Flooding causes substantial economic damage globally, particularly in the US. Climate change exacerbates this risk, necessitating effective and equitable flood adaptation measures. While existing programs like the FEMA's National Flood Insurance Program (NFIP) Community Rating System (CRS) aim to reduce flood losses, it is unclear whether their effectiveness varies across different communities and whether they exacerbate existing inequalities. The study's central research question focuses on understanding whether the effectiveness of flood adaptation policies varies across different community characteristics and how this variation relates to factors like income, race, and education. This research is crucial for ensuring climate justice and for the equitable allocation of limited resources for climate adaptation. It aims to prevent climate interventions from unintentionally perpetuating historical patterns of discrimination. The existing literature shows that low-income communities are disproportionately vulnerable to flooding and often lack adequate resources for adaptation. This study builds upon this existing work by quantitatively assessing the distribution of benefits from flood adaptation across various community types, offering a critical evaluation of the current CRS program as a case study.

Previous research has explored the effectiveness of the CRS in reducing flood losses, with mixed results. Some studies found reductions in flood claims while others did not. However, there's limited investigation into whether the program's effectiveness varies across different community types. Existing studies also highlight the disproportionate vulnerability of low-income communities to flood risks and the necessity for pro-poor climate adaptation initiatives. However, no prior work systematically quantifies the differential benefits of flood adaptation activities across various community characteristics.

This study utilizes the FEMA NFIP Redacted Claims dataset, containing approximately 2.5 million flood insurance claims, providing a quasi-experimental setup to compare flood losses in CRS participating and non-participating communities. The researchers use a novel data-driven method called CAUS.FLOW, which leverages deep generative models to measure the causal effect of flood adaptation policies, accounting for non-linear and complex relationships between community characteristics and outcomes. CAUS.FLOW estimates the conditional average treatment effect (CATE), the average treatment effect as a function of community covariates. The researchers define 27 distinct community typologies based on population, income, and racial composition (low, medium, and high for each characteristic), creating all possible combinations. For each typology, the CATE is computed for varying levels of average precipitation, flood risk, renter fraction, and educational attainment. Data on community characteristics are sourced from the U.S. Census Bureau American Community Surveys and the First Street Foundation dataset for flood risk scores. The outcome variable is the total insurance claims payments per policy. The authors describe CAUS.FLOW in detail, explaining how it uses normalizing flows models to estimate flexible transformations that map a complex target distribution (flood claims) to a simple base distribution (Gaussian), allowing for accurate estimation of CATE without strong assumptions about functional forms. The software for CausalFlow is publicly available.

The CRS program saves communities between $5,000 and $15,000 per household on average. However, this benefit is not evenly distributed. Savings sharply decline for low-income communities, especially in high-risk flood zones, contrasting with the trend observed for high-income communities. Even within low-income communities, predominantly white communities see significantly higher savings (>$6,000 per household) than predominantly minority communities. The effectiveness of flood adaptation measures increases with higher average precipitation for most community types, indicating the program’s effectiveness in mitigating flood losses. However, communities with high percentages of racial and ethnic minorities, high population, and mid-to-high income experience a decrease in savings with higher precipitation. These communities are mostly located in urban areas and have higher flood risk scores. High-population communities benefit from approximately $4,000 more in savings per policy than low-population communities, likely due to greater resources and capacity. Communities with higher educational attainment also see greater savings, possibly due to increased capacity to implement CRS activities. The study reveals a racial gap in savings exceeding $6,000 per policy in some cases, even when comparing low-income communities with similar characteristics except for race. These disparities suggest that existing flood adaptation programs may perpetuate institutional racism and inequalities.

The findings highlight the limitations of current flood adaptation programs, which, while effective overall, fail to provide equitable benefits across all communities. The uneven distribution of savings across communities underscores the need for tailored interventions that consider specific community characteristics and address systemic inequalities. The results show that a purely loss-reduction approach is insufficient; equitable distribution of benefits must be a central goal in designing and evaluating flood adaptation strategies. The observed disparities are linked to income, race, population, and education levels, illustrating the complex interplay of social and environmental factors affecting flood resilience. This research advocates for a shift towards a climate justice framework, emphasizing proactive measures to address existing patterns of inequality and discrimination.

This study demonstrates that while current flood adaptation practices reduce flood losses, the benefits are not equitably distributed. Future adaptation strategies must consider community-specific needs and resources, integrating equity priorities to ensure all communities benefit from climate adaptation investments. Future research should explore the impact of specific CRS activities, analyze the effects of uninsured communities, and further investigate the role of the outreach component in the observed disparities. The development of deep generative models offers opportunities for more nuanced causal inference in designing future policy interventions.

The study's analysis is based on households with flood insurance, excluding uninsured communities, who may experience even greater disparities in flood losses. The savings estimates may be conservative due to methodological corrections and the inclusion of communities in the control group that already meet minimal floodplain regulations. The study acknowledges the possibility of elite capture, where some households may be manipulating the system to their advantage, warranting further investigation.