Abstract
We combine household-level data on the choice to purchase flood insurance with experiment-based risk preference data and subjective risk perception data. The sample covers a wide geographic area (the entire U.S. Gulf Coast and Florida’s Atlantic Coast) and includes individuals exposed to varying levels of risk. This work represents one of very few analyses to do so. Results indicate that our experiment-based measure of risk aversion over the loss domain positively and significantly correlates with the decision to purchase a flood policy, as do perceived expectations of hurricane damage, eligibility for disaster assistance, and credibility of insurance providers. (JEL D81, R22)
I. Introduction
The National Flood Insurance Program (NFIP) was created by the U.S. Congress in 1968 to provide indemnity for flood hazard. At the time, there was widespread reluctance on the part of private insurers to issue policies for flood peril due to adverse selection, the catastrophic nature of flood hazard, and government’s tendency to provide disaster assistance in the wake of floods (Anderson 1974). In response to low participation rates (Kunreuther 1978), subsequent legislation required flood insurance for homeowners with a mortgage contract from a federally backed or regulated lender and whose home was located in a Special Flood Hazard Area (SFHA, also known as the 100-year floodplain, or area that faces a 1% chance of flooding each year according to the Flood Insurance Rate Maps [FIRMs] produced by the Federal Emergency Management Agency [FEMA]). Due to apparent lack of enforcement, mandatory purchase provisions were later strengthened, and a new program, the Community Ratings System (CRS),1 was created to provide incentives for local hazard mitigation projects.
Despite these efforts, NFIP participation rates remain relatively low,2 and participation is still not universal even among properties where mandatory purchase provisions apply. An obvious question, then, is what factors determine the decision to purchase a flood policy?
The standard model of behavior under risk assumes that individuals harbor subjective risk probabilities that are equivalent to those derived from objective calculations of risk. However, individuals often underestimate the risk (Camerer and Kunreuther 1989; Chivers and Flores 2002; Kunreuther 1984, 1996, 2006). If individuals underestimate the probability or potential magnitude of loss, insurance may appear unattractive, even at subsidized rates. Apart from risk perceptions, empirical work in other settings shows that risk preferences affect behavior under risk as well (Holt and Laury 2002; Kachelmeier and Shehata 1992; Lusk and Coble 2005). And although several papers make the case for both risk perceptions and preferences affecting flood insurance decisions (Kunreuther 1984, 1996, 2006; Braun and Muermann 2004; Kunreuther and Pauly 2004, 2006; Kousky, Luttmer, and Zeckhauser 2006; Smith et al. 2006; Kousky 2010a; Michel-Kerjan and Kousky 2010), there exists very little empirical work.
This research seeks to fill these gaps in the literature by examining participation in the NFIP along the Gulf of Mexico and Florida’s Atlantic Coast with survey data that includes experimentally derived risk preference information, subjective risk perception measures, and information on expectations of disaster assistance and credibility of insurance providers. Results indicate that the experimentbased measure of risk aversion (over the loss domain) positively and significantly correlates with the decision to purchase a flood policy, as do perceived expectations of hurricane damage, eligibility for disaster assistance, and credibility of insurance providers. Overall, this work represents one of very few analyses of flood insurance over a wide geographic area and including individuals exposed to varying risk levels using household-level data that accounts for both risk preference and subjective risk perceptions, as well as objective risk exposure.
II. Literature Review
Several studies have attempted to identify these factors using aggregate data. Using state-level data, Browne and Hoyt (2000) find that demand is decreasing in price and increasing in income, previous year’s flood damages, and, surprisingly, the amount of past federal disaster assistance (in possible contradiction to the charity hazard hypothesis discussed below). They also find evidence that demand is inversely related to the number of mortgages in the state (counter to what might be expected under federal law). Using censustract-level data, Kousky (2010b) finds higher market penetration in census tracts with more land in the 100-year and 500-year floodplains; lower market penetration in areas with levee protection and along major rivers; and that coverage is increasing in average home value, median income, and along major rivers. Dixon et al. (2006) use a stratified sample of communities to examine market penetration at the national level, but do not explore individual-level characteristics. Focusing on Florida, which represents about 40% of the total NFIP policies-in-force, Michel-Kerjan and Kousky (2010) analyze county-level and individuallevel flood insurance policy data to provide information on the characteristics of buyers and flood insurance contracts and examine patterns in claims and insurance costs.
Further information on factors affecting insurance choice, however, can generally be obtained only via household-level surveying. Work of this kind that has focused on flood insurance includes that by Kriesel and Landry (2004) and Landry and Jahan-Parvar (2011). Kriesel and Landry (2004) find that demand for flood insurance is positively influenced by the presence of a mortgage contract where flood insurance may be a requirement and the presence of community-level erosion protection projects, and negatively influenced by distance to the shoreline and the historical interval between hurricane events. Landry and Jahan-Parvar (2011) expand on their work and find that insurance coverage is also positively influenced by policy subsidies and objective measures of flood risk.
Although the above papers add to the body of knowledge regarding what factors influence flood insurance decisions, they do not address the role of subjective risk information. Rees and Wambach (2008) show that insurance demand depends upon the perceived likelihood of loss and the size of the conditional loss. A critical assumption of the standard insurance model (Mossin 1968; Smith 1968) is that these perceptions are equivalent to those derived from objective calculations of risk (likely to be used by insurance companies and government actuaries). If they are not equivalent, however, then behavior may deviate from theoretical predictions, and there is empirical evidence from a variety of insurance settings to support this claim: crop insurance (Shaik et al. 2008), earthquake insurance (Naoi, Seko, and Sumita 2010), homeowner’s insurance under wildfire risk (Talberth et al. 2006), as well as in an experimental setting (McClelland, Schulze, and Coursey 1993).
The only explicit empirical evidence of the role of risk information in flood insurance decisions is an article published 34 years ago by Baumann and Sims (1978). They find higher insurance uptake among homeowners who had suffered previous damage from a flood, had perceptions of increased severity of flood damage, and had higher education and income. Their analysis, however, consists of a series of individual chi-square tests (not multiple regression), so it is not clear whether the significant effects found would hold when controlling for other factors.
There are a handful of related empirical studies that provide additional guidance. Baker et al. (2009) analyze home purchase decisions under flood risk, focusing on the effect of perceived hurricane risk. Chivers and Flores (2002), also focusing on home purchase decisions, account for the role of perceived flood risk and perceived flood insurance costs. Botzen, Aerts, and van den Bergh (2009) analyze willingness of Dutch homeowners to carry out a variety of flood mitigation activities (not including purchasing flood insurance), and find a lower likelihood of mitigation among those that perceive less risk associated with floods and climate change. Sims and Bauman (1987) analyze the role of subjective risk information on flood mitigation activity, though it is not stated which activities are included in their analysis.
Other studies have used proxies for changes in flood risk information, such as housing price data (MacDonald et al. 1990; Harrison, Smersh, and Schwartz 2001; Dehring 2006; Bin et al. 2008; Bin, Kruse, and Landry 2008), deductible choice (Cohen and Einav 2007), flood disclosure laws (Troy and Romm 2004; Pope 2008), media coverage (Miles and Morse 2007), or widespread flooding events (Bin and Polasky 2004; Hallstrom and Smith 2005; Carbone, Hallstrom, and Smith 2006). While these studies provide some insight into general perceptions of the likelihood of loss, they make no direct connection between individual risk perception and behavior.
Although much of the literature focuses on perceptions regarding the risky event itself (e.g., a landfalling hurricane) and magnitude of damage, other related perceptions may play a significant role. Charity hazard, where an individual underinsures in anticipation of postdisaster aid, offers an additional explanation as to why individuals may fail to obtain insurance (Kaplow 1991; Raschky and Weck-Hannemann 2007). Raschky et al. (2010) provide some empirical evidence of charity hazard specific to flood insurance in Europe, and Botzen, Aerts, and van den Bergh (2009) find a lower likelihood of flood mitigation activities (not including insurance) in the presence of government disaster aid, as well as lower likelihood among those individuals that perceive disaster aid as a valid government responsibility. Additionally, perceived credibility problems of insurance providers and uncertainty about the likelihood of insurance payoffs, especially in the wake of Hurricanes Katrina and Rita, could also play a role in insurance purchase decisions (Kunzelman 2007; Sisco 2009). To our knowledge, the role of insurer credibility has not been empirically measured.
Finally, apart from the role of risk perceptions is the role of risk preferences. Several of the aforementioned papers discuss the role of risk preferences, and empirical work in other settings provides evidence regarding its role in behavior under risk (Holt and Laury 2002; Kachelmeier and Shehata 1992; Lusk and Coble 2005). To our knowledge, however, the role of risk preferences in flood insurance decisions has not been empirically examined. In sum, although the above papers provide some insight regarding what factors influence flood insurance decisions, they offer very little empirical evidence on the role of risk perceptions, and to our knowledge, none have addressed the role of risk preferences or insurer credibility empirically. This paper seeks to fill these gaps.
III. Conceptual Model
We assume individuals maximize their subjective expectation of utility. Subjective expectation of utility for individual i is given by
[1]where j = P, NP (representing the purchase and nonpurchase of flood insurance, respectively); represents the subjective perception of the likelihood of flooding; represents the subjective conditional expected loss in the event of a flood; λ represents the perceived likelihood of insurance payoff; H is expected disaster assistance, reflecting potential charity hazard; w is household wealth level; π(p)is the price of insurance, which depends upon observable objective risk factors, p; α indexes level of risk aversion, and m is a binary indicator for whether flood insurance has been deemed “mandatory” for individual i. There is evidence that properties in the SFHA subject to mandatory provisions sometimes forego flood insurance coverage (and we find this in our data). Our decision model in [1] is intended to provide a framework for understanding individual choice; while adverse selection (based on subjective risk perception measures) and moral hazard (based on expectations of disaster assistance) could be inferred from results, we do not explicitly model these phenomena within a market equilibrium framework. We also do not consider the choice of level of coverage in the event of purchase, but assume that this is chosen in a way that the individual perceives as “optimal” (which may or may not accord with first-order conditions of a conventional expected utility optimization problem). We have no reliable information on this endogenous variable, and it is excluded from the analysis.
For the NFIP, insurance premiums are set at the national level and reflect a limited set of observable risk factors, which we denote p. NFIP rates vary as to whether structures were built before the publication of FIRMs in their community, and owners of pre-FIRM structures pay explicitly subsidized insurance rates. Flood insurance premiums can also vary by flood zone, number of stories, structure type, and (for post-FIRM structures) elevation above base flood elevation (the estimated height of a 100-year flood). Since the set of factors that determines premiums is prescribed and limited, identifying the price elasticity of demand can be difficult because there is little variation in price independent of risk.
While some have used changes in rate schedules over time (U.S. General Accounting Office 1983) or exogenous variation due to explicit subsidization for pre-FIRM and other grandfathered structures (Landry and Jahan-Parvar 2011) to try to identify price elasticity, the correlation of price and risk make estimation of true price effects problematic (Bin et al. 2008). For example, Kousky (2010b) finds evidence of a counterintuitive positive price effect in her aggregate flood insurance data for Saint Louis, Missouri. Moreover, in order to produce accurate calculations of flood insurance premium, one must know very detailed structural information (such as elevation above base flood elevation) that is often not available.
This being the case, rather than including a proximate price variable that may be construed as a single measure of overall risk exposure, our empirical model includes observable risk determinants upon which premiums are based. We interpret the parameters for these variables as representing the risk effect net of expected flood insurance indemnity under the NFIP. For simplicity, we assume that the subjective expectation of utility can be approximated with a linear function:
[2]where β is a vector of coefficients for exogenous variables xi (which includes the arguments in [1] and other individual and spatial attributes), and εi is a random disturbance. Our simple behavioral model implies that flood insurance is purchased if . We assume the individual disturbances are independently and identically distributed normal with zero mean: ε~N(0,σ). Following Greene (2002), the log likelihood, ln L, is given by
[3]where yi = 1 is observed if , and yi = 0 otherwise; Φ(·) is the cumulative normal distribution function.
IV. Survey Design
A survey was designed for the primary purpose of collecting data on revealed behavior under environmental risk, specifically focusing on purchase of flood insurance. We collected individual-level information on covariates that the literature suggests should be important determinants of flood insurance purchase. These included (1) the expected number of future storms of a specific strength to strike an area, (2) the expected structural damage from a storm of a specified strength, (3) experience with flood hazard, (4) the perceived reliability of insurance providers (reflecting the likelihood of payout on insurance claims submitted), (5) expectations of disaster assistance, (6) measures of risk preference, and (7) wealth. In compiling this information, we intend to examine the influence of each of these factors on flood insurance demand in a multiple regression setting where we can control for various factors simultaneously.
Figure 1 presents the survey text for questions used to capture risk perceptions, the perceived reliability of insurance providers, and expectations of disaster assistance. We adhere to the standard “probability × consequence” definition of risk, using perceived frequency of landfalling hurricane and expected conditional storm-related flood damages as proxies for flood risk. Although floods can originate from others sources, we maintain that perceived hurricane frequency is a reasonable proxy for the particular area under study. Empirical evidence to support this claim includes that of Lecce (2000), who finds that for the Florida region—which makes up more than half of our sample—the majority of floods are in late summer and early fall, in other words, during hurricane season, and that flooding due to heavy precipitation is generally the result of tropical storms and air-mass thunderstorms. Additionally, between 1996 and 2005, approximately half of the flood-related fatalities along the Gulf of Mexico and Florida Atlantic Coast were related to hurricanes, tropical storms, downgraded tropical storms, or remnants of tropical systems (Ashley and Ashley 2008). Finally, Michel-Kerjan, Lemoyne de Forges, and Kunreuther (2012) report a “Katrina effect” in Louisiana and Mississippi, where the number of flood policies in force increased in 2006 at three to four times the growth rate observed in previous years, due to the number of historical flood claims from Hurricanes Katrina, Rita, and Wilma. Nevertheless, subjects’ expectations of damages conditional on a Category 3 storm will almost certainly include both wind and water damage, and thus what we have is not a direct measure of flood-related risk, strictly speaking.
Given possible complications with eliciting probabilities from respondents, and the difficulties incurred by respondents during pretesting with the concept of hurricane return interval, we asked respondents to report the expected number of hurricanes of a particular strength to directly strike their area over the next 50 years. We chose to focus on a Category 3 storm, as this hurricane strength seems to be a threshold for widespread flooding and wind damage. In order to measure subjects’ expectation of damage conditional on a Category 3 storm directly striking their community, we asked them to estimate the dollar magnitude of damage as a proportion of the total value of their housing structure. Information on flood experience includes the number of flood events experienced by the respondent and the number of years living on the Gulf or Atlantic Coast.
We employed ordered categorical response measures to gauge perceptions of credibility of insurance providers and likelihood of eligibility for disaster aid. Perceptions related to the reliability of insurance providers focused on respondents’ confidence in insurance providers’ intentions to pay storm damage claims in full. Because humanitarian disaster aid is typically available after most major storm events, we chose to focus our analysis of disaster aid on respondents’ perception of whether they would be eligible for disaster aid that compensated for storm damage to residential structures. This focus is appropriate for a test of the charity hazard hypothesis, as disaster aid to cover residential damage can clearly serve as a substitute for formal flood insurance. Both of these qualitative measures were transformed into binary data, where those with a score of at least 3 on the 1–5 scale were grouped together versus those with scores of less than 3. Thus, the dummy variable for insurance payout identifies those respondents that have at least some confidence that insurance companies will pay the full amount of claims, and the dummy variable for disaster assistance identifies those that perceive it as at least somewhat likely that they will be eligible for disaster assistance to cover flood damage to housing.
A measure of risk preference was obtained via a real-money experiment, following the approach of Holt and Laury (2002). At the end of the survey, each respondent was asked to make five pairwise choices between a low-variance and a high-variance risk of loss, and five pairwise choices between a low-variance and a high-variance chance of gain. Figure 2 shows some of the text from this section of the survey, as well as an example choice. Dollar values remained constant throughout the pairs; probabilities were set at 0.1/0.9 (as in the example shown), 0.3/0.7, 0.5/0.5, 0.7/0.3, and 0.9/0.1, respectively. Respondents were given $10 in start-up funds such that their participation in the loss gambles did not affect already-earned money that was used as an incentive for completing the survey. To ensure incentive compatibility in the risk experiments, respondents were told that one choice from each of the gain and loss sets would be randomly chosen, and the combination of individual choice and random outcome from the randomly selected items would be used to determine actual payoffs. Respondents were notified of their total payoff immediately following completion of the survey. We hypothesize that the probability of purchasing a flood policy is increasing in the level of risk aversion, and we examine risk preferences over both the gain and loss domains.
Proxies for wealth include the income category for the household and a dichotomous variable indicating whether the household owned additional, noncoastal real estate with value in excess of $100,000. In order to account for mandatory requirements for mortgaged homes, we combine information on mortgage status (collected in the survey) and information on flood zone, derived from the survey and FIRMs. Not all residents knew their flood zone. We addressed this by using resident street addresses to collect flood zone data from publically available digital FIRMs (DFIRMs), and from older scanned maps for non-DFIRM addresses. Where conflicts between the two sources arose, we gave precedence to that reported by the respondent.3
V. Survey Administration
We identified a sample frame consisting of all coastal counties in Alabama, Florida, Mississippi, Louisiana, and Texas, plus some additional noncoastal counties with a substantial level of land located in flood zones. This resulted in a total of 96 counties in the sample frame. We contracted with Knowledge Networks to administer the survey to subjects in their Knowledge Panel® who were 18 years or older and homeowners. Knowledge Networks boasts of having recruited the first (and only) online research panel that is representative of the entire U.S. population. Panel members are randomly recruited by probability-based sampling that covers both the online and offline populations in the United States. Households are provided with Internet access and hardware if needed. Knowledge Networks identified 1,536 panelists meeting our criteria, and of these, 1,070 agreed to participate.
Prior to administration of the survey, we conducted two focus groups of coastal residents in the New Orleans and coastal Mississippi and Alabama areas, and Knowledge Networks conducted a pretest of 25 panelists for further refinement prior to full implementation. In addition to the initial invitation, Knowledge Networks sent e-mail reminders to nonresponders on days 3 and 16 of the 21day field period. Participants were eligible to win an in-kind monetary inducement through a monthly Knowledge Networks sweepstakes and, for this particular survey, were offered an additional $5 cash incentive, as well as the chance to earn additional $10 cash (on average) by completing the Holt-Laury experimental questions at the end of the survey. We also requested panelists to allow us access to their street address for the purpose of collecting geolocation data. Not all panelists agreed; of the 1,070 original participants, 859 consented and completed the survey. Twenty, however, gave post office or otherwise unusable addresses and were dropped, and two additional observations were dropped because flood zone could not be determined (leaving 837). Item nonresponse on remaining survey questions used in the analysis resulted in dropping an additional 32 observations, 8 of which were on responses to the flood insurance (dependent variable) question, resulting in a total of 805 usable completes.
Figure 3 shows a map of respondents included in the sample. Four hundred ninety-five respondents were from Florida, 182 from Texas, 96 from Louisiana, and 32 from Alabama/Mississippi. The survey contained 41 questions (including experimental questions) and took about 20 minutes to complete. General demographic data were precollected by Knowledge Networks and provided to us. Table 1 reports a comparison of selected population and sample demographics. With the exception of slight overrepresentation of white, educated, and Internet-accessed respondents, the sample was representative of the overall population within the 96 coastal counties.
Additional data were amended to the survey dataset using respondent street addresses. This included home construction year, obtained from publicly accessible online county property tax assessor databases, and community FIRM publication dates (both of which were used to determine pre-/post-FIRM status); CRS score for resident community; flood zone (taken from publicly accessible FEMA DFIRM database for DFIRM communities and from scanned maps for non-DFIRM communities); and distance of home from the nearest shoreline.
VI. Results
Table 2 reports proportions of respondents with a flood insurance policy by state and SFHA status. Flood insurance uptake is much higher in SFHA zones (78%) versus non-SFHA zones (28%), where flood insurance is not required of mortgaged properties. In both SFHA and non-SFHA zones, Texas had the highest proportion of policyholders, followed by Louisiana, with Florida third (fourth), and Alabama/Mississippi fourth (third) among respondents in SFHA zones (non-SFHA zones). Table 3 presents variable descriptions, means, standard deviations, ranges, and hypothesized coefficient signs for variables included in the regression analysis. Thirty-six percent of the respondents held flood insurance, and 64% of the properties were currently mortgaged. Forty-two percent of the structures were built before the publication of FIRMs, and 15% were located in the SFHA.
Focusing on covariates of particular interest for our analysis, the average respondent expected 6.8 Category 3 storms to strike their area over the next 50 years, with a range of 0 to 90. Average expected damage from direct strike of a Category 3 storm was 33.8% of structure value, with a range of 0 to 100%. The average respondent had experienced 0.09 flood events in the past, with a range of 0 to 7, and has lived on the Gulf or Florida Atlantic coast for 28.5 years. The average level of perceived eligibility for disaster assistance to cover damage to housing is 2.72, or slightly likely, and the average level of confidence that insurance companies would make payouts that cover the full amount of claims (insurer credibility) is 3.02, or somewhat confident.
Following the method of Holt and Laury (2002), risk aversion was measured over the domain of both gain and loss gambles; the average number of low-risk choices in the gain (loss) domain was 2.96 (2.93) and ranged from 0 to 5. Thus, on average, our sample was slightly risk averse over both domains. Average household income category was 12.18 (ranging from 1 to 19), which translates into the $50,000-$59,999 range, and 6% of house-holds owned other noncoastal property valued in excess of $100,000; we use the latter as a proxy for wealth held outside of the coastal zone.
Table 4 presents the probit regression results.4 Several sets of variables were potential sources of multicollinearity (the gain and loss domain risk aversion variables; income, kids, and mortgage; income and mobile home; and expected damage and mobile home). However, we found no high correlations among these variables, and individual parameter results were not significantly affected by the dropping or adding of other variables. The model captures the effect of living in an SFHA zone with a dummy variable. We also interact the SFHA dummy variable with a mortgage contract dummy variable in order to measure the impact of the mandatory flood insurance purchase provisions. Results indicate that residence in the SFHA (100-year flood zone) has the second-largest impact on probability of holding flood insurance, and this effect is particularly strong for houses that are mortgaged; controlling for other factors, the average nonmortgaged household in the SFHA is 23.3% more likely to hold flood insurance in comparison with non-SFHA households. Mortgaged households in the SFHA, however, are 73.7% more likely to hold flood insurance relative to non-SFHA, nonmortgaged households. This is evidence that the mandatory purchase provisions stipulated under federal law have a significant impact on individual household participation in the NFIP, although coverage for mortgaged properties in the SFHA is not universal; 76% of SFHA mortgaged properties are covered by flood insurance in our dataset.
Individual perceptions of risk exhibit mixed results: the probability of holding flood insurance is greater for those with higher expected damages, with a marginal effect of 1.6% for a 1% increase in expected damages (as a proportion of structure value); expected number of future storms has a positive but insignificant effect. We estimate that additional past experience with floods (i.e., an increase of one past flooding event) increases the probability of holding flood insurance by 11.4%. This result is consistent with previous literature (Baumann and Sims 1978; Kunreuther 1978; Kriesel and Landry 2004; Carbone, Hallstrom, and Smith 2006) and suggests that experience with hazard events heightens sensitivity to risk. Our other indicator of flooding and hurricane experience, length of coastal residence, however, is statistically insignificant. Confidence in insurers’ willingness and ability to pay claims in their entirety has a positive and statistically significant effect on the probability of holding a policy, with a marginal effect of 6.7%. Respondents who perceive eligibility for disaster assistance to be at least somewhat likely were also more likely to purchase a flood policy, with a marginal effect of 7.8%. Thus, we find no evidence of charity hazard in our data; on the contrary, those that express greater expectations of disaster assistance are more likely to purchase flood insurance.
Risk aversion has a positive impact on the likelihood of holding a policy, but the effect is significant only for the loss domain. Those that choose an additional low-risk lottery over the loss domain in the Holt-Laury instrument are 2.7% more likely to hold flood insurance. Consistent with previous research (Baumann and Sims 1978; Browne and Hoyt 2000; Kunreuther 2006; Landry and Jahan-Parvar 2011), we find evidence of a positive income effect, with a one unit increase in income category (a $10,000 increase, on average) raising the probability of holding flood insurance by 1.6%. Our proxy for wealth held outside of the flood zone—the ownership of other noncoastal property with a value in excess of $100,000—was not statistically significant. Wealth effects on insurance demand are difficult to estimate, and this item remains a subject for future research. Lastly, consistent with the findings of Kriesel and Landry (2004), the likelihood of holding a policy decreases with distance from the shoreline.
The CRS of the NFIP credits voluntary flood hazard mitigation activities undertaken by local communities and provides discounts on flood insurance premiums for citizens in the jurisdictions in which flood hazard mitigation has taken place. Overall, our results indicate that market penetration is greater in communities that engage in greater flood hazard mitigation (as recognized by the CRS); a one-unit improvement in community rating (represented by a decrease in score), increases the likelihood of holding a policy by 3.0%. This result, however, is different for the state of Florida. Florida residency has the largest impact on the probability of holding a flood policy, reducing the probability of holding a policy by 50.4%. However, a statistically significant interaction effect with Florida residency and CRS score is found, which indicates that greater levels of flood hazard mitigation decrease NFIP participation in Florida; a one unit decrease in CRS score (indicating greater flood hazard mitigation) reduces the likelihood of a Florida resident holding a policy by 5.0%. Thus, we have evidence that within Florida, mitigation measures taken at the community level, reflected in improved CRS scores, may act as substitutes for individual flood insurance purchase. Other significant variables include housing type and race. Living in a mobile home reduced the likelihood of holding a policy by 14.3%, whereas Hispanics were found to be 9.2% more likely to hold a flood policy.
Although not reported here, we also estimated models over several subsamples of data, including SFHA versus non-SFHA parcels, nearshore versus distant (split at the median distance from the shoreline) locations, and high versus low CRS scores (split at a score of 7 or less for low and greater than 7 for high), to test for any structural differences across these distinct risk groups. The reasoning behind inclusion of the nearshore versus distant subsamples is that nearshore respondents should be more susceptible to flooding due to storm surge, whereas distant ones should be more susceptible to flooding due to rainfall. These subsample results yield subtle differences only. The effect of past flood events are significant among non-SFHA respondents, but not among SFHA respondents; and race (Hispanic) is significant among non-SFHA, but not among SFHA, respondents. Additionally, the effect of perceived damage from a storm is larger for SFHA homes than for non-SFHA homes. This may be an indication that our perceived risk measure captures both water and wind effects, and that non-SFHA respondents may focus less on flood damage relative to wind damage compared to SFHA respondents. Regarding the distance subsamples, the effects of perceived future damage, past flood events, race (Hispanic), and mobile home are significant among the distant subsample, but not among the nearshore subsample. The effect of eligibility for disaster assistance is significant among the nearshore subsample, but not among the distant subsample.
Regarding the CRS subsamples, several variables are significant among respondents in better-prepared communities (i.e., CRS ≤ 7), but not among the less-prepared ones (i.e., CRS >7). However, the marginal effects are consistent across the two subsamples. One variable, years on coast, took opposite signs, but marginal effects are very small in both cases. Overall, these models provide slightly more detailed results regarding which factors may be influencing some of the results reported in Table 4; however, the differences are minor, and we find the results reported to be consistent. These additional results are available from the authors upon request.
VII. Discussion
Our dataset describes flood risk perceptions, insurance and disaster assistance expectations, flood experience, risk preferences, income, wealth, and flood insurance holdings for a predominantly representative sample of Gulf of Mexico and Florida Atlantic Coast residents. This is the first dataset, to our knowledge, that compiles individual survey and geolocation data for a broad geographic region, in order to examine microeconomic factors influencing the choice to purchase flood insurance. Our results provide empirical evidence to support the propositions of expected utility and other theories of decision making under risk that subjective probability and magnitude of loss are important determinants of behavior.
Our data indicate that, after controlling for objective measures of risk (flood zone, housing type), risk preferences (measured as choice over lotteries in the Holt-Laury instrument), subjective perceptions (indicated by expected storm damage), past experience with flooding, and the existence of flood hazard mitigation significantly increase the likelihood of holding flood insurance. In other words, these additional factors “add on” to the likelihood of holding a policy over and above the effect of objective risk. As might be expected, we find a significant and positive relationship between past flood experience and perceived storm risk. Nonetheless, we find no interaction effect within the insurance choice model. Thus, the effects appear to be additive as modeled.
As hypothesized, experimentally derived measures of risk aversion have a positive and significant effect on the likelihood of holding flood insurance, indicating that individuals that display risk-averse behavior in an experimental context also manifest relatively more risk-averse behavior in their revealed choice. As with previous research (Lusk and Coble 2005), this provides evidence of the importance of risk preferences in decision making under uncertainty and the ability of the Holt-Laury instrument to provide useful information for measuring risk preference. Moreover, the significant effect of the loss lottery choices vis-à-vis the insignificant effect of the gain lottery choices suggests that individuals may exhibit different risk preferences over losses and gains, as posited by prospect theory (Kahneman and Tversky 1979; Laury and Holt 2005), though we include no explicit control for a reference point. This result highlights that the context of gain or loss matters in combining experimental and empirical data.
We find that individual expectations of future Category 3 hurricane strikes over the next 50 years have a statistically insignificant effect on the probability of holding flood insurance, but expected damage from Category 3 hurricanes has a positive and significant effect. This suggests that individual expectations of flood loss have a significant impact on the choice of purchasing flood insurance, though to the extent that unobserved risk factors (such as structure elevation) drive damage expectations and influence insurance price, our estimates could be biased. Since structures with greater elevation pay lower rates and should have less damage during flooding events, low expectations of damage could be correlated with lower risk and insurance price. Given that our model controls for objective/observable risk factors associated with the home, such as those factors used by FEMA to set rates, significance of the risk perception variable may suggest evidence of asymmetric information. If the homeowner’s perceptions accurately reflect the true risk, then this may also be evidence of adverse selection. On the other hand, the homeowner’s perceptions may be overblown, and the results simply reflect such misperceptions.
The significant effect of past experience is consistent with updating by Bayes’s law or behavioral phenomena such as availability bias (Tversky and Kahnemann 1973; Schwarz et al. 1991). Bayes’s law indicates that, for example, those with diffuse priors will adjust their subjective probability estimates based on new information such as the occurrence or nonoccurrence of a flood—as such, experience with flooding can increase subjective probability estimates leading to higher flood insurance demand. The availability heuristic posits that subjective estimates of probability reflect the availability of information or recall of specific events related to the uncertainty under consideration: increased experience with flooding may render available information that increases subjective probability of flooding and augments flood insurance demand. As both phenomena are consistent with our results, we can make no assessment as to whether probabilities are more likely to be updated by precise statistical procedures or whether they evolve by laws governing bounded rationality. The NFIP does account, somewhat, for previous flood experience by barring a Preferred Risk Policy (a policy with a substantial rate discount) to residents with multiple claims or Federal Disaster Relief payments, but this policy applies only to non-SFHA zones. As such, homes with multiple claims in SFHA zones suffer no such loss of benefits when multiple claims are made.
Regarding flood hazard mitigation activities, we find evidence of countervailing influences of two effects attributable to the CRS hazard mitigation programs: lower price of insurance (positive effect) and reduced likelihood or magnitude of loss (negative effect). In a study of factors affecting flood mitigation, Brody, Kang, and Bernhardt (2010) found that Florida communities have CRS scores two times lower than that of Texas, which they indicate as “evidence that Florida is far more prepared to mitigate floods than Texas” (179–80). Yet, our raw data indicate that flood insurance uptake rates are higher in Texas (see Table 2), and our model results indicate that flood insurance participation is lower in Florida. The regression results also indicate that flood mitigation activities that result in a lower CRS score are associated with greater levels of flood insurance purchase along the Gulf Coast, with the exception of Florida, where greater flood hazard mitigation is associated with lower levels of purchase. These results suggest that, despite the incentive structure of the CRS, which attempts to award lower insurance prices for mitigation activities that lower expected losses, mitigation may serve as a substitute for flood insurance, at least within the state of Florida.
Exploring this issue more deeply, Brody et al. (2009) find that Florida communities appear to disproportionately engage in incremental CRS point-earning activities—such as information provision, public outreach, and tightening of existing regulations, rather than those that require structural changes or largescale regulatory overhaul—as these types of projects are less expensive and more politically viable. Texas, on the other hand, appears to rely more heavily on structural approaches to flood mitigation (Brody, Kang, and Bernhardt 2010). Our results, then, suggest that the information provision activities encouraged by the CRS are not very effective at augmenting NFIP participation, whereas activities involving structural flood mitigation options are. It could also be that spatial indicators are capturing variations in geography and topology or differences in administration of flood insurance in Florida relative to other Gulf states.
Regarding expectations of disaster assistance, the charity hazard hypothesis states that such expectations can crowd out private demand for hazard insurance (Kaplow 1991; Browne and Hoyt 2000; Raschky and Weck-Hannemann 2007). The data indicate that perceived eligibility for disaster assistance has a statistically significant effect on the likelihood of holding a flood policy, but the estimated coefficient is positive (consistent with the state-level results of Browne and Hoyt 2000). In other words, we find evidence that expectations of eligibility for disaster aid are complementary to, rather than a substitute for, flood insurance. Policy makers should pay particular attention to this finding, because while charity hazard is widely hypothesized, our results suggest that the expectation of disaster assistance does not infringe, and may even increase, NFIP participation. We note, however, that expectations of eligibility are somewhat different than intentions to utilize disaster assistance as a sole source of support for recovery. Early efforts to increase individual participation in the NFIP sought to require flood insurance as a prerequisite for disaster aid, though this provision was never enforced. Quite the contrary, holding disaster insurance often precludes certain types of disaster aid. Thus, it is possible that our survey instrument is producing biased estimates of disaster assistance expectations if respondents exhibit concern over social acceptability. If those that actually rely on federal disaster assistance are ashamed to admit it, they may misrepresent their true beliefs in the survey. This hypothesis is consistent with the findings of Botzen, Aerts, and van den Bergh (2009) and Raschky et al. (2010), who find a lower likelihood of flood mitigation activities among individuals that perceive disaster aid as a valid government responsibility. Future research should focus on incentive compatibility and external validation of measures of expectations of disaster assistance in seeking to understand the influence of disaster aid on insurance and selfprotecting behavior.
Finally, our dataset is the first, to our knowledge, that accounts for the effect of perceived credibility of insurance providers on the choice to buy flood insurance at the individual level. We find that perceived insurer credibility has a positive influence on the probability of holding flood insurance. This effect could be a result of the wind-versus-water controversy and court battles between residents, states, and insurance companies that played out along the Gulf in the aftermath of Hurricanes Katrina and Rita (Kunzelman 2007; Sisco 2009). This result suggests that if the NFIP, other government entities, or private industry want to bolster demand for flood insurance, addressing concerns over settlement of claims in a timely and equitable manner, as well as enforcing credible regulations and tort law, could have fruitful consequences.
VIII. Conclusions
Prior to Hurricanes Katrina and Rita in 2005, the NFIP exhibited a cumulative deficit of $1.5 billion (2008 dollars) since its inception (Michel-Kerjan 2010). The lack of available reserves has led the NFIP to borrow over $19 billion from the U.S. Treasury to cover flood losses occurring between 2005 and 2008 (Michel-Kerjan and Kunreuther 2011). Crosssubsidies in the federal rate structure and high fees paid to private insurance intermediaries (amounting to approximately 40% of the collected premiums) are part of the reason for the program’s poor financial health (Michel-Kerjan 2010), but questions remain regarding the serviceability of FIRMs, problems of low market penetration and retention of policies, repetitive losses for some high-risk policies, and incentivizing individuals and communities to invest in flood hazard mitigation.
Our results indicate that objective measures of risk—such as presence in the SFHA or habitation of a mobile home—affect the probability of holding flood insurance. Location in the SFHA is one of the strongest indicators, and this effect is particularly strong for mortgaged properties (as would be expected under current regulations). Thus, it appears the mortgage requirements for flood insurance purchase in the coastal zone are more effective than previous results suggested (Kriesel and Landry 2004; Landry and Jahan-Parvar 2011). Still, coverage in the SFHA is far from universal, and the data indicate that the mortgage requirement is still circumvented by some households.
Further, our results indicate that systematic differences across households—such as perception of flood risk, risk preference with respect to monetary losses, past experience with flooding, and local hazard mitigation projects—help explain the variability in the choice to purchase flood insurance. Individuals that perceive greater flood damage and those that have experienced floods in the past are more likely to be covered. This suggests that, to the extent that risk perceptions can be influenced by information provision, market penetration for the NFIP could be augmented. This result supports some of the current efforts by FEMA (advertisements and promotional programs) to increase enrollment in the NFIP. Our results suggest that information that highlights potential magnitude of flood damage may be more effective than information on likelihood of flooding.
We find that flood hazard mitigation activities (as recognized by the CRS) generally increase market penetration for flood insurance, presumably due to premium discounts designed to address problems of moral hazard. For Florida, however, increased flood hazard mitigation credits are associated with lower flood insurance market penetration. One explanation could be that the reduction in risk associated with flood mitigation dampens demand for insurance (with an effect greater than that associated with increasing demand due to premium reductions). Brody et al. (2009), however, show that Florida communities focus most of their hazard mitigation activities on information provision and public outreach, which presumably would augment demand while doing less to lower community flood risk. The lower market penetration and chosen focus on information and outreach mitigation activities suggest that the state of Florida may be more vulnerable in its programmatic disposition within the NFIP. Other Gulf states, for example, Texas, have embraced structural flood mitigation approaches (e.g., acquisition and retrofitting) that should lower risk (Brody, Kang, and Bernhardt 2010), but also exhibit strong market penetration. This suggests that these other states may be better prepared to manage future flood hazards (at least to the extent that vulnerability is determined by CRS hazard mitigation activities and flood insurance market penetration).
Unfortunately, our results do not address other policy recommendations, such as issuance of multiyear flood insurance policies (Michel-Kerjan and Kunreuther 2011), catastrophe flood bonds, or the potential marketability of multihazard (e.g., wind and water) insurance products. The influence of these options on NFIP participation remains an important area for future research.
Acknowledgments
Special thanks to John Cartwright, Geosystems Research Institute, Mississippi State University for providing maps and much of the GIS-based data. We also thank two anonymous reviewers for their suggestions to improve the manuscript. This research was conducted under award NA06OAR4320264 06111039 to the Northern Gulf Institute by the NOAA Office of Ocean and Atmospheric Research, U.S. Department of Commerce.
Footnotes
The authors are, respectively, associate professor, Department of Agricultural Economics, Mississippi State University; associate professor, Department of Economics, East Carolina University, Greenville; and Giles Distinguished Professor, Department of Agricultural Economics, Mississippi State University.
↵1 The CRS awards premium discounts for all flood insurance holders in communities that undertake specific types of hazard mitigation, including provision of flood hazard information, disaster preparedness, and flood damage reduction. A lower score indicates a higher level of preparedness.
↵2 Market penetration of the NFIP across the entire United States was estimated at 26% of eligible parcels in 1997 (Pricewaterhouse Coopers 1999). Dixon et al. (2006) estimate that around 49% of residential properties in the SFHA held NFIP insurance, with an additional 2% to 3% of households holding private flood insurance, in 2004. They also find significant regional variation and higher market penetration in coastal areas. Landry and Jahan-Parvar (2011) estimate that 52% of residential properties in the nearshore coastal zone of the southeastern United States held flood insurance in 1999.
↵3 Models were estimated using flood zone based on both individual responses and GIS data. Both models worked equally well with no discernible differences in terms of covariate effects or levels of statistical significance.
↵4 A model was also estimated including a dummy variable for pre-FIRM construction. Houses constructed before the publication of FIRMs receive subsidized insurance rates but also can face greater risk because of less stringent building standards. As the expected effects of these two factors on insurance demand are countervailing, we have no prior expectation of the effect of pre-FIRM status. The effect of pre-FIRM was statistically insignificant in all models, and the coefficients of other covariates were generally unaffected by the inclusion of the pre-FIRM dummy variable. We exclude this variable because it allows us to use 76 additional observations (for which FIRM status could not be determined due to lack of data on year of construction). Results including pre-FIRM are available from the authors upon request.