Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Expected Utility and Risk Preferences

Abstract. Expected utility is a mathematical representation of preference over uncertain consequences. It links probability distributions, a utility index and an ordering of alternatives, and provides the foundational language for analyzing risk aversion, certainty equivalents, risk premiums, insurance and contingent transfers. This monograph reconstructs the historical and axiomatic foundations of expected utility, specifies the roles of objective and subjective probability, derives the main comparative measures of risk preferences and tests their implications with transparent numerical examples. It distinguishes the implications of a representation theorem from empirical claims about human decision making, and separates ordinary risk from ambiguity concerning probabilities. The analysis of Allais-type and Ellsberg-type choices, small-stakes calibration, background risk and heterogeneous elicitation shows that stable expected-utility preferences are a conditional modeling assumption requiring substantive verification. In insurance, the model explains the value of indemnity and risk sharing under explicit information, pricing and wealth conditions, while institutional demand additionally depends on premiums, liquidity, incentives, contract terms and beliefs. Original contributions of this dossier consist of a source-critical synthesis, formal comparisons and independently worked illustrations, rather than a proposed new preference theory. The resulting conclusion is methodological: expected utility supplies a rigorous benchmark whose analytical power increases when the state space, probability interpretation, utility domain, behavioral evidence and institutional constraints are stated together.

Keywords: Expected utility; von Neumann-Morgenstern utility; Bernoulli; Savage; preference axioms; independence axiom; subjective probability; risk aversion; Arrow-Pratt measure; certainty equivalent; risk premium; stochastic dominance; mean-preserving spread; insurance; risk sharing; ambiguity; Allais paradox; Ellsberg paradox; behavioral risk preferences; risk elicitation.

Introduction

A decision under risk compares feasible actions through the uncertain consequences they produce. The central question is how a preference ordering over such actions may be represented mathematically and how the resulting representation should be interpreted when probabilities, resources and institutional rules vary. In the conventional finite-state model, an action induces a lottery with outcomes and probabilities; expected utility evaluates that lottery through the probability-weighted utility of its outcomes. This structure is foundational in the axiomatic treatment of measurable utility by Herstein and Milnor (1953), and in the game-theoretic analysis associated with von Neumann and Morgenstern's Theory of Games and Economic Behavior (1944; commemorative edition 2004).

Risk preference is a relational property: a decision maker prefers one distribution of consequences to another under a specified information and resource structure. It may be represented by curvature when monetary consequences are mapped into a single wealth argument; it may also reflect probability weighting, ambiguity, time, reference points or state-dependent values. The term therefore covers more than one empirical mechanism. This dossier develops the expected-utility benchmark in enough detail to identify exactly which predictions follow from assumptions and which require new observations.

The primary research method combines reconstruction of original mathematical arguments, derivation from explicit assumptions, original counterexamples and numerical recomputation. Historical bibliographic claims are based on publisher, scholarly-journal, archival or library records. Where an article's complete original text was unavailable, claims are limited to its verified publication identity, published abstract or accessible extract. The monograph is an international work of theory and applied analysis; its insurance examples are stylized, with no purported estimates for a specific insurer or jurisdiction.

The scientific boundary is important. The economic analysis of rational insurance purchasing developed by Mossin (1968) concerns a particular application of preferences. Here the object is the general ordering of uncertain consequences and its mathematical validity. Insurance demand is used as a test of the general theory; detailed purchase optimization, underwriting and actuarial pricing remain subjects of the related Science and separate dossiers. The parent Science is Risk Theory, with additional scientific connections to Insurance Economics, Behavioral Insurance and Risk Perception, and Actuarial Science.

Bernoulli's Risk Problem

The St. Petersburg problem and the limits of monetary expectation

The St. Petersburg game pays a prize that doubles with the index of the first success in a sequence of independent fair coin tosses. A conventional specification pays 2^n monetary units when the first success occurs on toss n, which has probability 2^(-n) for n = 1, 2, .... The expected payment equals the divergent series sum_n 2^(-n) 2^n = sum_n 1. Its expected monetary value is infinite even though ordinary willingness to pay for participation is finite. The problem makes explicit the difference between valuing an uncertain monetary payoff by expectation and evaluating its consequences for a particular person's wealth.

Daniel Bernoulli's 1738 account, available in the 1954 English translation titled Exposition of a New Theory on the Measurement of Risk, proposes that marginal subjective value declines as wealth increases and develops a logarithmic valuation. Bernoulli's principle can be represented, for positive final wealth w, by u(w) = ln(w). Expected utility then evaluates E[ln(W)] rather than ln(E[W]). When prizes are added to positive initial wealth, the logarithm grows sufficiently slowly to make the classic St. Petersburg expected utility finite, despite the divergence of the monetary expectation. Thus the puzzle is resolved within a wealth-dependent valuation, rather than by assigning an arbitrary monetary ceiling to the prizes.

Formally, for initial wealth w0 > 0 and a St. Petersburg gain X, the series sum_{n>=1} 2^(-n) ln(w0 + 2^n) converges: ln(w0 + 2^n) grows on the order of n, whereas n/2^n is summable. The reasoning depends on expected utility of final positive wealth, the stated sequence of objective probabilities and the absence of additional participation or transaction effects. An infinite expected payout is therefore consistent with a finite expected log-utility valuation. There is no universal willingness-to-pay number independent of the chosen starting wealth and payoff conventions.

The conceptual contribution is broader than one paradox. A monetary amount has different marginal consequences depending on initial resources, and the utility function represents this dependence. A highly concentrated lottery may have a large mathematical expectation and a small welfare value to a vulnerable participant. This insight later became central to risk-sharing and insurance analyses, even though modern axiomatic expected utility and Bernoulli's historical account are mathematically distinct developments.

Historical continuity and the distinction between utility of wealth and lotteries

Bernoulli addressed valuation of risky wealth by treating the value of wealth as nonlinear. Von Neumann and Morgenstern established conditions under which preferences over probabilistic mixtures can be represented by expected utility, as described in their 1944 foundational book. Their utility index is derived from comparative lottery preferences, rather than simply posited as a psychological magnitude. This difference explains why a concave function of money is a substantive property of preferences in an expected-utility model, rather than one of the basic representation axioms itself.

Friedman and Savage (1948) examined the coexistence of insurance purchase and gambling in The Utility Analysis of Choices Involving Risk. Their proposed variations in utility curvature illustrate how one expected-utility specification can rationalize distinct choices over different wealth ranges. The broader lesson is that observing one insured loss and one lottery ticket does not immediately establish a single, globally concave utility function. The relevant payoff magnitudes, initial wealth, opportunity set, probability beliefs and background risks must first be made comparable.

The historical lineage also prevents a common category error. A representation of lottery preferences by expected utility does not establish that all people possess a stable logarithmic utility function, nor that a market premium equals an individual certainty equivalent. The first is an empirical claim about preferences; the second is an institutional claim about a price. They require separate evidence and assumptions.

Utility Axioms

Formal choice environment and the expected-utility representation

Let X = {x1, ..., xm} be a finite set of consequences. A simple lottery L assigns probabilities p_i >= 0 with sum_i p_i = 1 to consequences x_i. Let L(X) be a set of such lotteries closed under mixtures. A binary relation >=_P represents the decision maker's weak preference: L >=_P M means that lottery L is at least as preferred as lottery M. This is an ordering of lotteries, not necessarily of their expected monetary payouts.

The standard representation proceeds from completeness (every two feasible lotteries can be compared), transitivity (the ordering is internally consistent), continuity (no abrupt preference discontinuity across probability mixtures) and independence (if L >=_P M, then mixing both with the same third lottery N at the same probability preserves their ordering). Under suitable richness and nontriviality assumptions, there is a utility index u on consequences such that

U(L) = sum_i p_i u(x_i)

represents the preference ordering. Herstein and Milnor (1953) provide an influential axiomatic formulation. The axioms concern the structure of the preference relation; the linearity is in the objective probabilities, while u(x) may be nonlinear in monetary consequences.

Independence is especially consequential. If two lotteries differ only in part of their outcome distribution, adding an identical common probabilistic branch to both cannot reverse preference under expected utility. Algebraically, U(alpha L + (1-alpha)N) - U(alpha M + (1-alpha)N) = alpha[U(L)-U(M)]. The shared branch cancels for any 0 < alpha <= 1. This cancellation produces transparent comparative statics, but it also creates a testable restriction that some observed choices violate.

Completeness can be relaxed in richer models. Aumann's 1962 study of utility without completeness is a reminder that partial or indeterminate preference orderings may be mathematically coherent. A decision maker who declines to compare an unknown technological catastrophe with a familiar small financial gamble need not thereby lack all orderly preferences. Strict expected utility selects a complete representable ordering within its chosen decision domain.

Independence, continuity and the uniqueness of utility

An expected-utility index has positive affine uniqueness: if u represents preferences, then v(x) = a u(x) + b, with a > 0, represents exactly the same ordering of lotteries. Indeed, E[v(X)] = a E[u(X)] + b. This does not license arbitrary increasing transformations. For example, an increasing nonlinear transformation v = g(u) changes the expectation of transformed utility and can change the ranking of lotteries. The numerical size or unit of a von Neumann-Morgenstern utility level has no direct interpersonal interpretation.

An empirical implication follows. A dataset identifying lottery comparisons can potentially identify preference parameters only relative to maintained assumptions concerning probabilities, feasible stakes, outcomes, error or heterogeneity. It does not deliver an absolute cardinal quantity of happiness. Even when observed money choices can be rationalized by one utility index, that index can encode unobserved consumption constraints, external effects or reference dependence if the model's state description is incomplete.

Continuity and independence solve different problems. Continuity supports the construction of intermediate lottery equivalents; independence imposes an additive structure over probabilistic branches. A preference relation can remain complete, transitive and continuous while violating independence. The latter is precisely the kind of generalization examined by Machina (1982) in Expected Utility Analysis without the Independence Axiom. Empirical rejection of one axiom therefore has a more specific meaning than a global rejection of logically ordered preferences.

A useful methodological discipline is to classify a model claim before evaluating it: whether it concerns representation, comparative statics, behavior or welfare. The mathematical representation is conditional on axioms; comparative statics are conditional on representation and functional restrictions; behavioral validity requires observations; welfare inference additionally requires a defensible normative interpretation and often interpersonal aggregation assumptions. These are different inferential steps.

Subjective probability and competing notions of uncertainty

When the probability distribution is objectively specified by a known randomizing device, expected utility can take the probabilities as data. In many insurance problems, however, catastrophe probabilities, medical events and future behavior are partly unknown. Savage's The Foundations of Statistics, originally published in 1954 and later issued in a revised 1972 edition, constructs an influential theory linking coherent preferences over acts to subjective beliefs and utility under a sufficiently rich set of postulates. An act maps possible states of the world into consequences; probabilities become part of a representation of the decision maker's preferences.

Anscombe and Aumann's 1963 Definition of Subjective Probability distinguishes objectively randomized lotteries from acts with uncertain states, using the former as a device for eliciting the latter. Both frameworks clarify a key issue: inferred probability beliefs are not interchangeable with independently observed event frequencies. The mere ability to fit a model to decisions does not guarantee that elicited subjective probabilities correspond to actuarial loss frequencies.

The separation matters most where evidence is sparse. A change in insurance take-up after a disaster may reflect updated perceived probability, changing aversion to loss, liquidity, salience, or new contract availability. If one assigns the entire response to a fixed curvature parameter, the estimate confounds beliefs and tastes. A credible design elicits or bounds beliefs separately, records external probabilities when available, and specifies which expected-utility primitives can be identified from the observed decision set.

Ellsberg (1961) used Risk, Ambiguity, and the Savage Axioms to expose systematic difficulties with a unique subjective probability representation. For known probabilities, risk describes a distribution with stipulated chances. Under ambiguity, uncertainty about probabilities themselves affects decisions. An expected-utility calculation can be conducted using any single stipulated distribution, but selecting that distribution is an additional modeling step. Insurance against rare or structurally changing hazards regularly encounters this epistemic boundary.

Risky Wealth and Expected Utility

Concavity, certainty equivalents and risk premiums

Consider final wealth W = w + Z, where w is initial wealth and Z a risky net payoff. Under expected utility the decision maker prefers an available alternative A to B precisely when E[u(W_A)] >= E[u(W_B)], assuming that both expectations exist. Let the utility function be increasing and continuous over the feasible wealth domain. A certainty equivalent CE(W) solves u(CE(W)) = E[u(W)]. Since u is increasing, comparing a risky prospect with a sure amount is equivalent to comparing that amount with CE(W).

For a risk-averse expected-utility decision maker with concave u, Jensen's inequality gives E[u(W)] <= u(E[W]), and therefore CE(W) <= E[W]. Define the risk premium as pi(W) = E[W] - CE(W). It is nonnegative for concave utility and zero for linear utility or degenerate risk. With strictly concave utility and genuinely variable wealth, the inequality is strict, subject to the ordinary regularity conditions. A risk-neutral expected-utility decision maker has an affine utility index and ranks monetary prospects solely by expected wealth. A risk-seeking decision maker with convex monetary utility reverses the Jensen comparison for applicable risks.

Two interpretive limits follow. First, a risk premium is an amount of wealth the modeled person would surrender to replace a particular risk with its mean; it is not an insurance-company's quoted premium. Second, risk aversion is a property of preferences over final consequences, conditional on how those consequences are described. A person may have different utility functions for health, consumption, family wealth or future states, making a single monetary u(w) insufficient.

Pratt's 1964 Risk Aversion in the Small and in the Large derives operational comparisons of curvature and risk premia. This source and Arrow's subsequent influence gave rise to the standard Arrow-Pratt terminology. The essential contribution is to replace an unqualified claim that one curve “looks more concave” with a mathematically invariant local index and a global comparison of risk-bearing preferences.

Absolute and relative risk aversion

For a twice-differentiable, strictly increasing utility function, define absolute risk aversion at wealth w as

A(w) = -u''(w) / u'(w).

Relative risk aversion is R(w) = w A(w) for positive wealth. These ratios are invariant under positive affine transformations of utility. They are locally meaningful because a second-order Taylor approximation around wealth w yields, for a mean-zero small risk epsilon with variance sigma^2, the risk premium

pi(w + epsilon) approximately (1/2) A(w) sigma^2.

The approximation assumes sufficient differentiability, small risks relative to relevant wealth scales and suitably controlled higher moments. It should never be applied mechanically to a ruin-threatening or heavy-tailed loss. In the derivation, expand E[u(w+epsilon)] as u(w) + u''(w)sigma^2/2 + ... and u(w-pi) as u(w) - u'(w)pi + ..., then equate leading terms. The formula displays how the same variance implies a larger local risk premium when absolute risk aversion is larger.

The logarithmic utility u(w) = ln(w) for w > 0 has u'(w) = 1/w, u''(w) = -1/w^2, A(w) = 1/w and R(w) = 1. Its absolute risk aversion falls with wealth while its relative risk aversion is constant. Constant relative risk aversion can be represented by u(w) = w^(1-gamma)/(1-gamma) for gamma > 0, gamma != 1; the log case is the gamma = 1 limit. Constant absolute risk aversion can be represented by u(w) = -exp(-a w) for a > 0, with A(w) = a on an appropriate real-wealth domain. These functional forms are convenient modeling restrictions, and their plausibility depends on observed choice behavior and the relevant range of stakes.

There is an especially transparent exact result for normal monetary risks and exponential utility. If X is normally distributed with mean mu and variance sigma^2, then E[-exp(-a(w+X))] = -exp[-a(w+mu)+(a^2 sigma^2)/2]. It follows that the certainty equivalent of final wealth is w + mu - a sigma^2/2; the monetary risk premium is exactly a sigma^2/2. The equality depends on both CARA utility and normality. The same mean-variance formula generally becomes only an approximation for nonnormal risks, nonlinear wealth dependence or constraints.

Comparative risk aversion can be characterized without relying on one arbitrary utility scale. Under standard differentiability assumptions, an expected-utility decision maker is at least as risk averse as another over a common wealth domain if the first utility index is a concave increasing transformation of the second; this corresponds to an everywhere weakly larger Arrow-Pratt absolute risk aversion. The result compares all risky wealth prospects in that domain. It does not establish an order between two individuals who disagree about probabilities or face different outcomes and constraints. Pratt's comparative results provide the proper source for this restricted and economically useful claim.

Comparative risk, dominance and background uncertainty

Preference over lotteries and ordering of their distributions must be separated. First-order stochastic dominance (FSD) gives a ranking for every increasing utility function: distribution X first-order dominates Y when its cumulative distribution is never above that of Y, with strict inequality somewhere. It expresses a shift toward higher consequences without requiring risk aversion. Second-order stochastic dominance (SSD) ranks X above Y for every increasing concave utility when integrated cumulative distribution functions satisfy the appropriate ordering and integrability conditions. For equal-mean lotteries, a mean-preserving spread makes a prospect riskier in the conventional concave-utility sense.

Rothschild and Stiglitz's 1970 definition of increasing risk relates the spread notion, conditional-mean-preserving noise and concave-utility comparisons. The distributional distinction is analytically stronger than saying simply that one random variable has greater variance. A higher-variance lottery need not be ranked below a lower-variance lottery by every increasing concave utility when other features of the distributions differ. Conversely, a mean-preserving spread is rejected by every strictly risk-averse expected-utility evaluator to whom the comparison applies.

The conditional Jensen argument is particularly informative. If Y = X + epsilon and E[epsilon | X] = 0, then E[u(Y) | X] <= u(E[Y | X]) = u(X) for concave u, and hence E[u(Y)] <= E[u(X)]. This establishes a general welfare cost of conditionally fair added noise under the specified preferences. In insurance, an indemnity that removes a conditionally fair wealth fluctuation raises expected utility before accounting for its price. The inequality alone says nothing about the financing and incentive consequences of the transfer.

Background risk complicates comparisons. A household's risky income, volatile asset holdings and uninsured health expenditures alter the consequences of an otherwise identical insured loss. The sign and magnitude of an additional insurance purchase may depend on correlation between insured and background losses, the household's effective state-dependent marginal utility, and liquidity constraints. Analyses of multiple risks and dynamic decisions in The Economics of Risk and Time (Gollier, 2001) and Economic and Financial Decisions under Risk (Eeckhoudt, Gollier, and Schlesinger, 2005) place these interactions in a broader theoretical setting.

If two individuals face risks with the same marginal loss distribution but different correlations with income, their efficient coverage choices need not coincide. A common “risk-aversion coefficient” therefore fails to summarize the complete demand for risk transfer. Effective exposure and the timing of contingent payments matter alongside preferences.

Formal Choice Comparisons

Transparent numerical calculations

All examples in this section are constructed illustrations. They are computed from declared assumptions and are neither estimates of observed populations nor insurance price recommendations.

Consider a household with initial wealth 100 and a fair binary change of -50 or +50, each with probability 0.5. Under u(w) = ln(w), the expected utility is 0.5 ln(50) + 0.5 ln(150). Its certainty equivalent is the geometric mean sqrt(50*150) = 86.602540..., while expected wealth is 100. The implied risk premium is 100 - 86.602540... = 13.397460.... Linear utility gives certainty equivalent 100, and convex utility reverses the direction of the risk preference.

The log-utility result is sensitive to scale. Replace initial wealth 100 with 1000, preserving the same -50/+50 change. The certainty equivalent becomes sqrt(950*1050) = 998.749218...; the risk premium is only about 1.250782. This comparison is consistent with decreasing absolute risk aversion under logarithmic utility. It also exposes why estimating curvature solely from one small wager is unstable when the relevant wealth domain is unknown.

A compact table illustrates the main distinctions:

Prospect or model Distribution / assumption Expected wealth or payoff Certainty equivalent / ranking
Sure wealth 100 with certainty 100 100 under any increasing utility
Fair binary wealth 50 or 150, each with probability 0.5 100 86.602540 under log utility
Same fair change at higher wealth 950 or 1050, each with probability 0.5 1000 998.749218 under log utility
Risk-neutral model Affine utility of money Depends on distribution CE equals expected money
Strictly risk-averse model Strictly concave utility of wealth Depends on distribution CE below expected money for nondegenerate risk

Now consider an insured loss L = 400 occurring with probability p = 0.1 at initial wealth 1000, assuming one period, no taxes, no moral hazard, instantaneous full claim payment and otherwise identical final consumption opportunities. Without insurance, final wealth is 600 in the loss state and 1000 otherwise. Expected wealth is 960. The log-utility certainty equivalent is

CE_no = exp(0.1 ln(600) + 0.9 ln(1000)) = 1000*(0.6)^0.1.

A fully indemnifying contract with certain premium P yields final wealth 1000-P. The maximum log-utility willingness to pay for this full coverage, relative to the uninsured option, is P_max = 1000 - CE_no. The actuarially fair expected indemnity equals pL = 40. The difference between P_max and 40 measures the theoretical purchaser's risk-bearing benefit in monetary units, not the insurer's operating margin. Each step depends on the maintained complete-coverage, constant-wealth-domain assumptions.

A third comparison concerns contract loading. Let the insurer charge qI for indemnity limit I, with q = (1+lambda)p and no deductible or other friction. For one loss of L = 400, the two final wealth outcomes are W_0 = 1000-qI without loss and W_1 = 1000-L+(1-q)I with loss, where 0 <= I <= 400. With logarithmic utility, the objective is

EU(I) = (1-p)ln(W_0) + p ln(W_1).

The first-order condition for an interior optimum is

p(1-q)/(600+(1-q)I) = (1-p)q/(1000-qI).

For fair pricing, q=p=0.1, full coverage I=400 equalizes wealth across states and is optimal. For a 25% proportional loading, lambda=0.25 and q=0.125, the interior optimum is I = 182.857142... by direct solution, provided both state wealth levels are positive. The optimum is below the full loss because paying a loaded price for every additional unit of indemnity has a marginal cost even outside the loss state. This numerical result is model-specific: fixed fees, deductibles, utility of nonmonetary losses, investment opportunities and behavioral distortions alter the optimum. Concavity makes the stated first-order condition sufficient on the feasible interval when the optimum is interior.

These calculations connect three distinct concepts: actuarial expected loss, a consumer's certainty-equivalent risk premium, and an insurer's market loading. Conflating them produces false predictions. They agree only under special conditions, and realistic market prices also reflect risk capital, expenses, competition and contract restrictions.

Independence and ambiguity counterexamples

Allais (1953) formulated a celebrated challenge to the independence postulate in Le comportement de l'homme rationnel devant le risque. Consider four lotteries with money measured in millions: A pays 1 with certainty; B pays 1 with probability 0.89, 5 with probability 0.10, and 0 with probability 0.01; C pays 1 with probability 0.11 and 0 otherwise; D pays 5 with probability 0.10 and 0 otherwise. The widely discussed choice pattern prefers A over B and D over C.

Expected utility cannot generate both strict preferences with one utility index. A > B implies 0.11u(1) > 0.10u(5) + 0.01u(0). Conversely, D > C implies 0.10u(5) + 0.01u(0) > 0.11u(1). The inequalities contradict each other. This is a direct algebraic test of independence; it does not rely on choosing a particular concavity coefficient. A researcher who observes the paired pattern has evidence against that common expected-utility representation for those choices, subject to measurement error, attention and stable interpretation of outcomes.

Ellsberg's ambiguity thought experiment compares a known share of red balls in an urn with an unknown division of the remaining balls between two colors. A typical preference for betting on a known-probability color over an ambiguous color, coupled with the opposite preference for certain complementary events, conflicts with a single additive subjective probability model. The relevant issue is belief ambiguity, not merely dislike of high monetary variance. The original analysis is available in Ellsberg's 1961 article. The two classic paradoxes challenge different parts of the standard construction: common-consequence independence in one case, and a unique additive subjective-probability interpretation in the other.

Machina's independence-free expected utility analysis (1982), Yaari's Dual Theory of Choice under Risk (1987) and the cumulative formulation developed by Tversky and Kahneman (1992) establish that coherent but non-expected-utility decision models exist. Their mathematical mechanisms differ. Machina retains useful differentiable local analysis while relaxing independence; Yaari changes the mixture operation underlying the representation; cumulative prospect theory develops rank-sensitive decision weights and reference-dependent gains and losses. These alternatives should be compared on explicit observational and normative criteria rather than pooled as a single generalized “behavioral exception.”

Small-stakes calibration and elicitation

Expected utility with a globally concave wealth index imposes powerful restrictions across stakes. Rabin's 2000 calibration theorem demonstrates that appreciable aversion to modest monetary gambles, if attributed solely to expected utility of lifetime wealth, may imply extraordinarily strong aversion to larger favorable gambles. The finding is an internal model-calibration critique, not a proof that every locally concave expected-utility model is false. Its force depends on the maintained wealth domain, consistent preferences across stakes and the assumption that the same utility index explains the small-gamble choices. Rabin and Thaler's Anomalies: Risk Aversion (2001) argues that loss aversion and mental accounting can play important explanatory roles.

Empirical measurement poses a different problem: how to infer parameters from choices. Holt and Laury's Risk Aversion and Incentive Effects (2002) studies decisions in incentivized lotteries and highlights the relevance of incentives and stake size. Dohmen and colleagues' Individual Risk Attitudes (2011) relates survey willingness to take risks to an incentivized validation task and a range of real-world behaviors. The evidence supports studying heterogeneous, context-sensitive risk measures rather than simply transferring a parameter estimated in one task to every insurance decision.

Barsky, Juster, Kimball and Shapiro's risk-preference measurement study (1997) uses hypothetical survey choices to estimate preference heterogeneity and documents limited explanatory power for individual behavior. Such designs do not directly identify a universal constant-relative-risk-aversion coefficient. Elicitation results can be affected by background wealth, question framing, probability numeracy, stakes, hypothetical versus real rewards, domain and selection effects. Structural identification requires a measurement model and a specified choice environment.

A practical comparison is to distinguish four layers of evidence. Theoretical coherence tests check whether reported choices can be represented by the axioms; experimental validity checks stability across changes in incentives and presentation; behavioral prediction checks whether elicited parameters forecast relevant action; institutional prediction checks whether the model improves understanding after accounting for contract prices, liquidity, regulation and risk information. Passing one layer does not automatically establish the next. A theory of preferences gains practical usefulness from these layered tests.

Risk-Sharing Implications

Indemnity, premium loadings and optimal coverage

A standard insurance exchange replaces uncertain financial loss with a certain premium and a contingent claim payment. For a household, the central expected-utility benefit of fairly priced complete indemnity is the smoothing of final wealth across states. When premiums are fair, indemnity is certain and state-independent utility over wealth is strictly concave, a contract equalizing wealth across loss states maximizes expected utility within the stylized feasible class. This result follows directly from Jensen's inequality and the coverage calculation above. It is a benchmark for insurance theory rather than an empirical prediction of universal full insurance.

Mossin (1968) studied rational insurance purchasing under explicit economic assumptions. A loaded premium can make partial coverage optimal, as shown by the log-utility interior solution. Real contracts differ in deductible schedules, exclusions, caps, claim adjustment, credit exposure, taxes, capital requirements, distribution charges and time of payment. A model that calls one amount a “premium” without specifying how it is paid and which states it insures is too weak to support a purchase prediction.

The joint decision over insurance and prevention is still more involved. Ehrlich and Becker's Market Insurance, Self-Insurance, and Self-Protection (1972) distinguishes market indemnity from private actions that reduce loss magnitude or loss probability. For example, alarm systems can lower probability, fire-resistant construction can lower severity, and insurance can transfer remaining financial consequences. Their economic interactions depend on prices, technology and contract incentives. An expected-utility objective can organize the choice, but the feasible strategies and probabilities must be modeled explicitly. Holding risk probabilities constant while interpreting an insured party's prevention response can be untenable when insurance changes behavior.

Insurer solvency introduces a further state dependence. A contractual indemnity is valuable to the policyholder only to the extent that payment can be realized when needed. Correlation between insured losses, insurer failure and financial-market stress may cause nominal coverage to smooth wealth less effectively than the contract schedule suggests. Expected utility can incorporate this through additional states and outcomes, yet the required joint distribution may be difficult to estimate. The formal framework makes the missing inputs visible; it cannot supply them by assumption.

Income constraints can also dominate a local curvature calculation. A household with low current cash reserves might decline coverage at a price below its unconstrained willingness-to-pay because premium timing competes with essential consumption. Another household may buy coverage at a price above a simple one-period certainty-equivalent value because it protects future borrowing capacity or prevents a liquidity crisis. A model with only terminal monetary wealth omits such mechanisms. The disciplined response is to represent feasible payments, timing and liquidity contingencies within the consequences rather than to force all observed decisions into a single curvature statistic.

Mutual allocation, reinsurance and diversification

Consider two agents with aggregate resource endowment W(s) in state s, individual allocations c1(s) and c2(s), and increasing concave Bernoulli utilities u1 and u2. In a frictionless Pareto-efficient risk-sharing problem, maximize alpha E[u1(c1(s))] + (1-alpha) E[u2(c2(s))] subject to c1(s) + c2(s) = W(s) for each state and 0 < alpha < 1. Where probabilities are shared and the optimum is interior, the first-order condition is

alpha u1'(c1(s)) = (1-alpha) u2'(c2(s)).

This equates weighted marginal utility across participants within each state. Its content is allocative: an efficient treaty divides aggregate resources to balance marginal welfare rather than requiring equal nominal risk positions. Under common beliefs and exponential utility, risk-sharing rules can become linear in aggregate loss. With differing beliefs, limited liability, transaction costs or asymmetric information, the agreement and efficiency conditions change. The first-order statement alone says nothing about the bargaining division of the surplus or the legal enforceability of the transfer.

Karl Borch's 1962 Equilibrium in a Reinsurance Market provides a foundational insurance-market context for expected utility, mutual risk allocation and efficient reinsurance. His earlier work on reciprocal reinsurance treaties likewise treats treaty negotiation as a cooperative economic problem. These contributions distinguish the value created by pooling risks from the terms on which particular firms agree to exchange them. The same distinction applies to individual insurance: the social value of an efficient redistribution of risk is conceptually separate from the observed premium and distribution of bargaining power.

Pooling works particularly well when losses are imperfectly correlated. If n independent equal-sized risks have finite variance sigma^2, their average has variance sigma^2/n. This mathematical reduction becomes less powerful when common catastrophes or systematic macroeconomic events cause positive correlation. If pairwise correlations equal rho for otherwise identically distributed losses, the average variance is sigma^2[1+(n-1)rho]/n. As n grows, this approaches rho sigma^2, preserving a systemic component whenever rho > 0. The covariance formula is a straightforward derivation, but the correlation parameter must be empirically justified before it supports a real-world pooling estimate.

Expected utility connects this probabilistic diversification with welfare: less variable exposure of equal expected value ordinarily benefits a risk-averse participant if one distribution is a mean-preserving contraction of another. However, diversification can also change expected losses, liquidity and tail exposures, in which case an SSD claim requires proof rather than a variance-only comparison. The conceptual bridge between risk distribution and utility must be rebuilt for the actual portfolio being evaluated.

Information, incentives, welfare and institutional boundaries

Insurance is an information-sensitive exchange. A person may know more about private exposure or effort than the insurer, and the insurer may know more about coverage exclusions, settlement behavior or solvency. In a simple expected-utility formulation these asymmetries enter through beliefs, signal structures, feasible contracts and contingent actions. They cannot be removed by declaring the applicant “risk averse.” Premium differences may reflect actuarial differences, administrative costs, screening, regulation or market power; preference curvature is only one component of the outcome.

The interpersonal welfare use of expected utility adds an independent normative step. If two individuals possess utility functions unique only up to separate positive affine transformations, summing their utility levels without a justified normalization has no invariant interpretation. Harsanyi's Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of Utility (1955) makes visible the ethical and axiomatic choices underlying social-welfare aggregation. An individual's expected-utility insurance choice therefore cannot by itself establish the distributive justice or social desirability of a policy.

Risk-sharing institutions pursue several objectives that are not reducible to the private certainty equivalent: availability across income groups, continuity of essential services, prevention incentives, protection against insolvency and catastrophic fiscal exposures. A welfare analysis may explicitly value distributional objectives and public externalities, but these enter through additional welfare criteria or richer consequences. Calling an insurance market “efficient” requires specifying efficiency relative to which preferences, information constraints, feasible contracts and institutional rules.

The general model also distinguishes ex ante and ex post valuations. Before uncertainty resolves, an arrangement can increase expected utility. In a realized state, a particular participant may have an unfavorable monetary outcome relative to the no-trade counterfactual. This is not a logical inconsistency: ex ante preferences concern the distribution across possible states, while ex post comparisons concern a realized state. Stable legitimate contracts and the design of compensation mechanisms require separate institutional and legal analysis.

Critical Scope

Identification, evidence and theoretical alternatives

The expected-utility framework succeeds as a conditional representation. It is exact within its stipulated lottery domain and preference axioms. Its practical use depends on independent questions: what counts as a consequence, whether probabilities are objectively known, whether preferences are stable, whether stakes are correctly measured, whether constraints are observed and whether the utility index is suitable for the decision at hand. Overstating the model's reach weakens rather than strengthens its real analytical contribution.

Three distinct forms of uncertainty deserve separation. Sampling uncertainty concerns estimating parameters of a known statistical family; structural uncertainty concerns whether the selected loss-generation model describes the relevant events; preference uncertainty concerns the stability and identification of the evaluator's own choice criteria. Expected utility with a single probability distribution directly handles a specified random variable. Sampling or structural ambiguity requires additional modeling of the distribution or a range of plausible distributions. Preference uncertainty requires observations of choices and a defensible inference procedure. These layers can interact, especially under catastrophic risks and changing economic institutions.

Kahneman and Tversky's 1979 Prospect Theory presents systematic departures from descriptive expected utility in choices involving certainty and reference-dependent gains and losses. Their 1992 cumulative extension allows probability-weighting patterns and distinct responses to gains and losses. These are established theoretical alternatives with their own assumptions and empirical limitations. They do not license an inference that all risk aversion is caused by loss aversion. A critical research design should compare models using the same decision data and out-of-sample predictions, when possible, while disclosing sensitivity to parameter choices.

A robust empirical comparison could collect incentivized lottery choices at multiple stake levels; record wealth proxies and liquidity; elicit probability beliefs separately; vary descriptions of gains, losses and reference points; and compare resulting predictions with actual policy purchase or renewal. The resulting observations would permit separate tests of consistency, curvature, belief accuracy and context effects. They would still need careful attention to selection into the sample, measurement error, changes in opportunity sets and nonrepresentative experimental settings. This is a methodological research priority derived from the literature, not a new named theory or claim of experimental validation undertaken in this dossier.

The archival review of the present corpus reinforces a source-quality distinction. Publisher bibliographic pages establish author, year, title, venue, pages and DOI. Accessible abstracts justify limited claims about what a study investigates or finds. Theorems and detailed derivations are independently reconstructed here from stated assumptions where feasible, and any original text examined only in excerpts is not represented as exhaustively reviewed. This practice is essential for an encyclopedia whose claims may be retrieved independently of their original bibliographic context.

Integrated synthesis and research priorities

The literature supports four substantive conclusions. First, preferences over probability distributions can be represented as an expectation of an outcome-utility index when the necessary axioms hold. This yields rigorous derivations of choices, certainty equivalents, comparative risk aversion and efficient sharing. Second, the utility index is constrained by positive affine uniqueness and by its consequence domain; it cannot serve as a directly comparable universal scale of wellbeing. Third, insurance predictions require the addition of prices, timing, loss distributions, correlation, information and contract performance. Fourth, systematic behavioral violations and measurement instability make those primitives objects of empirical investigation rather than automatic facts.

These conclusions produce operational consequences for comparative insurance research. Source-to-claim analysis should record whether a claim arises from an axiom, a formal proposition, a numerical example, an experimental result or an institutional observation. A numerical illustration should document the initial resources and domain restrictions. An actuarial frequency should be distinguished from a subjective probability. An inferred preference parameter should carry the elicitation environment in which it was estimated. An apparent insurance-demand anomaly should first be tested against the full financial and contractual opportunity set.

The most important methodological distinction is between conditional explanatory power and unconditional behavioral generalization. Expected utility organizes risk analysis precisely because its assumptions and implications are explicit. The same precision makes its failures diagnosable. A preference reversal across common probabilistic branches questions independence; a disagreement over unknown probabilities raises ambiguity; extreme implications from small-stakes choices call for calibration analysis; inconsistent insurance uptake across contexts can motivate investigation of liquidity, beliefs and reference points. Each observation directs research toward a distinct proposition rather than toward a vague dismissal of rational choice.

No new authorial doctrine is established in this synthesis. Its original contribution lies in the integration of historical sources, formal derivations, counterexamples, a fully reproducible contract-choice illustration and a structured account of where economic inference must move beyond the bare expected-utility representation. The science of risk preferences advances when mathematical results, observed decisions and institutional conditions are treated as mutually relevant but separately testable components of explanation.

Conclusion

Expected utility remains a foundational analytical language for decisions under risk. Bernoulli established the importance of wealth-dependent valuation; von Neumann, Morgenstern and subsequent axiomatizations supplied a precise lottery-preference representation; Savage and Anscombe-Aumann clarified the representation of subjective beliefs; Pratt, Rothschild and Stiglitz developed powerful measures and orderings of risk. Together, these achievements justify a systematic study of certainty equivalents, risk premiums, relative and absolute risk aversion, stochastic dominance and efficient allocation.

The substantive results of this monograph are conditional and therefore verifiable. Concavity explains the welfare advantage of replacing a fair risk with its mean; the Arrow-Pratt index governs local risk premiums; normality and CARA yield an exact mean-variance certainty equivalent; independence imposes cancellation across common branches; and expected-utility optimization determines stylized insurance coverage when state wealth and premiums are specified. The independently worked log-utility comparisons show how differences in wealth and premium loading change choice even before behavioral mechanisms are introduced.

Allais, Ellsberg, Rabin and empirical risk-elicitation research establish the need to test the behavioral relevance of these conditions. Models with reference dependence, ambiguity sensitivity and alternative probability transformations offer additional descriptions, each with distinct implications. Insurance contracts further translate preference into choice through institutional prices, contingent performance, information and constraints. The general conclusion is that expected utility is a mathematically rigorous and practically valuable benchmark, and its proper scientific use requires explicit assumptions, source-specific evidence, sensitivity analysis and a clear distinction between individual preference representation and market or social outcomes.

This dossier is an original analytical synthesis of established scholarship. It introduces no named theory, doctrine, principle or conceptual system. Its numerical examples are hypothetical, and no empirical claims about the magnitude of preferences in any particular population or market are derived from them.

References

  • Anscombe, F. J., and R. J. Aumann. 1963. "A Definition of Subjective Probability." The Annals of Mathematical Statistics 34 (1): 199-205. Publisher DOI.

  • Allais, Maurice. 1953. "Le Comportement de l'Homme Rationnel devant le Risque: Critique des Postulats et Axiomes de l'Ecole Americaine." Econometrica 21 (4): 503-546. Journal DOI.

  • Aumann, Robert J. 1962. "Utility Theory without the Completeness Axiom." Econometrica 30 (3): 445-462. Journal DOI.

  • Barsky, Robert B., F. Thomas Juster, Miles S. Kimball, and Matthew D. Shapiro. 1997. "Preference Parameters and Behavioral Heterogeneity: An Experimental Approach in the Health and Retirement Survey." The Quarterly Journal of Economics 112 (2): 537-579. Journal DOI.

  • Bernoulli, Daniel. (1738) 1954. "Exposition of a New Theory on the Measurement of Risk." Translated by Louise Sommer. Econometrica 22 (1): 23-36. Journal DOI.

  • Borch, Karl. 1960. "Reciprocal Reinsurance Treaties Seen as a Two-Person Co-operative Game." Scandinavian Actuarial Journal 1960 (1-2): 29-58. Publisher DOI.

  • Borch, Karl. 1962. "Equilibrium in a Reinsurance Market." Econometrica 30 (3): 424-444. Journal DOI.

  • Dohmen, Thomas, Armin Falk, David Huffman, Uwe Sunde, Jürgen Schupp, and Gert G. Wagner. 2011. "Individual Risk Attitudes: Measurement, Determinants, and Behavioral Consequences." Journal of the European Economic Association 9 (3): 522-550. Publisher DOI.

  • Eeckhoudt, Louis, Christian Gollier, and Harris Schlesinger. 2005. Economic and Financial Decisions under Risk. Princeton, NJ: Princeton University Press. Publisher catalog.

  • Ehrlich, Isaac, and Gary S. Becker. 1972. "Market Insurance, Self-Insurance, and Self-Protection." Journal of Political Economy 80 (4): 623-648. Publisher DOI.

  • Ellsberg, Daniel. 1961. "Risk, Ambiguity, and the Savage Axioms." The Quarterly Journal of Economics 75 (4): 643-669. Journal DOI.

  • Friedman, Milton, and L. J. Savage. 1948. "The Utility Analysis of Choices Involving Risk." Journal of Political Economy 56 (4): 279-304. Publisher DOI.

  • Gollier, Christian. 2001. The Economics of Risk and Time. Cambridge, MA: MIT Press. Publisher DOI.

  • Harsanyi, John C. 1955. "Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of Utility." Journal of Political Economy 63 (4): 309-321. Publisher DOI.

  • Herstein, I. N., and John Milnor. 1953. "An Axiomatic Approach to Measurable Utility." Econometrica 21 (2): 291-297. Journal DOI.

  • Holt, Charles A., and Susan K. Laury. 2002. "Risk Aversion and Incentive Effects." American Economic Review 92 (5): 1644-1655. Publisher DOI.

  • Kahneman, Daniel, and Amos Tversky. 1979. "Prospect Theory: An Analysis of Decision under Risk." Econometrica 47 (2): 263-291. Journal DOI.

  • Machina, Mark J. 1982. "'Expected Utility' Analysis without the Independence Axiom." Econometrica 50 (2): 277-323. Journal DOI.

  • Mossin, Jan. 1968. "Aspects of Rational Insurance Purchasing." Journal of Political Economy 76 (4, part 1): 553-568. Publisher DOI.

  • Pratt, John W. 1964. "Risk Aversion in the Small and in the Large." Econometrica 32 (1-2): 122-136. Journal DOI.

  • Rabin, Matthew. 2000. "Risk Aversion and Expected-Utility Theory: A Calibration Theorem." Econometrica 68 (5): 1281-1292. Publisher DOI.

  • Rabin, Matthew, and Richard H. Thaler. 2001. "Anomalies: Risk Aversion." Journal of Economic Perspectives 15 (1): 219-232. Publisher DOI.

  • Rothschild, Michael, and Joseph E. Stiglitz. 1970. "Increasing Risk: I. A Definition." Journal of Economic Theory 2 (3): 225-243. Publisher DOI.

  • Savage, Leonard J. 1954. The Foundations of Statistics. New York: John Wiley & Sons; revised edition, New York: Dover Publications, 1972. WorldCat edition record.

  • Tversky, Amos, and Daniel Kahneman. 1992. "Advances in Prospect Theory: Cumulative Representation of Uncertainty." Journal of Risk and Uncertainty 5 (4): 297-323. Publisher DOI.

  • von Neumann, John, and Oskar Morgenstern. 1944. Theory of Games and Economic Behavior. Princeton, NJ: Princeton University Press; sixtieth-anniversary commemorative edition, 2004. Publisher anniversary edition.

  • Yaari, Menahem E. 1987. "The Dual Theory of Choice under Risk." Econometrica 55 (1): 95-115. Journal DOI.