Exposing disparities in flood adaptation for equitable future interventions in the USA

Flooding accounts for nearly a third of global disaster losses and is the costliest U.S. weather hazard, with >$5 billion in average annual losses. Climate change is projected to exacerbate flood damages dramatically. In response, communities have accelerated adoption of adaptation measures, yet it remains unclear whether these measures are equally effective across different community types. The U.S. FEMA National Flood Insurance Program (NFIP) Community Rating System (CRS), launched in 1990, incentivizes communities to implement recognized flood risk reduction activities (e.g., mapping, open space preservation, stormwater management, public information) in exchange for insurance premium discounts. This study asks whether the effectiveness of such flood adaptation interventions varies with community characteristics (population, income, race/ethnicity, educational attainment, flood risk, precipitation), and to what extent benefits are equitably distributed. The purpose is to quantify savings from adaptation and identify disparities to inform more just and effective policy design.

Prior work using the NFIP Redacted Claims dataset has assessed whether CRS participation reduces flood claims, with mixed findings across studies, though the preponderance of evidence indicates CRS reduces losses. However, few analyses have examined heterogeneity in effectiveness across community types. Broader literature shows low-income communities face disproportionate flood risks and often rely on coping mechanisms in the absence of protection; pro-poor adaptation and asset strengthening are recommended. Studies have also linked flood loss and vulnerability to socioeconomic and demographic characteristics and documented inequities in preparedness and recovery funding and exposure. Yet, quantitative estimates of how adaptation benefits (savings) distribute across communities by income, race/ethnicity, and other features remain scarce; this study addresses that gap.

Design: Quasi-experimental comparison of CRS participant (treated) and non-participant (control) communities using the FEMA NFIP Redacted Claims dataset (~2.5 million claims). Unit of analysis is ZIP code (community). Outcome Y is total insurance claims paid per policy (building + contents) per loss. CRS participation is assigned at the jurisdictional level and mapped to ZIP codes. For each ZIP-year of loss, the study computes average annual storm paid per policy. Covariates X include precipitation (monthly average during event month from PRISM), flood risk (First Street Foundation risk/vulnerability scores), median income, population, renter fraction, educational attainment (Bachelor’s or higher), and fraction of racial and ethnic minorities (from ACS, aligned to loss year). Causal estimation: The study estimates the conditional average treatment effect (CATE) = E[Y|X,T=1] − E[Y|X,T=0] using a deep generative modeling framework, CAUSAL.FLOW. Separate normalizing flow density estimators (Masked Autoregressive Flow) are trained for treated and control groups to learn p(Y|X,T). Monte Carlo integration of the learned conditional outcome distributions yields CATE at specified covariate values, within the support of observed data. This approach relaxes linearity and matching assumptions common in traditional methods. Community typologies: To interpret heterogeneity, 27 typologies are defined by low/mid/high bins of median income, population, and racial/ethnic minority fraction (using percentile-based thresholds; exemplary bin centers: income $40k/$60k/$90k; population 250/12,000/30,000; minority fraction 0.05/0.15/0.40). CATE is evaluated across these typologies while varying precipitation, flood risk, renter fraction, and educational attainment, holding other variables at fiducial values. Data scope: 14,729 unique communities across the continental U.S. Data sources: NFIP Redacted Claims (CRS status, losses, policies), First Street Foundation flood risk, PRISM precipitation, and U.S. Census ACS demographics. Code (causalCRS) and curated dataset are publicly available; FSF data access is licensed.

- Overall effectiveness: CRS participation reduces flood losses. Average savings per policyholder are typically $5,000–$15,000 per policy; in many typologies $8,000–$15,000, and for some communities savings exceed $20,000. Uncertainties on CATE estimates in representative cases are about $180–$300. - Precipitation: Despite higher precipitation generally worsening flood losses, CRS savings increase with average precipitation for >95% of community typologies, demonstrating strong mitigation effectiveness under wetter conditions. However, predominantly racial/ethnic minority, high-population, mid to high-income urban communities deviate, with savings decreasing as precipitation rises. - Population dependence: High-population communities realize about $4,000 more savings per policy than low-population communities. Combined with higher adoption likelihood in populous areas, program benefits skew toward larger communities. - Educational attainment: After joining CRS, effectiveness still depends on education levels, producing about a $2,000 savings gap between communities with higher vs lower educational attainment. Many CRS-creditable activities require technical capacity (e.g., warning systems, levees), amplifying this disparity. - Flood risk and income: For low-income communities, savings decline sharply in high flood-risk zones, whereas affluent communities maintain or increase savings at higher risk, suggesting expensive, effective measures are more attainable to high-income areas and that low-cost measures may only be effective at lower risk. - Racial disparities within income strata: Among low-income communities, predominantly white areas see over $6,000 higher savings per policy than predominantly racial/ethnic minority areas in some settings, indicating persistent inequities not overcome by CRS participation. - Conservatism of estimates: Savings likely conservative due to corrections for outreach components and the control group’s minimal floodplain regulation requirements that may already reduce losses slightly.

The study’s central question—whether flood adaptation benefits are equitably distributed—is answered by documenting significant heterogeneity in CRS effectiveness across socioeconomic and demographic dimensions. While CRS reduces losses overall and performs better under higher precipitation for most communities, benefits concentrate in high-population, higher-education, and higher-income or predominantly white communities, particularly in higher-risk areas. These patterns imply that current adaptation prescriptions can inadvertently reinforce existing inequities and institutional racism if not designed with equity at the core. Policy implications include tailoring requirements and resources to community capacity, providing technical support to less-populated and lower-education communities, prioritizing interventions that reduce surface imperviousness in urban minority communities where precipitation sensitivity undermines savings, and subsidizing or funding high-cost but effective measures in low-income, high-risk communities. Integrating climate justice principles—explicit consideration of income, race/ethnicity, and capacity—into program goals and evaluations is essential to ensure that adaptation investments do not replicate historical patterns of exclusion.

Using a deep generative causal inference framework (CAUSAL.FLOW) on nationwide NFIP claims, the study shows CRS flood adaptation reduces losses but yields widely varying savings across communities, with systematic advantages for populous, higher-education, higher-income, and predominantly white communities. To achieve equitable adaptation, programs must tailor requirements and support, reduce technical barriers, and embed equity, race, and inclusion in design and evaluation. Future work will move beyond binary participation to quantify the impacts of specific CRS activities and guide communities on selecting effective interventions given their characteristics. Advances in deep generative modeling and expanded data collection can further resolve complex causal relationships critical for policy design.

- Coverage bias: Analyses reflect households with NFIP flood insurance; uninsured communities—disproportionately lower income—are not observed, likely understating inequities and total potential benefits gaps. - Conservatism in effect estimates: Savings are corrected for outreach/public information components and compared against controls that still meet minimal floodplain regulations, both of which reduce estimated treatment effects. - Capacity and elite capture: Evidence suggests outreach may enable better claim filing and, at very high educational attainment, potential strategic behavior (“gaming”), complicating attribution of savings purely to hazard mitigation. - Generalizability within support: CATE estimates are valid within the covariate support of the data; extrapolation beyond observed ranges is not warranted. - Data constraints: First Street Foundation risk data are under license; interview data are non-public; some community characteristics are measured at ZIP and ACS interval resolutions, which may introduce aggregation timing mismatches.