New insights into US flood vulnerability revealed from flood insurance big data

Flooding is the deadliest and most costly natural disaster in the US and worldwide, and losses are projected to increase substantially due to climate change and continued development in flood-prone areas. Translating physical flood hazard (extent, depth) into economic loss typically relies on depth-damage functions that assume a monotonic increase in loss with water depth and are often applied by occupancy and construction type. Widely used regional functions (e.g., USACE, FIA) have been treated as off-the-shelf tools despite substantial uncertainty and scatter noted in prior work, making the vulnerability component a key bottleneck in flood risk estimation. Meanwhile, flood hazard modeling has advanced rapidly with high-resolution data and computational gains, yet converting hydrodynamic outputs to dollars of loss remains constrained by depth-damage functions. Existing verifications have been limited by lack of building-level pairing of depth and damage, or by reliance on modeled hazard, and small-sample studies have revealed large variability and prompted multivariate approaches beyond depth alone. Despite growing recognition of weak correlation between depth and damage and calls for probabilistic treatment, monotonic central-tendency depth-damage curves remain standard. This study uses more than two million NFIP claims with recorded inundation depth and loss to empirically evaluate vulnerability, finding poor agreement with standard curves and showing that relative damages follow bimodal distributions consistent with beta distributions that shift toward higher damage with increasing depth. The study examines geographic variability, structure value and age effects, and differences inside versus outside FEMA flood zones.

Standard depth-damage curves have been developed and applied in the UK (e.g., the Multi-Coloured Manual) and US (USACE/FIA), typically specifying structure/content damage as a percentage of value at 1-foot increments relative to the first occupied level. Prior literature highlights large uncertainty in depth-damage relationships and questions their reliability, noting the vulnerability module as a dominant source of uncertainty in risk assessments and cost-benefit analyses. Verification efforts have been hampered by data scarcity, spatial aggregation, and missing matched hazard-intensity measures. Comparative studies often test different vulnerability models using modeled hazard/exposure rather than validating depth-damage curves themselves. Detailed local studies show considerable variability and have led to multivariate damage models incorporating factors such as velocity, duration, contamination, warnings, precautionary measures, building quality, and socioeconomics. Multiple studies argue for probabilistic treatment of depth-damage. Concurrently, flood hazard models have improved markedly, but economic loss translation remains constrained by unverified and uncertain vulnerability functions.

Data: Analyzed 2,085,015 NFIP claims (1972–2014) obtained directly from FEMA, including total structural damages (actual cash value), total structure value, and inundation depth above the lowest occupied floor as recorded by certified loss adjusters, plus building characteristics and location (lat/long rounded to 0.1° for confidentiality). Positive, non-missing entries for value, damage, and depth totaled 976,363; analyses focus on the most common building type—one-story residential without basement—with all required fields available for 493,707 claims. Depth units exhibit artifacts where some depths appear recorded in inches (spikes at 6, 12, 18, 24, etc.); approximately 5% of records affected. To minimize artifacts, analyses emphasized low integer depths <8 feet where feet records dominate. Relative damage was defined as structural damages divided by structure value. Analytical approach: Constructed empirical distributions of relative damage by 1-foot depth increments (1–8 ft) for the focal building class. Compared empirical distributions and medians with multiple USACE/FIA depth-damage curves. Assessed fit of beta distributions to the empirical relative-damage distributions at each depth (via Supplementary materials), and contrasted against the one-to-one federal curve approach. Estimated damages using federal curves (primarily FIA/common federal function) given NFIP structure values and recorded depths, and compared to observed damages using coefficient of determination (CoD = 1 − Σ(Dfed − Dnfip)^2 / Σ(Dnfip − D̄nfip)^2) and mean absolute errors. Scenario-based analyses: Stratified analyses by geography (state and ZIP Code Tabulation Area means at each depth), by building value strata, by construction age (pre/post-1950, post-1980), by NFIP management context (pre-FIRM vs post-FIRM; within vs outside FEMA 100-year floodplain), and by specific events within USACE Districts (e.g., Chicago July 1996; Tropical Storm Allison 2001; Hurricane Ike 2008; Hurricane Isabel 2003). For event analyses, applied corresponding USACE District curves to estimate event totals and building-level errors. Evaluation criteria: Distributional shape (bimodality), beta distribution fit (R^2 in Supplementary), central tendency comparisons, aggregate bias (over/underestimation by depth), CoD for predictive power (positive indicates federal curves outperform empirical mean), and mean absolute error as a fraction of mean observed damage.

• Empirical depth-damage distributions are bimodal and better represented by beta distributions than by single central-tendency values. With increasing depth, the beta distribution shifts toward higher relative damage, increasing the likelihood of near-total losses.
• Central tendencies mask wide variability: for one-story, no-basement residences, at 1 ft inundation the median relative damage is 11% with interquartile range 4–30%; at 5 ft, the interquartile range spans 19–90%.
• US federal depth-damage curves are internally disparate for similar structure types and do not represent empirical central tendencies or variability. Using the FIA curve to estimate claim damages yielded systematic bias: shallow-depth damages overestimated by ~25% and deep-depth damages underestimated by ~25% (aggregate).
• Predictive performance of federal curves is poor at building level: CoD is negative across depths, meaning the empirical mean outperforms the curve; mean absolute building-level error equals ~84% of mean recorded damage across depths and 105% at 1 ft.
• Regional variation: Empirical data contradict some regional USACE curve differences (e.g., St. Paul vs Chicago Districts). While some event-level totals are approximately matched by regional curves (e.g., Hurricane Ike estimated $447.0M vs NFIP $450.0M), building-level errors are large (mean absolute error ~$34,213, ~75% of mean damage; CoD negative). Within-district ZIP-level variability is far larger than state/district averages.
• Possible east–west divide with greater vulnerability per foot west of the Mississippi; coastal versus inland differences are modest in these data (saltwater vs freshwater claims show fairly insensitive relative damages), so analyses did not separate them.
• Structure value effects: Relative damages are lower for higher-value homes. At 1 ft depth, 0–10% damage occurred in 85% of claims for structures >$750k vs 45% for <$150k. At 8 ft, 0–10% damage occurred 62% (> $750k) vs 12% (< $150k); 90–100% damage at 8 ft was <1% (> $750k) vs 38% (< $150k), indicating many repair costs are value-independent.
• Building age and FIRM context: Newer buildings suffer lower relative damages for a given depth. At 5 ft, post-1980 homes had 0–10% damage in 22% of claims vs 12% for pre-1950; at 1 ft, 52% (post-1980) vs ~43% (pre-1980) had 0–10% loss. Within FEMA 100-year flood zones, post-FIRM homes are more likely to have minimal losses than pre-FIRM (e.g., at 1 ft, 51% vs 43% with 0–10% damage). Outside flood zones, the pattern reverses: at 1 ft, pre-FIRM 47% vs post-FIRM 42% with 0–10% damage; at 8 ft, post-FIRM 52% vs pre-FIRM 33% have 90–100% damage.
• Notably, at least 27% of NFIP flood claims and 14% of claim dollars are outside FEMA’s 100-year floodplain, underscoring limitations of binary flood mapping and elevated vulnerability outside mapped zones, especially for newer (post-FIRM) structures.

Findings indicate that the common assumption of a deterministic, monotonic central-tendency relationship between depth and damage is invalid for large populations of structures. Rather than a single curve value per depth, damages are bimodally distributed and are better captured by a depth-conditioned beta distribution whose parameters evolve with depth, concentrating probability at low and high relative loss. This has major implications for preparedness, response, recovery, mitigation, and for cost-benefit analyses that currently rely on standard curves. Existing federal curves tend to overestimate shallow inundation losses (which are most frequent) and underestimate deep-inundation losses, and they perform poorly in predicting building-level outcomes, even when event totals sometimes align by chance. Geographic and management-context heterogeneity (age, value, in/out of mapped flood zones) materially affects vulnerability, with evidence that post-FIRM construction inside flood zones is more resilient, while outside zones newer homes are more susceptible—highlighting pitfalls of binary risk communication. Moving toward probabilistic, multivariate vulnerability models grounded in large empirical datasets can improve risk estimates and policy decisions. Enhanced data access (more complete NFIP claims with precise locations and attributes) would enable refined modeling, better mapping beyond single-probability zones, and more effective floodplain management.

This study leverages a large NFIP claims dataset to empirically characterize US flood vulnerability, demonstrating that relative damages at given depths are bimodal and well represented by beta distributions rather than monotonic central-tendency curves. Widely used US federal depth-damage functions fail to capture observed variability and bias shallow vs deep losses, yielding poor building-level predictions. Vulnerability varies substantially by geography, structure value, age, and floodplain management context; post-FIRM resilience gains are evident within mapped floodplains but not outside them. The work advocates for probabilistic, multivariate vulnerability modeling and for moving beyond binary flood-risk mapping. Future research should integrate additional building attributes (e.g., wall/foundation type), hazard characteristics (duration, velocity), and higher-resolution geospatial data, supported by broader sharing of detailed NFIP claims to improve the validity of flood risk assessments.

NFIP claims data contain known errors and artifacts, including misrecorded depth units (some entries likely in inches), with an estimated ~5% of records affected. To mitigate this, analyses focused on low integer depths (<8 ft) where feet records dominate. Many records lack complete fields; only claims with positive values for depth, damage, and structure value were used, and the primary analyses focused on one-story, no-basement residential structures, which may limit generalizability. Location information was coarsened (coordinates rounded to 0.1°) and other fields are redacted for confidentiality, constraining spatial precision and replication. Some variables relevant to damage (e.g., velocity, contamination, detailed construction) were not analyzed. While some claims indicate freshwater vs saltwater flooding, relative damages were found fairly insensitive to this distinction here and were not separated. Observed regional patterns may reflect event-specific dominance in some states and warrant rigorous multivariate, multilevel analyses to avoid confounding.