Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Structural Estimation of Insurance Demand and Costs

Abstract. Structural estimation of insurance demand and costs recovers economically interpretable relationships among contract choice, risk heterogeneity, expected insurer payments, prices, and market equilibrium. Its central empirical difficulty is that the population purchasing insurance changes with the premium and contract design. The conditional mean expenditure of insured persons consequently changes even when the underlying distribution of risks remains stable. This monograph develops the mathematical and empirical distinction between demand, average cost, marginal entrant cost, insurer production or administrative marginal cost, and equilibrium markups. It explains the identifying roles of exogenous price variation, claim and choice panels, contract-menu variation, exclusion restrictions, and joint estimation of private risk and preferences. The analysis examines standard competitive selection models, differentiated-plan markets, behavioral frictions, endogenous utilization, dynamic enrollment, and the boundaries of policy counterfactuals. Evidence from automobile insurance, employer health coverage, Medicare-related markets, and annuities demonstrates why structural estimates are market-specific. A core conclusion is that the welfare information contained in a reduced-form demand-and-cost representation depends on the policy margin, the contract environment, and the stability of behavior and costs under intervention. The monograph provides transparent derivations, falsification tests, an evidence-to-assumption audit, and a practical reporting standard for interpreting structural insurance estimates. It is an original analytical synthesis of established research, not the announcement of a new theory or an empirical estimate from newly collected data.

Keywords: Structural insurance economics; insurance demand; adverse selection; advantageous selection; average cost; marginal cost; risk heterogeneity; discrete choice; price endogeneity; moral hazard; insurer competition; market equilibrium; welfare counterfactuals; partial identification; actuarial expenditures; consumer information.

Introduction

Insurance is a jointly determined allocation of risk bearing, contingent payments, and financial resources. The price of a policy determines who enrolls; enrollment changes the risk pool; the risk pool influences expected payments and pricing; and the resulting contract menu may change subsequent exposure and utilization. Structural estimation addresses these reciprocal relationships by specifying the economic choices and information sets that could have generated observed demand, claims, prices, and contract offerings. The classic competitive analysis of Rothschild and Stiglitz (1976) places private risk information at the center of contract formation. The empirical program summarized by Einav, Finkelstein, and Levin (2010) turns observable choices and cost data into economic parameters and evaluable counterfactuals.

The adjective structural has a precise operational meaning. The researcher represents household preferences, loss processes, insurers' objectives, constraints, and equilibrium conditions through parameters that remain interpretable under a specified class of interventions. A coefficient is structural only relative to those maintained mechanisms and that policy class. A premium coefficient in a logit regression can predict take-up locally without identifying willingness to pay for protection, risk aversion, claim costs under a different deductible, or the reaction of rival insurers. Structural research makes those extra claims only after documenting their identifying assumptions.

The principal objects are heterogeneous willingness to pay, a distribution of expected losses conditional on enrollment, the cost of covering marginal consumers, and the supply response of insurers. The practical question is how much of each object can be learned from data and how much is imposed by functional form. The price-variation approach of Einav, Finkelstein, and Cullen (2010) provides a particularly transparent starting point: observed demand and expected claims can trace two schedules and identify welfare distortions in a specified selection environment. However, choice of one insurance tier among several, utilization responses to coverage, administrative loadings, and market power alter the interpretation of those schedules. This monograph derives these connections and explains the conditions under which each extrapolation is legitimate.

The international scope concerns transferable methodology, institutional contrasts, and the external validity of evidence. It does not presume that employer-sponsored United States health insurance, French automobile insurance, British annuity choice, or globally reported market aggregates obey a single empirical cost schedule. The OECD Global Insurance Market Trends 2026 reports broad comparable insurance-market aggregates, which describe institutional scale and performance but do not directly identify individual preferences or marginal entrant costs.

Methods consist of source-specific economic interpretation, symbolic derivation, comparison of observed research designs, and explicit sensitivity analysis. No confidential insurance data were accessed, no new dataset was estimated, and no source-specific welfare effect is presented as a universal parameter. A conceptual calculation below is labeled hypothetical wherever it illustrates, rather than measures, an economy.

Structural Market Model Definition

Agents, contracts, and states of the world

Let an insurance contract j be described by the premium p_j, coverage rules g_j, and associated administration and service attributes a_j. A consumer i has initial wealth w_i, private risk characteristic r_i, preferences theta_i, information I_i, and possibly liquidity or transaction constraints. The uninsured alternative is j = 0. Under a standard expected-utility representation, consumer choice maximizes expected utility of wealth after premium, loss, and indemnification, conditional on I_i. For example, for a binary loss X_i with claim probability pi_i and a deductible d_j, an illustrative payoff can be written as:

W_i(j, X_i) = w_i - p_j - X_i + max(X_i - d_j, 0)
U_ij = E[u_i(W_i(j, X_i)) | I_i] + b_ij - h_ij
j_i = argmax_j U_ij

Here b_ij represents nonfinancial service utility and h_ij transaction, search, or hassle cost. The example abstracts from limits, coinsurance, taxes, reimbursements, network constraints, and post-contract behavioral changes until these are introduced explicitly. Deductibles and utility curvature jointly shape contract ranking. Mossin's insurance-purchasing analysis establishes a foundational expected-utility benchmark, while Cohen and Einav's deductible-choice model shows how data on chosen deductibles and claims can help identify risk attitudes when private risk and risk aversion both vary.

A structural parameter vector may include theta = (beta, Sigma, gamma, phi, eta): beta for mean price and coverage preferences, Sigma for their heterogeneity, gamma for loss probabilities and severities, phi for induced utilization, and eta for pricing conduct or cost components. A model is empirically informative when each parameter is linked to an observable moment or a credible external restriction. Two structurally distinct mechanisms that generate identical observed plan shares cannot be distinguished merely by numerical convergence.

Population and choice margins

For consumer i, define the choice probability s_ij(p,g,z;theta) using available characteristics z and the actual feasible menu. Aggregate demand is Q_j = sum_i s_ij, appropriately weighted for sampling and market size. Both the extensive margin (whether to insure) and intensive margin (which plan or protection level to buy) matter. The two-margin insurance analysis of Geruso, Layton, McCormack, and Shepard (2023) demonstrates that policy changes directed at one margin can create trade-offs at the other. In a market with mandated coverage, the outside option may be restricted; estimating take-up among eligible choices is therefore different from estimating total insurance participation.

The econometrician observes only a portion of I_i and r_i. Claims after enrollment can reveal realized loss outcomes under the actual contract but are affected by coverage, reporting, utilization, and censoring. Nonenrollees' counterfactual insured claims are often missing. A clear structural study begins by drawing this observation boundary before estimating a probability distribution.

Accounting identity and causal architecture

Write total expected variable claims payments for an enrollment set E as T(E,g), and administration, distribution, capital, and other insurer costs as A(E,g). The expected accounting cost is C(E,g) = T(E,g) + A(E,g). If sorting can be summarized by coverage quantity q in a fixed contract environment, C(q) is the expected total cost associated with the marginal population induced to enroll. The corresponding marginal cost is dC(q)/dq. This is distinct from the physical or organizational marginal cost of issuing another policy to a risk type held fixed: the derivative includes the change in the composition of buyers. Consequently, marginal entrant costs and marginal underwriting costs require different conditioning statements.

Observed prices are endogenous because insurers may adjust rates for risk mix, selection, regulation, capital, competitors, and demand. The economic causality therefore runs in a loop: price to enrollment, enrollment to average claims, claims and operating costs to price, and price back to choice. An estimated price-demand association without a credible identification strategy blends all these pathways. The structural system must specify how the loop is broken statistically, by instruments, institutional timing, fixed effects, experiments, or jointly estimated equilibrium restrictions.

Demand, Willingness to Pay, and Contract Choice

Demand as the inverse willingness-to-pay distribution

Consider a fixed insurance product and a normalized eligible population of mass one. Let v_i be each individual's willingness to pay for that product relative to the outside option. At a uniform premium p, enrollment is Q(p)=Pr(v_i >= p), assuming consumers know the relevant features, can pay the premium, and make that comparison without other frictions. If the distribution is continuous and strictly monotone over the policy range, inverse demand D(q) is the premium at which q of the population enrolls. D(q) orders people by willingness to pay, not necessarily by health risk or loss probability. Risk aversion, expected loss, wealth, information, and service valuation all contribute to v_i.

A complete demand model must explain both why people select one deductible rather than another and why they buy insurance at all. For deductible d, the premium trade-off against a different deductible varies with the distribution of the total loss, probability of loss, and utility curvature. As Cohen and Einav (2007) explain through automobile-insurance evidence, jointly modeling realized claims and deductible choices allows risk and preference heterogeneity to enter separately, subject to assumptions about utility and claim behavior. Purely estimating price elasticities from alternative premiums may fail to identify those primitives.

Discrete choice and heterogeneous consumers

In a differentiated market, one useful specification is:

U_ij = delta_j - alpha_i p_j + beta_i' g_j + xi_ij + epsilon_ij
P_ij = Pr(U_ij >= U_ik for every feasible k)
Q_j = sum_i omega_i P_ij

The taste vector (alpha_i, beta_i) may vary across consumers. The unobserved quality xi_ij can correlate with premium because desirable plans charge more, generating endogeneity. The random utility shock epsilon_ij must have an interpretable distribution relative to what consumers know, and the outside option must be economically credible. Train's simulation methods provide the general machinery for mixed logit, simulated likelihood, and handling of price endogeneity in choice models. The differentiated-products demand-and-supply tradition of Berry, Levinsohn, and Pakes (1995) is methodologically important for recovering demand and supply parameters jointly; direct transfer to insurance requires modeling risk selection and contingent claim payments.

Willingness to pay versus experienced welfare

Revealed willingness to pay represents welfare under a model of informed and stable preferences. When decision errors, premium salience, inertia, or liquidity constraints affect choices, observed enrollment cannot automatically be inverted into a normative welfare ranking. In Medicare prescription drug coverage, Abaluck and Gruber (2011) document systematic differences between plan financial characteristics and their apparent choice weights. The health-insurance information-frictions study of Handel and Kolstad (2015) uses administrative and survey data to distinguish risk preferences from knowledge and hassle costs. Structural work should therefore define whether demand is behavioral demand, informed-preference demand, or a partially identified combination.

The distinction changes policy counterfactuals. A premium subsidy can increase enrollment through affordability, while a disclosure intervention can change enrollment through knowledge. A model calibrated only to a single observed price coefficient may predict identical enrollment shifts while assigning very different welfare meanings to them. An estimand must indicate whether it measures chosen-plan utility, expected material consumption, informational welfare, or a broader objective that values access and distribution. These meanings cannot be recovered from a plan-share equation alone.

Average Claims, Marginal Enrollees, and Economic Cost

Selection transforms the cost curve

For a single fixed insurance contract, let c_i denote the expected incremental resource cost of insuring consumer i under that contract, inclusive of expected claims payments where these are the transfer-equivalent cost relevant to the insurer. Sort prospective enrollees by descending willingness to pay v_i. At quantity q, define AC(q) as the mean expected c_i among the first q enrollees. The expected total payout schedule is T(q)=q AC(q), after normalizing eligible population size to one; hence:

MC(q) = dT(q)/dq = AC(q) + q dAC(q)/dq
AC(q) = (1/q) integral from 0 to q of MC(t) dt

This calculus identity requires an appropriately differentiable enrollment ordering and a fixed benefit design. It is often more useful than a claim that an average-cost regression directly reveals marginal risk. The derivative of total expected claims identifies the expected cost of consumers at the participation margin; the mean is a pooled statistic. The Einav, Finkelstein, and Cullen framework uses this distinction to measure selection and market-level welfare.

In adverse selection, consumers with higher expected claims tend to be willing to pay more, so the marginal entrant may be cheaper than those already insured; MC(q) is then commonly below AC(q). In advantageous selection, lower-risk consumers may be disproportionately eager to insure, and MC(q) may exceed AC(q). The sign is an empirical ordering statement, not an unchanging attribute of a country or line. The comparative account by Einav and Finkelstein (2011) and the review by Cohen and Siegelman (2010) emphasize variation in observed selection patterns across markets and insurance pools.

The difference between claims and insurer costs

Expected paid claims are not identical to the marginal economic cost of supplying insurance. A useful decomposition is:

C(q) = T(q) + V(q) + F + K(q)
MC_total(q) = dT(q)/dq + dV(q)/dq + dK(q)/dq

Here V(q) includes variable acquisition, underwriting, claims handling, fraud control, and administration; F is an allocated fixed cost; K(q) represents additional capital or reinsurance funding costs as a function of enrollment and risk concentration. A literal actuarial interpretation must account for claim development, incurred-but-not-reported losses, policy duration, risk sharing, recoveries, taxes, and the valuation horizon. The split avoids attributing an observed premium-minus-claims margin exclusively to market power. It can also reflect expenses and compensation for risk-bearing capital, although the size and efficiency of those components require independent evidence.

The welfare role of claims requires care. When health insurance pays providers, paid benefits can be associated with actual medical resource consumption, transfers, and responses of utilization to coverage. In a purely financial indemnity contract with externally fixed loss, an indemnity payment is a contingent transfer. The policy-specific social resource cost therefore need not equal the insurer's expected claim payout. Efficient enrollment measured by willingness to pay minus expected payments is a useful partial-equilibrium benchmark under its assumptions; a wider social welfare analysis must add administration, externalities, public finance, utilization, and provider costs where relevant.

Recovering cost of the marginal pool

Suppose an exogenous premium change reduces enrollment from q_0 to q_1. If risk exposure, contract terms, benefit rules, claims handling, and selection ordering stay fixed, the cost difference T(q_0)-T(q_1), divided by q_0-q_1, estimates average expected cost for the people who leave over that interval. It does not immediately identify the cost of each person or the marginal cost at other coverage levels. The division requires consistent denominators: policyholders versus covered lives, monthly versus annual cost, insured person-time, and pricing categories. The estimand changes if exit is concentrated among dependents, high-utilizers, or people switching to other plans.

Illustration, not an empirical estimate: assume q=0.5 of a normalized population enrolls and AC(q)=180-80q in arbitrary currency units for the fixed product. Then AC(0.5)=140, total expected payout T=70, and MC(0.5)=180-160(0.5)=100. A premium benchmark of 140 is associated with an entrant cost of 100; equating premium with the entrant's cost would overstate that marginal claim burden. The example illustrates a selection gradient only. It omits administrative loadings and supply equilibrium and has no observational status.

Cost heterogeneity and joint distributions

A marginal-cost curve is determined by the joint distribution of willingness to pay v and expected cost c. Even if researchers know the full marginal distributions of v and c, distinct dependence structures between them can generate different selection slopes. An estimated negative correlation between coverage and risk may reflect omitted preferences or reporting differences. The Medigap advantageous-selection study by Fang, Keane, and Silverman (2008) documents the importance of multiple observed attributes, including cognition and socioeconomic variables, for apparent favorable selection. Its empirical finding applies to the studied Medicare supplement market and cannot substitute for direct measurement in other populations.

For joint-risk structural work, the objective is to estimate f(v,c | z) or a sufficiently informative parameterization. The exact joint distribution is often only partially identified from one menu. Observing claims conditional on insurance and observing choices among several price-deductible alternatives add identifying content, but sampling and utilization changes can still frustrate extrapolation beyond the available support. An honest cost figure is therefore always accompanied by the population, coverage, time period, conditioning set, and model assumptions that define it.

Price and Coverage Variation: Identification

Why market prices are endogenous

An insurance premium reflects expected losses, contract generosity, product quality, loadings, insurer conduct, regulation, and anticipated clientele. A market-level regression of purchases on premiums can return a positive association even when individual demand slopes downward: high-risk markets may simultaneously have higher premiums and greater willingness to buy. Conversely, strong competitors can select risk pools and undercut prices. Formally, with a demand equation Q=Q(p,x,xi), the unobserved demand determinant xi can correlate with the insurer's price-setting equation p=P(Q,c,regulation,competition). Holding observed x fixed does not eliminate that simultaneous causation.

Valid identifying variation must change the price faced by consumers without independently changing their preferences, latent health, product quality, or future expected claims through a direct channel. Employer contribution formulas, administratively set price schedules, timing rules, or credible quasi-experimental shifts can sometimes supply such variation. The 2010 welfare-estimation study illustrates the combined use of price, enrollment, and claims data in one employer environment; it is not evidence that every employer price schedule is exogenous.

An instrumental-variable strategy must identify the relevant first stage, exclusion restriction, monotonicity or treatment-effect interpretation, and the market equilibrium response. Cost shocks can be invalid instruments for demand when policyholders anticipate the same underlying risk shock. Rival insurers' prices can also be endogenous when common market disturbances affect rivals simultaneously. Geographic variation alone does not guarantee exogeneity if geography proxies for risk exposure, provider prices, income, or insurance mandates.

Price variation and the composition effect

Let p be an instrumented premium and write expected aggregate paid claims as T(p)=Q(p)AC(p). Differentiating produces dT/dp=(dQ/dp)AC+Q(dAC/dp). If dQ/dp differs from zero and all other assumptions hold, the ratio (dT/dp)/(dQ/dp) identifies cost of marginal enrollment along the price-induced participation path. The calculation is path-specific when changes also shift contract type, quantities insured, or medical utilization. It is particularly important to state whether T(p) is measured on a fixed eligible cohort or only among observed insured persons.

Claims databases often exclude rejected applicants, claim-free insured persons, never-enrolled individuals, and claims outside reporting windows. Those missing data generate selection in observed outcomes beyond the market selection under study. Researchers should track the enrollment population from a stable eligibility file and link claims to exposure periods, even when missingness requires weighting or bounds. Denominators and timing can dominate apparent precision in estimated average costs.

Coverage generosity and contract-menu experiments

The premium difference between a high-deductible and low-deductible plan reveals how much consumers pay for reduced out-of-pocket exposure only after adjusting for differences in perceived networks, claims administration, inertia, and service attributes. A single premium difference cannot separately identify the distribution of claim probability, risk aversion, and loss severity without additional data or restrictions. The structural exercise in Cohen and Einav (2007) combines claims and contract choices. The UK annuity analysis by Einav, Finkelstein, and Schrimpf (2010) estimates private mortality-related selection and preference heterogeneity across guarantee periods, illustrating a different contract and outcome structure.

Coverage variation also changes actual use of insured services. Health expenditure observed under generous insurance may exceed expenditure that the same consumer would generate under higher cost-sharing. The study of selection on moral hazard by Einav and coauthors (2013) shows why heterogeneous utilization responses can themselves influence plan choice. Thus, a fixed-risk cost curve extrapolated across deductibles requires a clear assumption about utilization, while a joint model estimates selection and behavioral response together.

Identification conditions and their diagnostics

Point identification requires at least four distinguishable ingredients. First, sufficient support exists in prices, menus, and consumer risk characteristics to identify the policy-relevant margin. Second, the specified choice and loss models map parameters to distinct observable distributions. Third, identifying variation satisfies its exclusion or institutional assumptions. Fourth, the outcome measured under each contract corresponds to the modeled cost concept. Strong functional-form restrictions can produce point estimates despite weak data; that computational fact does not demonstrate empirical identification.

Useful diagnostics include first-stage strength, overlap and positivity, price-shock balance tests, placebo timing, stability of contract composition, tests for pre-trends where relevant, and comparisons across alternative instruments. The reviewer should record which assumptions are tested with observations and which remain maintained restrictions. A credible study can report a range of welfare effects when the necessary exclusion restriction or tail distribution is uncertain. Precision conditional on a structural model and robustness to the model's central identifying assumptions are distinct achievements.

Latent Risk and Preference Heterogeneity

Two private dimensions

The standard two-dimensional insurance-selection problem separates underlying risk r from willingness to pay for risk transfer, which includes preference parameter a. A riskier household may value insurance more because expected claims are higher. A more risk-averse household may also value it more while engaging in preventive behavior that lowers expected cost. Their covariance determines whether observed enrollment tilts toward high or low cost. Fang, Keane, and Silverman (2008) provide a concrete empirical case in which factors beyond health risk and risk aversion matter to sorting.

A tractable latent distribution might represent log claim intensity and risk-aversion parameters as correlated random effects conditional on age, income, exposure, and other observed characteristics. A parametric copula links the margins, or finite mixtures permit distinct latent classes. In either case, the choice of joint-distribution family is a substantive identifying assumption. A good researcher compares alternatives and avoids treating a chosen copula's tail behavior as directly observed evidence.

Claims likelihood with insurance exposure

For a binary accident event, observed claim Y_i over exposure t_i may follow a Bernoulli or survival formulation conditional on risk r_i. A claim-count model might assume Poisson counts with exposure offset, while claim severity requires a separate distribution and censoring rules. A structural deductible-choice likelihood integrates over unobserved risk and preferences:

L_i(theta) = integral [
  Pr(choice_i | r, a, menu_i, theta)
  * f(claims_i | choice_i, r, contract_i, theta)
  * f(r, a | z_i, theta)
] dr da

This expression is a schematic likelihood: insured claims must be interpreted as contract-conditional observations, and missing counterfactual costs for the uninsured do not become observed merely by integrating. Panels, claims histories, risk-adjuster scores, and observed changes in plan terms can strengthen inference. However, claims reported only after enrollment may be endogenous to benefit generosity and demand behavior. An insurer's proprietary risk score can itself incorporate previous coverage and care, weakening its status as an exogenous proxy.

Learning, expectations, and information

Consumer information can evolve through claim experience, diagnosis, employer communications, and market education. A structural model with fixed lifetime risk beliefs may misinterpret learning as stable preference heterogeneity. Insurers likewise update underwriting information, sometimes with regulatory restrictions on permitted rating. Selection can occur on private information, private preferences, or anticipated behavioral responses; each has a different welfare mechanism. The cross-market adverse-selection review explains why a positive coverage-loss association alone does not establish one unique mechanism.

Latent heterogeneity should be assessed against out-of-sample behavior. An estimated random-effect distribution that fits current deductibles may predict poorly when a new option or information intervention becomes available. The most informative validation involves distinct choice sets or future enrollment periods rather than a randomly split sample in which the same incentive structure appears in both subsets. Where data cannot distinguish risk preferences from mistakes or uncertainty about future health, the model should report those competing explanations.

Supply, Competition, and Equilibrium Prices

Average-cost pricing and the selection equilibrium

Consider the uniform-contract benchmark in which consumers are ordered by willingness to pay and insurance provision is competitively priced at average total variable cost. The stylized equilibrium solves D(q*)=AC(q*) if such a fixed point is feasible and stable. The socially efficient enrollment quantity in the simple partial-equilibrium model instead satisfies D(q^eff)=MC(q^eff). This comparison assumes the social value of coverage is represented by willingness to pay, that risk-bearing transfers and costs are appropriately measured, and that administrative, fiscal, behavioral, and market externalities are absent or accounted for.

When MC(q) lies below AC(q) over the relevant range and inverse demand decreases with q, a stable price-equals-average-cost equilibrium generally produces less enrollment than the simple welfare benchmark. When MC lies above AC, the inequality can reverse. The graphical argument follows the insurance selection framework discussed by Einav and Finkelstein (2011). It depends on the particular pricing institutions: a regulated community-rated market, an experience-rated employer plan, a risk-adjusted marketplace, and a competitive insurer market with differentiated contracts use different price equations.

A welfare calculation can be written:

W(q) = integral from 0 to q [D(x) - MC_total(x)] dx
W(q_eff) - W(q_star)
  = integral from q_star to q_eff [D(x) - MC_total(x)] dx

This expression defines the insurance-allocation surplus under the stated assumptions; it does not include an automatic conclusion that all risk transfers or indemnity payments consume social resources. A credible welfare report says which part of claims is a real resource use, which part is a transfer, and how the benchmark treats risk aversion and public financing.

Market power and differentiated insurers

In many real markets an insurer sets a premium by maximizing expected profit rather than mechanically equating price to average claims. For one product, a simplified first-order condition is:

Profit(p) = [p - c(p)] Q(p) - F
dProfit/dp = Q(p) + [p - c(p)] Q'(p) - c'(p) Q(p) = 0

The term c'(p) Q(p) captures changes in average cost induced by the price-dependent enrollment mix. Omitting it treats buyer composition as invariant and can bias the apparent markup. With several products, a firm internalizes substitution among its products; the system of first-order conditions uses a demand-derivative matrix and a cost response for each product. Premiums also may depend on regulation, insurer capacity, solvency constraints, and strategic conduct.

The Medigap pricing study of Starc (2014) finds that brand preferences and low demand elasticity matter for pricing above cost in its studied US Medicare supplement setting. The selection-and-market-power analysis of Mahoney and Weyl (2017) shows theoretically and through calibrated applications that the interaction of selection and imperfect competition can reverse the effects of policies that seem desirable when each distortion is considered alone. Neither result establishes one universal markup or a welfare sign for all insurance markets.

Estimating insurer marginal cost

Recovering a marginal cost from observed premiums and a fitted demand model requires a supply model: ownership groups, pricing decision units, marginal revenue conditions, and an explicit definition of the cost function. When expected insurance payouts vary with the selected pool, the estimated structural marginal cost combines insurer operations and selection unless components are separately measured. Researchers can increase credibility by comparing cost estimates inferred from pricing first-order conditions with claims-based payout data and audited expense categories. Disagreement may signal unmeasured loading, misspecified conduct, unobserved product quality, capacity costs, or invalid instruments.

A joint demand-supply equilibrium can be viewed as a fixed-point map from premiums to demand, risk pools, costs, and revised premiums. Solving it numerically is meaningful only after checking whether the equilibrium exists, is unique over the relevant parameter set, and respects the insurer's actual institutional constraints. Multiple equilibria and nonexistence are legitimate theoretical outcomes in adverse-selection settings, as the foundational Rothschild-Stiglitz analysis makes clear. A solver finding one root does not certify uniqueness or economic relevance.

Risk adjustment and imperfect competition

Risk-adjustment transfers can change incentives to enroll and price consumers of different expected expenditure while leaving consumer-facing premiums partly regulated. If the transfer formula depends on diagnoses recorded under existing insurance, it may also affect coding and utilization. Its welfare impact depends jointly on who enters, which plan they choose, and how competing firms react. The two-margin study and imperfect-competition model make the general point that demand, risk mix, and supply response must be evaluated together. A policy simulation that changes risk adjustment while freezing insurer prices is a short-run conditional scenario, not a full market-equilibrium prediction.

Estimation, Computation, and Statistical Inference

A research design before a likelihood

The first step in a structural insurance study is to choose the estimand, observation unit, and counterfactual. Typical units include policy-year, household-year, employee-plan-year, insurer-product-market, and annuity-contract cohort. These units create different definitions of exposure, claims, and switching. Researchers should define enrollment eligibility, menu availability, pricing basis, contract renewal rules, claims runout, and the population for which welfare is reported. Choice probabilities can be estimated at one unit while claims are aggregated at another, but the bridge needs explicit weights and linkage identifiers.

The next step is to map observables to mechanisms. Premiums and plan menus identify choice margins given exogeneity; claims reveal realized costs conditional on enrollment and coverage; longitudinal changes reveal switching and responses; firms' financial statements and administrative records constrain expenses and fixed costs; and regulatory institutional details define feasible contracts. A researcher should distinguish independent external restrictions from parameters estimated solely because a parametric form makes a likelihood numerically tractable.

Maximum likelihood, moments, and simulation

A mixed discrete-choice model typically integrates over heterogeneous preferences and risk. Maximum likelihood multiplies or sums the log of household-specific conditional choice probabilities integrated over latent types. Simulation approximates those integrals by draws. GMM uses moment restrictions such as observed shares, price-response moments, and cost moments. Both methods require valid exclusion assumptions and an identified mapping between population moments and parameters. Train's second edition explains simulation error, maximum likelihood, mixed logit, and related estimation techniques.

It is important to preserve the difference between prediction and structural identification. A flexible machine-learning predictor can improve holdout predictions of plan purchase or claims while leaving the counterfactual response to a never-seen premium, mandate, or contract uncertain. Bajari and coauthors (2015) study machine-learning-based demand prediction outside insurance; their work is useful for prediction design but does not itself validate insurance welfare counterfactuals. Predictive performance is one diagnostic among several.

Separation of statistical and structural uncertainty

Model-fit uncertainty arises from sampling variation in observed choices and claims; structural uncertainty arises from competing preference distributions, cost technologies, excluded variables, market-conduct assumptions, and welfare objectives. Confidence intervals conditional on a chosen model quantify only a subset of total uncertainty. Standard errors should reflect the actual variation unit: consumers nested in employers, geographic markets, insurers, or policy years often generate clustered dependence. If prices vary at employer or insurer level, person-level independent standard errors substantially overstate effective sample size.

Bootstrap procedures should reproduce the original identification structure, including the price variation and estimation of the cost curve. Two-step estimators need to propagate uncertainty from the first stage through simulated welfare measures. Simulation draws should be sufficiently numerous and held fixed under parameter comparisons where suitable to reduce numerical noise. An optimization routine should report convergence tolerance, numerical gradients or verified analytic derivatives, starting values, local maxima, and feasible bounds. A point estimate that changes sharply with starting values warrants reporting the instability.

Counterfactual computation as a distinct stage

Once a model is estimated, an intervention changes one or more inputs: premium subsidy, permitted rating, risk adjustment, deductible menu, enrollment default, information, or firm conduct. The analyst must re-solve consumer choice and, for equilibrium counterfactuals, insurer decisions and the distribution of enrolled claims. Calculating demand at a lower premium while keeping the previous cost pool is an intentionally restricted simulation. Its value depends on clearly naming that restriction.

An adequate implementation includes: baseline reproduction; a scenario table of altered primitives; solution of enrollment, cost and premium conditions; policy budget and transfer accounting; welfare by population subgroup; and stability checks across functional forms and identification assumptions. If a scenario exceeds the observed premium or contract support, its outputs are model extrapolations and should be labeled accordingly. The main lesson from Beyond Testing is that insurance-market models can move beyond detecting selection to quantifying consequences, provided their economic assumptions are disciplined by data.

International and Cross-Line Evidence

Contradictory selection signs are informative

Structural insurance research cannot adopt one universal sign for the association between coverage and claims. In French automobile data, Chiappori and Salanié (2000) report no evidence of asymmetric information in the market and sample studied using their specified tests. In US Medigap, Fang, Keane, and Silverman (2008) report advantageous selection associated with socioeconomic and cognitive variables. In a specific US employer health-insurance application, Einav, Finkelstein, and Cullen (2010) find adverse selection but small estimated welfare implications in that setting. These are empirical findings about different contracts, populations, and observation mechanisms. They should be placed side by side as evidence of conditional heterogeneity, not averaged into a global selection coefficient.

The meta-analytical review by Cohen and Siegelman (2010) explains that observed correlation between coverage and subsequent events can emerge from adverse selection or moral hazard; absence of that correlation need not prove the absence of all private information. Structural identification must establish the precise economic mechanism and the timing of information relative to coverage choice.

Automobile insurance

Automobile deductibles are analytically useful because realized accidents, liability and damage claims, contract terms, and annual premiums can often be linked at policy level. The Cohen-Einav automobile model estimates risk aversion with unobserved accident risk, illustrating the role of a joint latent distribution. Its transfer to another jurisdiction requires compatible contract language, liability rules, accident reporting, compulsory coverage, and claims development. No global inference follows about average risk aversion or price elasticity from one insurer's portfolio.

The French result of Chiappori and Salanié also clarifies the empirical distinction between a test for conditional correlation and a full structural estimate of demand, cost, or equilibrium. A market can pass a particular correlation test while still having meaningful heterogeneity in driving exposure, claim reporting, deductible valuation, and premium response.

Employer and individual health insurance

Employer-based health choice often offers administratively measured premium contributions and integrated claims data; it also raises questions of eligibility, wage-offset incidence, and the restricted menu. The Alcoa employee application demonstrates the possibility of tracing demand and cost selection with price variation in a single employer's benefits setting. The estimated welfare distortion is conditional on the company's choices and model, not a national estimate of health insurance inefficiency.

Individual and multi-insurer health markets add insurer competition, risk-adjustment transfers, external subsidies, provider networks, and provider-price differences. The two-margin study of Massachusetts data highlights the importance of analyzing enrollment and plan sorting jointly. A meaningful multinational comparison must align the measurement of premiums, taxes, public subsidies, policyholder costs, provider payments, and actual risk-bearing liabilities.

Medicare supplement and prescription drug insurance

The Medigap setting studied by Starc (2014) connects differentiated-firm pricing, consumer brand preference, and intermediary rents. The related but distinct Fang-Keane-Silverman selection analysis examines enrollment and expenditure patterns. These studies are complementary precisely because demand, supply, and selection can differ in their data and causal identification strategies.

The Medicare Part D evidence creates an important contrast between choice error and learning. Abaluck and Gruber (2011) find patterns consistent with distorted weighting of premiums and expected out-of-pocket costs. Ketcham and coauthors (2012) document improvement in plan choices over time within their studied sample. A structural model of choice that ignores information acquisition and switching costs can misread dynamic learning as stable preferences, which changes welfare conclusions.

Annuity markets and mortality selection

Annuities exchange premiums or accumulated retirement wealth for streams of future payments, making longevity, bequest preferences, and guarantee periods central. Einav, Finkelstein, and Schrimpf (2010) estimate an annuity-choice model for the UK, including private mortality information and preference heterogeneity, and evaluate welfare and mandate counterfactuals in that institutional setting. The benefit-payment horizon and survival-selection mechanism are fundamentally different from medical utilization or automobile accidents. A pooled insurance-cost model can borrow the structural logic but must replace the specific mortality, payment, and choice equations.

International aggregate statistics and limits of comparability

The OECD 2026 market report collects comparable 2025 insurance market indicators for more than 50 OECD and non-OECD jurisdictions. Aggregates can motivate research questions about premiums, claims, insurance penetration, investment results, and institutional differences. They do not, by themselves, supply individual-level exogenous premium variation, joint choice-and-claims records, or a valid instrument. A meaningful international structural comparison therefore needs each country's contract, taxation, public social-insurance, regulation, reporting, and exposure definitions, plus harmonized microdata where possible.

The joint lesson is that structural estimators are portable as methods, while their identifying variation and estimated parameters are bound to institutional mechanisms. The responsible international claim is a comparative explanation of what data and assumptions are needed across markets. A globally universal demand elasticity, marginal cost slope, or policy welfare effect cannot be inferred from the cited market-specific studies.

Dynamic Enrollment, Information, and Moral Hazard

Insurance changes subsequent behavior

In health and some property and liability insurance, protection can alter utilization, care-seeking, maintenance, mitigation, claim submission, or reporting behavior. When behavior changes after contract choice, the counterfactual expected claims for consumer i depend on the contract j: c_i(j), not a fixed personal risk scalar. Premium variation then changes the selected risk population and, if associated with different coverage, can change each member's utilization. Structural work must distinguish the selection gradient holding contract terms fixed from the change in contract-conditional expenditure.

The selection-on-moral-hazard research by Einav, Finkelstein, Ryan, Schrimpf, and Cullen (2013) shows that consumers may sort on anticipated responsiveness of their medical utilization to insurance. As a consequence, assigning the same utilization response to everyone can distort predictions for a new deductible. For interventions that change benefit generosity, a consistent model jointly simulates risk, utilization, reimbursement, and demand. When only enrollment prices vary with coverage held constant, the simpler cost-of-entrant interpretation may remain informative.

Renewals, inertia, and switching

A dynamic insurance consumer chooses while facing contract renewal, past enrollment, information acquired from claims, switching costs, and future options. An illustrative state may include prior plan j_(t-1), current information I_t, expected future health, and account balances or accumulated wealth. A dynamic value function satisfies a Bellman equation under assumed discounting and transition probabilities. A static logit applied to current enrollment can mistake the cost of switching for a strong intrinsic preference for an insurer.

Handel's inertia study uses a major change in employer insurance provision to identify inertia and considers interventions that reduce it. Within that setting, simplifying individual choices can intensify adverse selection, with adverse aggregate welfare consequences. The result follows from a joint response of consumers and the risk pool: better individual matching to preferences need not imply higher market-level surplus when premiums respond endogenously.

Handel and Kolstad (2015) differentiate information frictions and hassle costs using combined administrative and survey evidence. Meanwhile, Ketcham and colleagues (2012) observe changes in Medicare Part D plan-choice performance over time. Together these findings justify testing whether changes across policy years reflect learning, stable taste, evolving health, menu revisions, or insurer repricing rather than presuming one fixed behavioral mechanism.

Experience rating and selection over time

The profitability of a contract may depend on a sequence of renewal premiums, claim experience, lapse probabilities, and changing composition of insured cohorts. Ex post claims are not a direct proxy for the ex ante expected cost seen by the insurer, particularly if pricing was based on a contemporaneous risk score. Estimation of long-run equilibrium should model new business, renewal, lapse, aging, and claims development on consistent time horizons. Where contracts guarantee renewability or constrain rating, the value of insurability itself changes willingness to pay.

A dynamic welfare counterfactual should account for transitional and steady-state effects separately. For example, information disclosure may shift selection immediately but improve understanding and lower switching costs over several renewals. A one-period estimate can describe the short-run effect while omitting adjustment dynamics; treating it as permanent requires an additional stability assumption.

Policy Counterfactuals and Welfare

Appropriate counterfactual questions

A structural model earns its policy relevance by making assumptions explicit and comparing internally consistent counterfactual allocations. Possible interventions include premium subsidies, tax treatment, coverage mandates, risk adjustment, limitations on risk-based pricing, standardized menus, higher deductibles, default enrollment, transparency, and changes in insurer competition. Each intervention changes one or more parts of the structural system. The welfare answer depends on whether it is modeled as partial equilibrium, full insurer equilibrium, or a broader public-finance setting.

For a uniform fixed plan, a subsidy s paid per policy changes the consumer-facing price from p to p-s. If consumer demand is Q(p-s) and insurer cost is C(Q), fiscal spending is approximately sQ, excluding administration and interactions with taxes or other benefit programs. A full welfare calculation counts who pays the subsidy, any financing distortion, changes in surplus from risk protection, real administrative costs, and changes in claims-related resource use. It does not count both subsidy spending and the corresponding premium transfer as independent social resource losses.

Mandates and choice restrictions

Mandating enrollment alters the feasible outside option and can change equilibrium prices, risk composition, and coverage quality. The annuity analysis by Einav, Finkelstein, and Schrimpf (2010) studies guarantee choices and welfare under particular UK annuity-market conditions. Its estimated intervention effects are findings about that market and the model's assumptions. Compulsory automobile liability insurance, compulsory health coverage, and required annuitization have different contractual objects and public-policy goals, and their welfare functions must reflect those differences.

A compulsory product with heterogeneous willingness to pay may impose costs on households whose perceived benefit is low. Whether that is socially desirable depends on third-party externalities, public risk sharing, redistribution, and enforcement costs. An insurance-demand model based only on private contract utility cannot supply all of these values. The correct practice is to state the omitted components instead of treating the insurance surplus curve as a complete social welfare function.

Rating regulation and redistribution

Community rating can pool different risks at one price, whereas risk classification or experience rating differentiates premiums subject to legal constraints. The selection effect of a rating rule depends on whether low-risk consumers leave, whether risk adjustment transfers offset expected cost, and how incumbents and entrants respond. Distributional consequences also matter: a change can increase aggregate surplus while reducing affordability for a vulnerable group. Structural models should report enrollment, price, claims and fiscal effects by risk, income, age, or other permissible analytical categories, observing jurisdiction-specific data and legal safeguards.

The interaction with insurer market power makes one-dimensional recommendations unreliable. Mahoney and Weyl (2017) show that interventions aimed at adverse selection may have adverse consequences in calibrated imperfect-competition environments. Their results motivate a joint evaluation of markup and selection, rather than a policy rule mechanically extrapolated from the sign of an observed risk gradient.

Information and default design

Improving a consumer's information can change perceived plan generosity and prices even when the contractual premium does not move. In their health-insurance choice study, Handel and Kolstad (2015) treat knowledge and hassle costs as separate determinants of demand. The normative effect of reducing those frictions depends on whether insured persons choose more appropriate plans, how their risk types sort, and how insurers reprice. Empirical evidence about revealed decisions does not establish the complete welfare function unless the model specifies what counts as an informed choice.

A robust reporting format includes a matrix of direct, composition, and equilibrium effects. For a subsidy, direct effects are altered net premiums; composition effects are changing enrollee risk; equilibrium effects are premiums or plan offerings adjusting. For information, direct effects are changed beliefs or choice cost; composition effects are sorting; equilibrium effects are repricing or entry. For risk adjustment, direct effects are insurer transfer schedules; composition effects are enrollment and coding; equilibrium effects are competitive price and product response. This is an organizational synthesis of established mechanisms, not a new named theoretical model.

Beyond the insurance firm's balance sheet

The insurer's profit and the welfare of insured persons are different accounting concepts. Policyholders value financial protection under uncertainty; insurers incur claims, administrative, and capital costs; public authorities may redistribute resources; and healthcare or repair services use real resources. In an efficient-allocation exercise, premiums can be transfers while insurance administration and treatment resources are costs, although the allocation of risk can create genuine utility gains. For risk-neutral consumers with predetermined losses, protection has a different welfare meaning than for risk-averse consumers. Every calculation must declare the welfare unit, aggregation weights, perspective, and time horizon.

An evaluation of catastrophic or long-tail insurance also requires distributional and tail-risk analysis, since expected cost alone conceals insolvency and correlation exposure. Such extensions are substantively relevant but cannot be recovered from the one-dimensional participation schedule without additional information. They delimit the applicability of simple demand-and-cost welfare diagrams.

Robustness, Replication, and Scientific Limits

Identifying assumptions audit

Before approving a structural estimate, the editor should ask whether the model separately identifies willingness to pay, claims under the chosen contract, marginal cost of entrants, supply loadings, and behavioral responses. A single premium and enrollment cross section almost never identifies all of them without strong restrictions. The available source must describe genuine instruments or price assignments, the population covered by data, how contract comparability is maintained, and the precise interpretation of cost. This audit is substantive rather than a checklist of software output.

An identification table can organize the review:

Economic object Minimum informative variation Common threat Stronger validation
Demand response Exogenous premium or contribution variation Correlation with private risk or product quality Institutional first stage; balance and placebo analyses
Choice of deductible Menu differences plus claims or credible risk information Joint unobserved risk and preference heterogeneity Joint likelihood; alternative risk distributions
Average payout cost Claims linked to a defined insured cohort Claims censoring and enrollment selection Exposure and runout checks
Marginal entrant cost Price-induced change in enrollment and aggregate claims Coverage-induced utilization and changing eligibility Fixed-plan comparisons; utilization sensitivity
Operating marginal cost Expense or supply-side information Fixed cost allocation and capital loadings Audited cost decomposition
Markup or pricing conduct Demand elasticities, ownership, supply objective Misspecified conduct and omitted expenses Alternative supply models, observed accounting comparison
Welfare Identified valuations and resource costs Frictions, transfers, and unsupported extrapolation Alternative welfare objectives and partial-identification bounds

Sensitivity to functional form

For risk-averse expected utility, changing the assumed utility curvature changes inferred willingness to pay from deductible data. Assuming a normal random effect versus a mixture distribution can change cost-tail predictions even if both fit average enrollment. Imposing a fixed moral-hazard response can bias costs in a new high-deductible scenario. An insurer supply model that assumes price setting by single-product competitors can misstate markups when firms own multiple brands. Credible structural research shows whether its key policy ordering survives these alternative specifications.

A clear sensitivity report gives estimates under a baseline and several empirically defensible alternatives, plus the assumptions that produce each difference. Model fit should be reported separately for choice shares, claims distributions, risk gradients, prices, renewals, and predicted policy margins. Passing one set of fit diagnostics does not validate all other mechanisms. Functional-form robustness is particularly important when identification relies heavily on data outside the observed support.

Source-to-claim integrity

The literature contains contradictory market-specific findings that should remain distinguishable. The French automobile finding in Chiappori and Salanié (2000) is a negative result for their asymmetric-information test, not a proof that selection is absent in every French policy segment. The employer-based Einav, Finkelstein, and Cullen result documents adverse selection with a modest welfare effect for a particular application, not a universal cost of selection. The Medigap advantageous-selection finding does not license reclassifying other supplemental insurance markets without evidence. The insurer-pricing study reports a specific Medigap competition and markup mechanism, not a global insurer margin.

An evidence ledger should record the precise claim supported by each source, its publication identity, population, identification strategy, and limits. The sources in this monograph are linked directly through descriptive phrases; the complete bibliographic identities follow below. The analysis distinguishes primary theoretical and empirical contributions from methodological texts and international descriptive statistics. Secondary interpretations and original synthesis are labeled in the narrative rather than attributed to studies that did not make those arguments.

Replication and external validity

A replicable structural insurance study should preserve anonymized specification materials, instrument construction, source data access conditions, parameter transformations, optimization procedures, standard-error code, policy scenario rules, and a documented table connecting raw observations to model variables. Where proprietary insurer claims cannot be disclosed, the full code and an adequately documented synthetic dataset can still validate computational logic without pretending to reproduce the confidential empirical result. Audit trails should distinguish independently executed checks from checks described only in the source literature.

External validity is best assessed mechanism by mechanism. Expected utility and endogenous sorting are general conceptual tools; the empirical distributions of risk, preferences, information, market power, and service costs are contingent. Institutional differences in mandatory insurance, public subsidies, underwriting regulations, provider payment, intermediary compensation, and contract enforceability can make numerical transfer inappropriate. OECD cross-jurisdictional aggregates help document market structure but do not solve microeconomic identification.

Research contribution and its limits

This dossier contributes an integrated analytical account of the relationships among insurance demand, entrant-cost selection, supply pricing, dynamic behavior, and welfare, with a concrete audit of what each empirical literature can establish. It derives the average-cost-to-marginal-cost identity, distinguishes the cost of the marginal risk pool from insurer operations, and presents a framework for interpreting alternative counterfactual scopes. These are explanatory derivations and an original synthesis of established economics. No new named theory, general equilibrium result, or novel empirical treatment effect is claimed.

The monograph's scientific limits are equally explicit: its symbolic derivations are conditional; its worked numerical example is hypothetical; its cited empirical effects remain tied to their observed markets; and a direct independent reanalysis of proprietary insurance datasets has not been undertaken. Any future dataset-based extension would require its own data access, ethical and regulatory basis, reproducible estimation, and dated scientific editorial record.

Integrated Conclusions

Structural estimation of insurance demand and costs explains market outcomes through the interaction of consumers' contract choices, heterogeneous risks, insurer expenses, pricing conduct, and equilibrium restrictions. The most consequential identification step is to separate willingness to pay from expected insured cost and to recognize that the expected cost of those who enter at a price margin differs from the average cost of all current enrollees. The accounting identity linking average and marginal entrant cost supplies a transparent empirical bridge, but it carries policy meaning only when the relevant product, enrollment population, and behavior are stable over the studied variation.

The empirical evidence establishes conditional diversity. Automobile insurance, employer health coverage, Medigap, prescription drug plans, and annuities differ in selection, preference heterogeneity, information, behavior, contract design, and insurer competition. Their results accumulate into a coherent research program precisely because each study isolates a specific market margin. Cross-country insurance statistics map institutional variation but cannot replace identified individual-level demand and claims data.

Structural evaluation becomes most informative when it states the decision environment, the source of identification, the observed coverage and risk variables, the nature of insurer costs, and the equilibrium response. Welfare counterfactuals require explicit social accounting of risk protection, real resource costs, financial transfers, public finance, information, and distributional effects. A policy recommendation is only as strong as the assumptions validated for the proposed intervention.

The editorial consequence is a disciplined standard of inference: preserve exact source-to-claim attribution; report preference, cost, and supply uncertainty separately; test alternative behavioral and risk distributions; validate against distinct menus and periods; and distinguish estimates from extrapolations. Within those boundaries, structural insurance economics supplies a rigorous basis for comparing allocation and pricing under uncertainty. The present monograph consolidates these principles as a completed source-based analytical synthesis suitable for further scientific and editorial assessment before an authorized batch release.

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

  • Bajari, Patrick, Denis Nekipelov, Stephen P. Ryan, and Miaoyu Yang. 2015. "Machine Learning Methods for Demand Estimation." American Economic Review 105 (5): 481-485. https://doi.org/10.1257/aer.p20151021

  • Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. "Automobile Prices in Market Equilibrium." Econometrica 63 (4): 841-890. https://doi.org/10.2307/2171802

  • 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

  • Cohen, Alma, and Peter Siegelman. 2010. "Testing for Adverse Selection in Insurance Markets." The Journal of Risk and Insurance 77 (1): 39-84. https://doi.org/10.1111/j.1539-6975.2009.01337.x

  • Einav, Liran, Amy Finkelstein, and Jonathan Levin. 2010. "Beyond Testing: Empirical Models of Insurance Markets." Annual Review of Economics 2: 311-336. https://doi.org/10.1146/annurev.economics.050708.143254

  • Einav, Liran, Amy Finkelstein, and Mark R. Cullen. 2008. Estimating Welfare in Insurance Markets Using Variation in Prices. NBER Working Paper 14414. Cambridge, MA: National Bureau of Economic Research. (Working-paper predecessor of the 2010 journal article.) https://doi.org/10.3386/w14414

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

  • Einav, Liran, Amy Finkelstein, and Paul Schrimpf. 2010. "Optimal Mandates and the Welfare Cost of Asymmetric Information: Evidence from the U.K. Annuity Market." Econometrica 78 (3): 1031-1092. https://doi.org/10.3982/ECTA7245

  • Einav, Liran, Amy Finkelstein, Stephen P. Ryan, Paul Schrimpf, and Mark R. Cullen. 2013. "Selection on Moral Hazard in Health Insurance." American Economic Review 103 (1): 178-219. https://doi.org/10.1257/aer.103.1.178

  • Einav, Liran, and Amy Finkelstein. 2011. "Selection in Insurance Markets: Theory and Empirics in Pictures." Journal of Economic Perspectives 25 (1): 115-138. https://doi.org/10.1257/jep.25.1.115

  • Fang, Hanming, Michael P. Keane, and Dan Silverman. 2008. "Sources of Advantageous Selection: Evidence from the Medigap Insurance Market." Journal of Political Economy 116 (2): 303-350. https://doi.org/10.1086/587623

  • Geruso, Michael, Timothy J. Layton, Grace McCormack, and Mark Shepard. 2023. "The Two-Margin Problem in Insurance Markets." The Review of Economics and Statistics 105 (2): 237-257. https://doi.org/10.1162/rest_a_01070

  • 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

  • Ketcham, Jonathan D., Claudio Lucarelli, Eugenio J. Miravete, and M. Christopher Roebuck. 2012. "Sinking, Swimming, or Learning to Swim in Medicare Part D." American Economic Review 102 (6): 2639-2673. https://doi.org/10.1257/aer.102.6.2639

  • Mahoney, Neale, and E. Glen Weyl. 2017. "Imperfect Competition in Selection Markets." The Review of Economics and Statistics 99 (4): 637-651. https://doi.org/10.1162/REST_a_00661

  • Mossin, Jan. 1968. "Aspects of Rational Insurance Purchasing." Journal of Political Economy 76 (4, Part 1): 553-568. https://doi.org/10.1086/259427

  • Organisation for Economic Co-operation and Development (OECD). 2026. Global Insurance Market Trends 2026. Paris: OECD Publishing. 59 pages. Published November 2, 2026. https://doi.org/10.1787/c8be6585-en

  • Rothschild, Michael, and Joseph Stiglitz. 1976. "Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information." The Quarterly Journal of Economics 90 (4): 629-649. https://doi.org/10.2307/1885326

  • Starc, Amanda. 2014. "Insurer Pricing and Consumer Welfare: Evidence from Medigap." The RAND Journal of Economics 45 (1): 198-220. https://doi.org/10.1111/1756-2171.12048

  • Train, Kenneth E. 2009. Discrete Choice Methods with Simulation. 2nd ed. Cambridge: Cambridge University Press. https://doi.org/10.1017/CBO9780511805271