Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Structural Identification of Behavioral Insurance Preferences

Abstract. Insurance contracts make behavior under risk observable, but they do not automatically reveal the psychological and economic primitives that generate it. A deductible choice combines perceptions of claim frequency, preferences over uncertain consumption, the pricing and composition of the offered menu, expectations about insurer performance, the probability that an alternative was actually considered, and circumstances that constrain the household's ability to pay. Structural identification asks which of these latent components can be recovered from a specified joint distribution of choices, prices, risks, covariates, and outcomes under stated restrictions. This monograph examines that question across expected-utility, rank-dependent, reference-dependent, random-utility, limited-consideration, and partially identified models. It derives deductible choice inequalities and a transparent observational-equivalence example, compares point and set identification, evaluates the role of claims, multiple insurance lines, subjective-belief measurements, price variation, and experiments, and separates descriptive model fit from causal and welfare validity. Primary insurance studies demonstrate both substantial departures from simple expected-utility interpretations and the sensitivity of inferred preferences to choice-set assumptions. A source-critical synthesis establishes the decisive condition for meaningful structural estimates: the data and maintained restrictions must jointly distinguish the object called a preference from subjective beliefs, exposure, decision frictions, and the consequences of the insurance contract. The result is a disciplined account of what existing empirical designs establish, what remains dependent on auxiliary assumptions, and how counterfactual claims can be qualified without inventing an additional theory.

Keywords: Structural identification; behavioral insurance; deductible choice; risk aversion; subjective probabilities; probability weighting; rank-dependent utility; heterogeneous preferences; insurance demand; partial identification; revealed preference; choice sets; limited consideration; information frictions; welfare counterfactuals; insurance econometrics.

Introduction

The history of empirical insurance economics contains a persistent temptation: to read a purchase of expensive protection as a direct observation of high risk aversion. Insurance choices are informative precisely because the buyer selects a contract whose payments vary across states of the world. Yet the same observed contract can be preferred for different reasons. A household may have a highly curved utility function, believe an insured loss to be unusually likely, overweight a small objective probability, anticipate troublesome claims administration, or simply fail to notice a cheaper option. In an insurance market, contract design determines both the signal provided by choice and the admissible interpretation of that signal.

Structural identification is the mapping from an observable probability distribution and explicit behavioral restrictions to a parameter, a parameter distribution, or an identified set. A model is point identified when different admissible parameter values cannot generate the same distribution of observables. It is partially identified when the data and restrictions restrict the admissible values to a set. A numerical maximum-likelihood estimate is not by itself an identification result; identification is a property of the model and observation process, while estimation and uncertainty quantification address finite samples. Manski's account of identification problems clarifies this distinction.

The focus of the present Research Dossier is structural identification of behavioral preferences in insurance purchasing. Its central object is the inferential relation among preferences, beliefs, consideration, insurance prices, and outcomes. The neighboring question of estimating market equilibrium demand and insurer costs belongs to the distinct dossier Structural Estimation of Insurance Demand and Costs. This text uses market equilibrium and welfare where needed to test the meaning and transportability of identified preference parameters, while preserving that division of scientific scope.

The canonical empirical foundation includes Cohen and Einav's deductible-choice estimation, Barseghyan, Molinari, O'Donoghue, and Teitelbaum's evidence on probability distortions, and the subsequent methodological progression to partial identification across contexts, limited consideration, and heterogeneous choice sets. These studies pose substantively different questions despite sharing deductible data as an empirical environment. Their differences are part of the evidence rather than a problem to be averaged away.

The methods here are formal derivation under transparent stylized assumptions, critical comparison of original empirical papers and their publication records, and a source-to-claim synthesis of identification conditions. All numerical examples below are hypothetical calculations. No restricted insurer records, independent survey respondents, new randomized experiment, or pooled meta-analysis were assembled for this monograph. The studies cited retain their original samples, periods, jurisdictions, and inferential limits. A scientific conclusion will be described as identified only relative to the assumptions and observables that actually identify it.

Insurance Choice Structural Models

The observational object and the parameter object

A basic observation contains the contract selected by household i, the menu offered, the premiums and deductibles attached to every available alternative, and measured characteristics. Depending on the dataset, it may also contain subsequent losses, past claims, expenditure, stated beliefs, search histories, renewal decisions, or administrative switching records. These variables are not interchangeable. An offered menu is an institutional opportunity set. A considered menu is the subset the individual represents in the decision. A purchased policy is an outcome of that decision. Subsequent realized claims reflect a stochastic process conditioned on exposure and sometimes affected by the selected contract.

Let the observed contract index be c_i in a known menu C_i. Let p_i denote the true claim probability under a specified horizon and loss definition, q_i a subjective probability, a_i a measure of utility curvature, and w_i(.) a weighting function. Let A_i be the unobserved consideration subset of C_i, k_i any search or switching cost, and m_i the financial state. The broad empirical mapping is

Pr⁡(ci=c∣Ci,zi)=∫Pr⁡(ci=c∣Ci,p,q,a,w,A,k,m,zi) dF(p,q,a,w,A,k,m∣zi).\Pr(c_i=c\mid C_i,z_i) =\int \Pr(c_i=c\mid C_i,p,q,a,w,A,k,m,z_i)\,dF(p,q,a,w,A,k,m\mid z_i).

This equation is a decomposition of the analyst's inferential problem, not a claim that each latent variable can be learned from the same dataset. The integral makes apparent that one observed discrete outcome can depend on multiple joint latent distributions. Identification is achieved through informative variation and restrictions on the integral, not through naming the components.

Three data distinctions are particularly consequential. First, offered prices must be the prices faced by the chooser, not just a retrospective insurer tariff. Second, the measured insured event must correspond to the event whose cost changes with the deductible. Third, observing the full administrative menu does not establish that every alternative received attention. Barseghyan, Molinari, and Thirkettle's limited-consideration model estimates preference heterogeneity while allowing heterogeneity in considered alternatives, making the third distinction an identifying component rather than a residual behavioral story.

Deductible menus, risk states, and economic environments

Consider a simplified insurance contract with annual premium P_j and deductible D_j. For an insured loss sufficiently large to exhaust either deductible, and one binary claim indicator X, the consumer's monetary cost is C_j=P_j+D_jX. Assume X takes value one with objective probability p and zero otherwise. For two contracts L and H, let D_LP_H; L is the lower-deductible, higher-premium contract. Define the premium difference ΔP=PL−PH>0\Delta P=P_L-P_H>0 and the deductible saving ΔD=DH−DL>0\Delta D=D_H-D_L>0.

Under linear utility and correct beliefs, expected cost minimization selects L when

ΔP≤pΔD.\Delta P\leq p\Delta D.

The inequality is an exact result under the stated binary-loss, fixed-coverage, risk-neutral setting. It is not a general statement about real insurance policies. Real losses may lie below one or both deductibles, involve several events, limits, exclusions, coinsurance, waiting periods, taxes, and claim-handling services. The more complete indemnity payoff is a function of the loss and the contract, and the expected incremental payout must then be integrated against the entire loss-severity distribution. For a policy with a deductible and a covered monetary loss L, the individual's retained cost is typically min⁡(L,Dj)\min(L,D_j) only under a particular reimbursement rule and absence of other limits.

For this reason a premium-to-deductible ratio is a useful local statistic but not an omniscient measure of risk preferences. The analyst must model the relevant claim threshold, loss frequency, deductible applicability, and interactions with other coverages before interpreting the observed choice. Cohen and Einav (2007), in their Israeli automobile insurance study, use individual-specific price menus and modeled claim risk rather than assuming that a raw deductible histogram isolates curvature.

The contract may also change losses. Moral hazard can alter reporting, effort, maintenance, or the probability a marginal claim is filed. Contract choice can select heterogeneous risk types. These two channels have different empirical meanings. The French auto-insurance test of Chiappori and Salanié (2000) investigates asymmetric information using contracts and accidents within its observed institutional setting; its reported absence of evidence is a finding for that sample and design, not a universal absence of private risk information.

Structural likelihoods and the meaning of exogeneity

A parametric model often assigns utility to each plan and allows additive choice noise, with a logit or mixed-logit probability for each plan. Conditional on a latent vector t_i and observed z_i, a familiar representation is

Pr⁡(ci=j∣ti,zi)=exp⁡(Vij(ti,zi)/σ)∑k∈Ciexp⁡(Vik(ti,zi)/σ).\Pr(c_i=j\mid t_i,z_i)= \frac{\exp(V_{ij}(t_i,z_i)/\sigma)} {\sum_{k\in C_i}\exp(V_{ik}(t_i,z_i)/\sigma)}.

The scale σ\sigma and the deterministic utility level cannot generally be separately recovered without normalization. A mixed model integrates this expression over a distribution of t_i. Allowing a flexible latent distribution is not equivalent to identifying the separate interpretations of its dimensions. Observational equivalence can survive increasingly precise estimation.

The analyst must specify why offered price variation is informative rather than a proxy for the household's risk, wealth, insurer screening, geography, or information environment. Insurer prices commonly depend on observable rating variables and may vary with unobserved firm-side signals. A valid price instrument shifts one aspect of the feasible choice comparison while satisfying the necessary exclusion and relevance conditions. Administrative changes, discontinuities, randomized menus, and within-market premium variation are candidates only when their assignment mechanisms have been examined. A cost-shifter that also changes plan salience or quality may identify a composite response rather than pure price sensitivity.

An observational likelihood that fits premiums and choices can remain systematically misspecified if the true consideration sets differ across purchasers or if realized claims enter as predetermined covariates despite contract-sensitive claim filing. Consistency of an estimator is a statement conditional on its data-generating assumptions. It cannot repair a missing exclusion restriction.

Historical development of insurance preference estimation

Expected utility supplies an internally consistent benchmark for risky consumption. Kahneman and Tversky's prospect-theory analysis documented behavioral regularities motivating alternative descriptions, and Tversky and Kahneman later developed cumulative prospect theory, using rank-sensitive decision weights. Yaari's dual theory of choice provides another treatment in which attitudes toward probabilities play a central role.

Insurance studies subsequently linked these theoretical distinctions to real contract menus. Sydnor's home-deductible evidence reports insurance against modest financial risks at prices that imply implausibly high curvature under a standard expected-utility reading. That result is a diagnostic challenge for a narrow curvature-only model. It does not select one uniquely correct alternative mechanism.

Barseghyan and colleagues' field-risk-preference survey examines the identification and estimation of expected and nonexpected utility using individual and aggregate market data. The research trajectory is therefore an expansion of observable conditions and competing behavioral channels: from treating deductible variation as a signal about risk aversion toward formally distinguishing parameters that can be separated and components that remain observationally equivalent.

A useful organizing question is not which behavioral label produces the closest narrative description of a purchase. It is what additional observation would make rival explanations yield different predictions under the same contractual environment. That question guides the derivations that follow.

Identification of Risk Aversion

Expected-utility inequalities and a verifiable benchmark

Let a household have starting wealth W, no other differences between two plans, and Bernoulli utility u over final wealth. Under objective expected utility, its value from plan j is

EUj=(1−p)u(W−Pj)+pu(W−Pj−Dj).EU_j=(1-p)u(W-P_j)+pu(W-P_j-D_j).

The lower-deductible plan is chosen if EUL≥EUHEU_L\geq EU_H. With constant absolute risk aversion and u(x)=−exp⁡(−ax)u(x)=-\exp(-ax), where a>0a>0, minimizing the certainty-equivalent cost is equivalent to maximizing expected utility:

CEj=Pj+1alog⁡(1−p+pexp⁡(aDj)).CE_j=P_j+\frac{1}{a}\log\left(1-p+p\exp(aD_j)\right).

Taking a→0a\to0 gives CEj→Pj+pDjCE_j\to P_j+pD_j. A Taylor expansion about a=0a=0 yields

CEj=Pj+pDj+a2p(1−p)Dj2+O(a2).CE_j=P_j+pD_j+ \frac{a}{2}p(1-p)D_j^2+O(a^2).

Hence, to first order in a, the lower deductible is selected when

ΔP≤pΔD+a2p(1−p)(DH2−DL2)+O(a2).\Delta P\leq p\Delta D+ \frac{a}{2}p(1-p)(D_H^2-D_L^2)+O(a^2).

The derivation shows why risk aversion can rationalize paying more than the actuarial incremental payout, while clarifying the role of the absolute stakes and event probability. It also shows why a precisely measured price premium alone does not recover curvature when the risk probability remains unknown.

For a purely illustrative case, set p=0.02, D_L=250 dollars, D_H=1,000 dollars, and ΔP=30\Delta P=30 dollars. The risk-neutral value of the deductible saving is 0.02×750 = 15 dollars. A risk-neutral consumer with accurate beliefs prefers H under the binary model. A sufficiently risk-averse consumer may prefer L, but the same choice also follows from a belief that p exceeds 0.04, without any curvature. These are arithmetic implications of the stated hypothetical setting and are not estimates from real policyholders.

The same logic permits inequality-based bounds. If each member of a household's available menu was considered and the chosen contract maximized expected utility, the observed choice yields one or more comparisons EUci(a,p)≥EUk(a,p)EU_{c_i}(a,p)\geq EU_k(a,p). For known p and menu values, their intersection restricts a, often to an interval. Replacing known p by an estimated or bounded probability widens or shifts the interval. Abandoning full consideration may remove inequalities for alternatives never evaluated.

Heterogeneous risks, observed claims, and bounded curvature

Insurance data frequently contain subsequent claim outcomes that seem to provide an objective anchor for p. Claims are valuable, but a realized claim is an observation of a stochastic process, not a direct measurement of a person's subjective probability, and claims can be censored by reporting thresholds. Under a model in which repeated claim events occur at a constant rate λi\lambda_i over exposure t, a Poisson count likelihood has the form

Pr⁡(Ni=n∣λi,ti)=exp⁡(−λiti)(λiti)n/n!.\Pr(N_i=n\mid \lambda_i,t_i) =\exp(-\lambda_it_i)(\lambda_it_i)^n/n!.

The implied probability of at least one event is 1−exp⁡(−λiti)1-\exp(-\lambda_it_i), provided the Poisson model describes the event definition and exposure period. If a claim record contains only indemnified claims, a deductible may influence the observable count by altering reporting. The analyst must distinguish the structural loss probability from the administratively recorded claim probability.

Cohen and Einav's structural treatment of adverse selection allows unobserved heterogeneity in both claims risk and risk aversion. The paper's joint modeling reveals how a claim distribution can inform the marginal risk distribution, while deductible choices and maintained restrictions help infer risk preferences and their association with risk. The interpretation is a model-based joint inference, not an assumption-free measurement of every individual's psychological trait.

Unobserved risk and unobserved preference can be correlated. This matters because the price offered to an individual, the contract selected, and the subsequent claim may all reflect latent types. Conditioning directly on realized claims as though they were an exogenous proxy for subjective beliefs can introduce post-choice conditioning bias. A more coherent design specifies the temporal order: underwriting signals and menu assignment precede choice; loss and reporting are realized after choice; updating of beliefs precedes subsequent decisions.

Repeated choices across different menus can sharpen inference. A single deductible choice may reveal that a lies above a threshold under a particular p. When the same person selects across several coverage lines, one can intersect the compatible sets, provided preference stability is sufficiently credible and each line's risk and payoff technology is modeled.

Cross-context stability and partial identification

Barseghyan, Prince, and Teitelbaum's test of contextual stability uses household choices across automobile and home coverage to assess the restrictions of objective expected utility with stable preferences. The authors report rejecting their tested stability hypothesis and find evidence that estimated home-coverage risk aversion can exceed auto-coverage risk aversion. That is evidence against the stated joint specification and its invariance restrictions. It does not establish that every individual's psychological curvature literally changed whenever the policy category changed.

The later stability-based inference study by Barseghyan, Molinari, and Teitelbaum (2016) develops household-specific parameter bounds from revealed preference and uses restrictions across three insurance lines to narrow them without imposing a parametric distribution of unobserved preference heterogeneity. Their method illustrates an important alternative to fitting a single preference distribution: report the compatible region and the exact assumptions that tighten it.

Suppose a household's auto choice implies a compatible parameter set A_auto and its home choice implies A_home. A stable common parameter must belong to Aauto∩AhomeA_{\mathrm{auto}}\cap A_{\mathrm{home}}. If that intersection is empty, at least one maintained assumption is violated: stable preference, correctly specified loss probabilities, considered choices, utility form, or accurate menu observation. An empty intersection does not by itself identify which assumption failed. If the intersection contains a broad interval, a point estimate of the preference parameter requires additional structure.

The general value of partial identification is interpretive transparency. A stable interval that excludes zero curvature can support a bounded positive-curvature conclusion conditional on the model; an interval containing zero and many positive values supports a weaker statement. Parametric point estimates may be useful for prediction or policy simulation, but their precision should not conceal the restrictions that created uniqueness.

Statistical inference versus identification

A finite-sample confidence interval measures sampling uncertainty conditional on a specified statistical procedure. An identification region describes observationally admissible parameters even with an arbitrarily large sample. A narrow reported standard error around a misspecified, nonidentified target can give a false impression of precision. This distinction is particularly acute in insurance because observed choice cells may be sparse, many products have similar financial payoffs, and the relative price spread can be strongly correlated across alternatives.

A sensible empirical report separates (i) the point or set identified by the maintained model, (ii) sampling uncertainty around that object, (iii) uncertainty about generated risk probabilities or losses, and (iv) variation across plausible alternative behavioral specifications. A confidence set for a partially identified parameter should account for the geometry of the inequalities and not simply take the uncertainty around one convenient point in the set.

Risk aversion is itself scale sensitive. Absolute risk aversion is dimensioned in inverse monetary units; relative risk aversion depends on wealth and on the utility specification. The behavioral implications of curvature estimated from modest insurance stakes should therefore be reconciled with independent economic restrictions rather than mechanically transplanted to all choices. Chetty's labor-supply-based risk-aversion bounds illustrate that alternative domains can discipline implausibly large curvature estimates, although transporting any parameter between domains is itself a substantive assumption.

The methodological conclusion is precise. Deductible choices generate informative utility inequalities, and claim data can help identify objective risk distributions. Separating risk aversion from beliefs and other choice mechanisms additionally requires valid restrictions and observable variation. Every statement about an individual's estimated preference belongs to that complete identification argument.

Probability Weighting versus Belief Distortion

The fundamental observational equivalence

In the simplified binary-loss, linear-value model, suppose the decision maker behaves as if the event receives decision weight w(q)w(q), where q is the subjectively believed event probability. The incremental willingness to pay to lower the deductible is w(q)ΔDw(q)\Delta D. If the only choice-relevant quantity is the product w(q)w(q), then choices identify that composite weighting, subject to the usual price and consideration assumptions. They do not separately determine whether the person believes the loss to be likely or transforms a correctly understood probability.

An explicit counterexample exposes the problem. Consider a high-deductible and low-deductible contract whose difference in deductible is 200 dollars and whose difference in premium is 15 dollars. The threshold ratio is 0.075. A household with q=0.02 and decision weight w(0.02)=0.10 will favor the lower deductible under this linear binary illustration. So will another household with a subjective probability q=0.10 and identity weighting w(q)=q. Both generate an effective event weight of 0.10. Every binary price variation that only tests this composite can leave the two explanations indistinguishable. The numerical values are illustrative and are not behavioral estimates.

The general observational-equivalence relation is

w1(q1)=w2(q2)⟹the same binary choice ranking under identical payoffs,w_1(q_1)=w_2(q_2) \quad\Longrightarrow\quad \text{the same binary choice ranking under identical payoffs},

provided payoffs enter linearly, the objective event partition and reference point are fixed, and no other utility channel differs. With nonlinear value functions, the same result applies to an appropriately defined effective event weight or to combinations of weight and outcome valuation; the number of unidentified components can increase rather than shrink.

Barseghyan, Molinari, O'Donoghue, and Teitelbaum (2013) explicitly develop the weighting-versus-misperception distinction. Their study is methodological and illustrates the identification strategy using simulated insurance choices. The illustration should not be reported as a separate empirical finding that actual consumer beliefs were measured and separated.

A critical consequence follows: the claim that an estimated decision weight exceeds an objectively estimated p is not sufficient by itself to label the difference psychological probability overweighting. That interpretation depends on the model's restrictions on subjective beliefs, reference effects, other preferences, and choice attention. The term probability distortion is often a safer description of the reduced-form wedge.

Multi-outcome lotteries and ranking variation

Binary lotteries are restrictive because many formally distinct preference representations collapse into a single effective probability. With more than two outcomes, cumulative rank-dependent models attach weights to probability tails and their differences, while subjective probability distortions change event probabilities independently of the outcome ranking. When the same underlying events enter lotteries whose outcome rankings change, rank-specific weighting and fixed subjective beliefs can imply different preferences.

The 2013 field-data identification strategy specifies two salient conditions: lotteries with more than two outcomes and changes in outcome ranking across lotteries. These conditions help distinguish rank-dependent probability weighting from systematic probability misperception in the paper's model. They are sufficient ingredients within its maintained restrictions, not a promise that any large list of insurance options delivers full nonparametric identification.

For example, imagine two policies covering two loss categories. Under one policy the financially worst outcome is a severe flood; under another, a payment structure makes a fire loss the worst outcome. If subjective event probabilities are common across policies but rank-dependent weights attach to cumulative tail probabilities, the change in rank can alter choice in a manner that cannot be reproduced by a fixed transformation of each event probability under otherwise identical structure. The example is conceptual. Actual identification still requires contract data revealing the ranking, correct joint-loss states, and exclusion of changing reference points or perceived claim-service differences.

Insurance severity distributions create further complexity. A household may misunderstand the frequency of catastrophic losses, the severity conditional on a claim, or both. A model with a single perceived probability for any claim compresses these dimensions. Multiple deductible thresholds can generate useful variation because different deductibles alter payments in different loss intervals. Yet such variation only distinguishes mechanisms when the distribution of losses, financial indemnities, and information available to buyers are sufficiently measured.

Reference dependence and alternative probability distortions

Prospect theory and cumulative prospect theory allow valuations relative to a reference point, different treatment of gains and losses, and nonlinear decision weights. Tversky and Kahneman's cumulative formulation connects weights to ranked outcome distributions. For insurance the choice of reference point matters: the consumer may frame premiums as certain losses, uncovered deductibles as future losses, or coverage as avoiding a salient reference loss. Such framing alters the interpretation of a utility parameter extracted from a premium-deductible comparison.

Kőszegi and Rabin's reference-dependent preference model makes expectations relevant to the reference point. Because expectations can respond to available contracts, the same menu change may affect behavior through reference dependence as well as relative prices. This channel requires more than relabeling a residual as loss aversion. Researchers must specify how the reference point forms, when expectations are measured, and whether they change with choice or enrollment.

Yaari's dual theory of risk provides a conceptual contrast to curvature-based utility because probability transformations can generate risk attitudes with a different representation. Barseghyan and colleagues' insurance preference estimates consider standard risk aversion alongside generic probability distortions and find distortions important in their U.S. property-insurance setting. Their study also shows why its estimated distortion pattern cannot be explained by the specific tested Kőszegi–Rabin loss aversion or disappointment-aversion mechanisms alone. That is a comparison among particular formal models, not a blanket rejection of all reference dependence.

A warning about overinterpretation is warranted. A decision-weight curve recovered under assumed correct claim beliefs is an estimated curve under those assumptions. A subjective belief distribution obtained from self-reports is itself vulnerable to scale interpretation, misunderstanding of the event, rounding and elicitation incentives. The same nominal probability can refer to annual loss incidence, reported claims, paid claims, or losses exceeding a deductible. These event definitions must be reconciled before placing psychological meaning on the gap.

Eliciting beliefs and separating measurement error

The most direct way to add identifying information is to observe subjective beliefs separately from insurance choice. Prospective probability questions, incentivized belief elicitation, structured forecasting, and experiments with transparent probabilities can all inform the distinction. However, stated beliefs are measurements with their own error process. In a latent-variable model one might write a survey response as

Bi=qi+ηi,B_i=q_i+\eta_i,

where BiB_i is the reported probability and ηi\eta_i is elicitation or reporting error. This additive representation is a local approximation and must be adapted to bounded probabilities and possible heaping at zero, round percentages, or one. Assuming ηi\eta_i is independent of q_i or insurance choice is itself a restriction that requires justification.

Repeated belief questions can help characterize reliability, while external risk indicators and later claims can help anchor an objective event definition. Neither solves the entire problem without assumptions. A person can sincerely report an accurate probability while treating that probability differently in a high-dread context, and the same person may report a rounded probability that does not match a latent continuous belief.

Experimental manipulations that change objectively communicated probabilities independently of prices offer another route. A properly designed experiment can test whether choices respond to a controlled probability change in a way consistent with a particular weighting function. Yet communicating a risk may also shift salience and trust. The causal effect of the communication identifies the treatment as implemented; attributing its entire effect to belief change requires an exclusion claim.

The defensible inference hierarchy is therefore ordered by observables. A binary deductible choice alone often identifies only an effective risk-value wedge. Multiple ranked outcomes can distinguish selected weighting and misperception representations. Measured subjective beliefs can constrain that interpretation with a specified measurement model. Randomized probability or information variation can supply causal leverage when co-interventions are evaluated. No one stage should be presented as if it automatically delivered all later stages.

Hidden Preferences and Heterogeneity

Preference distributions and heterogeneous choice sets

Individuals can be heterogeneous along several dimensions even when offered apparently identical contracts. Some households may have different objective risk, some different utility curvature, and others different probability transformations, reference points, liquidity, knowledge or search patterns. There is also variation in what alternatives actually enter the decision. These dimensions can be correlated. Consequently, a population model with one random coefficient on deductible cost may fit aggregate shares while blending very different mechanisms.

A fundamental decomposition is

Pr⁡(ci=j∣Ci,zi)=∑A⊆Ci:j∈A∫Pr⁡(j maximizes value on A∣θ,zi) dF(θ,A∣zi).\Pr(c_i=j\mid C_i,z_i) =\sum_{A\subseteq C_i:j\in A} \int \Pr(j\text{ maximizes value on }A\mid \theta,z_i) \,dF(\theta,A\mid z_i).

Here θ\theta represents preference and belief parameters, while A represents consideration. The choice probability is averaged over both. If the researcher sets A=C for every individual, the estimated preference distribution must absorb the choice share consequences of that restriction. A person who never notices a cheaper higher-deductible plan can look similar, in the observed choice alone, to someone who considers it and strongly values protection.

In their 2021 limited-consideration study, Barseghyan, Molinari, and Thirkettle provide conditions for semi-nonparametric point identification of heterogeneity in standard risk aversion and consideration sets from risky discrete choices. The method is tailored to observable variation satisfying specific assumptions about alternatives and covariates, and it includes an application to property insurance. This is a substantive advance in what the data can reveal when the researcher treats consideration as an explicit random object.

Barseghyan, Coughlin, Molinari, and Teitelbaum's heterogeneous-choice-set analysis instead derives sharp partially identified regions under weaker restrictions on choice sets. In their auto collision data, choices can be rationalized by expected utility with relatively low risk aversion and heterogeneous restricted consideration; the article reports that more than three quarters of households require limited choice sets under its analyzed framework. The proportion is a sample-and-model finding. It is neither a universal rate of insurance inattention nor evidence that all low-deductible purchasing is mistaken.

Both approaches demonstrate that identification strength and behavioral flexibility move together. Strong variation and conditions may support point identification, while weaker assumptions can make a set-valued result more reliable. Describing one approach as definitively superior without reference to data and restrictions would confuse the object of inference with a preference for estimator convenience.

Information, attention, inertia, and dominated plans

Insurance decisions are frequently embedded in complex administrative procedures and multidimensional plans. Premiums, expected out-of-pocket payments, networks, drug formularies, provider access, future claims service, and switching time can enter the choice. In such environments a parameter estimated from premium responsiveness may combine monetary preferences and decision costs.

Handel and Kolstad's health-insurance choice study combines administrative enrollment and claims records with survey measures of consumer knowledge and hassles. It finds these information-friction measures important for plan choices and for the estimated preference parameters. The study illustrates that additional measurement can alter the underlying economic interpretation of an otherwise well-fitted demand model.

Handel's analysis of health-plan inertia uses a major change at a firm to estimate switching behavior and examine policy interventions. A central result is that reducing inertia can improve some individual choices yet worsen selection and aggregate welfare in the particular modeled employer insurance market. The finding is a direct warning against assuming that correcting a behavioral friction always creates a positive equilibrium welfare effect.

Abaluck and Gruber's Medicare Part D choice analysis documents substantial deviations between premiums, expected out-of-pocket costs, and risk-reduction attributes in observed prescription-drug-plan choices. Their quantified welfare implication is a partial-equilibrium result under an explicit benchmark of rational selection; its magnitude should not be generalized to other insurance systems.

Bhargava, Loewenstein, and Sydnor's dominated-plan evidence examines health plan decisions by 23,894 employees at a U.S. firm and finds selections from a menu where many lower-deductible options were financially dominated by otherwise comparable higher-deductible plans. Dominance creates a stronger diagnostic than uncertain curvature alone when the dominance conditions genuinely hold, but administrative simplicity must still be distinguished from perceived differences in access, service, or nonfinancial attributes.

Ericson and Starc's Massachusetts standardization study shows that a change in plan standardization and presentation affected the valuation of plan features and the chosen generosity and brands. This observation undercuts the assumption that price-menu interventions isolate prices while leaving preferences, salience, and product definitions unchanged. In a structural design, the presentation of alternatives can be a treatment variable and must be measured as such.

Joint selection of risk and insurance intensity

Insurance demand is determined jointly by risk and preferences. If two latent types both purchase generous coverage, one because of high objective risk and the other because of a strong taste for insurance, aggregate coverage-risk correlations can conceal important selection. Finkelstein and McGarry's long-term-care evidence demonstrates such offsetting dimensions of private information. A lack of positive aggregate correlation does not establish that informational selection is absent.

The distinction between adverse and advantageous selection matters for identification of preferences. A consumer can be privately low risk and highly insurance-averse or privately high risk and not especially risk-averse. A model that treats unobserved claim risk as independent of preferences can distort the inferred joint distribution and the interpretation of marginal demand. To identify that joint distribution, researchers need valid cost and claim information alongside observed menus and enough restrictions on the dependence structure.

Insurance contract behavior can additionally change risk outcomes. Higher deductibles may reduce small claims filed with an insurer, while differences in precaution could affect actual losses. A model of observed claims therefore needs an observation equation; simply equating administrative claims with exogenous physical loss incidence may contaminate estimation.

An identification report should distinguish at least four objects: risk type distribution before contract choice; choice-conditional risk distribution after self-selection; realized loss distribution under a contract; and the insurer's recorded claim distribution. Each can be estimated from different observations and restrictions. Conflating them can lead to policy conclusions that reverse when transferred to another deductible structure or underwriting regime.

Comparing property, health, and long-term-care contexts

Property-deductible studies offer comparatively transparent monetary tradeoffs, especially when the insurer records individual premium schedules and large-loss probabilities. Health insurance contains additional quality, provider, medical utilization, information, and timing channels. Long-term-care insurance combines distant risk, complex benefit triggers, mortality and lapse risks, long horizon discounting, and selection on information and risk attitudes. The fact that all are called insurance does not make their latent behavioral parameters exchangeable.

The Barseghyan–Molinari mixture analysis explicitly allows both expected-utility and dual-theory preference types, heterogeneous consideration sets, and bundle choices across two property-insurance contexts. Its semi-nonparametric identification argument relies on sufficient covariate variation across contexts and specific regularity assumptions. The paper also analyzes welfare from a hypothetical change combining insurance lines. It should be treated as a context-dependent model with an identifiable mechanism under its conditions, rather than proof that two primitive psychological types exhaust the population.

This cross-study comparison yields a useful caution. An apparent preference difference between health and auto insurance may represent a genuine domain effect, but it may equally represent differences in payoff complexity, observed state space, beliefs, claims coding, or search. Tests of stability are informative only after those differences have been incorporated into the joint model. When a single common parameter cannot explain choices, the scientifically appropriate next question is which maintained restriction was empirically contradicted.

Counterfactual Validity

Structural transport and policy interventions

A fitted model can predict observed choice shares and nevertheless fail under a policy that changes the menu, information environment, or prices. Counterfactual validity asks whether the mechanism that identifies preferences remains invariant under the contemplated intervention. The most familiar structural argument assumes that an intervention alters the environment while leaving certain primitive preferences or technologies fixed. In behavioral insurance, the primitives held fixed must be named; menu presentation, reference points, beliefs, consideration, and market selection may move with the intervention.

Suppose a regulator requires a uniform deductible menu. If a researcher estimated aversion from existing choices while assuming full consideration, replacing the menu can alter both financial incentives and what plans consumers notice. A counterfactual holding the estimated curvature constant and changing only prices may misrepresent enrollment. The studies of choice-set heterogeneity and plan standardization supply concrete reasons to model such changes.

Transport between insurers raises similar issues. Different rating algorithms produce different price spreads for an otherwise similar consumer. Insurers can vary in claim service, brand reputation, contract complexity, and observable or latent quality. A choice coefficient estimated in one product universe can be an equilibrium response to that product universe, not a stable consumer invariant. Counterfactual transfers require an explicit statement about products, exposures, information, and the distribution of consumers at the destination.

Statistical external validity and structural invariance should be reported separately. External validity concerns whether an empirical estimate applies to another population or setting. Structural invariance concerns whether the modeled behavioral relationship survives a specified intervention within or beyond the original setting. A study can have a representative sample and still use an incorrect structural mapping for an intervention; it can also identify a coherent local parameter that is not representative internationally.

Welfare under distorted choices

Welfare analysis requires a criterion for evaluating outcomes, and behavioral choices make that criterion especially consequential. A revealed-preference welfare calculation treats an individual's chosen plan as the maximizing option under the model's normative objective. If the model includes mistakes, inattention, distorted beliefs, or present bias, the observed choice may not represent the individual's best attainable welfare according to an alternative criterion. This creates a distinction between choice-predictive utility and the welfare measure adopted for policy.

Consider a household selecting a low deductible because the alternative was not considered. A welfare estimate that treats this selection as evidence of very high curvature can assign a large consumption value to the extra coverage. A model that treats the same behavior as incomplete consideration can instead assign a welfare cost to the forgone cheaper alternative. Both may fit the original observed choice. The welfare difference reflects identification and normative assumptions, not merely estimation error.

Handel and Kolstad's information-friction welfare analysis directly studies how these alternative microfoundations affect counterfactual allocations. Abaluck and Gruber's Part D benchmark evaluates outcomes under an explicitly specified optimization standard. Both underscore the need to name the welfare objective and the relevant feasible alternatives.

A transparent welfare protocol states: whose welfare is counted; whether utility reflects subjective or objective risks; whether beliefs are treated as autonomous values or correctable misperceptions; whether unconsidered alternatives were actually feasible; how consumers trade premiums, risk, liquidity, and quality; and how insurer costs, cross-subsidies, and fiscal effects enter the calculation. A single monetary compensating variation from a fitted choice model should not obscure these decisions.

Equilibrium feedback and selection

Changes in individual plan choice can shift the insurer's cost pool. An informational intervention may cause low-risk consumers to reduce coverage disproportionately, raising average claims cost among those remaining insured. Conversely, it might lead high-risk consumers to buy appropriate coverage while changing premiums faced by all participants. The net welfare effect depends on market structure, insurer pricing, the risk distribution, and the policy mechanism.

Einav, Finkelstein, and Cullen's price-variation welfare framework links insurance demand and cost curves to welfare measurement under selection. Their empirical design explains why both the quantity response to price and the cost of marginal enrollees matter. This equilibrium analysis complements but differs from identifying behavioral primitives: a model can estimate welfare-relevant demand and cost responses without uniquely identifying each psychological component, while a psychologically detailed model may need additional assumptions to forecast equilibrium premiums.

Handel's inertia and adverse-selection result is particularly instructive because the direction of a policy's direct effect on individual choices differs from its overall equilibrium welfare result in the studied setting. It gives a concrete reason to keep individual-level preference identification distinct from aggregate policy evaluation.

Barseghyan and Molinari's bundle-insurance counterfactual likewise connects heterogeneous preference types and consideration to a hypothetical product intervention. The lesson for interpretation is that welfare predictions must use the same choice-set and preference mechanisms that were identified, together with any additional assumptions governing changes under the intervention.

Evidence boundaries and research ethics

A practical welfare or policy claim should be traceable backward from the proposed counterfactual to the original data and exclusion restrictions. What variation identified the demand response? Did any survey measure the buyer's knowledge? Were claims defined before selection? Were premium changes exogenous? Did the data observe the rejected alternatives? Were the estimated parameters identified as levels, rankings, bounds, or a distribution?

The contextual evidence also includes the limits of privacy and reproducibility. Insurance administrative data can expose sensitive financial and health information. Some original studies use proprietary records; the public literature can validate reported methods and published results without providing access to individual observations. It would be inappropriate to imply a fresh replication or independent numerical verification of unpublished data unless access and computation actually occurred.

Evidence review is strongest when it separates source validation, mathematical checking, and empirical replication. Publisher records can verify titles, bibliographic facts and study abstract claims. Author-hosted articles can permit closer scrutiny of identification assumptions. Supplemental code may establish what routines were implemented. None of these acts alone proves that every numerical inference in a confidential dataset was independently reproduced.

The scope of this monograph is international in comparative ambition, but its documented empirical examples originate chiefly in Israeli automobile insurance, U.S. property and health insurance, French automobile insurance, and selected institutional health-policy settings. The cited works are valuable precisely because their institutional conditions are specified. Their findings should not be elevated to global population constants.

Model Comparison and Robustness

Nested and nonnested specifications

A coherent comparison begins by stating the competing representations and their identifiable implications. Useful baseline classes include objective expected utility with common curvature, objective expected utility with heterogeneous curvature, subjective expected utility, rank-dependent probability weighting, reference-dependent preferences, random utility with full consideration, and risk preferences jointly estimated with restricted consideration. Some classes nest others under parameter restrictions; other classes are nonnested or can be observationally equivalent on a specific menu.

Likelihood fit can inform the choice among fully specified models, but it is not enough to establish that the model's behavioral interpretation is true. A more flexible model can fit more choice patterns by construction. Penalized fit, held-out predictive likelihood, and cross-context validation help compare forecasting performance. Identification tests, dominance restrictions and excluded variation answer a different question: which interpretations survive the data under explicit restrictions.

The findings of Sydnor's modest-risk insurance study provide an example of a diagnostic moment: standard expected utility requires implausibly strong curvature for some observed choices. The results of The Nature of Risk Preferences show that probability distortions can improve explanation in a more general structural model. The 2021 heterogeneous-choice-set study shows that allowing restricted consideration changes what can be concluded about curvature. These findings should be interpreted cumulatively: each broadens or tests a maintained restriction, and the fact that multiple mechanisms can rationalize choices makes independent distinguishing information valuable.

A population mixture model brings its own identifiability conditions. Barseghyan and Molinari's two-type specification gains identification from choices in multiple contexts and covariate variation under formally stated assumptions. Arbitrarily adding more latent types without corresponding variation may create multiple parameterizations that generate nearly identical choice shares.

Identification stress tests

A research-grade identification audit can be implemented as a documented sequence of analytic questions. First, reconstruct the offered price-deductible menu for every observed chooser and confirm that the menu corresponds to the decision date. Second, define the loss event and distinguish claim incidence, claim reporting, and paid claims. Third, assess whether pricing variation depends on observed and unobserved risk. Fourth, derive the exact choice inequalities for each model. Fifth, inspect whether any latent components enter those inequalities only through a composite expression. Sixth, vary consideration and subjective risk restrictions and report resulting changes in identified sets.

This sequence is an editorial synthesis of established econometric concerns, not a newly named theory or a claim of original mathematical discovery. Its force comes from the examples: a binary effective weight cannot separate beliefs and weighting; a consideration set of size one can rationalize a selected option regardless of the preferences over unseen alternatives; and a post-choice claims measure can misrepresent ex ante risk if reporting depends on the chosen contract.

Robustness checks should also interrogate assumptions that are easily hidden in code. The researcher should test sensitivity to alternative risk prediction methods, Poisson versus other claim-frequency assumptions, the size and shape of heterogeneity distributions, dependence between risk and preferences, wealth measurement, starting points of optimization, and the scale of random utility noise. A quantitative result that survives many specifications is more informative when the specifications actually challenge the economic mechanism rather than merely change an optimizer or a minor covariate.

Cross-validation must prevent leakage from future claims into ex ante predictors if the purpose is to forecast or infer beliefs at the time of purchase. A carefully held-out period can assess predictive stability after product or pricing changes. A held-out insurer can test some forms of transport, but differences in risk composition and contract presentation must remain part of the analysis.

Replication and reporting standards

The minimum reporting unit for structural insurance preference research is more detailed than an estimate and its standard error. It includes the policy event and exposure definition; the complete offered menu; the assumed consideration process; the model of claims and losses; the utility or weighting function; the normalization of latent parameters; the nature of endogenous prices; the identifying variation; the estimand; numerical estimation procedures; and the uncertainty attached to the identification region and estimator.

Replication evidence should disclose which data are public, restricted, or confidential and provide code, synthetic test fixtures, or sufficient algorithmic descriptions when legal and privacy conditions constrain release. The published limited-consideration research record points to supplemental materials and a replication package. The authors also list companion MATLAB programs among their public research software. The availability of computational materials must be distinguished from access to confidential policyholder data; the latter is governed by the original research conditions.

All empirical claims in this text remain attributed to their specific research reports. The interpretive and mathematical discussion was developed for this monograph, but no new dataset was introduced, no external peer review was commissioned, and no publication status is inferred from the existence of a source manuscript. A scientifically meaningful archive preserves the actual checked sources and their relationship to exact claims rather than treating a collection of DOI strings as a substitute for review.

Integrated findings

The first finding is that a deductible choice reveals a ranking of contracts under the person's actual opportunity and consideration sets. Recovering curvature, probability weighting, or subjective probability separately requires further restrictions. This conclusion follows directly from the expected-utility inequalities and the behavioral observational-equivalence example.

The second finding is that claims observations are essential to the risk side of many models but are not direct readings of subjective beliefs. They should enter through an explicit loss and observation process with consistent temporal ordering. Selection, reporting, and moral hazard determine whether observed claims identify objective exposure.

The third finding is that multiple lines of insurance create more restrictions and potentially more identification, provided cross-context stability is evaluated rather than presumed. The progression from Barseghyan, Prince, and Teitelbaum (2011) to the 2016 partial-identification analysis shows why a violated common-parameter restriction calls for diagnosing the model rather than mechanically assigning the difference to psychology.

The fourth finding is that limited consideration can be a fundamental source of otherwise apparent high risk aversion. The 2021 Econometrica analysis and 2021 AER model provide distinct formal treatments: partial versus sufficient-condition point identification. Their results demonstrate how a preference parameter's interpretation changes with the assumed opportunity-to-attention mapping.

The fifth finding is that consumer mistakes and information frictions must be separately measured when welfare is at stake. The human-centered health-insurance estimates, Medicare plan choices, and dominated-menu evidence illustrate different ways observed decisions can depart from an ideal full-information benchmark. Each study's empirical setting and definition of mistake remain distinct.

The sixth finding is that counterfactual welfare requires more than identified taste parameters. It also needs a welfare criterion, valid intervention invariance, and, where relevant, price and selection feedback. The insurance welfare framework and inertia equilibrium result show why improved individual decision making and improved market welfare need not coincide.

These findings amount to an original synthesis of established literatures and formal distinctions, with no new named model, proprietary estimate, or unsupported novelty claim. Their value lies in forcing each substantive inference to remain attached to its identifying variation and defensible domain.

Conclusion

The structural identification of behavioral insurance preferences begins with a mathematically definite object: the joint relationship between contract opportunities, observed selections, risks, outcomes, and latent decision mechanisms. In a simple deductible menu, purchasing protection establishes a choice inequality. It can bound risk aversion if risk, payoffs, beliefs, and consideration are restricted. When the effective decision weight combines a subjective belief and a nonlinear weighting function, the same observation admits more than one behavioral explanation. When a cheaper option was never considered, the inference shifts again.

The empirical literature supports this layered interpretation. Cohen and Einav established a structural approach to using deductible menus and risk heterogeneity. Barseghyan and collaborators investigated probability distortions, stability, partial identification, and attention-dependent choice sets. Health-insurance evidence introduced direct measurements of information frictions, inertia, plan complexity, and dominated alternatives. Welfare research then showed that individual choices, aggregate selection and insurer costs must be analyzed together for policy predictions. Each contribution identifies a different component under a different set of restrictions.

A durable scientific conclusion follows from their comparison. The credibility of an insurance preference estimate rests on how the identifying variation discriminates among competing explanations. The parameters of a model acquire an empirical meaning through observable exclusions, measurement assumptions, and stability conditions; their numerical precision is secondary to that logical structure. Set-valued conclusions, explicit observational equivalence and carefully bounded counterfactuals can therefore be stronger scientific outcomes than a falsely unique psychological coefficient.

This Research Dossier provides a source-grounded international analytical synthesis of those identification questions. Its equations are illustrative or algebraically derived from declared assumptions; its literature findings remain attributable to their original studies; and its discussion specifies where additional evidence would change the inferential result. The work is prepared as a research-source monograph for later authorized batch publication, without assigning a first-publication date or public research identifier.

References

  • Abaluck, Jason, and Jonathan Gruber. 2011. "Choice Inconsistencies among the Elderly: Evidence from Plan Choice in the Medicare Part D Program." American Economic Review 101 (4): 1180–1210. https://doi.org/10.1257/aer.101.4.1180.

  • Barseghyan, Levon, Maura Coughlin, Francesca Molinari, and Joshua C. Teitelbaum. 2021. "Heterogeneous Choice Sets and Preferences." Econometrica 89 (5): 2015–2048. https://doi.org/10.3982/ECTA17448.

  • 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, Francesca Molinari, Ted O'Donoghue, and Joshua C. Teitelbaum. 2018. "Estimating Risk Preferences in the Field." Journal of Economic Literature 56 (2): 501–564. https://doi.org/10.1257/jel.20161148.

  • Barseghyan, Levon, Francesca Molinari, and Matthew Thirkettle. 2021. "Discrete Choice under Risk with Limited Consideration." American Economic Review 111 (6): 1972–2006. https://doi.org/10.1257/aer.20190253.

  • Barseghyan, Levon, Francesca Molinari, and Joshua C. Teitelbaum. 2016. "Inference under Stability of Risk Preferences." Quantitative Economics 7 (2): 367–409. https://doi.org/10.3982/QE511.

  • Barseghyan, Levon, and Francesca Molinari. 2023. "Risk Preference Types, Limited Consideration, and Welfare." Journal of Business & Economic Statistics 41 (4): 1011–1029. https://doi.org/10.1080/07350015.2023.2239949.

  • 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.

  • Chetty, Raj. 2006. "A New Method of Estimating Risk Aversion." American Economic Review 96 (5): 1821–1834. https://doi.org/10.1257/aer.96.5.1821.

  • Chiappori, Pierre-André, and Bernard Salanié. 2000. "Testing for Asymmetric Information in Insurance Markets." Journal of Political Economy 108 (1): 56–78. https://doi.org/10.1086/262111.

  • 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.

  • Einav, Liran, Amy Finkelstein, and Mark R. Cullen. 2010. "Estimating Welfare in Insurance Markets Using Variation in Prices." Quarterly Journal of Economics 125 (3): 877–921. https://doi.org/10.1162/qjec.2010.125.3.877.

  • Ericson, Keith M. Marzilli, and Amanda Starc. 2016. "How Product Standardization Affects Choice: Evidence from the Massachusetts Health Insurance Exchange." Journal of Health Economics 50: 71–85. https://doi.org/10.1016/j.jhealeco.2016.09.005.

  • Finkelstein, Amy, and Kathleen McGarry. 2006. "Multiple Dimensions of Private Information: Evidence from the Long-Term Care Insurance Market." American Economic Review 96 (4): 938–958. https://doi.org/10.1257/aer.96.4.938.

  • 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.

  • Handel, Benjamin R., and Jonathan T. Kolstad. 2015. "Health Insurance for 'Humans': Information Frictions, Plan Choice, and Consumer Welfare." American Economic Review 105 (8): 2449–2500. https://doi.org/10.1257/aer.20131126.

  • 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.

  • Kőszegi, Botond, and Matthew Rabin. 2006. "A Model of Reference-Dependent Preferences." Quarterly Journal of Economics 121 (4): 1133–1165. https://doi.org/10.1093/qje/121.4.1133.

  • Manski, Charles F. 2003. "Identification Problems in the Social Sciences and Everyday Life." Southern Economic Journal 70 (1): 11–21. https://doi.org/10.1002/j.2325-8012.2003.tb00553.x.

  • 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: 297–323. https://doi.org/10.1007/BF00122574.

  • Yaari, Menahem E. 1987. "The Dual Theory of Choice under Risk." Econometrica 55 (1): 95–115. https://doi.org/10.2307/1911158.

  • Barseghyan, Levon. 2026. "Publications and Software." Cornell University, author-maintained publication listing. https://barseghyan.economics.cornell.edu/publications.html.