RESEARCH DOSSIER
Optimism Bias and Comparative Risk Assessment
Abstract. Optimism bias in comparative risk assessment concerns the relation between how a person judges exposure to adverse events and how that person's exposure compares with a specified population. The scientific problem is one of identification: a favorable self–other judgment can arise from genuine differences in exposure, asymmetric information, reference-group selection, question framing, motivated beliefs, or statistical features of the event. This monograph develops a source-critical analysis of these mechanisms and their consequences for insurance and protective decisions. It distinguishes absolute probability error, comparative belief error, relative placement, and behavioral response, and examines primary experimental work, methodological criticism, empirical flood-risk studies, insurance-choice experiments, and research on the effect of risk communications. A comparison with actual hazard exposure is necessary for individual calibration; group-level evidence requires explicit assumptions about reference classes and scale use. Insurance take-up is governed jointly by beliefs, perceived losses, contract terms, affordability, incentives, and institutions, so low demand cannot be attributed automatically to optimism. The principal analytical conclusion is that optimism bias should be evaluated as an empirically testable component of a broader insured-risk decision process, with identifiable comparison objects, consistent time horizons, exposure-sensitive benchmarks, and appropriately qualified causal claims.
Keywords: Unrealistic optimism; comparative risk judgment; insurance demand; subjective probability; perceived vulnerability; reference class; risk calibration; flood insurance; behavioral insurance economics; measurement validity; risk communication; heterogeneity.
Introduction
An insurance decision is made about an uncertain future loss, under a contract that redistributes some financial consequences across states of the world. Consumers rarely know the probability of every insured contingency. They estimate vulnerability from experience, locations, information, emotion, social comparison, and the terms presented to them. The insured event may be a familiar illness, an automobile collision, a house fire, a flood, or a rare systemic catastrophe. The same person can hold coherent perceptions about one risk and poorly calibrated perceptions about another, and can act conservatively on a low probability despite believing that personal exposure is lower than other people's.
Comparative risk judgment adds a specific question: how likely is an adverse event to affect me, relative to people such as me? In the classic unrealistic-optimism experiments, Weinstein (1980) found that student respondents frequently placed their prospects more favorably than those of classmates across a range of future events. The result initiated a lasting research program, but the interpretation of a self–other difference depends on the event, the group, the question, and what is known about actual exposure. The later four-form conceptual review by Shepperd, Klein, Waters, and Weinstein (2013) is especially important because it separates absolute from comparative optimism and individual from group-level judgments.
The central aim here is to establish what can be known about optimism bias when the decision concerns insurable exposure. The analysis addresses three connected questions. First, when is a favorable comparative judgment evidence of inaccurate belief rather than accurate recognition of differential vulnerability? Second, which cognitive, motivational, informational, and institutional explanations fit the observed measurements? Third, what inferences about insurance uptake, loss prevention, and welfare follow from the strongest available evidence? The monograph uses selective source-critical synthesis rather than claiming a new experiment, a new meta-analysis, or a new theory.
The scientific material consists of historically foundational experiments, methodological critiques, review articles, insurance-choice experiments and observational studies of flood-risk beliefs and insurance demand. Source-specific statements are supported by contextual links to the verified original article, journal record, scholarly archive, or publisher record. Estimates drawn from study abstracts are identified as such; no proprietary insurer portfolio, household microdata, or individual prediction model has been independently reanalyzed. The resulting arguments are original analytic interpretations of established evidence, with empirical findings separated from deductions and proposed research designs.
This dossier belongs principally to Behavioral Insurance and Risk Perception. Its cross-disciplinary relevance extends to catastrophe risk, health insurance, inclusive insurance, consumer protection, and insurance economics, while each of those subjects retains its independent scientific treatment. The specific object here is comparative personal exposure and its relation to insured decisions, rather than the whole field of subjective probability, social cognition, or hazard perception.
Comparative Risk Judgments: Concepts and Research History
The claim embodied in a self–other comparison
A comparative statement has at least four constituents: a person, an event, a time horizon, and a comparison population. To say that one's probability of an accident is below that of a typical driver may be accurate for a driver with less mileage, safer roads, and fewer violations. The same statement may be mistaken if those factors actually increase exposure. Without specifying the reference class and verifying relevant conditions, the adjective “unrealistic” describes an interpretation rather than an established result. The difference between reported beliefs and objective vulnerability is therefore central to behavioral inference.
Weinstein's (1980) first study asked more than 200 students to compare their prospects for 42 future events, finding favorable mean judgments for positive outcomes and adverse-event vulnerability. A second study suggested that making people consider factors applicable to others could reduce, though not abolish, the pattern. The original findings establish an experimental comparative-judgment phenomenon in a particular study population. They do not establish that every participant who expressed a favorable comparison was actually less accurate about future personal losses. The direct study record and design support that measured distinction.
In a follow-up health-hazard experiment, Weinstein (1982) studied 100 students comparing their vulnerability to 45 health- and life-threatening problems, reporting significant optimistic comparative patterns for 34. Perceived controllability, lack of prior experience, and beliefs about when a condition first occurs contributed to variation. A subsequent community-wide risk survey of 296 central New Jersey respondents examined 32 hazards and found optimistic tendencies linked to perceived preventability and experience. These studies broaden the empirical basis beyond one set of future events while retaining a particular historical geography and set of participants.
The history also contains an important caution about the word “bias.” Harris and Hahn's (2011) critical methodological analysis examines why population averages and event frequencies can generate misleading evidence when participants give comparative ratings. Shepperd and colleagues (2013) clarify that forms of unrealistic optimism defined against observable outcomes or individualized actuarial predictions face a different evidentiary problem from a group average of directly elicited relative judgments. Taken together, these works require a distinction between a phenomenon in response distributions and a confirmed mistake about the probability of personal harm.
Absolute and comparative optimism
Absolute optimism concerns an estimate for the self relative to a defensible benchmark: for example, a household gives a five-year flood probability of 1% when a well-validated, exposure-specific estimate is 6%. The comparison is between subjective and modeled probability, whose uncertainty must be reported. Comparative optimism concerns the difference between self-risk and risk attributed to a reference person or group. Someone may accurately know that personal hazard probability is 6% but believe the typical neighbor's probability to be 12%, even when the objective neighbor probability is 4%; comparative beliefs may then be misleading despite a calibrated personal estimate. The reverse combination is also possible.
Shepperd et al.'s taxonomy of unrealistic optimism identifies individual versus group and absolute versus comparative manifestations. A group mean below a nominal “average risk” response can be suggestive, but aggregation cannot by itself recover how different every participant's actuarial circumstances are. Insurance science needs the person-level analysis whenever the downstream question is why a given household accepts or rejects protection. This implies that questionnaire design, individual exposure data, and observed choice should be kept conceptually and statistically distinct.
A third distinction concerns confidence rather than direction. An individual can place herself favorably while admitting wide uncertainty; another can predict her personal risk with excessive precision despite a correct mean. Moore and Healy's analysis of overconfidence differentiates overestimation, overplacement and overprecision in a broader decision-judgment context. Applied cautiously to insurable risk, these are separate questions: whether estimated personal hazard frequency is accurate, whether the ranking among peers is accurate, and whether intervals describing uncertainty cover plausible frequencies. Their concurrence is an empirical hypothesis, not part of the definition of comparative optimism.
The insurance object of comparison
Loss probability is only one dimension of insured exposure. Expected uninsured financial harm can be represented as the probability of an event multiplied by a loss distribution conditional on its occurrence, adjusted for deductibles, limits, exclusions and mitigation. Two people can face identical probabilities of a flood yet very different potential insured and uninsured losses because one property has a lower floor, more valuable contents, or exclusions that affect recovery. A consumer may say “I am safer” while referring to lower expected severity, faster recovery, broader insurance coverage, or a lower chance of an event. Comparative risk assessment must determine the referent of the statement before treating it as a probability estimate.
That ambiguity becomes more consequential when insurance itself changes the perceived object. A homeowner with a comprehensive policy may answer that personal “risk” is low because financial risk is buffered, even while geographic hazard is unchanged. A question about the probability of flood damage and a question about the probability of unaffordable uninsured flood damage measure different constructs. Insured-risk research therefore specifies the event, the event's severity, the reference population, the horizon and the coverage state in every elicitation. These specifications are analytic necessities of interpreting judgment, not a new theory of insurance.
Personal and Population Exposure: Explanatory Mechanisms
Self-information, reference persons, and cognitive asymmetry
The individual usually knows more about personal behavior than about the behaviors of a randomly selected peer. This information imbalance can produce systematic asymmetry even when people make no effort to protect their self-image. Chambers, Windschitl and Suls (2003) demonstrated in three comparative-judgment studies that event frequency matters because respondents give greater weight to their own event likelihood than to the comparison target's. Their experiments support a nonmotivational mechanism in which common events can elicit comparative pessimism and uncommon events comparative optimism. This mechanism is relevant to rare insurance losses, where a person can rely on concrete evidence about a home or vehicle but possesses only an undifferentiated image of the average household.
Moore and Small (2007), in their comparative-judgment experiments, argued that greater information about oneself than others can create both better-than-average and worse-than-average judgments. The work concerns task comparisons and is used here as an explanatory analogy rather than direct evidence that insurance claimants behave identically. The implication is methodological: a shift in apparent optimism following additional reference-group information may reflect correction of a peer estimate, correction of a self-estimate, or a change in which dimension receives attention. Without measuring these pathways separately, a communication experiment cannot identify which judgment was modified.
The review of moderators by Helweg-Larsen and Shepperd (2001) provides direct grounds for this separation. Some variables, including experience and perceived control, more strongly influence judgments about oneself; the chosen target group can predominantly influence estimates about others. If a questionnaire changes the referent from “the average person” to “the average owner of a similarly situated property,” a smaller favorable self–other gap may arise without changing a respondent's perception of the physical hazard at home. It would be incorrect to record that movement as established objective risk recalibration without an independent exposure benchmark.
Experience, available examples, and controllability
Personal experience is both information and a psychological event. An individual who has never flooded may infer that a home is safe, although the observation period is short and rare events are expected to be absent for most households. Conversely, experiencing a flood can elevate immediate estimates because an otherwise abstract event has become imaginable. The New Jersey community study found that extrapolations from event-free personal histories and perceived preventability were associated with favorable vulnerability judgments. The result suggests one route by which low-frequency insurable losses may be discounted, while the actual validity of such an inference depends on the property's exposure and the history's duration.
Research on judged causes of death makes a complementary point. In the lethal-event frequency experiments, Lichtenstein and colleagues (1978) reported systematic distortions related partly to availability and memorability, including disproportionate attention to some rare events. This means that a vivid disaster can also cause probability overestimation. The direction of a salience effect depends on which representations the person retrieves, the objective base rate and the elicitation procedure. Insurance-risk communication should therefore treat availability as a potential source of directional error rather than assuming it always drives complacency.
Perceived control poses a different problem. A well-maintained building, safe driving or preventive healthcare can reduce risk; the belief that “my choices matter” can therefore be partly accurate. When control is overgeneralized to hazards beyond personal agency, such as a widespread flood or a common infrastructure failure, favorable relative judgments may become detached from exposure. The distinction is causal: an effective action alters the loss process; a belief in control without effective action alters the assessment of the process. For research, observable prevention measures should be collected independently rather than treated as an unverified explanation in respondents' narratives.
Emotion, severity, and decision weighting
People evaluate the seriousness of the loss along with its likelihood. Slovic's (1987) analysis of risk perception emphasizes that public judgments include dimensions beyond a technical frequency and can affect the reception of policy interventions. The insurance researcher should distinguish subjective probability from perceived loss severity, dread, trust, and capacity to recover. A person may correctly perceive a low annual probability but attach such high subjective cost to the outcome that insurance is purchased; another may recognize high exposure but face a severe affordability constraint.
The literature on decisions under risk provides a separate account of how probabilities are converted into choices. Tversky and Kahneman's (1992) cumulative prospect theory treats decision weights as distinct from objective probabilities and identifies a fourfold pattern of risk attitudes in their experimental setting. Thus the same stated probability can lead to different coverage decisions. Comparative optimism is a property of a belief or judgment; probability weighting is a property of choice under uncertainty. An insurance purchase cannot identify either psychological process in isolation. Connecting the two without additional evidence conflates what the consumer thinks will happen with how the consumer values the gamble.
Why apparently conflicting findings can coexist
The literature's combinations of optimism, pessimism and accurate estimates cease to be paradoxical when underlying measures are separated. An uncommon event can elicit favorable comparative placement because self-information is salient; its catastrophic consequence can elicit fear; media exposure can elevate estimates for a short period; and insurance affordability can still determine whether coverage is feasible. In another setting, common exposures may elicit comparative pessimism despite low perceived severity. Different results across hazards, populations and question formats are expected when reference groups, information conditions, event frequencies and action constraints change. An explanatory analysis that accounts for these variables is stronger than assigning all outcomes to one stable personality trait.
Optimistic and Pessimistic Beliefs: Measurement and Identification
The estimand and its benchmarks
A defensible study begins by specifying what a reported comparison is intended to estimate. Let p_i(T) denote the best available external estimate of person i's probability of a precisely defined event within horizon T, conditional on measured exposure. Let p_R(T) denote the estimated probability for a clearly defined reference population R. Let q_i(T) and q_R(T) denote the person's reported subjective probabilities for self and reference, elicited on the same scale and over the same horizon. The reported comparative gap is G_i = q_i(T) − q_R(T). The estimated exposure gap is B_i = p_i(T) − p_R(T). An analytic descriptive quantity is D_i = G_i − B_i. For an adverse event, D_i below zero describes a comparative belief more favorable to self than the selected benchmark. Its interpretation as unrealistic comparative optimism depends on the validity and uncertainty of the exposure estimates, the elicitation, and the identity of R.
This is bookkeeping for existing measurement distinctions, not a proposed psychological law or a named model. It makes transparent why an apparently favorable G_i is not equivalent to a negative personal calibration error q_i − p_i. Algebraically, D_i = (q_i − p_i) − (q_R − p_R). One can obtain a large negative D_i because personal risk is underestimated, because peer risk is overestimated, or through both terms. A regulator, insurer or researcher concerned with household decisions should know which component is present. The decomposition follows directly from the definitions; it is not an empirical result.
A purely illustrative numerical example clarifies the stakes. Suppose a household's annual hazard probability is externally estimated at 4%, while the appropriately weighted peer reference is 2%. The respondent reports personal risk of 1.5% and peer risk of 3%. The perceived comparative gap is −1.5 percentage points while the exposure-based gap is +2 percentage points; D_i is −3.5 percentage points. The self estimate is 2.5 points below its benchmark, and the peer estimate is 1 point above its benchmark. This example establishes arithmetic only. Real annual flood probabilities are uncertain, may vary sharply by parcel, and cannot be transferred from this hypothetical illustration to policyholder decisions.
There are several possible benchmarks. Historical realized frequencies can calibrate a homogeneous repeatable event over sufficiently long observation, but estimating individual rare-event probability from a short history is unstable. A geographic hazard model can provide an exposure-specific distribution, yet its assumptions about climate, land use and defenses must be understood. Insurer pricing may incorporate expenses, capital and regulation and is therefore an unsuitable direct proxy for pure event probability unless those components are removed. An insured claim frequency measures reported, covered and processed losses, not all damaging events. Benchmarks should be documented alongside their known selection effects and uncertainty, and absence of a validated benchmark should limit the claim to comparative belief or placement, not established inaccuracy.
Direct and indirect elicitation
A direct format asks a respondent to rate personal likelihood relative to a reference person, often on an ordered scale from “much lower” to “much higher.” An indirect format asks for personal and reference likelihood separately and computes their difference. These instruments can produce different distributions because participants interpret the comparative scale, the meaning of the average person, and absolute frequency differently. The 2013 methodological synthesis discusses how distinct measures support distinct claims and why criticism of group-level direct comparisons does not automatically negate all forms of unrealistic optimism.
For insurance research, the indirect format provides two estimates for analysis. Even so, asking about the reference group may change personal judgments, and a probability estimate without a time interval is ambiguous. Questions such as “What is the chance of flooding?” should be replaced by a stated event definition and horizon: “What is the chance that water from river overflow enters the ground floor of this address during the next five years?” The comparison target should correspond to a meaningful group, such as addresses within the same floodplain, rather than an unspecified “average American.” In survey design, the narrower statement can improve construct validity while introducing cognitive burdens that must be tested.
Numerical probabilities are also challenging when events are rare. Respondents may round 0.1% to zero, report impossibility for a very low likelihood, or use qualitative anchors that differ across languages. An ordered response category measures relative intensity rather than a cardinal probability difference. Researchers should distinguish a rank order from an absolute probability and avoid interpreting a movement of one Likert point as a fixed actuarial change. Where feasible, consistent frequency representations—such as events per 1,000 comparable properties over a stated horizon—can be tested against percentage wording. Format comparisons must randomize wording order and check whether numbers are understood.
Group-level inference, heterogeneity, and bounded scales
A group mean is not an individual diagnosis. If high-risk and low-risk people are mixed, a mean favorable self-assessment can reflect actual heterogeneity, unbalanced group membership, or measurement artifacts. When everyone responds relative to an “average” that is not an actual member of the sampled population, a mathematical inference from the mean to individual error can be invalid. The cautionary note on unrealistic optimism by Harris and Hahn (2011) makes these inferential vulnerabilities particularly salient. Appropriate analysis must declare whether it tests group-level relative responses, person-level accuracy, or the effect of experimental manipulation.
Statistical regression toward the mean may influence the inferred risk of others when respondents know their own circumstances more precisely than peer circumstances. Moore and Small's (2007) self–other information analysis demonstrates how such asymmetry can generate favorable and unfavorable comparisons across task difficulty. For rare insured events, floor effects add a special complication: probabilities approaching zero cannot be revised downward by as many points as high probabilities, and coarse categories may hide differences between a very low hazard and a practically impossible event. The scale's support and event base rate should be reported alongside the observed direction of comparison.
Analyses should inspect whether a self–other gap changes when peer descriptions become more specific, participants learn objective exposure factors, event horizons shift, and the comparison is expressed in probabilities rather than ordinal categories. A robust result across these tests warrants a more confident behavioral explanation than one survey question. Robustness does not supply a ground truth for every household; it helps distinguish method-induced patterns from persistent judgments. The distinction is especially important when an intervention will affect insurance costs or decisions about protective expenditure.
A practical identification sequence
An internally coherent research design can proceed in a series of explicit operations. Researchers first define the covered or uncovered adverse event, its severity, horizon and reference class. They then obtain self and peer beliefs before giving corrective information, collect objective exposure indicators and past experience separately, and randomize the order or format of the comparative questions when feasible. A validated external benchmark is estimated with its uncertainty. Personal calibration, peer calibration and comparative placement are then computed separately. Finally, insurance and mitigation choices are measured with prices, coverages, budget limits, mandatory requirements and previous policies observed.
The aim of this sequence is to distinguish descriptive measurement from causal explanation. If random assignment changes the comparison information and shifts the self–other gap, it identifies an effect of the information on the reported gap for the studied population. It does not automatically identify a change in real insurance purchase unless purchasing is independently observed. Likewise, a regression relating comparative optimism to nonpurchase can be confounded by wealth, geographic exposure, premium variation, education, financial literacy, previous losses, contract complexity and insurer availability. Where transactions are unavailable, hypothetical choices should be labeled hypothetical.
The methodological review of insurance experiments by Jaspersen (2016) provides an appropriate general warning: experimental and stated-preference designs vary in their realism and in the choices they elicit. The value of a controlled laboratory question is identification of a mechanism under stated assumptions, while a field observation offers ecological relevance with more complex confounding. Strong conclusions about the effect of comparative optimism on actual premiums paid or claims borne require connecting these forms of evidence rather than treating one design as universal.
Misclassification risks in administrative data
Insurance records can make measurement appear precise while introducing new errors. A paid claim requires insured damage, notification, coverage eligibility, and insurer administration; its probability differs from an objective weather event and from the conditional probability of financial hardship. A property with insurance may appear safer in uninsured-loss data because some losses are reimbursed. A household without insurance may have no observable claims in a carrier's records, creating an especially misleading association between coverage and reported historical loss. Attrition, deductible differences, exposure changes and postdisaster mobility can further alter observed frequency.
Researchers comparing subjective beliefs to modeled insured losses should account for the contractual layer explicitly. A flood map describes hazard conditional on defenses and geography; an insurance rate adds expected covered loss, administrative cost, catastrophe capital and sometimes subsidies or cross-subsidies; a catastrophe declaration measures a political or administrative threshold. These are different constructs. The denominator of any comparative claim must be declared—persons, addresses, policies, policy-years, covered exposures or insured sums—because the same phrase “risk of loss” can otherwise change meaning between data sources.
Comparative Risk in Insurance: Experiments and Field Evidence
Insurance laboratory and survey findings
Behavioral insurance experiments show that identical economic choices may be presented in frames that elicit different responses. Johnson, Hershey, Meszaros and Kunreuther (1993), in insurance-framing experiments, examined hypothetical consumer decisions and reported framing effects and distortions in evaluations of premiums, losses and probabilities. This is evidence that the wording and context of insurance choices matter. It is not direct evidence that all such subjects displayed a specific measured optimism bias against an objectively defined peer group.
A review of this wider literature is supplied by Jaspersen's (2016) insurance-demand methodology survey, which inventories differing experimental structures rather than treating a generic insurance purchase as a transparent behavioral measure. A probability supplied to a participant, an individually elicited belief, an externally modeled exposure and a financially consequential decision each answer a different scientific question. In a hypothetical experiment, a participant may misunderstand the mechanism, simplify a complex policy, respond to the investigator's framing, or act differently because no real premium is due. An observational study can avoid some of these features but must explain selection into the sample and the market.
The early risk-frequency judgment experiments and the risk-perception synthesis help contextualize why heterogeneous lay beliefs arise. They do not themselves establish a causal effect of comparative optimism on insurance take-up. A sound scholarly account preserves this separation and avoids attributing every form of probability distortion, availability effect, or affective worry to comparative optimism. The relevant mechanisms may coexist and deserve independent measurement.
Geographic exposure and flood beliefs in the Netherlands
Flood risk provides a useful observational setting because physical factors can be externally characterized, albeit imperfectly. Botzen, Aerts and van den Bergh (2009) investigated about 1,000 Dutch homeowners in a geospatial flood-risk study. Their measurements included perceived flood probability, judgments relative to an average resident, expected damage, knowledge and socioeconomic characteristics; objective geographical indicators included elevation, proximity to river systems and protection by dikes. The study reported generally low flood-risk perceptions, higher beliefs in some objectively more exposed locations, and underestimation among certain homeowners outside dike-protected areas.
The Dutch case matters in three separate ways. It demonstrates that individual risk beliefs can respond meaningfully to real exposure differences, so low subjective risk cannot be called wrong merely because another region has high risk. It shows that some objectively vulnerable groups may hold beliefs below what the environmental evidence would justify. It also illustrates that an identifiable protective infrastructure changes both physical risk and what people think about that risk. The conclusion that better knowledge may raise awareness is connected to a specific national flood-protection setting and a survey design. It cannot determine the size of the same effect under other defense systems without additional evidence.
In the study, the environmental variables proxy for objective exposure rather than establish a certain prediction for any particular future flood. Geospatial risk estimates remain sensitive to river management, elevation data, defense reliability, changing precipitation, time horizon and model assumptions. A researcher should therefore resist treating the modeled reference as infallible. Even when the subjective ranking is correlated with physical indicators, the household may still misjudge absolute frequency or severity. The substantive result is an imperfect alignment between beliefs, geography, protection and knowledge, not a single uniform nationwide psychological error.
Flood events and changing insurance demand in the United States
Gallagher's (2014) flood insurance panel study examined nationwide U.S. flood events and insurance policies. The study found that flood insurance take-up rose after a flood and subsequently declined, while nonflooded communities in the same television media markets showed a smaller response. Its interpretation highlights learning and the decay or incompleteness of remembered risk information. These findings are consistent with a time-varying risk-information process; they do not measure the sign or calibration of every household's comparative peer judgment.
A separate analysis by Kousky (2017) provides a critical institutional correction. In the hurricane insurance study, the author used policy data from U.S. Atlantic and Gulf states between 2001 and 2010 and found that apparent postdisaster insurance increases were partly attributable to federal assistance rules requiring recipients to maintain coverage. In the reported models, being hit by a hurricane in the preceding year was associated with a 7.2% increase in net purchases; when aid-related requirements were incorporated, the estimated increase in voluntary purchases attributed to hurricanes was much smaller, at 1.5%. The precise percentage figures come from the source's reported study results and should not be interpreted as general worldwide elasticities.
Together these studies establish an important interpretive boundary. Postdisaster policy increases may arise from changed perceived vulnerability, new information, direct experience, mandatory protection linked to aid, altered premiums, or changes in financial capacity. A temporal increase followed by decay is consistent with psychological recency and learning, yet it is insufficient to identify a prior self–other optimism error. The institutional rules must be read with the beliefs and observed behavior. Without such controls, a study may describe a change in insurance sales while misidentifying why it occurred.
Low-frequency risks and the cost of probability discovery
Kunreuther and Pauly (2004), in their disaster-insurance analysis, offer a theoretical explanation for insurance neglect that does not require all consumers to be dispositionally optimistic. An individual may incur explicit or implicit costs in learning the probability of an infrequent event and decide not to spend the resources needed to make a finely calibrated estimate. The account complements evidence about comparative judgments because limited information can produce low take-up even where the consumer has no pronounced favorable peer comparison.
This distinction has direct implications for policy design. A one-page explanation of actual location-specific exposure may reduce information costs; a subsidy may alter affordability; a simplified contract may reduce comprehension costs; a requirement attached to financial assistance may directly change the choice set. Such interventions operate on different mechanisms. If they raise coverage, the result cannot be attributed automatically to removal of optimism bias. Behavioral insurance research should vary interventions independently when possible and measure which psychological and economic channel actually changed.
Evidence boundaries and the selection problem
The cited flood studies are stronger than a narrative built solely from illustrative insurance decisions, but their outcomes and institutions remain specific. The Dutch survey links self-reported beliefs to geographically estimated risk at one historical point. The U.S. policy studies observe insurance take-up and experience over time, with separate concerns about mandates and public assistance. Neither design alone reveals all components of comparative optimism: personal belief, reference belief, exposure-based probability, and financial choice at the same household level. A causal bridge requires data or experiments that observe those variables jointly.
A related selection problem is that available insurance policy data cover purchasers, and surveyed property owners may differ from renters, informal residents, uninsured households or people who move after a disaster. Attrition may correlate with loss exposure. People with high perceived risk may self-select into contracts that expand coverage, while protected individuals may later describe their “financial risk” as lower. Any quantitative inference therefore needs sampling weights, an explicit risk population, stable definitions of insured loss, and a stated strategy for selection into observation. The evidence here supports research priorities and qualified interpretations rather than an unsupported numerical estimate of worldwide optimism bias.
Protective Action, Insurance Choice, and Causal Interpretation
The decision channels between belief and insurance
The relationship between perceived vulnerability and insurance purchase has several stages. A consumer first constructs a representation of an adverse event and its likelihood, then judges what losses would result, considers available financial and physical protections, encounters a premium and policy wording, and decides under income and liquidity constraints. Comparative optimism may influence the first stage, yet it cannot mechanically determine the last. The causal relevance of comparative belief depends on whether the consumer uses a self–other comparison when evaluating the protection. A person can think that everyone faces a small risk and still insure because loss would be financially catastrophic. Another can correctly perceive a high personal risk and remain uninsured because coverage is unavailable or unaffordable.
For a simple benchmark, let an individual face a loss of magnitude L with probability p, and let an insurance policy offer full indemnity at a fixed premium q with no deductible or exclusion. Without insurance, terminal wealth takes the values W and W−L. With insurance it is W−q in either state. Under expected utility with increasing concave u, the individual buys insurance when u(W−q) is at least (1−p)u(W)+p u(W−L). This formal comparison identifies the channels through which perceived p, wealth W, potential loss L, premium q and risk preferences interact. It does not prescribe the decision under real contracts, behavioral probability weighting or uninsured indirect losses. It also shows why comparing p to the perceived risk of a peer is insufficient: the individual's own valuation and the policy's financial consequences enter directly.
The purchaser may instead act as if probability were a subjective value p̂, with other elements held fixed. A lower p̂ can reduce the value attributed to insurance. If the belief concerns only a relative ranking—“my house is safer than the typical house”—and the consumer already knows the relevant personal probability, comparative optimism may have no additional effect. When the peer estimate serves as an anchor for personal risk, the comparison can affect p̂ and the purchasing decision. Distinguishing those cases requires measuring personal probability, peer probability, the actual offer and observed take-up rather than labeling all lower-than-peer ratings a cause of underinsurance.
An insurance premium also carries loads for expenses, risk capital, taxes and distribution; it may be subsidized or regulated. A household can reasonably decline a contract even under accurate probability beliefs if the premium is high relative to expected benefit, available reserves, credit access and risk tolerance. Conversely, a household can purchase coverage for regulatory or mortgage compliance despite assigning little probability to the insured event. The disaster insurance-neglect account of Kunreuther and Pauly (2004) emphasizes how information acquisition and limited attention affect decisions about very low-frequency, high-severity hazards. Its explanatory contribution broadens the analysis beyond motivational optimism.
Risk appraisal and preventive behavior
A favorable self–other belief could affect actions that reduce the underlying probability or severity of harm, including floodproofing, safer driving, vaccination or preventive medical care. Yet action depends on whether the individual perceives the recommended response as effective, feasible and worth its direct cost. A message that raises threat perception without explaining a workable response may elicit worry without inducing protective behavior. The distinction is empirically important because protection may be the intended policy outcome rather than the survey answer itself.
Brewer and colleagues' (2007) vaccination-risk meta-analysis examined 34 studies with a combined 15,988 participants. The reported associations between risk perceptions and behavior varied by construct, including approximately r=0.26 for likelihood and r=0.16 for severity; these are associations in a particular health setting and must not be interpreted as causal effects on insurance. The review also discusses the possible two-way relationship: perceptions can affect behavior, while behavior can subsequently change perceived risk. This reciprocity is particularly relevant to insured populations, where prior protection may alter the financial outcome being judged.
A broader experimental meta-analysis of risk appraisals by Sheeran, Harris and Epton (2014) examined whether interventions that heighten risk appraisal change intentions and behavior. The study found that changes in perceived risk could increase intentions and behavior, with effects moderated by perceived efficacy of the response. The exact intervention, outcome and participant population matter. A study about health protection tests the plausibility of a behavioral mechanism; using it as a numerical estimate of flood insurance demand would substitute one outcome domain for another without evidence.
Risk communication for insurance accordingly has at least three independently testable components: information about exposure, information about the potential magnitude and distribution of loss, and an explanation of available protective and financial responses. A person who learns that a property's annual flood probability is higher than previously believed may raise an absolute estimate while maintaining a favorable comparison to still more exposed neighbors. Measuring only the change in relative rating would miss the central correction. A separate assessment of whether the person understands deductibles and exclusions may determine whether knowledge becomes a suitable purchase decision.
Probability, severity and financial loss
Low-frequency insurable events illustrate why probability alone is an incomplete risk descriptor. Suppose two hypothetical households have annual event probabilities of 0.5% and 2%, and the same uninsured loss of $100,000 if an event occurs. Their simplified annual expected losses are $500 and $2,000, respectively. The first may genuinely face one-quarter of the second household's annual event probability. A survey response placing the first household below the comparison group could be accurate; treating that response as unrealistic optimism without checking exposure would be a classification error. Even if the first household overestimated the second household's hazard, a policy choice still depends on expected utility, costs and the risk of wealth depletion.
For a second hypothetical comparison, suppose the event probability is the same at 1% but the conditional uninsured losses are $20,000 and $300,000. Their annual expected uninsured losses are $200 and $3,000. The risk ranking changes depending on whether respondents compare event frequency or expected loss. In catastrophe settings, tail severity and the possibility of correlated neighborhood damage also matter. Expected values omit the distribution of financial hardship, especially when reserves or borrowing are limited. This arithmetic is presented only to clarify definitions; the figures are hypothetical and imply no real underwriting rate or actuarial estimate.
An observable difference in expected-loss beliefs may also be compatible with a psychological distinction between event risk and perceived financial vulnerability. A fully insured person can judge expected personal uninsured loss to be low because indemnity will respond, although the objective probability of hazard is unchanged. Researchers should specify whether baseline beliefs are elicited before or after a description of coverage, and whether respondents hold in mind premitigation physical risk or residual postinsurance financial risk. A durable research design can therefore record physical exposure, expected severity, coverage terms and perceived financing capacity as separate measures.
Probability weighting and alternative behavioral mechanisms
Behavioral decision theory provides explanations that coexist with accurate or inaccurate comparative beliefs. In cumulative prospect theory, Tversky and Kahneman (1992) distinguish decision weights from objective probabilities and show why choices may depart from expected utility. A person's willingness to pay for a lottery or an insurance product cannot simply be inverted into a unique belief about event frequency when subjective valuation and probability weighting are both unknown. A participant who declines insurance may overweight a small risk but dislike the premium and contract terms; a participant who buys may be responding to loss aversion rather than to a realistic self–other placement.
The framing and insurance experiments of Johnson and colleagues (1993) further show why consumer responses must be interpreted within the presentation of the option. The label attached to a financial product, reference prices, deductibles and contextual comparisons can affect hypothetical willingness to buy. Such variation is relevant to insurer disclosure, but it should not be collapsed into the single mechanism of optimism bias. Experimentally separating comparative-risk elicitation from choice framing helps identify the incremental role of relative belief.
Social comparison can nevertheless enter the decision economically. A consumer may regard coverage as unnecessary because neighbors with comparable properties do not purchase it, or may interpret the neighbor's policy as evidence of severe exposure. These inferences are especially fragile when residents differ in subsidies, mortgage status, wealth, risk preferences and property elevation. A household's comparison to a friend is informative only to the extent that the friend's exposure and policy circumstances are relevant. The appropriate empirical test records whose behavior is visible and which characteristics the respondent believes are shared.
Causal identification and feedback
An observational finding that people who report low comparative vulnerability buy less insurance is compatible with several causal structures. Comparative optimism may reduce take-up, but low-risk households may accurately perceive their advantage and buy less coverage. Existing insurance can change definitions of financial risk, and previous losses can simultaneously increase perceived susceptibility and demand. Income and financial sophistication can influence both survey responses and contracts. Selection into a panel and attrition after severe events can compound these relations. A regression coefficient controlling only for age and income cannot resolve all of these alternatives.
A stronger experimental design can randomly assign an objectively justified comparison message while holding premiums, coverage and information about absolute risk constant. It can separately randomize the presentation of absolute hazard probabilities, because supplying only a relative statement may change both personal and peer beliefs. Measuring each belief before and after information clarifies the mechanism. Observing a binding insurance purchase or a real mitigation expenditure would make the resulting behavioral evidence more policy-relevant than a stated intention alone. Such a trial must still evaluate comprehension, consent and heterogeneous effects; randomization of a message does not mean the message is truthful for every address unless the external benchmark is accurately matched.
A field design can combine geocoded property exposure, repeated belief elicitation, actual policy offers and administrative policy histories. Event studies around floods should distinguish physical damage from exposure to media coverage, mandatory purchase requirements, premium changes and eligibility for aid. Gallagher's (2014) insurance-learning evidence and Kousky's (2017) posthurricane contract analysis demonstrate why the timing of experience and institutional incentives must be included when assessing uptake. Their studies address different mechanisms and should be used together as a check on monocausal explanations, not added together to manufacture a combined causal estimate.
Welfare and the interpretation of corrected beliefs
The welfare question is whether a correction improves the match between choices, actual exposure and the person's legitimate priorities. A movement toward an external probability estimate is evidence of improved numerical calibration only when that estimate is sufficiently reliable and the event definition matches the respondent's question. A change in the direction of policy purchase may be beneficial for some households and less suitable for others. A sound welfare analysis compares expected utility or a declared social objective under the full menu of available insurance and protective options, including distributional effects of policy prices and public guarantees.
A communication strategy that induces universal pessimism may increase demand for products beyond the level justified by actual risk and contractual value. The appropriate objective is accurate, decision-relevant understanding. Insurers and public agencies can provide specific hazard information and readable contract descriptions while leaving room for heterogeneous preferences. When the probability of severe harm is uncertain and climate or land-use conditions are changing, a range with clear assumptions is more defensible than spurious precision. The work of Shepperd and colleagues on the limits of comparative inference is therefore directly relevant to the ethics and science of communicating insured risk.
Cultural, Institutional, and Distributional Boundaries
Reference classes across countries and cultures
Comparative judgments are elicited in socially situated language. “Average person,” “typical neighbor,” “someone of your age,” and “owner of a similar home” identify different populations. In a country with extensive flood defenses, a respondent may implicitly compare personal exposure with all households nationwide; elsewhere the reference may be others in the same high-risk informal settlement. Translating a comparison scale into another language preserves neither the imagined group nor necessarily the same probability semantics. International replication must explicitly define the reference class, test cognitive comprehension and evaluate measurement invariance before interpreting differences in mean ratings as differences in the prevalence of optimism bias.
The foundational Weinstein student comparisons, the New Jersey community sample and the Netherlands flood survey supply evidence from distinct populations and historical contexts. They establish recurring questions and some empirical parallels, not a statistically representative map of the world's cultures. The strength of a cross-country inference is therefore limited by whether comparable events, objective risk benchmarks, questionnaire formats and sampling frames were used. A difference between two studies may reflect their hazard geography as well as their cultural environments.
Cultural explanations should be tested against social and institutional alternatives. Trust in authorities, collective response arrangements, religious or community practices, familiarity with insurers and the importance of household networks could influence both perceived vulnerability and the value of private protection. Each factor requires appropriately measured evidence. This monograph treats such pathways as researchable hypotheses, not as established country rankings or attributes of national character. A cross-cultural study gains explanatory value by measuring the mechanisms rather than assuming them from a country label.
Insurance institutions and the comparison baseline
Insurance arrangements affect the meaning of a risky event. A mandatory national protection scheme can make event exposure salient while reducing residual uninsured financial exposure. A voluntary private market may require individuals to infer available protection from intermediaries, advertising and family experience. Mortgage-related flood insurance requirements can turn an apparent consumer choice into contractual compliance. Postdisaster assistance can influence expectations about who will bear losses. These features alter decisions even if objective hazard and cognitive judgments remain unchanged.
Kousky's (2017) U.S. assistance-rule analysis is particularly instructive because separating compulsory from voluntary uptake materially changes the empirical interpretation of a posthurricane insurance increase. The lesson extends beyond the United States: cross-national comparisons of household take-up need a classification of mandatory coverage, public guarantees, subsidies, exclusions and market access. A low rate of voluntary purchase may be appropriate in a jurisdiction offering comprehensive automatic protection, and a high purchase rate may reflect legal mandates. Neither statistic directly measures comparative optimism.
The distinction also matters for the comparison phrase “more protected than other people.” A person can feel less financially vulnerable because a public mechanism provides partial compensation, even when personal physical risk remains high. Survey designers should separately elicit expected government support, understanding of eligibility and the expected proportion of loss borne by the household. When beliefs about public aid are objectively mistaken, that problem calls for institutional information as well as probability calibration. Treating an inaccurate belief about payment rules as an inaccurate hazard forecast would merge distinct scientific objects.
Financial constraints and heterogeneous vulnerability
A person's capacity to reduce or transfer risk is shaped by income, property tenure, collateral, credit, liquidity and local insurer availability. Low-income households may know their exposure accurately while facing greater financial losses relative to wealth. Conversely, a high-income household might accept considerable uninsured loss because it can self-finance. The same measured comparative confidence can therefore support different decisions at different wealth levels. The expected utility benchmark illustrates that the marginal value of wealth and the scale of loss enter the choice, even with identical probabilities.
Risk information can be unequal. Highly technical hazard maps, complex disclosure statements and digital-only insurance enrollment may be accessible to some respondents and inaccessible to others. The remedy may involve clearer frequency presentations, local-language explanation, assistance interpreting contracts and realistic mitigation options. Because reference populations themselves can be stratified, a poorly designed comparative message may compare a precarious household with a geographically similar but structurally reinforced property. A comparison that omits housing quality, floor elevation, drainage, employment consequences or displacement risk may systematically misrepresent actual vulnerability.
The study of risk perceptions should therefore record exposure characteristics that bear causally on harm and the constraints that bear on protective action. Where possible, results should be disaggregated by income, age, tenure, disability-related accessibility needs and geographic exposure without turning protected characteristics into simplistic causal substitutes. These practices improve generalization and avoid blaming uninsured populations for purported optimism when a binding financial or institutional constraint is responsible. Such interpretive care is consistent with the broader methodological attention to insurance decision experiments in Jaspersen's review.
Time, experience and changing hazards
A judgment can become inaccurate without any change in the respondent's psychological disposition. Flood infrastructure can be reinforced or deteriorate; construction can alter drainage; migration can change exposure; climate-related processes can change event frequency and severity; and insurance contracts can change definitions of covered loss. The object of calibration is consequently indexed to time and information. A statement that was accurate at one date may require revision after new hazard evidence. Longitudinal analysis needs contemporaneous objective benchmarks rather than retroactively assigning a later hazard map to earlier belief data.
Following a severe event, experience can increase salience, information seeking, perceived risk and demand, but population movement and insurer withdrawal may also change the set of observed households. The U.S. flood learning study shows time variation in uptake and points toward a dynamic information process. It is compatible with gradual forgetting or decreasing salience; the exact mechanism needs direct measurement if it is to be presented as established. A longitudinal survey can test this by measuring exposure beliefs and peer comparisons at repeated intervals, alongside changing prices, contracts and public policy.
Different future horizons deserve special attention. A one-year probability and a thirty-year cumulative probability are related through assumptions about temporal dependence and stationarity. Under a purely illustrative independent annual event model with constant annual probability p, the probability of at least one event in T years is 1−(1−p)^T. If p=1%, the thirty-year probability is approximately 26.0%; a respondent who reports “one in a hundred each year” and “roughly one in four in thirty years” is internally coherent under those assumptions. Presenting the second number without explaining the horizon can appear alarming and distort a self–other comparison. Under changing hazards or correlated annual events, the simple expression requires adjustment rather than silent extrapolation.
Consumer protection and responsible disclosure
A decision-support intervention should distinguish personal from population risk and label model uncertainty. Explanations that compare an address to a relevant group must report the basis of comparison, not rely on unsubstantiated claims that a particular individual is “safer than most people.” Consumer-facing insurance communications should identify the coverage period, insured event, material deductibles and exclusions, and whether the displayed probability reflects the likelihood of an event or the probability of an insured claim. These disclosure priorities follow analytically from the measurement distinctions developed in this dossier.
An insurer using individualized hazard models faces a further governance problem: the model may be predictive in aggregate while miscalibrated for vulnerable subgroups, or it may mix risk determinants with proxies for socioeconomic status. A responsible comparison should therefore be auditable for data quality, exposure definition and distributional effects. The information has value when it improves comprehension, but it can mislead when framed as a definitive ranking without confidence intervals, comparable denominators or an accessible explanation of relevant drivers. This question concerns the reliability and communication of risk estimates, not an authorization to disclose private underwriting scores.
Public policy can compare several intervention types under controlled evaluation: accurate hazard information, enhanced contract disclosure, navigation assistance, insurance affordability support, physical mitigation and institutional reforms. Each intervention changes a different decision margin. The same observed purchase can have different welfare meanings depending on subsidies and contract exclusions. The relevant evaluation reports comprehension, calibrated beliefs, actual take-up, mitigation, net financial protection and heterogeneity of effects. An increase in premiums collected is not by itself a measure of improved consumer welfare.
A research agenda derived from the evidence
The strongest next empirical contribution would combine three measurements usually separated in the literature: objectively estimated personal and peer exposure, directly elicited personal and reference-class beliefs, and actual insurance decisions under observed constraints. Representative sampling would include insured and uninsured households; repeated waves would distinguish experience from the passage of time; preregistered questionnaire variants would test how the reference class and number format affect responses. Where feasible, ethically designed randomized information provision could identify changes in belief and choice. Such a design would test established competing explanations without creating a new theoretical system by declaration.
Researchers should predefine which outcomes are primary: absolute calibration error, comparative calibration error, risk ranking, policy take-up, contract quality, or physical mitigation. Statistical analysis should estimate uncertainty in both the survey response and the modeled exposure benchmark, report heterogeneity, and handle missingness and selection. Replication across relevant jurisdictions should standardize constructs while preserving locally meaningful institutional variables. Results could then establish whether comparative optimism contributes incremental explanatory value after geographic hazard, affordability, contractual requirements, information acquisition and general probability weighting have been taken into account.
Conclusion
Optimism bias in comparative risk assessment is a question about the accuracy and consequences of a particular class of self–other judgments. The foundational experiments demonstrate that favorable relative assessments occur under identifiable elicitation conditions. Subsequent research establishes that their interpretation is sensitive to event frequency, egocentric information, sample heterogeneity, the imagined comparison population, response format and the availability of an objective benchmark. Those are scientific conditions for identifying error. A descriptive mean difference is an observed psychological result; a claim of unrealistic personal exposure requires additional evidence about the person and the comparison class.
The original synthesis advanced in this monograph separates four inferential stages: the content of the comparative belief, its calibration against relevant exposure, the causal process through which the belief may influence behavior, and the economic consequences of the resulting protection decision. This ordering is a method of organizing the existing literature, rather than a new formal model or doctrine. It clarifies why a household can be accurately below average on physical hazard while inaccurate about a financial loss, or can hold a biased comparison while buying a suitable insurance policy. A valid research conclusion should identify the stage supported by data rather than use the label “optimism bias” as a substitute for an explanation.
The strongest insurance-specific observational evidence reviewed here concerns flood risk and uptake. The Dutch homeowner study shows that perceptions and physical exposure can be compared within a specified setting and that knowledge and protection structures matter. The American flood and hurricane studies reveal adjustments in policy uptake after experience, together with the need to distinguish learning from requirements linked to disaster aid. These results support the relevance of risk perception and institutions; they do not supply a universal causal coefficient translating favorable peer comparisons into an insurance-demand shortfall.
Evidence about risk communication and protective behavior also requires bounded interpretation. Findings from vaccination and other health-related interventions suggest that perceived likelihood and coping efficacy can affect intentions and actions. Their outcomes and effect sizes cannot be transferred numerically to insurance contracts without independent verification. Behavioral insurance experiments help identify framing and decision weighting, while observational policy histories help reveal market and institutional constraints. A cumulative explanation needs both kinds of evidence and should preserve their distinct identification strengths.
For international insurance and risk science, the principal practical implication is to design comparative-risk research around explicit reference groups, relevant horizons and exposure-sensitive objective benchmarks. Evaluators should observe the conditions of the insurance decision: premiums, available coverages, mandatory requirements, income, information access and anticipated public relief. Where a specific optimism error is demonstrated, accurate risk communication may improve decisions, especially when paired with understandable and feasible protective responses. Where other constraints dominate, policy must address those constraints directly.
The broader scientific conclusion is that comparative optimism is measurable and substantively relevant, yet its direction, size, accuracy and behavioral significance are conditional on the specified risk and population. Research moves forward by making those conditions observable, testing competing mechanisms and reporting calibrated limits. The resulting knowledge supports more precise household risk assessment, clearer insurance communication and better evaluation of protection outcomes without turning a heterogeneous empirical literature into a universal psychological claim.
References
Botzen, W. J. W., Aerts, J. C. J. H., & van den Bergh, J. C. J. M. (2009). Dependence of flood risk perceptions on socioeconomic and objective risk factors. Water Resources Research, 45, W10440. https://doi.org/10.1029/2009WR007743.
Brewer, N. T., Chapman, G. B., Gibbons, F. X., Gerrard, M., McCaul, K. D., & Weinstein, N. D. (2007). Meta-analysis of the relationship between risk perception and health behavior: The example of vaccination. Health Psychology, 26(2), 136–145. https://pubmed.ncbi.nlm.nih.gov/17385964/.
Chambers, J. R., Windschitl, P. D., & Suls, J. (2003). Egocentrism, event frequency, and comparative optimism: When what happens frequently is “more likely to happen to me.” Personality and Social Psychology Bulletin, 29(11), 1343–1356. https://pubmed.ncbi.nlm.nih.gov/15189574/.
Gallagher, J. (2014). Learning about an infrequent event: Evidence from flood insurance take-up in the United States. American Economic Journal: Applied Economics, 6(3), 206–233. https://benny.aeaweb.org/articles?id=10.1257%2Fapp.6.3.206.
Harris, A. J. L., & Hahn, U. (2011). Unrealistic optimism about future life events: A cautionary note. Psychological Review, 118(1), 135–154. https://pubmed.ncbi.nlm.nih.gov/21058872/.
Helweg-Larsen, M., & Shepperd, J. A. (2001). Do moderators of the optimistic bias affect personal or target risk estimates? A review of the literature. Personality and Social Psychology Review, 5(1), 74–95. https://journals.sagepub.com/doi/10.1207/S15327957PSPR0501_5.
Jaspersen, J. G. (2016). Hypothetical surveys and experimental studies of insurance demand: A review. Journal of Risk and Insurance, 83(1), 217–255. https://onlinelibrary.wiley.com/doi/10.1111/jori.12100.
Johnson, E. J., Hershey, J., Meszaros, J., & Kunreuther, H. (1993). Framing, probability distortions, and insurance decisions. Journal of Risk and Uncertainty, 7(1), 35–51. https://business.columbia.edu/faculty/research/framing-probability-distortions-and-insurance-decisions.
Kousky, C. (2017). Disasters as learning experiences or disasters as policy opportunities? Examining flood insurance purchases after hurricanes. Risk Analysis, 37(3), 517–530. https://onlinelibrary.wiley.com/doi/10.1111/risa.12646.
Kunreuther, H., & Pauly, M. (2004). Neglecting disaster: Why don't people insure against large losses? Journal of Risk and Uncertainty, 28(1), 5–21. https://doi.org/10.1023/B:RISK.0000009433.25126.87.
Lichtenstein, S., Slovic, P., Fischhoff, B., Layman, M., & Combs, B. (1978). Judged frequency of lethal events. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 551–578. https://scholarsbank.uoregon.edu/items/6e84075e-f186-4dfb-bbb1-380229907c3f.
Moore, D. A., & Healy, P. J. (2008). The trouble with overconfidence. Psychological Review, 115(2), 502–517. https://pubmed.ncbi.nlm.nih.gov/18426301/.
Moore, D. A., & Small, D. A. (2007). Error and bias in comparative judgment: On being both better and worse than we think we are. Journal of Personality and Social Psychology, 92(6), 972–989. https://pubmed.ncbi.nlm.nih.gov/17547483/.
Sheeran, P., Harris, P. R., & Epton, T. (2014). Does heightening risk appraisals change people's intentions and behavior? A meta-analysis of experimental studies. Psychological Bulletin, 140(2), 511–543. https://pubmed.ncbi.nlm.nih.gov/23731175/.
Shepperd, J. A., Klein, W. M. P., Waters, E. A., & Weinstein, N. D. (2013). Taking stock of unrealistic optimism. Perspectives on Psychological Science, 8(4), 395–411. https://pmc.ncbi.nlm.nih.gov/articles/PMC4451212/.
Slovic, P. (1987). Perception of risk. Science, 236(4799), 280–285. https://pubmed.ncbi.nlm.nih.gov/3563507/.
Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5(4), 297–323. https://doi.org/10.1007/BF00122574.
Weinstein, N. D. (1980). Unrealistic optimism about future life events. Journal of Personality and Social Psychology, 39(5), 806–820. https://doi.org/10.1037/0022-3514.39.5.806.
Weinstein, N. D. (1982). Unrealistic optimism about susceptibility to health problems. Journal of Behavioral Medicine, 5(4), 441–460. https://pubmed.ncbi.nlm.nih.gov/7154065/.
Weinstein, N. D. (1987). Unrealistic optimism about susceptibility to health problems: Conclusions from a community-wide sample. Journal of Behavioral Medicine, 10(5), 481–500. https://pubmed.ncbi.nlm.nih.gov/3430590/.