Abstract
In this paper, we empirically examine whether the Federal Emergency Management Agency’s Public Assistance (PA) program, which targets postdisaster cleanup and infrastructure rehabilitation, affects household purchases of flood insurance. Using the fixed-effects model with instrumental variables to address the endogeneity of disaster aid, we find that increased PA grants reduce a county’s flood insurance take-up rates, thereby driving down its total insurance coverage and premiums paid. Our findings provide empirical evidence on the crowding-out effect of public disaster programs, and shed light on their implicit social costs and increased federal financial exposure to natural disasters and climate change. (JEL Q54, Q58)
1. Introduction
Natural disasters cause substantial social and economic losses across the United States. Among all, flooding is the most prevalent hazard and has resulted in the largest amount of monetary damage second only to hurricanes (Perry 2000; Kunreuther 2001; Kunreuther and Rose 2004; Brody et al. 2007; Miller, Muir-Wood, and Boissonnade 2008). The costs of weather-related hazards continue to increase, as a result of population growth and climate change, and managing disaster risks has become an important policy concern of both policy makers and researchers. This further raises additional questions of how to coordinate public risk management with private protection actions, and how to incorporate predisaster mitigation, postdisaster response and relief, as well as risk pooling and transfer mechanisms into a comprehensive disaster policy.
In this paper, we examine the impact of federal disaster grants on household purchases of flood insurance from the National Flood Insurance Program (NFIP). It is widely believed that the ex post disaster aid reduces the private incentive for ex ante risk mitigation, because it leaves the recipients with limited financial liability for their actual losses and establishes the expectation that they can rely on the external assistance provided by third parties such as the government. Economists call this problem the “Samaritan’s dilemma” (Buchanan 1975), or the “charity or moral hazard” (Ehrlich and Becker 1972; Shogren and Crocker 1991; Coate 1995; Kim and Schlesinger 2005; Lewis and Nickerson 1989). Others argue that governmental disaster aid is perceived as a substitute for private insurance and crowds out purchases of disaster insurance (Lewis and Nickerson 1989; Kaplow 1991; Kelly and Kleffner 2003; Kousky, Michel-Kerjan, and Raschky 2018).
The charity hazard hypothesis, particularly pertaining to the effect of postdisaster aid on disaster insurance purchases, has been empirically examined in only a handful of studies so far (Kousky, Michel-Kerjan, and Raschky 2018; Deryugina and Kirwan 2017; Raschky and Weck-Hannemann 2007; Raschky et al. 2013; Raschky and Schwindt 2009; Petrolia et al. 2015). Our paper builds on this line of inquiry but is distinguished from previous studies in several ways. First, while most other studies focus on the governmental disaster relief provided directly to affected households (e.g., in the form of cash payments), our paper is one of the first few studies that examine the crowding-out effect of disaster aid received by local governments for rehabilitation purposes (Petrolia et al. 2015). In particular, we examine the influence of the Federal Emergency Management Agency’s (FEMA) Public Assistance (PA) program, which provides aid to state, local, and tribal governments to fund debris removal, emergency protective measures, and restoration of public infrastructure such as roads and bridges following a major disaster shock.
Conceptually, the PA grant could affect household demand for disaster insurance through two channels. First, the PA grant could serve as a signal of government bailout, just like private disaster relief, and may reinforce individuals’ expectation of such aid even if they are not directly compensated for their disaster losses. Second, the PA-funded mitigation-related projects might help enhance the community-level resilience to future shocks and reduce the perceived need to purchase disaster insurance.1 However, it is unclear whether the crowding out, if there is any, is socially optimal or leads to underinvestment in private protection, since the public mitigation spending may promote a “false sense of security” (Stefanovic 2003; Kousky, Luttmer, and Zeckhauser 2006; Kousky and Olmstead 2010; Burby 2006).
By using a national sample of NFIP-participating counties between 1998 and 2010, our empirical analysis tests whether increased PA program aid flowing to a jurisdiction following a major flood event reduces household purchase of flood insurance in the community, all things being equal. We also control for FEMA’s Individual Assistance (IA) program that directly provides housing and other need-based financial assistance to affected households for their uninsured losses. To address the endogeneity of disaster relief (both PA and IA aid), we employ a panel fixed-effects model with multiple instrumental variables (IVs) that capture a county’s political importance and states’ representation in FEMA’s oversight congressional committees (which we detail in the methods section).
Considering that recent disaster shocks may update the risk perception and raise demand for flood insurance (Kousky 2017; Gallagher 2014), we control for the exogenous flooding shocks by using objective precipitation data. Our research is one of the first few that disentangle the impact of postdisaster aid on insurance purchases from the learning effect of recent disaster shocks (Kousky 2017). In addition, we also account for housing mortgage loans, because the NFIP requires flood insurance for homes located in 100-year floodplains that are financed by federally backed mortgage companies. Notably, most prior empirical studies fail to account for mortgage loans, which could be a confounding factor for estimating the effect of federal disaster relief.
To preview our results, we find that the PA grant a county receives has a negative effect on its household flood insurance purchases, which is statistically significant primarily at the extensive margins (measured by the total number of policies in force and total amount of coverage and premiums paid). For a 10% increase in PA project spending, a county’s total number of policy holders declined by 1.5%, and the total insurance coverage and premiums decreased by 1.4% and 1.2%, respectively. On the other hand, we find that the IA grant increases the number of policy holders, largely because of its mandate for the aid recipients to purchase and maintain flood insurance. Overall, our findings suggest that public disaster aid targeting disaster response and public infrastructure rehabilitation crowds out insurance purchases and may result in social welfare losses, which warrants attention in the ongoing discussion on NFIP reform. Although the NFIP insurance premiums are intended to generate sufficient revenues to cover losses, the large debts incurred after the 2005 and 2012 hurricane seasons adversely affected the program’s financial soundness.2 Moreover, low take-up rates still remain a concern across many communities exposed to frequent floods and inundation risk (Dixon et al. 2006; Kriesel and Landry 2004; Kousky and Michel-Kerjan 2012). Considering the negative impact of the PA grants on flood insurance purchases, policy makers should reevaluate the actual costs of federal disaster aid by accounting for the implicit losses associated with the flood insurance dropout.
2. Policy Background
The National Flood Insurance Program
The NFIP was established by the federal government as part of the 1968 National Flood Insurance Act. As a voluntary partnership between federal agencies, communities, and private insurers, the program was intended to address the growing cost of flood incidents and the unavailability of flood insurance in the private insurance market, and to provide an alternative to federal disaster assistance (FEMA 2002). The NFIP develops flood insurance rate maps (FIRMs) for participating communities, in which areas with a 1% annual chance of flooding, also known as 100-year floodplains (including both inland areas and those exposed to storm surge), are identified as special flood hazard areas (SFHAs). Although communities can voluntarily join the program, they are required to adopt minimum floodplain management regulations (e.g., elevate newly constructed or substantially renovated buildings to or above the base floodplain elevation level) after participating in the NFIP. In particular, homeowners whose properties are located in an SFHA and have a mortgage from a federally backed or regulated lender must buy flood insurance, as required by the Flood Disaster Protection Act of 1973. Private companies partner with the NFIP to write insurance policies and process claims, but the risk is fully borne by the federal government.
The NFIP sets flood insurance premiums based on a community’s flood zone maps, hydrological modeling, and characteristics of the structure. Therefore, the NFIP premiums do not change in response to individual flood events (i.e., property owners pay the same amount of premium regardless of whether their home is flooded). Recent disaster events such as hurricanes Katrina, Ike, and Sandy have resulted in substantial debts to the program. There have also been growing concerns that the NFIP fails to price the premium at actuarial rates that reflect the true flooding risks and ignores the heterogeneity of flood exposure across jurisdictions (Czajkowski, Kunreuther, and Michel-Kerjan 2013; Kousky and Shabman 2012).
The NFIP participation rate was initially very low. A number of reforms and amendments have been introduced to encourage community participation as well as to boost the number of insurance policies in force for participating communities. For example, as part of the Flood Disaster Protection Act of 1973, communities that are located in flood-prone areas but do not participate in the NFIP are not eligible for certain categories of federal disaster assistance (e.g., acquisition of property). FEMA also provides supplementary financial assistance in the form of hazard mitigation grants to NFIP-participating communities, enabling them to develop hazard mitigation plans and implement risk-mitigating measures.
Currently, more than 22,000 communities participate in the NFIP program and more than 90 private insurance companies partner with the federal government on writing policies and processing claims nationwide (Kousky 2017; Kousky and Michel-Kerjan 2015). Although the number of insurance policies has increased over time (reaching more than 5.58 million policies as of December of 2010, the last year in our sample), the take-up rates among participating communities still remain low (Kousky and Michel-Kerjan 2012). In Table 1, we list the top 10 states in terms of the number of policies in force. Florida and Texas together make up more than half of the national total, whereas the other coastal states still fall short in terms of flood insurance penetration. In terms of take-up rates per capita, Florida remains the top state with 11.28% penetration, followed by Louisiana and South Carolina with 10.76% and 4.39% take-up rates, respectively.
National Flood Insurance Program Policies in Force, by State
FEMA’s Public Assistance Program
In the United States, the federal disaster aid policy is governed by the Robert T. Stafford Disaster Relief and Emergency Assistance Act enacted in 1988, which is an amendment to the 1974 Disaster Relief Act. The act authorizes the president to issue declarations for disasters that cause large-scale damage and overwhelm the local capacity to respond. Once the declaration has been issued, FEMA provides disaster relief through the Disaster Relief Fund (DRF), which is a “no-year” account financed primarily through ad hoc supplemental appropriations (McCarthy 2011; Lindsay and McCarthy 2012; Lindsay 2014). Currently, the DRF is used to fund three major programs, including PA, IA, and the Hazard Mitigation (HM) program (funding disaster prevention and mitigation projects that lessen the adverse impacts of future disasters). It is estimated that PA program accounted for almost half of total DRF funding between fiscal year 2000 and fiscal year 2013, whereas IA and HM spending was altogether 30% (Brown and Richardson 2015).
Our focus in this paper is on the PA program, which provides grants to local, tribal, and state governments and certain types of private nonprofit organizations for immediate disaster response (e.g., debris removal), restoration of public infrastructure, and hazard mitigation measures (FEMA 2010).3 Therefore, PA projects have a public good nature with the goal of assisting communities to quickly respond to and recover from a disaster event. In terms of cost distribution, the federal share of a PA project is typically no less than 75% of its eligible cost, and the remaining non-federal cost share is often split between the grant applicants and subrecipients. Over the period 2000–2013, the average annual federal obligations for PA grants was estimated at 3.9 billion, while the highest annual spending was $17.1 billion in 2005 due to Hurricane Katrina (Brown and Richardson 2015).
3. Relevant Literature
Many recent studies have explored the determinants of disaster insurance purchase (e.g., Browne and Hoyt 2000; Zahran et al. 2009; Michel-Kerjan and Kousky 2010; Petrolia, Landry, and Coble 2013), and some have examined the influence of postdisaster aid (e.g., Petrolia et al. 2015; Kousky, Michel-Kerjan, and Raschky 2018; Deryugina and Kirwan 2017; Kousky 2017). Conceptually, household purchase of disaster insurance, as an economic decision, can be affected by household income, the price of insurance, values of the properties exposed to risk, and the perceived losses from future disaster events. The extant empirical literature has provided support for some of these determinants. For example, Browne and Hoyt (2000) use aggregated statelevel data to show that the number of flood insurance policies in a state correlates positively with household incomes and negatively with prices. Microlevel studies including those by Kousky (2011), Kriesel and Landry (2004), and Landry and Jahan-Parvar (2011) also find a similarly positive relationship between income and flood insurance demand. Petrolia, Landry, and Coble (2013) use survey data to show that expected hurricane damage and risk aversion can significantly drive private demand for flood insurance.4
One important issue arising from this literature is that the occurrence of a disaster event can trigger changes in risk perception and disaster policy interventions simultaneously, both of which may influence private insurance decisions. Specifically, a major flooding event can provide new information to update the perception of future flooding risks, thereby driving flood insurance purchases. A few recent studies have examined learning from disasters and found that past shocks can cause behavioral biases in risk-related decision-making. For example, Gallagher (2014) finds a major flooding event triggers a significant, yet temporary, spike in the community-level flood insurance take-up rates, which he attributes to the Bayesian updating process incorporating forgetting. Other recent studies (e.g., Carbone, Hallstrom, and Smith 2006; Hallstrom and Smith 206; Bin and Landry 2013; Atreya, Ferreira, and Kriesel 2013; Atreya, Ferreira, and Michel-Kerjan 2015; Kousky 2017) also provide similar evidence on the learning dynamics by examining the impact of recent disasters on housing prices.
Disaster aid policy can influence the demand for flood insurance in different ways. The “charity or moral hazard” hypothesis proposes that ex post disaster relief (which could come from the government or charitable donations) reduces purchases of disaster insurance ex ante, because it establishes individuals’ expectation of receiving such aid and thereby affects their perceived losses from future shocks. Several recent papers have examined the causal impact of postdisaster aid provided by federal government on insurance purchases. For example, using U.S. zip code–level data, Kousky, Michel-Kerjan, and Raschky (2018) find that the receipt of IA grants significantly decreases the average coverage of flood insurance purchased in the following year, while another type of disaster aid, the subsidized disaster loans provided by the Small Business Administration, has little impact on insurance.
Raschky et al. (2013) compare the disaster relief programs in Austria and Germany based on survey data. Their study indicates that expected disaster relief can significantly lower the willingness to pay for flood insurance coverage, and this effect is particularly larger for government relief programs with higher levels of certainty. Using an experimental stated preference design, Botzen et al. (2009) also find a negative relationship between government disaster aid and individual mitigation measures. Although not directly related to flood insurance purchases, Petrolia et al. (2015) find that federal flood mitigation grants at the county level significantly reduce household purchases of wind coverage, suggesting the crowding-out effect of disaster aid may spillover to risk management of other types of disasters. Deryugina and Kirwan (2017) examine the link between agriculture disaster aid and crop insurance at the U.S. county level over the 1990–2010 period. They find a significant and negative effect of the federal crop disaster payments on farmers’ out-of-pocket insurance expenditures, which also supports the charity hazard hypothesis.
In addition to the channel of moral hazard, disaster aid could also affect insurance purchases through institutional requirements. For example, in the U.S. context, households receiving IA grants are required to purchase and maintain flood insurance coverage for their properties postdisaster. If later they decide to opt out of flood insurance, they are not eligible for future federal disaster aid. For example, Kousky (2017) finds that the federal mandate for IA aid recipients to purchase flood insurance explains most of the increase in insurance take-up rates in the aftermath of major hurricanes.
Overall, examining the impact of governmental disaster aid on private risk management behaviors faces two empirical challenges. Given the ex post nature of disaster relief, it is critical, but often difficult, to isolate its influence on insurance purchases from the disaster learning effect and other policy interventions such as the aid requirement. It is worth noting that unlike with the IA program, the receipt of PA grants does not impose any requirement to purchase flood insurance. Therefore, our study should offer valuable insights into the effect of government aid that provides protective public goods and recovery resources on voluntary insurance decisions. We also include IA grants in this research because (1) as the governmental disaster relief directly provided to affected households, these grants could also crowd out purchases of disaster insurance; and (2) the institutional requirement imposed on the receipt of IA aid could drive more households to purchase flood insurance. These two channels could counteract each other and leave the actual effect of IA grants ambiguous.
Another empirical challenge is to address the endogeneity of disaster relief (both PA and IA), which could arise from two sources (Kousky, Michel-Kerjan, and Raschky 2018). First, the allocation of disaster aid might be influenced by a locality’s insurance penetration and coverage level, which leads to the reversed causality or simultaneity problem (Dixon et al. 2006). Second, there might be unobserved time-variant characteristics in a county that correlate with both the federal disaster aid it receives and local insurance purchases (e.g., the expectation of ex post disaster aid), which cannot be fully accounted for in a typical fixed effects model.5
Several papers have attempted to address the endogeneity problem using political ideology and election variables to instrument for disaster relief (e.g., Kousky, Michel-Kerjan, and Raschky 2018; Deryugina and Kirwan 2017). Their justification of IVs is that federal aid is partially motivated by political considerations and can be used to mobilize swing voters’ support.6 The political economy of postdisaster aid was initially highlighted by Garrett and Sobel (2003), who find that federal disaster relief is significantly higher during presidential election years, and political swing states receive more presidential disaster declarations. However, we note that not all factors related to political ideology are strictly exogenous to insurance decisions. For example, Botzen et al. (2016) use survey data to show that Democrats perceive a higher risk of flood damage than Republicans, and moreover, significantly more Democrats than Republicans expect to receive federal disaster relief after a major flood.
Finally, we note that our study also adds to a small but important literature on evaluating the efficacy of governmental disaster assistance programs. For example, Healy and Malhotra (2009) find that $1 of spending on disaster preparedness and mitigation can reduce future disaster damage by $15, whereas disaster relief spending has little impact on disaster damage, thereby suggesting significant welfare losses associated with ex post federal relief programs. Davlasheridze, Fisher-Van-den, and Klaiber (2017) show that the returns on FEMA’s hazard mitigation projects are twice that of its disaster recovery programs, suggesting significant social gains from diversifying spending across different programs. However, their research does not tease out whether public disaster aid projects complement or substitute private investment in risk mitigation and disaster insurance. From this perspective, our paper informs a better understanding of the implicit costs associated with the PA program.
4. Data
We collected the data for this study from multiple sources. The NFIP data, including the number of policies in force (residential policies), total premiums paid, and total dollar insurance coverage for every participating county were obtained from FEMA through a Freedom of Information Act (FOIA) request. The data on PA and IA grants were obtained from FEMA’s website (FEMA 2016a) and a FOIA request, respectively. Because the raw data indicate the amount of approved grants by county and types of disaster, we calculated the total amounts of PA and IA grants for each county-year observation for flooding and hurricane-related disasters only.
In this paper, we use precipitation data to construct the county-level annual rainfall anomalies to control for the exogenous flooding conditions. By doing this we can separate the effect of disaster aid on insurance purchases from the learning effect through the updated risk perception. Notably, most of the previous studies modeling flood insurance purchase behavior do not control for the exogenous flooding conditions. Some use the presidential disaster declarations of flooding or hurricane events (e.g., Gallagher 2014), but these may not accurately measure the severity of flooding damage and are also subject to the endogeneity concern, because it is more likely for disasters to be declared for underinsured communities that lack the resources necessary to cope with them.
Our precipitation data come from the National Climate Data Center Global Historical Climatology Network.7 We calculate the rainfall anomaly measuring the proportional deviation of annual precipitation (Std_RAINt) in year t from a county i’s long-run average annual precipitation (Mean_RAINi) over the 1950–2000 period, as indicated in equation [1].8 Thus, positive values of anomaly indicate excessive rainfall in a county in a given year.
[1]
Using the relative deviation measure, we are able to account for local adaptation (i.e., localities that often experience intense precipitation should be better adapted to such situations and would suffer fewer losses, as compared to localities less exposed to extreme rainfall).9 In the Appendix we show that this variable is highly predictive of county-level flooding damage and flood-related presidential disaster declarations, which include hurricanes, coastal storms, and surge incidences.10
For other control variables measuring a county’s socioeconomic characteristics, we use the personal income and population series from the Bureau of Economic Analysis Regional Economic Account Systems.11 Since flood insurance purchase is a mandate for properties located in flood plains and financed by federally backed mortgages, we control for a county’s mortgage activity using the residential mortgage data from the Federal Deposit Insurance Corporation (FDIC),12 which provides information about financial liabilities and assets by institutions. Using the FDIC unique number, we identify all relevant residential loan categories and aggregate them to create a variable measuring the total dollar amount of residential mortgages in a given county-year.13
With respect to the data for our IVs, we use presidential election data from Dave Leip’s Atlas of U.S. Presidential Elections14 to identify swing counties. Following the approach of Kousky, Michel-Kerjan, and Raschky (2018), we define a county as a “swing county” if the winning margin between the two main political parties is less than or equal to 5% in the most recent presidential election (there were four presidential elections held, in 1996, 2000, 2004, and 2008, relevant to our sample time frame) and the county is located in a swing state.15 We use the biannual editions of Almanacs of American Politics (Barone and Ujifusa 1988, 1990, 1992, 1994; Barone, Ujifusa, and Cohen 1996, 1998; Barone et al. 2000; Barone, Cohen, and Cook Jr. 2002; Barone and Cohen 2004, 2006, 2008) to create two variables measuring a state’s representation in the Senate and House subcommittees overseeing FEMA’s operation.
Table 2 reports the summary statistics of our main variables. Our sample is based on an unbalanced panel of a total of 2,094 NFIP-participating counties for the years 1998–2010, and the time span is dictated by the availability of the PA and IA grants data. The sample average number of NFIP policies in force is approximately 1,699, corresponding to approximately a 1.1% take-up rate on average. The maximum number of policies is approximately 434,000, accounting for 80% of insurance penetration. The mean value of total coverage is $333 million per county, which is $138,707 coverage per policy holder. On average an NFIP-participating county pays a total of $888,000 for insurance premiums, corresponding to approximately $607.5 per policy holder. The average PA grant received by a county in our sample is $817,950. This variable equals zero for years when no disasters are reported, and the maximum value is $3.7 billion approved for Orleans Parish after hurricane Katrina. All variables with dollar values are adjusted to 2010 real prices using the urban consumer price index.
Summary Statistics
5. Method
To examine the effects of the PA grants on household insurance purchase, we estimate a fixed effects panel data model specified in equation [2] as follows:
[2]
The dependent variable, yc,t, measures different aspects of household insurance purchases in county c for a given year t, including the total number of insurance policies in force, policies per capita (i.e., take-up rates), total dollar amount of insurance coverage, and total amount of premiums paid. These insurance policies cover only residential properties. The number of policies captures the extensive margin, while the total amount of coverage as well as the total premiums paid reflect a combination of both the intensive and extensive margins.
PAc,t–1 is our key variable of interest, which denotes the total FEMA-approved PA grants related to flooding disasters received by county c in year t – 1. We log transform this variable and add one to avoid losing the observations with a value of zero.16 IAc,t–1 measures the total FEMA-approved IA grants related to flooding incidences and is also log transformed. To account for recent flooding experiences, we include a distributed lag of rainfall anomalies (Rainc,t) in the current year as well as over the past three years.17 Xc,t–1 denotes a vector of variables measuring a country’s socioeconomic characteristics including per capita personal income, size of population, and total dollar amounts of residential mortgage loans (all three variables are logged transformed). λt denotes the year fixed effects, which control for the common shock to all counties in the same year (e.g., policy changes in the national flood insurance program, such as the requirement of minimum coverage and premiums, and changes of other federal disaster policies). λc is the county-specific fixed effects, which control for the time-invariant unobserved county characteristics (e.g., the baseline flood risks such as spatial flood hazard areas, geography). Finally, εc,t denotes the error term.
We cluster standard errors at the congressional district level because FEMA’s oversight committee members represent different congressional districts within a state. If a county crosses the boundaries of multiple congressional districts, we assign the county to the district that covers the majority of the county area.
Identification Strategy
As we discussed earlier, both the PA and IA variables are endogenous due to the potential reversed causality and omitted variable bias, which may still cause our fixed effects estimators to be biased. To address the endogeneity concern, we take an IV approach, specifying the first-stage model in equation [3]:
[3]
where the endogenous variables (Endog), including total PA (PAc,t) and total IA grants (IAc,t), are respectively a function of the same set of control variables as in equation [1] and the IVs, denoted as Zc,t. Our choice of IVs is motivated by the political economy of governmental disaster aid (Bea 2005; Reeves 2011; Husted and Nickerson 2014; Garrett and Sobel 2003; Sobel, Coyne, and Leeson 2007; Sobel and Leeson 2006). It has been widely observed that the president and state elected officials can exercise their discretionary power to use disaster aid to buy out votes in politically important jurisdictions (Dixit and Londregan 1996; Lindbeck and Weibull 1987; Dahlberg and Johansson 2002; Mayer 2007). Drawing upon the prior literature, we create three IVs that capture the political influence on disaster relief, assuming that they exogenously affect the aid decision while exerting no direct effect on disaster insurance purchases.
First, we follow the approach of Kousky, Michel-Kerjan, and Raschky (2018) and instrument for disaster aid, using the interaction term between a binary variable indicating whether the county is a “swing county” and is located in the swing state, and another binary variable indicating whether the previous year was a presidential election year. The rationale for this IV is based on Garrett and Sobel’s (2003) finding that FEMA’s disaster payments are significantly higher during presidential election years and higher for states with greater political importance.
The other two IVs capture a state’s representation (i.e., number of seats) in the House and Senate subcommittees that oversee FEMA’s operation. They are also motivated by the finding by Garrett and Sobel (2003) that congressional oversight can influence the allocation of disaster aid in a way that states holding seats in one of the many congressional subcommittees overseeing FEMA could leverage their influence to allocate more aid for their own constituencies during disasters and emergencies. Following Garrett and Sobel (2003), in Table 3 we list the relevant Senate and House subcommittees, categorized as Stafford and non-Stafford according to the aspects of FEMA’s operation they oversee. Stafford subcommittees directly oversee FEMA’s spending related to presidential disaster declarations, whereas non-Stafford subcom-mittees oversee FEMA’s operations related to other nondisaster incidents (e.g., agricultural and food emergencies, public health emergencies, and oil and hazardous materials spills) (Lindsay 2014).
Senate and House Subcommittees Overseeing FEMA’s Disaster and General Operations
Following Davlasheridze, Fisher-Vanden, and Klaiber (2017), our IVs include a state’s total number of seats in the Stafford subcommittees and a dummy variable indicating whether a state has a senator serving in one of the non-Stafford subcommittees.18 We expect that membership in these committees affects flood insurance only through increased leverage on FEMA disaster program spending, and they are uncorrelated with households’ decision to insure.19
In Table 4, we report the first-stage estimation results of equation [3]. PA grants are the dependent variable in column (1), and column (2) reports results for IA grants. Our IVs are strong, since the first-stage F-statistics in both models are above the threshold level of 10.20 As expected, we show that swing counties receive significantly more PA and IA grants during the presidential election years, and the aid from both these programs increases if a state is represented in one of the non-Stafford subcommittees. We also find that states with more seats in the Stafford subcommittees receive significantly more federal aid from the IA program, all else being equal. However, this variable is not statistically significant for predicting PA grants.
First-Stage Results
6. Results
In this section, we first report the fixed effects model results in Table 5 without instrumenting for the endogenous PA and IA variables. We show that PA grants are insignificant for explaining household flood insurance purchases, whereas IA grants have a positive and significant effect on insurance decisions at almost all margins (i.e., receipt of a larger IA grant would increase the total number of insurance policies, total dollar coverage, premiums paid, as well as the take-up rate per capita in a county).
Fixed Effects Model: Impact of Public Assistance Grants on Household Flood Insurance Purchases
Baseline Estimates from the Fixed Effects IV Model
Table 6 reports the estimation results from the fixed effects IV model for the same set of outcome variables. We report the Hansen J statistic for the overidentifying restriction, with associated p-values, which shows that for all specifications, we fail to reject the null hypothesis that our IVs are uncorrelated with the error term. Notably, we find that using the IVs has changed the sign and significance of the PA coefficients in most cases, demonstrating the importance of accounting for the endogeneity of federal disaster relief. Our estimates suggest that for a 10% increase in flooding-related PA grants a county receives, the number of policy holders and take-up rates decrease by approximately 1.5% in the following year, while the total coverage and premium payments decrease by 1.4% and 1.2%, respectively.
Fixed Effects Instrumental Variable Model: Impact of Public Assistance Grants on Household Flood Insurance Purchases
It should be noted that the total amount of insurance coverage and insurance premium payments capture both the extensive and intensive margins of aggregated flood insurance purchases. Since the changes in these variables are quantitatively similar to those in the take-up rates, we believe that the impact of PA aid is primarily concentrated on the extensive margin.21 Overall, our results provide empirical evidence on the crowding-out effect of the PA program on household purchases of flood insurance.
With respect to IA grants, we show that accounting for endogeneity has more than quadrupled the magnitude of IA coefficients, although their direction and significance remain largely the same. The estimated coefficients suggest that increasing IA grants by 10% would increase the number of policy holders and take-up rates by about 1.1%, and total coverage by 0.9% in the following year. Unlike PA grants, the positive effect of IA grants on household purchases of flood insurance could be attributed to the program requirement that individual recipients must purchase flood insurance. This result is consistent with the finding by Kousky (2017) that the majority of the flood insurance take-up spike after hurricanes is driven by the IA aid requirement.
As for the other control variables, we find that recent flooding experiences (up to year t – 3) have a significant and positive impact on private insurance purchases. Such an effect is most pronounced one year following a rainfall shock, and it gradually decays over time, meaning that the spike in flood insurance purchase is only temporary. This finding resonates with prior research (e.g., Gallagher 2014; Kousky 2010, 2017) in suggesting that disasters trigger learning and behavioral bias (e.g., availability heuristic) in insurance purchases. In Appendix Table A2, we show that the effect becomes insignificant or only marginally significant beyond three years’ lag.
As expected, our results also show that the more residential mortgage loans there are in the county, the higher the total amount of insurance coverage. The more populated and wealthier the county, the higher the flood insurance penetration.22
We note that despite the statistical significance, the magnitude of the crowding-out effect or the economic significance of PA grants appears to be relatively small. However, this does not mean that we should not be concerned with the negative implication of the PA program on disaster insurance penetration, particularly when we take into account the large size of PA program spending. To put our estimates into perspective, a county receiving an average amount of PA grants in the sample would lose approximately 255 insurance policy holders in the following year. In terms of total coverage, for every 1% increase in PA grants, which corresponds to approximately $8,888 of the average PA grant in our sample, the total insurance coverage and premium payments would decrease by $452,000 and $1,079, respectively. On average, these losses correspond to $45.2 million in coverage and $107,915 in premiums per county, implying approximately $313.4 million in lost revenue from premiums for the NFIP in the year following the receipt of a PA grant. The insurance dropouts due to postdisaster governmental relief should warrant policy-makers’ attention, because they not only reduce revenue to the NFIP, but may result in additional outlays through federal programs to cover future disaster losses.
We also note that our estimated effect of the IA grant differs from the finding by Kousky, Michel-Kerjan, and Raschky (2018), as the latter suggests that the receipt of an IA grant reduces the average coverage (i.e., intensive margin) of flood insurance purchases by $4,000 to $5,000, while it has little impact on the take-up rate (i.e., extensive margin). Nonetheless, one possible explanation for the different findings is that Kousky, Michel-Kerjan, and Raschky (2018) examined only the policies that were purchased voluntarily (i.e., they excluded policies that were purchased as a requirement of receiving IA aid). Since our data do not separately identify insurance policies that are required for IA, we did not take this approach. Therefore, we interpret the positive effects of IA as being driven by the aid requirement. Another factor that may be driving different findings is that Kousky, Michel-Kerjan, and Raschky (2018) do not control for PA grants, which we show are also important for explaining the demand for flood insurance.
The Effect of PA Grants by Type of Project
Since a certain proportion of the PA-funded projects involve restoration of public infrastructure and disaster mitigation, they might help enhance the resilience of affected communities and, therefore, lower the perceived risk and incentive to insure. Few studies have attempted to examine the effect of public flood mitigation projects on insurance purchases, and the empirical evidence is mixed. For example, Kousky (2011) finds that the take-up rates of flood insurance in Missouri’s St. Louis County declined with levee protection. Browne and Hoyt (2000) show that FEMA’s spending on emergency preparation, planning, and mitigation programs is statistically insignificant for explaining statelevel flood insurance purchases. In their study of counties in Georgia, Atreya, Ferreira, and Michel-Kerjan (2015) find that mitigation programs have little effect on the number of insurance policies.23 Another two studies (Zahran et al. 2009; Frimpong, Petrolia, and Harri 2017) examine the local mitigation activities, measured by the Community Rating System (CRS), and both find a positive effect of the CRS on flood insurance purchases in states including Florida, Alabama, and Mississippi.24 It should be noted that we do not account for the effect of CRS participation in this paper because a very small proportion of counties in our national sample participate in the CRS program and there is insufficient within-county variation to identify its effect in a panel fixed effects model.
To examine whether PA grants affect flood insurance through the risk mitigation mechanism, we divide PA projects into two broad categories: (1) emergency response projects (including debris removal and emergency protective measures), and (2) permanent rehabilitation works (covering a broader range of activities such as roads and bridges, water control facilities, public buildings and equipment, utilities, parks, recreational areas and other facilities, and fire management) (FEMA 2010). The underlying idea of testing their separate effects on insurance decisions is that if the PA grants crowd out flood insurance purchase primarily because they are expected to improve local resilience to future shocks, then households should be more responsive to grants financing permanent rehabilitation projects than to spending on emergency responses. In our sample, approximately 62% of total PA grants were allocated to permanent work projects, with the remaining 38% allocated to emergency response and cleanup activities.
In Table 7, we report our estimation results from two separate regressions including each of the two types of grants, along with the same set of regressors in the baseline fixed effects IV model. We find that both types of projects have a statistically significant and negative effect on flood insurance purchases, which is consistent with our baseline results. We note that the coefficients on the permanent works are slightly higher than those on the emergency response projects. However, the difference between the two sets of coefficients is not statistically significant. These findings may suggest that private agents do not necessarily distinguish different components of PA grants related to their specific functions (e.g., long-term mitigation vs. immediate cleanup). One possible explanation is that they do not have adequate information about how the federal money is distributed across different programs, and the implication of these programs on risk reduction. It is also likely that households are more responsive to public projects that are more visible and more aligned with their preferred risk management strategies.
Fixed Effects Instrumental Variable Model: Impact of Public Assistance Grants (by Type) on Flood Insurance Purchases
To further examine this issue, we estimate a model in which total flood-related damage in a county is regressed on the PA and IA grants it received in the previous year, along with other control variables (see Appendix Table A4). The PA coefficient is negative and statistically significant, which is suggestive of the efficacy of this public program in enhancing local resilience. However, we compare the marginal effect of PA grants (in terms of a $1 increase) on damage reduction with its marginal effect on decreasing insurance coverage (when setting all other variables at their sample means) and find that the latter estimate is much larger than the former in terms of magnitude.25 This finding suggests that the crowding-out effect of PA grants on household flood insurance purchases exceeds its potential risk-mitigating effect. Even if part of the reduced flood insurance take-up could be associated with perceived resilience enhancement because of the receipt of PA aid, the latter is not sufficient to offset the underinvestment in flood insurance.
In sum, while we show that PA grants significantly reduce flood insurance take-up rates, we cannot conclude with clear evidence on the mechanisms of such crowding-out effect (i.e., whether it is through signaling federal bail-out or enhancing local resilience), as well as its net welfare effect. As we noted above, it is very likely that people may underinvest in insurance or other protective measures ex ante (i.e., below the socially optimal level) if they consider government-funded flood protection as a substitute for private protection. Such reliance may likely promote increased exposure and vulnerability to future natural disasters (Kousky, Luttmer, and Zeckhauser 2006; Kousky and Olmstead 2010; Burby 2006). For example, in a study examining migration following natural disasters, Boustan, Kahn, and Rhode (2012) find that the levee systems funded by the federal government during the early 1930s were responsible for the limited outmigration from flood-prone areas. The catastrophic losses in New Orleans caused by Hurricane Katrina are another example of how federal disaster policies targeting structural protection led to increased vulnerability in flood-prone areas, known as the “safe development paradox” (Burby 2006). Future research should devote more attention to the impact of disaster aid on the recipients’ perception of local risk and resilience.
7. Conclusion
Given their uncertain and destructive nature, natural disasters often trigger responses and interventions at different levels of government. In the United States, the federal government has been playing an important role in assisting affected jurisdictions in disaster responses, recovery, mitigation, and risk transferring, with most of its policies involving substantial welfare redistribution. Meanwhile, there have been growing concerns with the welfare and efficiency implications of the federal postdisaster relief and recovery programs, as well as their financial viability. In this paper, we examine the effects of FEMA’s PA grants on household flood insurance purchase behavior, employing a national sample of NFIP-participating counties. Our research generates several important findings and policy implications.
First, we find consistent evidence of the crowding-out effect of the PA aid on flood insurance purchases, largely concentrated on the extensive margin. In particular, when we employ the IV strategy to account for the endogeneity of postdisaster grants, we estimate that for a 10% increase in PA project spending, a county’s total number of policy holders declines by 1.5%, and the total insurance coverage and premiums decrease by 1.4% and 1.2%, respectively. We also find that an increase in IA grants actually has a positive effect on flood insurance take-up rates. This suggests that the institutional requirement of the disaster aid policy may offset its potential negative impact on insurance purchase (because of the charity hazard problem).
Second, our findings shed light on the implicit social cost of the PA program when accounting for insurance drop-outs and revenue losses to the NFIP (e.g., lost premium payments). Our estimate suggests that a county that receives average PA grants (in our sample) would lose approximately 255 insurance policy holders in the following year, corresponding to over $313 million in lost revenues for the NFIP. Furthermore, the reduction in disaster insurance purchase may imply that future disaster losses will likely be covered by additional outlays of federal disaster programs, such as individual housing assistance.
Given the scientific consensus that climate change will likely increase the frequency and intensity of extreme weather events such as flooding (Allan and Soden 2008; Emanuel 2005, 2013; Knutson et al. 2013; Field et al. 2012), the economic cost of natural disasters is expected to grow in magnitude, which would further strain federal and local government resources. The current federal disaster policies, which mostly involve postdisaster assistance and are expansionary in the size of spending, generate concerns regarding the increased federal financial exposure to natural hazards. Our research suggests that the PA program incurs a higher implicit cost (compared to its face value) because of its dampening impact on flood insurance purchases. To address these problems, members of Congress and the Government Accountability Office have proposed a new policy scheme concerning deductibles for the PA program (FEMA 2016b). The new reform, if enacted, will require states that re-quest the federal PA grants to commit funding for disaster response as well as for mitigation and resilience enhancement. Although this initiative may not completely eliminate the reliance of state and local governments on federal disaster aid, it should help transfer the disaster risks back to the subnational governments and encourage them to increase investment in ex ante hazard mitigation and preparedness as part of long-term disaster recovery.
Moreover, if disaster deductibles become a requirement, it may also make the existing insurance incentive programs such as the CRS more attractive for many localities that currently are not part of such programs. The CRS, which is intended to bridge public and private mitigation and protection, recognizes participating communities for their mitigation efforts and in return rewards their residents with discounts in insurance premiums (Fan and Davlasheridze 2016; Brody et al. 2009; Landry and Li 2011; Sadiq and Noonan 2015). If states are required to commit funds in order to receive federal PA grants in the future, their communities would be more likely to participate in the NFIP and CRS and increase their mitigation efforts in exchange for premium discounts. Understanding how local-level mitigation policies, including those defined by CRS creditable activities, interact with federal disaster aid would be one future extension of this research. Finally, it is also important to acknowledge that in this research we do not incorporate the federal home mortgage program spending because we do not have access to that data. To enable a more comprehensive assessment of the welfare implications of federal disaster programs, future research should also account for the potential effect of federal mitigation grants and projects.
Acknowledgments
This research is being supported by the NSF PIRE-Coastal Flood Risk Reduction Program: Integrated, multiscale approaches for understanding how to reduce vulnerability to damaging events (award #1545837).
Footnotes
↵1 Improving protective measures and infrastructure is an important component of the PA program. For example, in 2006, FEMA granted $8.6 million as part of PA grants to the Alabama State Port Authority to restore the levee system protecting the eastern shore of Mobile Bay. For more details see https://www.fema.gov/news-release/2006/01/26/fema-grants-86-million-public-assistance-restore-gaillard-island-levee/berm.
↵2 Insurance claims from the 2005 and 2012 hurricanes forced the NFIP to borrow money from the Treasury Department. As of December 31, 2014, FEMA owed the Treasury Department $23 billion, up from $20 billion from November 2012 (Government Accountability Office 2015).
↵3 Brown and Richardson (2015) estimate that about 24% of the total PA grants between FY 2000 and FY 2013 was spent on public buildings, 21% on protective measures, and 19% on debris removal.
↵4 Their study also finds that respondents who express greater expectation of disaster assistance are more likely to purchase flood insurance, which seems to contradict the charity hazard hypothesis.
↵5 One possibility, as discussed by Kousky, Michel-Kerjan, and Raschky (2018), is that better-educated households may have higher demand for insurance and are better able to navigate the bureaucratic red tape to receive federal aid.
↵6 For example, Kousky, Michel-Kerjan, and Raschky (2018) use the interaction term between the timing of presidential election years and a swing county indicator. Deryugina and Kirwan (2017) use percentage voting for the third-party candidate to instrument for agricultural disaster aid.
↵7 Note that the raw dataset reports daily precipitation at individual weather stations. Following the approach used in other climate economics research (e.g., Deryugina and Hsiang 2014), we map the weather stations to counties based on their latitude and longitude and compute the annual total rainfall for a given county-year observation. For counties with multiple stations, we take the average of their annual sum.
↵8 The main reason for using the average over the 1950– 2000 period is because precipitation patterns in more recent years appear to deviate more from the historical normal, presumably due to climate change. As such, the historical mean is less obscured by the recent abnormalities. Specifying this variable in this fashion to assess adaptation responses to extreme weather events is a common approach in the literature (e.g., see Miao and Popp 2014; Davlasheridze and Geylani 2017).
↵9 It is also important to acknowledge that excessive rainfall is only one of the natural factors that could lead to floods. For example, storm surge associated with strong winds and hurricanes is another major cause of coastal flooding. However, by showing the strong link between rainfall anomalies and actual flooding damage, we are confident that in using rainfall shocks we account for the exogenous flooding condition.
↵10 In Appendix Table A1 we report the estimated results from a fixed effects model in which we regress a set of variables measuring flood severity (total flood-induced damage, total NFIP payments, total number of flood events, and total number of declared flood and hurricane disasters and emergencies) on the contemporaneous rainfall anomaly, income per capita, and population. We show a significant and positive effect of the rainfall anomaly across all specifications.
↵11 See https://apps.bea.gov/iTable/iTable.cfm?acrdn=7& isuri=1&reqid=70&step=1#reqid=70&step=1&isuri=1 (retrieved September 1, 2016).
↵12 See https://www5.fdic.gov/idasp/advSearch_warp_download_all.asp?intTab=2 (retrieved September 1, 2016).
↵13 Four different categories of loan amounts are included in mortgage loan calculations: (1) 1–4 family residential loans that include total secured loans by 1–4 family residential properties; (2) loans secured by 1–4 family first liens; (3) loans secured by 1–4 family junior liens; and (4) all other adjustable rate closed-end loans secured by 1–4 family residential properties, secured by first liens (FDIC 2016).
↵14 See https://uselectionatlas.org.
↵15 A swing state is defined in a similar fashion as a swing county, using state-level election results data also available from Dave Leip’s Atlas of U.S. Presidential Elections (https://uselectionatlas.org).
↵16 Because the dependent variables and PA variable are all in log terms, we can interpret the coefficients in the form of elasticity.
↵17 As a robustness check, in Appendix Table A2 we also provide results from models in which rainfall lags extend over the past five years.
↵18 The main argument for specifying non-Stafford committee as a dummy variable instead of keeping it as a count similar to the Stafford committee is that since non-Stafford committees oversee FEMA’s minor programs, a state’s overall congressional representation should be sufficiently important to exert some influence over FEMA disaster spending. It is less likely that having one more representative would help the state obtain additional disaster aid from the federal government. Nonetheless, we want to note that the models in which both IVs are given in counts provide similar estimation results, which are reported in Appendix Table A3.
↵19 There could be a potential problem using Stafford or non-Stafford committee membership, if one argues that they comprise self-selected state representatives who are concerned with their states’ flooding risk and work to seek more federal disaster aid to please their constituents, with the sole interest to be reelected. However, this argument seems to have received little support from the empirical research. On the contrary, it shows that such committee preferences are the exception rather than the rule. In fact, the research suggests that the majority of committee members put institutional and party priorities above electoral preferences and needs of individual legislators (Krehbiel 1990, 1991, 1994; Gilligan and Krehbiel 1990; Groseclose 1994; Maltzman 1995; Adler and Lapinski 1997; Overby and Kazee 2000; Overby, Kazee, and Prince 2004; Battista 2004; Prince and Overby 2005).To further examine this possibility, we have estimated several auxiliary models in which the dependent variables are the number of state seats in Stafford and non-Stafford committees, respectively (we also run the model with dummies corresponding to state committee representation). We regress the membership variables on flooding damage, rainfall anomaly, state GDP, and population, all lagged by one year, along with the state and year fixed effects. We employ different estimation strategies to estimate these models, including linear regression, Poisson, and logit regressions. Results, available upon request, show that neither flooding damage nor rainfall anomalies had strong statistical power to explain a state’s representation in these committees. We should also point out that Stafford/non-Stafford (sub)committees listed in Table 1 are much broader in the policy areas they oversee than just natural disaster policy. Hence, we believe that it is less likely that the state legislators’ membership in these committees is solely driven by constituents’ concerns related to natural disasters.
↵20 In Table 7, we report the Hansen J statistics of our IVs, which are all statistically insignificant.
↵22 Since PA and IA grants are dispersed at the county level, while the insurance purchase happens at the individual level, to identify the average causal effect one could argue that population-weighted models should be used. General issues associated with weights are discussed by Solon, Haider, and Wooldridge (2015), where they show that average treatment effects under exogenous stratification are not always recovered by population weighting. However, the authors still recommend comparing the results of weighted with unweighted estimates for possible model misspecification issues arising from the failure to model the heterogeneous effects. In Appendix Table A6, we present results from the IV regression in which we apply weights to reflect counties’ population shares. Our estimates from weighted and unweighted models are very consistent in terms of magnitudes as well as signs and significance.
↵21 In Appendix Table A5 we report the results from the model in which we use the average coverage per policy holder as an outcome variable measuring the intensive margin only, and find that PA grants have no statistically discernable impact on the decision regarding how much to cover against floods.
↵23 It should also be noted that neither of the two studies accounts for the endogeneity of disaster mitigation programs, making their estimates subject to potential bias.
↵24 The CRS was created as part of the NFIP to reward local communities that undertake flooding mitigation measures that exceed the NFIP’s minimum floodplain management standards. Communities with higher CRS scores are able to receive more insurance premium discounts. Therefore, the CRS can also lead to more purchases of flood insurance through the policy price reduction.
↵25 We estimate that a $1 increase in PA grants would reduce flooding damage by $1.2 and insurance coverage by $14.
