RESEARCH DOSSIER
Disaster Salience and Insurance Uptake
Abstract. Catastrophes make remote losses concrete, transmit new information about hazard, activate public relief institutions, and alter the financial capacity to purchase insurance. A rise in insurance purchases after an event consequently records the combined effects of risk learning, attention, media exposure, policy requirements, product availability, pricing, and household liquidity. This monograph examines when an event induces take-up, how the response travels beyond damaged locations, whether it persists, and how empirical designs can distinguish voluntary behavioral change from institutional compliance. The evidence centers on U.S. flood insurance panel studies and incorporates research from Japan, China, and European flood-affected households. Gallagher (2014) identifies a postflood purchase increase and an effect among geographically undamaged communities within the same television market. Kousky (2017) finds that postdisaster aid-linked insurance obligations explain much of the policy response to U.S. hurricanes. Choi, Diffenbaugh, and Burke (2024) estimate a positive but temporary response to disaster-scale flooding in a voluntary-coverage sample, while Petkov and Ortega (2025) find more persistent learning in exposed areas outside the designated high-risk floodplain. Experimental evidence from Chinese agricultural insurance establishes that simulated disaster experience and explicit payout-probability information can change actual purchases, although the mechanisms and contracts differ from U.S. property insurance. This dossier develops an original source-grounded synthesis organized around event information, transaction mechanisms, competing interpretations, and identification constraints. It shows why a postevent rise in policies cannot, by itself, identify an availability heuristic, prove durable protection, or establish welfare improvement. The defensible unit of evaluation is a specified contract outcome for a specified at-risk population under a specified institutional regime and observation horizon.
Keywords: Disaster salience; insurance uptake; flood insurance; catastrophe learning; media spillovers; hazard perception; voluntary insurance; National Flood Insurance Program; policy renewal; disaster assistance; causal inference; insurance affordability; behavioral economics; risk communication.
Introduction
Insurance against low-frequency natural hazards requires a household to transfer money today in exchange for a contractual promise concerning a loss that may be remote, uncertain, or previously unexperienced. A catastrophic event rearranges that decision environment. Floodwaters expose a physical hazard, news reports increase attention, social networks transmit stories of loss, and insurers, lenders, and public agencies contact households. Insurance sales may rise at precisely the moment that a household's disposable funds have been depleted or a carrier's willingness to underwrite has deteriorated. Each of these effects matters to interpretation.
The flood-insurance evidence of Gallagher (2014) establishes the enduring empirical puzzle: take-up rises after major events and then recedes. A distinct analysis of posthurricane purchases demonstrates that disaster-assistance requirements can themselves generate new policies. Research that merely labels the entire increment as a psychological response attributes institutional action to household belief. The task of this dossier is to determine what each available research design actually identifies and to make explicit the conditions under which results can be compared.
The relevant outcome has several possible meanings: an application submitted, an offer accepted, a premium paid, a policy bound, an effective contract, a change in insured limits, an active policy at a census date, a renewal, or a self-reported willingness to insure. These quantities answer different questions. The distinction becomes particularly consequential in the National Flood Insurance Program (NFIP), where standard waiting-period rules ordinarily delay the effective date of a new policy by 30 days and where an exception applies in certain mortgage and mapping circumstances. A surge in purchases ahead of or after a forecast does not mean that the same event will be insured.
The distinction between an initial policy purchase and coverage selection is illustrated by Kousky’s St. Louis County insurance study, which examines policy take-up, insured limits, deductibles, and exposure classification. These measurements capture different dimensions of financial protection.
The geographical object is also important. A disaster may damage one street, trigger a declaration across a county, appear on television throughout a region, and become visible to distant acquaintances online. Those populations experience different information and institutional treatment. The distinction between direct physical exposure and mediated attention is central to the interpretation of a media spillover. Similarly, designated high-risk floodplains influence insurance mandates and premium signals, so a postdisaster comparison of mapped and unmapped locations is also a comparison of institutions.
This research uses four forms of evidence. First are peer-reviewed insurance demand and policy-tenure studies using insurer or public-program administrative records. Second are randomized interventions that affect observed insurance purchases. Third are household surveys, property-price studies, and adaptation studies, used to clarify mechanisms while preserving their outcome boundaries. Fourth are original program rules and data documentation that establish what a policy count means. The synthesis is documentary and analytical; it does not claim to reestimate confidential microdata, replicate numerical coefficients, or introduce a new named theory.
The principal claims are straightforward. Recent disasters often increase immediate insurance activity, but magnitude and duration depend on exposure, coverage mandates, media reach, affordability, product design, and the definition of uptake. Direct experience and wider media markets can transmit information simultaneously. A temporary sales response can coexist with rational learning about a changed hazard assessment, with salience effects, and with postdisaster regulatory obligations. The strongest policy inference is therefore to measure protection over the full contract life cycle and to separate an event's informational consequences from its institutional and financial consequences.
Disaster Information and Salience
What a disaster makes salient
A disaster can modify an objective loss distribution through physical changes such as impaired infrastructure, erosion, or altered drainage. It can also reveal previously unknown characteristics of an existing distribution. In addition, an event changes the cognitive accessibility of that distribution without necessarily changing the underlying probability. These are analytically separable claims. The classic availability heuristic of Tversky and Kahneman (1973) concerns judgments influenced by how readily an example is retrieved from memory. Applied to catastrophic insurance, a vivid recent loss may affect a household's perceived likelihood or consequences of a future event even when the hazard has not changed.
Risk updating is not reducible to this heuristic. A flood can supply credible evidence about local vulnerability, actual drainage capacity, an insurer's claims service, or a map classification that a household previously misunderstood. Gallagher's national panel evidence is most consistent, in the author's comparison, with a Bayesian learning specification incorporating forgetting or incomplete retention of information. This should be read as comparative model fit, rather than as direct observation of household Bayesian calculations. The data identify temporal patterns in insurance behavior; the proposed belief mechanism remains an interpretation mediated by the model.
There are additional nonpsychological pathways. A major event raises the expected benefit of learning about product terms because households discover that an existing homeowner policy does not necessarily transfer flood losses. A policyholder may have had inadequate coverage limits, not merely inadequate hazard beliefs. Similarly, an event may introduce a newly active agent, insurer solicitation, a lender notice, or a public financial-assistance condition. All may produce a purchase even when perceived probabilities remain unchanged.
An economic explanation for persistent underinsurance begins before the disaster. In Kunreuther and Pauly's (2004) analysis, the cost of discovering the probability of a rare event can inhibit insurance purchase. A recent hazard may lower that informational cost by making probability and loss scenarios visible. But an event also can increase the cost of search: households must repair homes, replace documents, negotiate claims, and find accommodation. The net effect on attention depends on whether postevent learning is more accessible than the competing burdens of recovery.
For research purposes, event salience can be separated into measurable exposures: property-level damage, local warning, personal flood experience, geographic proximity, a presidential or equivalent disaster declaration, broadcast-market membership, social-media contact, and timing relative to the purchase decision. None is a perfect substitute for a directly measured subjective probability. A severe hazard experienced at a distance may be more salient than a minor loss at home; an administrative declaration may provide a powerful public signal without any physical household damage.
Event chronology and information chronology
The relevant dates include forecast issuance, impact, media circulation, public declaration, assistance eligibility notice, policy application, premium payment, binding, policy commencement, and renewal. Treating these as a single event date leads to misclassification. For a forward-looking purchase mechanism, the proper question is whether the information could plausibly have reached the decision-maker before the purchase. For a welfare question, the relevant variable is whether coverage became effective before the insured loss occurred.
The NFIP purchase guidance documents the standard 30-day wait and exceptions, including eligible mortgage transactions and certain map changes. This rule illustrates why observing a purchase in the aftermath of a headline storm and observing insured protection for that storm are different events. In temporal studies, a one-year postflood policy increase can measure prudent preparation for future flooding even when it did not insure the historical flood.
The interpretation of official map categories also varies. A household outside a designated Special Flood Hazard Area may infer that flooding is implausible or that protection is unnecessary. A flood reaching such property can change that inference sharply. In contrast, a household already in a heavily signaled high-risk zone may receive little genuinely new information from a further event. This distinction becomes empirically important in newer studies of the floodplain periphery.
Salience and observed preference
The same behavioral pattern can have several underlying causes. A purchase following an event could reflect updated expected losses, fear, easier access to insurance advice, assistance conditions, or a temporary discount. A nonpurchase could reflect unchanged beliefs, risk tolerance, unaffordable premiums, a missing product, household debt, or unwillingness to trust an insurer. It is therefore methodologically incorrect to infer psychological states from purchase status alone. Surveys of beliefs can help, but risk perception questions themselves may be affected by news attention and wording.
An analytic account must also differentiate information on hazard probability from information on an insurance contract's payoff. The experimental study by Cai and Song (2017) demonstrates why this matters. Their intervention with Chinese rice farmers provides a disaster-experience simulation and, separately, information about the probability of insurance payment. The researchers report measurable changes in actual purchases, with the information treatment contributing strongly to take-up. A policy encouraging the household to envision an event and a policy explaining when insurance pays are causally different instruments.
Flood and Catastrophe Attention
Evidence from U.S. flood events
The central long-panel evidence comes from Gallagher (2014). Combining regional flood history and flood insurance policy records, the study finds that purchases rise in the year following a major flood and decline afterward. The published AEA abstract additionally reports an increase among communities that were not flooded but shared the television media market of affected communities, at about one-third of the response observed in flooded communities. This supplies evidence that event knowledge can travel beyond direct inundation. It does not make television the only possible shared-market channel: economic conditions, geographic networks, and disaster governance may co-vary with broadcasting geography.
A subsequent hurricane policy analysis by Kousky (2017) reports a 7.2% increase in net purchases following a hurricane in the previous year and an effect that dissipates within approximately three years. More revealing is the institutional decomposition: disaster declarations and grant-linked insurance obligations explain much of the response. When required purchases are accounted for, the paper estimates a much smaller 1.5% increase in voluntary policies following hurricanes. The observed behavioral signal in headline take-up therefore becomes substantially smaller after separating compelled and voluntary contract formation.
This does not imply that a compelled policy has no protective value. An obligation may deliver a contract to a household that would otherwise remain uncovered. It establishes that the immediate purchase cannot be used as unmediated evidence of spontaneous belief updating. The welfare and distributional questions—whether coverage is appropriate, affordable, maintained, and paid after loss—remain substantive.
The newer U.S. adoption study by Choi, Diffenbaugh, and Burke (2024) focuses on locations where flood insurance is not mandated through the high-risk-zone mortgage rule. The authors assemble panel information on insurance, flood events, disaster declarations, and property exposure, and report an average increase of approximately 7% in county-level take-up following a disaster-scale flood in the preceding year. Effects weaken with elapsed time. The 7% is a relative change in a constructed take-up measure, not an increase of seven percentage points in the share of all households insured. The paper reports heterogeneity by coastal or inland exposure, primary residence and premium burden. Its data and construction give stronger relevance to voluntary transactions than a series mixing mandatory and voluntary policies, while its reliance on county-level declarations still leaves indirect exposure within the treated county.
These studies are complementary rather than numerically interchangeable. Gallagher analyzes large regional floods and broadcast-market spillovers; Kousky studies Atlantic and Gulf Coast hurricane-related policies with explicit aid mandates; Choi and coauthors select voluntarily insured areas and compare disaster-scale with other flood events. The event definitions, control groups, risk-map coverage, time intervals and dependent variables differ. Reporting their coefficients as a pooled mean would manufacture an effect with no coherent common estimand.
Persistence at the floodplain periphery
The study by Petkov and Ortega (2025) links building footprints and inundation exposure following a large flood event. Its findings challenge a universal account of rapid decay: increased take-up disappears within three years in the Special Flood Hazard Area but is highly persistent beyond that zone's boundary. A plausible explanation is that flooding outside the mapped high-risk zone reveals genuinely new risk information to households whose previous classification encouraged a low-risk inference. Persistence can thus depend on the difference between event information and the preevent public signal.
Persistent uptake in this setting remains an outcome conditional on the studied event, properties, and map structure. It is not evidence that all novel hazards will generate durable insurance attachment. A household's income, access to a mortgage, and ability to find a policy still determine whether learning becomes protection. The sharp empirical contribution is to show that the spatial boundary of an administrative flood designation can organize the informational response.
Evidence across different institutions
The Japanese prefectural study by Shao, Hoshino, and Nakaide (2022) finds a significant but short-lived, approximately one-year relationship between flood experience and insurance demand, with heterogeneous relationships across socioeconomic characteristics and different dimensions of flood harm. The paper's aggregate prefectural design cannot be read as identifying each homeowner's beliefs or purchase decision. Its value is in showing that an event's measurable attributes and institutional insurance arrangements alter the mapping from disaster to demand.
The insurance product itself differs across jurisdictions. Japanese residential flood protection is offered through private arrangements with distinctive policy packaging. U.S. NFIP policies are embedded in a federal insurance and mortgage regime. Chinese weather insurance for rice farmers in Cai and Song's field purchase experiment pays according to a product-specific weather-related design and subsidy setting. A percentage increase in farmer take-up and a percentage increase in flood policy count do not measure the same economic change. An intellectually disciplined comparison preserves the institutional variables explaining each result.
Cai and Song (2017) report that a simulated insurance game increases real agricultural-insurance take-up by 9.1 percentage points, approximately 46% relative to its roughly 20% baseline, within the experiment's setting. Participants observe hypothetical shocks and payouts before the actual purchase decision. The published experiment results also identify a substantial response to explicit payout-probability information. This evidence establishes that structured experience and contractual information can affect actual purchasing even without a real flood striking the participant. It does not prove that simulations are welfare-enhancing for every insurance product or transferable to all disaster markets.
Studies of adjacent choices help identify the boundary of the question. Kousky (2010) documents changes in property values around the 1993 Missouri flood, including price effects outside narrow mapped flood zones; Atreya, Ferreira, and Kriesel (2013) find that a flood-risk property-price discount in a Georgia setting disappeared over several years. These studies address the capitalization of perceived or experienced risk into real estate prices, not insurance purchases. They support the hypothesis that risk attention can evolve over time but cannot be converted into insurance take-up coefficients.
Purchase and Renewal Dynamics
A stock-flow distinction
A policy-in-force count is a stock. New purchases, renewals, lapses, cancellations and contract expirations are flows. The net change in contracts between two dates can be written as an accounting identity:
where N denotes genuinely new contracts, R reactivated contracts, L lapses, C cancellations, and A other administrative changes within a consistently defined contract universe. The symbols represent a descriptive decomposition, not a newly proposed behavioral theory. When renewal records are counted as transactions, the analyst must avoid mistakenly adding standard renewals to the change in active policy stock. Product changes, ownership transfer, duplicate contracts and boundary changes in the eligible population require explicit conventions.
The NFIP data documentation explains that the public policies data are transactional and that successive renewals generate separate records. For a snapshot, a policy must have an effective date no later than the target date and a termination date after it. Dataset boundaries also matter: the same NFIP guidance notes that complete public policy-level records are available from 2009, while earlier periods have incomplete retention. The federal NFIP data catalog describes the privacy-redacted transactional archive and its analytical limitations. Studies using pre-2009 records therefore require special data access, reconstruction, or narrower claims of comparability.
A rigorous event study should report at least three outcomes: first-time purchase per eligible uninsured unit, active policy prevalence over time, and continuation conditional on an existing contract. Policies with different limits should also be weighted by coverage amount or insured value where feasible. A market can show an increase in contract count even while aggregate uncovered value rises because property values or exposure grow faster. Conversely, a stable count can conceal substantial policy upgrading.
The problem of temporary attachment
A catastrophe can motivate a purchase at renewal time, but the contract's benefits depend on continuation. The NFIP tenure study by Michel-Kerjan, Lemoyne de Forges, and Kunreuther (2012) analyzes program records from 2001–2009 and reports a median tenure for newly issued policies of roughly two to four years. The study also finds that claim size relates differently to policy duration: policyholders experiencing small claims tended to retain coverage longer, while those experiencing large claims were more likely to lapse sooner. The finding contradicts a simple premise that a more severe realized loss necessarily generates stronger future persistence.
A lapse can reflect recovery costs, mortgage repayment, a move, perceived unfairness in claims, rising premiums, or a changed estimate of future risk. This is why memory decay should be treated as one candidate mechanism rather than a residual label attached to every policy that disappears. The Choi–Diffenbaugh–Burke analysis reports temporary event-related adoption effects in nonmandatory areas, but those reduced-form results do not individually decompose lapses into attention, price and income effects.
Coverage becomes effective on a contractual calendar
The NFIP insurance manual and the current consumer purchase guidance set out the ordinary waiting-period structure and documented exceptions. The effective date is a distinct object from first contact with an agent or premium commitment. A hazard forecast, purchase date and eventual payout eligibility should be preserved as separate variables. In particular, recent transactions cannot automatically be classified as timely indemnification for the event that prompted them.
Renewal decisions may also respond to premium changes unrelated to disaster salience. The NFIP's Risk Rating 2.0 guidance describes the transition to property-specific pricing beginning for new policies in October 2021 and affecting renewals from April 2022. Research spanning these dates must account for the introduction and phase-in of altered prices. A nationwide price-policy change does not disappear as an interpretive issue merely because a model controls for common year effects: heterogeneous premium adjustments can be correlated with local risk and purchasing incentives.
What persistence means for protection
Protection persistence is properly measured at several horizons. The first is contract activation; the second is continuation through one full policy year; the third is renewal over multiple hazard seasons; the fourth is retained adequate limits in relation to changing reconstruction cost and hazard. The fifth is actual claims performance if a covered loss occurs. The relevant public interest is sustained access to a comprehensible, enforceable risk-transfer contract at a premium the household can pay. A short-lived postdisaster sales increase may still benefit those covered during a later event, but it is weak evidence that the long-run insurance protection gap has closed.
Media Exposure and Spillovers
The geography of information
Disaster information travels through broadcast television, local newspapers, social networks, public authorities, insurance agents and direct observation. These channels expose people to different combinations of severity, frequency, credibility and emotional vividness. Gallagher's television-market spillover result provides one of the most influential observations: communities without direct inundation nevertheless increased policy uptake when they belonged to an affected media market. The comparison supports a role for shared information. It does not identify how much of the effect was caused specifically by television viewing rather than other shared exposures.
The distinction between a media market and an individual information treatment matters for causal interpretation. Television market boundaries may overlap with shared river basins, economic activities and emergency authorities. People living in an undamaged community may know injured relatives, encounter closed roads, receive insurer advertising or experience nearby evacuation. A broadcast-market instrument becomes persuasive when these rival channels are examined; its underlying geographical assignment is not a randomized television intervention.
Media can influence the perceived probability of loss, the anticipated severity of damage, knowledge of insurance exclusions, and the apparent prevalence of claims disputes. These are different messages. Footage of high water might increase perceived risk; a clear explanation of contract boundaries might improve product understanding; sensational attention to claim denials might reduce trust. The net purchase response depends on the information bundle and the population receiving it.
The direct-experience gradient
Direct flood exposure is typically accompanied by financial loss and a vivid personal account. Exposure through a neighbor can share local hazard information without identical loss. Distant media coverage may carry the message that disasters occur but provide less credible information about a particular property's vulnerability. These categories should be measured separately rather than pooled into a single binary disaster-experience indicator.
The floodplain-periphery analysis offers a further refinement. Risk communication by maps establishes a preevent signal; inundation beyond the mapped high-risk boundary directly contradicts it. A household's response is thus conditioned by what the event disconfirms. Media coverage may increase attention in both zones, while the amount of new location-specific risk information differs.
This leads to a practical interpretive distinction. If purchase grows most among individuals with direct losses, both perceived-risk revision and recovery institutions are plausible explanations. If purchase grows among undamaged residents within a broadcast market, mediated information is more plausible, though residual geographical confounding remains. If change persists chiefly where official maps understated experienced hazard, durable information updating becomes more plausible. None of these patterns alone directly identifies a cognitive bias; each guides the design of additional tests.
Cross-border institutional communication
Insurance literacy also depends on what households are told about who carries catastrophic loss. The 2025 study by Endendijk and colleagues compares 719 flood-affected homeowners in Germany, Belgium, and the Netherlands following the 2021 European floods. Its principal outcomes concern adaptation actions and intentions. Insurance compensation is associated with postdisaster adaptation, while institutional arrangements and perceived responsibility differ across the countries. These observations reveal that flood experience is interpreted through an insurance system and through public expectations about protection. They do not identify a comparable causal event effect on new insurance take-up in each country.
National institutional arrangements affect the information content of public communication. In one country flood damage may be included within a standard property product; in another it requires optional additional coverage; in a third certain flood exposures may lack conventional insurance access. A message to buy insurance therefore has a different feasible meaning across systems. EIOPA's protection-gap dashboard explicitly distinguishes economic losses, insured losses, current risk and country insurance arrangements, underscoring the need to separate insufficient coverage from psychological inattention.
For policy experiments, a credible design would randomize a standardized disclosure within eligible jurisdictions and distinguish local event imagery, an actuarial probability statement, an insurer payout illustration, and a neutral statement of existing coverage. The research question would be whether each message changes verified contract purchases and subsequent renewals, not merely immediate willingness to pay. The design should also measure whether a message causes misplaced confidence or overinsurance.
Behavioral versus Budget Constraints
Risk perception meets the budget
A household may understand its flood risk and still be unable to pay the required premium. The timing of a disaster adds financial stress: earnings may be interrupted and repairs become urgent. In other cases a claim payment or public assistance grant creates liquidity that permits mitigation or a later premium payment. A model treating every nonpurchaser as inattentive misses this demand constraint. It also misses supply-side exclusions, deductibles, and underwriting capacity that determine whether a real contract is offered.
The European cross-border flood study finds that expected or received compensation is associated with household adaptation after flooding. Its study sample is composed of surveyed flooded homeowners and the dependent variable is protective adaptation, not randomized insurance adoption. The association is relevant to the postevent liquidity mechanism and institutional experience, but should not be labeled an estimated effect of compensation on purchasing new coverage.
Affordability is also a selection mechanism. Lower-income properties may face proportionally burdensome premiums, and households with mortgage mandates may maintain coverage even when the burden is high. In the voluntary-area panel study, Choi and coauthors investigate whether cost burden mediates postdisaster responses. They find that those with a lower relative burden respond more strongly in relevant comparisons. As the study reports observational interactions, the interaction should be interpreted with its controlled covariates and measurement definitions; it is not itself a randomized price experiment.
An event can also change insurance supply. A carrier may tighten underwriting, revise deductibles, withdraw from a catastrophe-exposed market or raise premiums. Under those conditions, a small increase in sales is consistent with very large latent demand but limited market capacity. Conversely, heavily subsidized postevent offers could produce a surge without lasting willingness to pay at the actuarial price. A behavioral account should measure the availability and price of actual offers alongside purchases.
Assistance, mandates and measured demand
One of the clearest confounders is a public assistance requirement. In Kousky's (2017) hurricane analysis, disaster-aid rules account for a large share of the observed increase in policy purchases. The implication for empirical research is not that event experience is irrelevant; it is that the same event activates an institution that directly changes the decision set. A credible claim about voluntary demand requires separately recording the legal or administrative mandate that produced coverage.
Related evidence addresses the amount of protection rather than the number of contracts. Kousky, Michel-Kerjan, and Raschky (2018) combine federal disaster assistance records with flood insurance data and estimate that assistance grants reduce average insured coverage the next year by around $4,000–$5,000 in their sample. The authors report a response on the intensive margin and no corresponding effect on the take-up margin, a distinction they relate in part to aid-related insurance purchase requirements. Their identifying assumptions include fixed-effects and instrumental-variable strategies addressing endogenous grants. The estimate is context-bound and cannot be interpreted as a universal reduction caused by any dollar of assistance.
This result is essential for reading coverage totals. Public assistance may preserve or create the extensive margin—whether a contract exists—while reducing the insured limit chosen. A study relying only on contract count could find no crowd-out; a study measuring limits could find meaningful substitution. Both outcomes may be correct because their estimands differ. Public recovery funds, subsidized insurance, and private risk transfer form an institutional system whose components influence each other.
Decisions about protection and protection substitutes
Households also mitigate loss through elevation, barriers, relocating valuables, and moving. Those actions can reduce expected loss and thereby change optimal insurance choice. They can be complements if a lender or insurer requires mitigation or if protection makes rebuilding affordable; they can be substitutes if mitigation decreases the insured loss exposure enough to reduce desired coverage. Osberghaus (2017) uses German longitudinal household and insurance information to estimate effects of flood experience on mitigation, finding stronger responses for behavioral than technical measures. The study supports the proposition that disaster experience affects an array of protective decisions, but those mitigation effects cannot be mechanically counted as insurance purchases.
Welfare analysis further requires asking whether the offered product actually covers the perceived hazard and whether its deductibles, waiting periods and exclusions are understood. A household that purchases an inappropriate policy immediately after a disaster may have responded to salience without improving protection proportionately. Conversely, an informed household may rationally prefer self-protection, savings or public arrangements when insurance terms are unfavorable. The appropriate benchmark is risk-sensitive welfare under feasible contracts, not maximum take-up as an end in itself.
Low-probability neglect and heterogeneity
In their theoretical account of neglected disasters, Kunreuther and Pauly emphasize the cost of obtaining reliable probability information. This helps explain why a rare event may interrupt ordinary inattention. The Robinson and Botzen (2019) experiment examines probability neglect, emotional responses and willingness to pay for flood insurance. It provides a psychological mechanism in an experimental context, rather than evidence that every real postdisaster purchase reflects probability neglect. Distinct emotional and financial factors can operate in opposite directions within the same population.
The practical implication is to allow heterogeneous effects by prior hazard knowledge, flood-map status, family experience, education, income, mortgage status, insurance agent access, loss severity and previous claims. A pooled average can hide groups for whom a disaster triggers lasting learning, groups whose purchase is compelled and groups whose financial capacity collapses after loss. The Japanese evidence of Shao and colleagues is consistent with such heterogeneity, although aggregation at prefecture level limits individual inference.
Causal Evidence and Limitations
Defining the estimand before interpreting coefficients
The scientifically relevant causal question is the effect of a clearly defined exposure on a clearly defined insurance outcome for a defined eligible population over a defined horizon. An example is the effect of a directly observed severe flood, in the absence of a new insurance mandate, on first-time purchases among previously uninsured eligible homeowners during the following twelve months. A different question is the effect of a federal disaster declaration on aggregate policy prevalence. The first concerns exposure and voluntary demand; the second includes declaration-linked eligibility, relief and institutional responses. Both are legitimate, but their answers are not interchangeable.
Let Y_i(t;1) and Y_i(t;0) represent the potential verified contract outcome at horizon t for eligible unit i with and without a specified disaster-information exposure. The average difference across a defined population expresses the causal quantity of interest. Administrative disaster exposure is seldom randomized. To interpret an event-study coefficient as this average effect, researchers must justify why unobserved time-varying determinants of insurance demand do not systematically differ between exposed and comparison populations, or otherwise use credible instrumental, discontinuity, natural-experiment or experimental variation. This formal statement is a conventional causal-inference notation, not a new authorial model.
A methodological comparison of the strongest studies
Gallagher (2014) combines national flood and policy data to estimate time patterns and tests broadcast-market spillovers. The strength is broad event and geography coverage; the principal interpretive limitation is separating information learning from correlated local institutions and selection into exposure. Kousky (2017) connects storms, declarations and federal assistance obligations; its special contribution is distinguishing observed purchases from those linked to the policy response. Choi, Diffenbaugh, and Burke (2024) restrict attention to voluntary settings outside designated high-risk flood zones and distinguish major declarations from smaller events; its approach strengthens the voluntary-uptake estimand but exposure is still assigned at a spatial level above the home. Petkov and Ortega (2025) exploit building-level inundation and mapping boundaries after a major event, illuminating distinct persistence near the floodplain boundary; its findings are conditioned by a particular event and market.
Shao, Hoshino, and Nakaide (2022) offer cross-jurisdictional evidence on disaster attributes and Japanese uptake, but prefecture aggregation poses ecological inference limitations. Cai and Song (2017) provide the most direct experimental manipulation of disaster-related information among the sources reviewed, and measure subsequent actual farmer purchases; external validity to residential flood products is limited by commodity, contract, subsidy and institutional differences. Kousky, Michel-Kerjan, and Raschky (2018) examine the role of federal aid in coverage choice using an instrumental-variable approach; their question concerns aid-induced substitution, not the autonomous salience effect of a flood. The 2001–2009 NFIP tenure study then reveals why even a genuine first purchase cannot be assumed to persist.
Across studies, statistical significance does not imply large absolute protection. A 7% relative increase starting from 0.6% baseline coverage represents a much smaller change in percentage points than the same proportional increase from 10% coverage. Researchers must report baselines, absolute effects, confidence intervals and sample eligibility. Likewise, a difference in policy counts may reflect a denominator shift if homes are destroyed, households relocate or the mapped population changes after a disaster.
Threats to identification
A first threat is correlated shock. Flooding can simultaneously destroy assets, close employers, change credit markets and stimulate government rebuilding. County and year fixed effects remove fixed county differences and shared annual shocks, but they do not automatically remove postevent local income changes. A second threat is policy co-treatment: declarations and assistance rules affect purchase incentives at the same time as information. A third is supplier response: insurers may change marketing, capacity, rates or deductibles. A fourth is spatial contamination: an apparently unexposed control area may be in the same information market, violating a no-interference assumption.
A fifth threat is measurement. Flood incidence in storm databases may describe a jurisdiction rather than individual properties; disaster declarations depend on administrative processes as well as hazard severity. Policy files can include renewals, nonresidential contracts, private-market coverage outside the NFIP, delayed data corrections, or multiple transactions for the same property. The official NFIP data FAQ specifies how effective and termination dates should be used and notes abnormal records can remain in released data. Inadequate matching of policy and exposed-property denominators is therefore a substantive bias risk, not merely a formatting issue.
A sixth threat is survivorship. Severely damaged households may move or lack the finances to buy coverage; a postevent sample restricted to residents remaining in place omits displaced households. A seventh is reverse causality or anticipation: flood warnings and forecasts can prompt purchases before impact, complicating the definition of the preevent period. An eighth is policy-year misalignment: a delayed effective date, annual renewal calendar and retrospective aggregation can distribute a single response across more than one calendar year. These problems require event-time and transaction-time data rather than a single before/after comparison.
Evidence table: outcomes and permitted inferences
| Study and setting | Observed outcome | Supported reading | Important inference limit |
|---|---|---|---|
| Gallagher (2014), U.S. regional floods | Policy uptake and television-market geography | Event-related increase, decay and undamaged-market spillover | Broadcast-market association does not isolate each household's media exposure |
| Kousky (2017), U.S. hurricanes | Net purchases and assistance-related obligations | Much postevent growth is associated with mandated coverage | Total purchases overstate a purely voluntary behavioral effect |
| Choi et al. (2024), U.S. voluntary flood-insurance areas | Constructed active-policy take-up rates | Disaster-scale events yield a positive, temporary average response | Administrative disaster exposure differs from direct home inundation |
| Petkov and Ortega (2025), flooded U.S. property periphery | Take-up following mapped and unmapped exposure | Persistence differs across floodplain status | Single-event and geographically specific external validity |
| Cai and Song (2017), Chinese agricultural insurance | Actual take-up after randomized game and probability information | Disaster simulation and product information can cause purchases | Agricultural contract and subsidy setting differ from home flood coverage |
| Shao et al. (2022), Japanese prefectures | Aggregated flood insurance purchase response | Heterogeneity and a brief event association | Prefecture-level estimates cannot establish individual cognition |
| Kousky et al. (2018), U.S. disaster assistance | Coverage limits and take-up | Assistance can reduce amount insured without lowering policy count | Grant assignment and institutional mandate condition the estimate |
| Michel-Kerjan et al. (2012), NFIP | Policy tenure and lapses | Initial uptake may not survive later renewals | Historical U.S. portfolio, not a causal event-attention experiment |
A reproducible research agenda
A stringent future empirical analysis would predefine the target population using geocoded, eligible residential properties and an explicit rule for compulsory coverage. It would construct event time from high-resolution inundation or hazard data rather than relying only on declarations; assign news-market exposure separately from damage; record premium quotations and insurer availability; and use the full transaction history to classify first purchase, lapse, reactivation, coverage change and renewal. It would incorporate indicators for disaster assistance, lender mandates, property turnover, income shocks and contemporaneous premium reforms. Pretrends would be plotted; spatial spillovers and cluster dependence would be examined; and alternative postevent horizons would be reported with multiplicity and sensitivity considerations appropriate to the design.
Where experiments are feasible, a randomized communication trial could compare risk-map explanation, plain-language payout illustration, probabilistic hazard statement and neutral control, while tracking actual purchases after the usual cooling or waiting period and subsequent renewal. Ethical design would use accurate hazard information, avoid fear-based deception, and ensure that an offer is genuinely available to all randomized groups. Measured outcomes should distinguish comprehension, application, bound coverage and subsequent claims experience. The experiment would estimate the effect of information on transactions without assuming that more insurance automatically increases welfare.
Replication of a published finding requires using the same event population, denominator, contract universe and policy-year conventions. The open Choi data archive (2023) supplies supporting panels for the associated published analysis, although access to an archive alone is not a claim that this monograph has rerun its model. Authors should report publicly reproducible code, precise historical data versions, legal regimes, anonymization limits and the date on which all records were acquired.
Integrated interpretation and policy implications
The evidence supports three mutually compatible observations. First, catastrophic loss information affects behavior at the time of event, including among people without direct damage. Second, a considerable share of observed purchase change can arise from public requirements and insurer-market arrangements. Third, the persistence of change depends on what the event taught, whether protection remains affordable, and whether a contract can be maintained. These observations specify a bounded explanatory account grounded in existing research, not a new theoretical doctrine.
Risk communication can be improved by separating the probability of an event from the consequences to a particular property and from the terms governing indemnification. Authorities should document flood risks beyond a mapped high-risk boundary and communicate insurance obligations without implying that being outside the boundary means no physical hazard. Lenders and aid programs should specify the durability and scope of any required insurance. Insurers should explain waiting periods, exclusions, deductibles, premium changes and renewal steps in a form households can use before an event. Regulators evaluating success should report verified coverage continuity and adequately insured value, rather than a headline number of policies sold after a disaster.
The international EIOPA dashboard illustrates why the underlying public objective is the protection gap: insured and uninsured losses can evolve separately from the public salience of one event. Better event awareness can reduce this gap only when it translates into suitable, effective, sustainable coverage. Market access, pricing and public financial protection can be more binding than communication for some populations. A well-supported policy package therefore assesses information, contract design and capacity together, while retaining separate causal accounts of each mechanism.
Conclusion
Disaster salience and insurance uptake are connected through a sequence of information, institutional response and actual contract decisions. Catastrophes can make rare hazards cognitively accessible and reveal previously underestimated physical exposure. Media markets extend this information beyond the site of damage. Yet the effect measured in insurance data is the product of belief revision, attention, public rules, financial capacity, insurance offers and renewal practices operating at the same time.
The most credible U.S. field evidence documents a postevent increase that often weakens after a few years. Gallagher (2014) provides an influential comparison of direct and media-market exposure. Kousky (2017) demonstrates why disaster-assistance requirements must be deducted conceptually from claims about spontaneous demand. Choi, Diffenbaugh, and Burke (2024) confirm a temporary average response in voluntarily insured settings, while Petkov and Ortega (2025) establish an important qualification: policy attachment may persist where a flood reveals risk beyond the high-risk map boundary. Cai and Song (2017) show experimentally that simulated experience and information about expected insurance payoffs can each affect actual purchasing within a Chinese agricultural setting. These findings form a consistent account only when their outcomes, institutional conditions and population denominators are kept distinct.
The central analytical conclusion is that a catastrophe-driven policy sale is an observable transaction, whereas salience is a proposed informational or cognitive mechanism. Establishing the latter requires a design that separates event exposure from relief requirements, income and premium changes, supplier reactions and other shared shocks. Evaluating protection requires a still longer horizon: effective commencement, adequate limits, continuation and claims performance. This distinction prevents temporary heightened attention from being mistaken for durable risk transfer and prevents a household's budget constraint from being relabeled a failure of risk perception.
For the encyclopedia, the enduring lesson is methodological. Disaster salience belongs in insurance science as an empirical subject with identifiable behavioral channels, precise transaction outcomes and demonstrable institutional interactions. Scientific interpretation is strongest when each measured effect is attached to its original event, market, time horizon and causal design. A catastrophic event can initiate insurance learning, but sustained protection depends on the contractual, financial and institutional environment that allows the learning to remain effective.
References
Atreya, A., Ferreira, S., & Kriesel, W. (2013). Forgetting the flood? An analysis of the flood risk discount over time. Land Economics, 89(4), 577–596. https://doi.org/10.3368/le.89.4.577
Cai, J., & Song, C. (2017). Do disaster experience and knowledge affect insurance take-up decisions? Journal of Development Economics, 124, 83–94. https://www.sciencedirect.com/science/article/pii/S0304387816300657
Choi, J. (2023). The effect of flood exposure on insurance adoption among US households [Data set]. Zenodo. https://zenodo.org/records/8306902
Choi, J., Diffenbaugh, N. S., & Burke, M. (2024). The effect of flood exposure on insurance adoption among US households. Earth's Future, 12, e2023EF004110. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023EF004110
Endendijk, T., Rodriguez Castro, D., Dillenardt, L., Guntu, R. K., Botzen, W. J. W., de Moel, H., Thieken, A. H., Kreibich, H., Dewals, B. J., & Aerts, J. C. J. H. (2025). Flood experience and access to insurance contribute to differences in homeowners’ post-disaster adaptation in a cross-border region of Western Europe. Communications Earth & Environment, 6, Article 426. https://www.nature.com/articles/s43247-025-02385-z
European Insurance and Occupational Pensions Authority. (2025). Dashboard on insurance protection gap for natural catastrophes [Dashboard; updated December 5, 2025]. https://www.eiopa.europa.eu/tools-and-data/dashboard-insurance-protection-gap-natural-catastrophes_en
Federal Emergency Management Agency. (2024). National Flood Insurance Program flood insurance manual (April 2024). https://www.fema.gov/sites/default/files/documents/fema_nfip_flood-insurance-manual_042024.pdf
Federal Emergency Management Agency. (n.d.). Buy a flood insurance policy [National Flood Insurance Program consumer guidance]. https://www.floodsmart.gov/get-insured/buy-a-policy
Federal Emergency Management Agency. (n.d.). Frequently asked questions about NFIP data [Data documentation]. https://agents.floodsmart.gov/flood-maps-and-data/faqs-about-nfip-data
Federal Emergency Management Agency. (n.d.). FIMA NFIP Redacted Policies (OpenFEMA) [Data catalog entry]. Data.gov. https://catalog.data.gov/dataset/fima-nfip-redacted-policies-openfema
Federal Emergency Management Agency. (2021). They have questions, you need answers: Helping your clients understand Risk Rating 2.0 [Program guidance]. https://agents.floodsmart.gov/articles/they-have-questions-you-need-answers-helping-your-clients-understand-risk-rating-20
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://www.aeaweb.org/articles?id=10.1257%2Fapp.6.3.206
Kousky, C. (2010). Learning from extreme events: Risk perceptions after the flood. Land Economics, 86(3), 395–422. https://www.rff.org/publications/journal-articles/learning-from-extreme-events-risk-perceptions-after-the-flood/
Kousky, C. (2011). Understanding the demand for flood insurance. Natural Hazards Review, 12(2), 96–110. https://ascelibrary.org/doi/10.1061/%28ASCE%29NH.1527-6996.0000025
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
Kousky, C., Michel-Kerjan, E. O., & Raschky, P. A. (2018). Does federal disaster assistance crowd out flood insurance? Journal of Environmental Economics and Management, 87, 150–164. https://www.sciencedirect.com/science/article/pii/S0095069617303479
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
Michel-Kerjan, E., Lemoyne de Forges, S., & Kunreuther, H. (2012). Policy tenure under the U.S. National Flood Insurance Program (NFIP). Risk Analysis, 32(4), 644–658. https://onlinelibrary.wiley.com/doi/10.1111/j.1539-6924.2011.01671.x
Osberghaus, D. (2017). The effect of flood experience on household mitigation—Evidence from longitudinal and insurance data. Global Environmental Change, 43, 126–136. https://www.sciencedirect.com/science/article/pii/S0959378016302394
Petkov, I., & Ortega, F. (2025). Learning from experience: Flooding and insurance take-up in the flood zone and its periphery. Journal of Risk and Insurance, 92(2), 312–356. https://onlinelibrary.wiley.com/doi/10.1111/jori.70002
Robinson, P. J., & Botzen, W. J. W. (2019). Determinants of probability neglect and risk attitudes for disaster risk: An online experimental study of flood insurance demand among homeowners. Risk Analysis, 39(11), 2514–2527. https://pubmed.ncbi.nlm.nih.gov/31245870/
Shao, J., Hoshino, A., & Nakaide, S. (2022). How do floods affect insurance demand? Evidence from flood insurance take-up in Japan. International Journal of Disaster Risk Reduction, 83, 103424. https://www.sciencedirect.com/science/article/pii/S2212420922006434
Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5(2), 207–232. https://www.sciencedirect.com/science/article/pii/0010028573900339