.

Selection makes the inverse problem harder. Deductible options can influence claim filing and preventive behavior, so claim records following purchase are jointly determined by physical exposure and contractual incentives. Low expected loss and a willingness to retain losses may be positively associated; high-risk individuals may seek low deductibles, inducing adverse selection. The Cohen and Einav study explicitly permits correlation between unobserved claim risk and inferred risk aversion. A credible study must document which claims identify physical risk, how reporting varies by deductible, and what information the consumer had when making a selection.

Distinguishing mechanisms through complementary measurements

A useful empirical design combines four information sets. First, it obtains the exact personalized premium-deductible menu and contractual payoff rules at the time of selection. Second, it estimates objective risk through independent exposure variables, eligible loss events, claims histories, and severity information, checking for deductible-sensitive reporting. Third, it elicits consumer beliefs and contract comprehension in ways that allow forecast calibration to be evaluated. Fourth, it records behavior under exogenous menu or presentation changes and, where possible, links repeated choices across products and time. These sources do not necessarily appear together in a commercial dataset, but the gaps can be stated precisely.

For probability weighting and subjective probability misperception, the decisive distinction requires variation that affects the rank-dependent valuation of lotteries differently from a simple change in belief. The field identification strategy formalizes the importance of more than two outcomes and changing outcome rankings for distinguishing the mechanisms under its model. Alternatively, independently elicited probability forecasts can measure subjective beliefs, though forecasts may be noisy and survey responses may not coincide with decision beliefs. Experiments can randomize credible risk information while holding the menu constant; a remaining response to payoff framing then has a different interpretation from a response to changed objective frequencies.

For comprehension and attention, identify options that are financially dominated across states under the fully specified cost-sharing rules. Persistent purchase of such options is more diagnostic than a willingness to pay above expected indemnity, provided the dominance calculation includes all real plan attributes and consumer-relevant transaction costs. The Bhargava–Loewenstein–Sydnor evidence is particularly useful in this regard. Comprehension quizzes, payoff displays, and randomized plan explanations can test whether choices improve when the decision-relevant consequences become visible. Crucially, a plan with a higher expected expenditure is not automatically dominated: dominance requires weakly higher consumer cost in every relevant state and strictly higher cost in at least one, holding all other attributes equal.

For inertia, observe renewals and changes in menus or defaults. Persistence after unanticipated plan reconfiguration, controlling for objectively comparable alternatives, is stronger evidence than an unchanged deductible by itself. The Handel employer-market analysis and the Croes–Schut–Varkevisser panel illustrate distinct ways of estimating persistent choice. Separating true state dependence from persistent unobserved household characteristics requires additional design or model assumptions. A preferred option can remain optimal over many years, and longevity alone is not an error.

A practical source-to-claim evidence hierarchy

Research results should be grouped by the strength of the inferential link they actually establish. The most elementary evidence is the revealed frequency of deductible selections, conditional on observed menus. A stronger descriptive result estimates the price paid for incremental indemnity and its actuarial expected value using credible loss distributions. A further step establishes within-household or across-policy associations after accounting for risk and demographics. Structural identification of latent utility and probability transformations requires maintained functional and distributional assumptions. Causal evidence that presentation, information, defaults, or prices affect choices requires defensible exogenous variation. Welfare analysis requires an additional normative specification of utility, belief accuracy, dynamic selection, and market consequences.

This sequence is an editorial classification of established inferential requirements, not a newly proposed theory or measurement scale. It prevents a common overextension: reporting a selection frequency as evidence of a stable risk preference, or reporting better comprehension as proof of a higher welfare outcome. The studies reviewed here occupy different levels. Sydnor documents the price-to-protection puzzle; Cohen and Einav and Barseghyan et al. estimate conditional structural preference parameters; Bhargava and coauthors identify financial dominance within a menu; Kunreuther and Pauly test choice changes under experimentally altered defaults; Handel examines market-equilibrium consequences of reducing inertia. Their results are related but not interchangeable.

Heterogeneous households, risk states, and external validity

Observed heterogeneity can arise from demographics, resources, private information, experience, decision process, and changes in product rules. Structural preference distributions can be skewed because a small group strongly values protection, because a model allocates unmeasured beliefs to utility curvature, or because some consumers make near-deterministic default choices. Heterogeneity across insurance lines can reflect the genuinely different states against which households insure. Consequently, a distribution of estimated coefficients is not yet a distribution of invariant psychological traits.

Out-of-sample prediction offers an important discipline. Estimate a model on one set of households, periods, or product menus and evaluate its probability forecasts for choices in a distinct set. Then compare models that include only prices and objective risk, those adding wealth-based curvature, and those adding measured beliefs or attention. A more elaborate model should demonstrate meaningful predictive improvement and transparent parameter stability. If coefficients change when a menu, framing, or survey prompt changes, the researcher should report that dependence rather than reinterpret it as a property of all future insurance decisions.

Cross-jurisdiction comparison also needs a common economic outcome. The same nominal $500 or €500 deductible has different meaning under different premium concessions, purchasing-power levels, risk rates, policy limits, and health systems. A more defensible comparative measure is the incremental premium divided by the expected additional indemnity, supplemented by the distribution of possible out-of-pocket costs and liquidity exposure. Even that ratio does not fully normalize differences in default enrollment, subsidies, market structure, claim-handling practices, or background risks. The current literature supports a cross-market qualitative synthesis with explicit contextual boundaries rather than a single pooled international risk-aversion estimate.

Limits of the evidence reviewed

Several foundational empirical articles use a single U.S. insurer or employer, with nonrandom sample restrictions and proprietary data. Some outcome distributions are estimated rather than observed for each household. Data on subjective beliefs, liquidity, claims that were never filed, or the precise explanation presented by an agent are often unavailable. Experimental insurance choices may differ from financially consequential market choices, while field datasets often identify choices better than mental processes. The Netherlands and Switzerland have distinctive statutory health insurance institutions. Historical rates or deductible levels from those studies cannot be treated as contemporary offers.

The present dossier does not pool estimates, reproduce proprietary datasets, or independently rerun every replication package. It verifies study identity, reported main findings and the limits supported by accessible primary materials. These limitations reduce the scope of quantitative generalization; they do not prevent analytic derivation of basic contract identities or evidence-based distinctions between study designs. The correction for multiple plausible mechanisms is especially important: a descriptive residual from an expected-utility model establishes model misspecification relative to the observed choice data, whereas the identification of a unique psychological mechanism requires a stronger design.

Interpretation of Household Behavior

What can be concluded from a low deductible?

A low deductible is evidence that the selected contract was preferred, actively or passively, within the consumer's observed or perceived feasible set at a particular time. With an accurate complete menu and verified payoff equivalence, it can constrain a willingness-to-pay interval for greater contingent reimbursement. If the incremental premium exceeds objective expected reimbursement, a risk-neutral expected-cost benchmark favors a higher deductible. Additional explanations include risk-averse final-wealth utility, subjective overestimation of the claim probability, nonlinear weighting of a small probability, liquidity protection, reassurance, misunderstanding, or a default choice. Each is a candidate with distinct testable implications.

A high deductible is correspondingly compatible with greater tolerance for retention, lower perceived claim risk, liquidity resources, price sensitivity, distrust of insurance, or an inability to pay the lower deductible's premium. Inferring generalized risk seeking from the high deductible is no more justified than inferring generalized risk aversion from the low one. The empirically informative object is a conditional comparison: for a particular menu and modeled loss distribution, what range of valuations can account for the decision, and which additional observations narrow that range?

One key asymmetry arises when the lower deductible option is not financially dominated but simply expensive in expectation. The buyer may purchase meaningful insurance against an unpleasant state, making expected-cost calculations insufficient to declare an error. When a selected plan is strictly dominated after all relevant attributes are accounted for, no conventional preference for more protection at the same or greater cost can explain it. The evidence then shifts toward misunderstanding, limited attention, absent consideration, or preferences for attributes outside the researcher's measured model. This interpretive discipline prevents unwarranted judgments about household competence.

Behavioral preference inference and consumer protection

The consumer-protection relevance of deductible research begins with understandable displays of premiums, likely and worst-case out-of-pocket obligations, and comparisons across the same risk scenarios. A table giving the change in annual premium, the reduction in a covered claim payment, and the event frequency required for expected-cost break-even is more informative than an isolated deductible label. Consumers may benefit from a transparent picture of annual aggregate health expenditure or claim severity, especially when copayments, coinsurance, limits, and maximum out-of-pocket rules operate together.

Information interventions should be evaluated rather than presumed sufficient. Biener and Zou's Swiss experimental findings suggest that improving information or narrowing the menu does not eliminate all consequential errors. The dominated-plan evidence indicates that misunderstanding can persist in substantial decision settings. A practical intervention may include consequence graphs, plain-language scenarios, explicit price differences, and reminders to revisit a previous selection. Research evaluation should measure comprehension, realized costs, take-up, and distributional effects rather than focus only on whether the average deductible rises.

Default design creates an important tradeoff. An automatic renewal may protect people who would otherwise fail to maintain insurance, yet it may also perpetuate a poorly matched deductible. A prompt to reconsider a menu may improve individual fit but alter selection across risk groups. In Handel's welfare analysis, reduced inertia can worsen market adverse selection. Thus an insurer or regulator evaluating choice architecture should consider individual consumer outcomes together with premium equilibrium, cross-subsidization, access, and coverage continuity. The design of a useful reminder is a separate issue from the generality of a psychological preference claim.

Implications for insurance pricing and model governance

Pricing models may be tempted to use deductible choice as a proxy for latent risk aversion or private risk information. This is statistically attractive because the choice is recorded in administrative transactions. It is also sensitive to measurement and governance. A deductible can encode product knowledge, agent recommendations, liquidity, access to alternatives, and historical defaults as well as risk tolerance. Underwriting or consumer segmentation that treats the coefficient as a pure preference can misclassify households whose choices were driven by constraints. Model documentation should therefore specify whether the parameter predicts claims, predicts choice conditional on menu, or purports to measure a psychological trait.

In a market with asymmetric information, price changes can alter the composition of insured households. A population's historical deductible distribution is not a fixed demand curve once premiums, deductibles, subsidies, and choice framing change. The original risk and preference heterogeneity model and the later health-market equilibrium study show, in different empirical environments, why risk selection matters to insurer economics. Counterfactual predictions should update who chooses each contract, how loss reporting responds, and what new premium schedule can be sustained.

Price optimization also requires institutional and legal information about allowable rating factors, disclosure, consumer rights, and contract enforceability in each jurisdiction. This dossier supplies scientific mechanisms and research interpretation; it does not evaluate a specific country's current regulatory requirements or provide individualized insurance purchasing advice. A defensible research design is transparent about the distinction between descriptive estimates and policies that might be adopted under particular legal and distributional goals.

Research priorities implied by existing evidence

The primary opportunity for progress is joint measurement of actual contract costs, accurate event exposure, subjective beliefs, liquid resources, consideration, and repeated behavior. More informative samples would record alternative quotes, not only the chosen policy; losses that were incurred but not reported; real-time probability forecasts; and the person's understanding of the payoff schedule. Institutional changes or credible randomized messages could help estimate the causal effects of information and defaults. Cross-market linkage would allow tests of preference stability without pretending that deductible definitions are identical.

Another priority is uncertainty reporting. A study may estimate a claim frequency precisely but remain uncertain about loss severity or unfiled claims; it may measure an information intervention's average effect while knowing little about long-run selection; it may estimate a probability-distortion function while being unable to distinguish weighting from inaccurate beliefs. Each inferential limit should be documented at the location where a parameter is interpreted. The objective is a better-supported account of real choices, not the multiplication of names for residual effects.

Finally, welfare analysis should specify the counterfactual and normative criteria. A low-deductible option may provide valuable state-contingent liquidity; a dominant high-deductible plan may save money for every relevant state; a default may preserve coverage despite some inefficient persistence; and a more informed individual choice can produce market-level selection consequences. The quality of an intervention is measured against these identifiable outcomes and distributional consequences. The same observed movement toward a higher deductible can represent improved comprehension, a premium shock, reduced perceived risk, or loss of affordability. No direction of deductible change alone defines success.

Conclusion

Deductible choices reveal an economically significant relationship between certain premiums and retained contingent losses. Their interpretation as behavioral risk preferences is conditional on the menu, the loss process, observed and unobserved risks, contract understanding, liquidity, and mechanisms of choice. The first foundational finding is that a low deductible can carry a premium far above expected additional reimbursement. Sydnor's homeowner evidence and related structural studies make that discrepancy empirically important. The second is that standard expected-utility curvature can require implausible cross-scale risk parameters when used as a single explanation for intense aversion to modest retained losses. Rabin's calibration result clarifies the theoretical pressure, and subsequent insurance studies show why broader specifications are necessary.

The third finding is that nonlinear probability valuation, subjective risk beliefs, and non-price behavioral factors can generate observationally similar deductible decisions. The Barseghyan and coauthor studies identify a substantial effective probability distortion in their U.S. property and automobile data while distinguishing the available statistical identification from the deeper psychological interpretation. The fourth is that observed risk-taking is heterogeneous and context dependent. Cross-line comparisons and multi-domain evidence support both a shared component of risk evaluation and important departures from a globally stable single parameter.

The fifth finding concerns choice quality. Health-insurance studies demonstrate that dominated selections, misunderstanding, and renewal inertia can be central. Controlled experiments do not find a universal low-deductible preference, and interventions that improve one aspect of a household's choice may have incomplete or adverse effects elsewhere. Flood evidence shows that deductible and limit selections can probe different loss scales within the same policy. U.S., Swiss, and Dutch results are informative precisely because their institutional differences are retained in the interpretation.

The original synthesis of this dossier is methodological and inferential. An observed deductible determines a conditional choice relation; the economic cost of reducing it depends on the full distribution of incremental indemnity; a psychological preference is inferred only after risk, beliefs, menu design, resources, and choice process have been addressed. Where an intervention is contemplated, individual and market-level welfare must both be specified. These conclusions follow from explicit contract algebra and the cited evidence without proposing an additional named theory, doctrine, or conceptual system. The resulting research standard is disciplined attribution: measure the contract, reconstruct the feasible alternatives, examine the event distribution, test competing explanations, and state the exact level of inference that the evidence supports.

References

  • Barseghyan, Levon, Francesca Molinari, Ted O'Donoghue, and Joshua C. Teitelbaum. 2013a. "Distinguishing Probability Weighting from Risk Misperceptions in Field Data." American Economic Review 103 (3): 580–585. https://doi.org/10.1257/aer.103.3.580.

  • Barseghyan, Levon, Francesca Molinari, Ted O'Donoghue, and Joshua C. Teitelbaum. 2013b. "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.

  • Barseghyan, Levon, Jeffrey Prince, and Joshua C. Teitelbaum. 2011. "Are Risk Preferences Stable across Contexts? Evidence from Insurance Data." American Economic Review 101 (2): 591–631. https://doi.org/10.1257/aer.101.2.591.

  • 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.

  • Biener, Christian, and Lan Zou. 2024. "More Options, More Problems? Lost in the Health Insurance Maze." Journal of Risk and Insurance 91 (1): 5–35. https://doi.org/10.1111/jori.12461.

  • 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.

  • Collier, Benjamin L., Daniel Schwartz, Howard C. Kunreuther, and Erwann O. Michel-Kerjan. 2017. "Risk Preferences in Small and Large Stakes: Evidence from Insurance Contract Decisions." NBER Working Paper 23579 (revised 2018). https://doi.org/10.3386/w23579.

  • Collier, Benjamin L., Daniel Schwartz, Howard C. Kunreuther, and Erwann O. Michel-Kerjan. 2022. "Insuring Large Stakes: A Normative and Descriptive Analysis of Households' Flood Insurance Coverage." Journal of Risk and Insurance 89 (2): 273–310. https://doi.org/10.1111/jori.12363.

  • Croes, Ramsis R., Frederik T. Schut, and Marco Varkevisser. 2025. "Adverse Selection and Consumer Inertia: Empirical Evidence from the Dutch Health Insurance Market." European Journal of Health Economics 26: 641–651 (first online 2024). https://doi.org/10.1007/s10198-024-01725-8.

  • Einav, Liran, Amy Finkelstein, Iuliana Pascu, and Mark R. Cullen. 2012. "How General Are Risk Preferences? Choices under Uncertainty in Different Domains." American Economic Review 102 (6): 2606–2638. https://doi.org/10.1257/aer.102.6.2606.

  • Handel, Benjamin R. 2013. "Adverse Selection and Inertia in Health Insurance Markets: When Nudging Hurts." American Economic Review 103 (7): 2643–2682. https://doi.org/10.1257/aer.103.7.2643.

  • Johnson, Eric J., John Hershey, Jacqueline Meszaros, and Howard Kunreuther. 1993. "Framing, Probability Distortions, and Insurance Decisions." Journal of Risk and Uncertainty 7 (1): 35–51. https://business.columbia.edu/faculty/research/framing-probability-distortions-and-insurance-decisions.

  • Kahneman, Daniel, and Amos Tversky. 1979. "Prospect Theory: An Analysis of Decision under Risk." Econometrica 47 (2): 263–291. https://doi.org/10.2307/1914185.

  • Kunreuther, Howard, and Mark Pauly. 2022. "Do People Have a Bias for Low Deductible Insurance?" Journal of Risk and Uncertainty 64 (1): 1–17. https://doi.org/10.1007/s11166-022-09368-x.

  • Rabin, Matthew. 2000. "Risk Aversion and Expected-Utility Theory: A Calibration Theorem." Econometrica 68 (5): 1281–1292. https://doi.org/10.1111/1468-0262.00158.

  • 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.

  • Tversky, Amos, and Daniel Kahneman. 1992. "Advances in Prospect Theory: Cumulative Representation of Uncertainty." Journal of Risk and Uncertainty 5 (4): 297–323. https://doi.org/10.1007/BF00122574.