RESEARCH DOSSIER
Flood Depth Damage Functions and Property Vulnerability
Abstract. Flood depth-damage functions translate inundation intensity into estimated physical damage and monetary loss. Their apparent simplicity conceals substantial scientific and institutional questions: what precisely constitutes water depth; which value forms the denominator of a loss ratio; how building configuration, elevation, material, contents, duration, contamination, velocity, flood origin, and precaution modify damage; and how an engineering damage estimate becomes an insurance claim under a particular contract. This monograph critically reconstructs the foundations, estimation methods, validation evidence, applications, and limits of depth-damage functions through international official documentation and peer-reviewed empirical research. It compares the United States' USACE and FEMA Hazus approaches, the European Commission Joint Research Centre's global functions, German and Dutch multivariable research, and large-scale U.S. flood insurance claims evidence. It shows why average conditional damage functions remain useful for aggregate decisions while heterogeneity, selection, uncertain valuations, and shifting building stocks limit single-property prediction. Two transparent hypothetical accounting illustrations separate damage ratios, replacement values, and coverage transformations. The central scientific conclusion is that defensible flood-vulnerability assessment requires agreement among the hazard reference level, asset inventory, valuation basis, empirical sample, model estimand, and intended insurance decision. This is an original critical synthesis of existing science; it establishes no new theory, proprietary damage model, or causal estimate.
Keywords: Flood depth-damage functions; flood vulnerability; property insurance; inundation depth; stage-damage curves; building elevation; first-floor height; contents damage; FEMA Hazus; USACE; Joint Research Centre; flood insurance claims; loss ratio; exposure valuation; model validation; uncertainty; mitigation; transferability; flood risk management.
Introduction
Flood risk assessment connects a distribution of possible water conditions with the assets exposed to those conditions and the consequences of inundation. A flood hazard model may specify water surface elevation, inundation extent, velocity, duration, or annual probability. An exposure model identifies buildings and their characteristics. A vulnerability model describes damage conditional on water and asset characteristics. An economic model attaches prices to affected components. An insurance model applies policy terms to eligible loss. A single estimate of “flood damage” can conceal all of these operations.
The intellectual history of the field explains the importance of the damage function. Hydraulic flood maps can be exceptionally detailed, yet an economically consequential decision often depends on how a depth increment translates into repair cost. Merz, Kreibich, Schwarze, and Thieken (2010) demonstrated in their review of economic flood damage that data scarcity, inconsistent economic definitions, incomplete validation, and neglected uncertainty can dominate interpretation. The review remains a methodological starting point because it treats damage assessment as a substantive scientific problem rather than an automatic extension of hydrology.
The European Commission's global functions (Huizinga, de Moel, and Szewczyk, 2017) seek comparability across regions, especially where local data are absent. The USACE generic relationships (2000) and FEMA's Hazus flood manual (2025) embody distinct U.S. engineering and loss-estimation traditions. Their value lies in explicit asset categories and documented calculation rules. Their transfer requires attention to source populations, program purposes, and function vintages.
Empirical findings challenge the temptation to interpret a mean depth-damage curve as a deterministic law for a particular building. Wing, Pinter, Bates, and Kousky (2020), in their analysis of U.S. flood claims, examined 2,085,015 NFIP claim records from 1972–2014; 976,363 contained the variables required for their main depth-related analysis. They found broad, often bimodal conditional damage distributions and important departures between widely used curves and observed claim-based relationships. These observations imply that interpretation must distinguish conditional population averages from the full distribution of outcomes and from the progression of damage at one fixed property as water rises.
This dossier addresses three research questions. First, how are depth-damage relationships defined and estimated, and which data and assumptions support them? Second, how do building elevations, contents, and asset characteristics alter physical and monetary loss? Third, how should model uncertainty, validation, engineering mitigation, and contractual transformations be evaluated when results guide insurers, public authorities, and property owners?
The method is comparative examination of named first-party technical documents, legal texts, and original scholarly publisher records, combined with mathematical reconstruction and explicitly hypothetical numerical examples. The evidence record distinguishes accessed institutional manuals and peer-reviewed articles from illustrative calculations. No restricted insurer dataset, proprietary catastrophe model, individual claim file, or newly collected field survey was accessed. All claims about empirical samples belong to the cited authors and their documented observation periods; their numerical findings are not presented as replicated calculations.
Definitions, Historical Context, and Primary Source Foundations
The measurement object
A depth-damage function relates a defined water-depth variable to a damage measure for a defined asset population. An absolute function produces a monetary amount per building, floor area, unit, or land-use area. A relative function produces a fraction of a particular asset value. The function can be tabulated, piecewise linear, parametric, empirical, expert-elicited, or produced by a multivariable statistical procedure. A graph without units, reference elevations, value definition, and building class is incomplete as a scientific object.
Let H denote the event water-surface elevation relative to a stated vertical datum, and let F denote the elevation of a building's relevant floor above the same datum. For a simple above-floor exposure, internal water depth is d = max(0, H − F). This identity is a bookkeeping convention, not a structural flood simulation. It is inappropriate where groundwater, basement ingress, unequal water levels, barriers, drainage, or pressure-related failure change the relevant path of inundation. Exterior depth above surrounding ground, depth above the first finished floor, and depth above the lowest occupied floor are not interchangeable.
The standard relative-loss expression is D = V × r(d), where D is direct damage cost, V is the valuation base, and r(d) is the estimated damage fraction for the same type of asset and valuation convention. For separate building and contents components, D = V_B r_B(d) + V_C r_C(d) applies only when both r terms are fractions of their respective values. A different convention appears in USACE EGM 01-03: its specific residential content damage functions express contents loss as a percentage of structure value. Applying such a function again to contents value would create a denominator error. This distinction is consequential wherever functions are exchanged between software systems or converted to different inventories.
Damage may mean repair cost, replacement cost, depreciated asset loss, lost production, restoration expense, or a claims settlement. For a damaged building, replacement or repair values can be measured at pre-event, event, or reconstruction prices; costs can include or omit debris, professional fees, taxes, upgrades, and additional mitigation. The Merz and colleagues assessment documents how spatial, temporal, and economic boundaries influence the meaning of flood-loss numbers. A valid model must report these boundaries together with the monetary output.
Historical development of standardized curves
Early depth-damage approaches offered practical rules for converting mapped flood elevations into local loss estimates. They were often assembled from observed flood losses, post-event surveys, claims records, or expert judgments for common building classes. Generic functions were attractive because direct inspection of every vulnerable building is expensive, especially for regional planning.
A precise historical anchor is the U.S. Army Corps of Engineers' December 4, 2000 Economic Guidance Memorandum 01-03. Its documentation identifies U.S. flood losses collected in 1996–1998, differentiates one-story, multistory, and split-level residential buildings without basements, and reports that quadratic and cubic depth-only relationships were efficient for its underlying regression data, even though additional variables had been investigated. The memorandum also states that certain cleanup, unpaid labor, and emergency prevention costs were outside the generic functions. These are properties of the documented dataset and application, not an empirical claim that depth is everywhere the only important causal determinant.
Later research expanded the explanatory dimensions. Thieken, Müller, Kreibich, and Merz (2005), analyzing 1,697 German household interviews after the August 2002 flood, found relationships involving depth, duration, and contamination, as well as building characteristics and precaution. Wagenaar, de Jong, and Bouwer (2017) used a Dutch flood dataset to study multivariable supervised learning under data constraints. The progression illustrates an empirical refinement: identifying additional covariates may improve prediction, while simultaneously increasing data, validation, and portability requirements.
International institutional frameworks
The JRC 2017 technical report constructs geographically broad damage relationships using normalized curves and country-specific maximum damage values. Its methodological rationale explicitly acknowledges that national models differ in source assumptions and are not automatically comparable. Normalized damage curves and country-valued maxima address different portions of the same calculation. Continent-level grouping is a practical approximation in the absence of fully local functions; it is not proof of invariant construction vulnerability within a continent.
The Hazus 7.0 Flood Model Technical Manual (FEMA, June 2025) specifies damage functions and economic-loss calculations for baseline aggregated building inventories, user-supplied inventories, essential facilities, and other categories. Functions are assigned using the hazard and properties such as occupancy, stories, and foundation type. The 2025 document is a dated methodological source and must be identified as Hazus 7.0, not represented as a timeless or automatically current software release.
Flood-risk mapping and planning are embedded in distinct legal frameworks. The European Union's Floods Directive 2007/60/EC establishes risk assessment, mapping, and management planning obligations in its legal scope. It does not prescribe a universal international depth-damage formula. De Moel, van Alphen, and Aerts (2009), in their study of European flood maps, documented methodological and availability differences among national mapping practices. Legal obligations concerning flood-risk management and scientific assumptions concerning asset losses must therefore be recorded as separate propositions.
In U.S. federal flood insurance, the NFIP Dwelling Form is a defined insurance contract rather than an engineering damage standard. Its operative language identifies insured buildings, personal property, direct physical loss, limitations, deductibles, and settlement rules. The building and contents coverage provision confirms that these types of coverage can be purchased separately. The contract establishes which part of physical damage can give rise to insured benefits. International applications require the corresponding contract actually governing the risk, since a U.S. federal policy is not a global insurance form.
Distinguishing the economic layers
The following chain is analytically essential: hydrologic conditions → exposure → physical alteration → repair and recovery cost → contractually covered loss → assessed insurer obligation → actual payment. Each arrow represents a set of measurements or rules and may alter both magnitude and classification. For example, temporary displacement can generate substantial economic consequences even when direct building damage is modest. In another case, severe physical damage can exceed an insurance limit. A claim dataset may reflect exclusions, deductions, reporting and adjustment practices rather than all physical losses.
The definition of the estimand should precede calibration. A function fitted to insured paid claims estimates a relationship between recorded payments and included predictors for an observed, selected insured population. A function fitted to engineer-verified reconstruction estimates may target direct physical repair cost. Neither function can automatically substitute for the other without accounting for coverage conditions, measurement, and sample selection.
Depth-Damage Curve Estimation
Data architecture and observational units
Construction of a depth-damage curve begins with selection of an observational unit. Building-level records preserve specific elevations, construction, and contents; neighborhood or census-block records enable broader coverage while averaging heterogeneous structures. Land-use-class estimates are sometimes the only feasible format for international or continental assessment. Each aggregation level changes the possible validation design and interpretation of error.
A defensible building-level dataset records the event identifier; georeferenced building location or secure linkage key; hazard source, duration, and vertical datum; exterior and internal water elevations; first-floor and basement geometry; construction, occupancy, age, stories, foundation and utilities; damage status; inspected damaged components; cost estimate and pricing date; contents values; mitigation; and insurance coverage or settlement fields where used. Missingness indicators and source provenance are part of the dataset, especially when administrative inputs were collected for purposes other than research.
Observed records may be produced by engineering inspections, household surveys, municipal recovery programs, insurance adjusters, contractors, or administrative payments. Each process selects a different subset of events and damaged assets. The use of actual claims is valuable but does not eliminate ascertainment effects. Wing and colleagues (2020) report data-quality and unit anomalies in the historical NFIP material, including some depth fields apparently recorded in inches while labeled as feet. Before fitting a function, the dataset needs unit checks, bounded-value checks, depth-elevation reconciliation, duplicate resolution, and a documented exclusion log.
Empirical, synthetic, and hybrid derivation
An empirical curve uses observed outcomes and an estimation method. In its simplest form, observations are assigned to depth bins and a mean or median damage fraction is calculated per bin. The bin estimate is conditional on the observed portfolio composition, reporting process, denominator, and event sample. Narrow bins may be noisy; broad bins can smooth physical transitions such as water entering living spaces or reaching electrical equipment.
A parametric representation may fit a polynomial, spline, generalized additive model, or another restricted function. It requires control of extrapolation, physical plausibility, and the valid depth domain. A cubic fitted to observed depths can behave implausibly outside that interval. A constrained curve may enforce bounds between zero and unity for a consistently defined replacement-value ratio; its interpretation remains conditional on the chosen denominator and whether repair costs include components exceeding replacement value.
A synthetic curve estimates component vulnerability and costs through engineering assumptions or expert elicitation. It can be useful where recorded losses are absent, but assumptions should be explicit for floor finishes, wall systems, electrical systems, mechanical equipment, contents movement, repair thresholds, and post-flood cleanup. Expert functions should be checked against independent inspection records when available. Combining empirical survey evidence, component estimates, and expert review forms a hybrid method with its own traceable assumptions.
In the United States, USACE's EGM 01-03 provides an actual example of empirical regression converted into standardized planning relationships. In contrast, the JRC global methodology combines a literature survey and standardized normalized curves with values for asset and land-use classes. A researcher comparing their outputs must first align whether each source estimates a fraction of structural value, replacement cost per area, or a combined asset amount.
Conditional mean, median, and distribution
A common but insufficient interpretation is that a building flooded to depth d “suffers” exactly r(d) of its value. The statistically appropriate interpretation is often that r estimates an average or typical conditional loss for a stated population. A conditional mean and conditional median can differ substantially when damage is skewed, zero-inflated, or near-total for a subset of buildings.
The NFIP claims analysis by Wing et al. (2020) demonstrates this distinction directly. At an inundation depth of five feet, the reported interquartile range of relative damage was approximately 19% to 90%. These are the authors' empirical statistics for the analyzed claims and their specific valuation measure, not a transferable predictive interval for every insured building worldwide. The same study reports multimodal conditional distributions and highlights mismatches with several standard functions for nominally similar residential classes.
Monotonicity requires careful statement. For an individual fixed building with unchanged contents and without intervening actions, additional inundation can reasonably be expected to enlarge the set of damaged components under a comparable event. Across observed buildings at increasing depths, however, building value, composition, materials, mitigation, location and claim selection change. A cross-sectional estimated mean need not rise monotonically in observed data. A population comparison and an individual counterfactual answer different scientific questions.
The use of a probability distribution conditional on depth and assets can represent this dispersion. Standard statistical choices include mixtures, bounded continuous distributions, and separate models for the probability and severity of damage. These are established modeling families, not a newly proposed doctrine of flood vulnerability. The evidence for choosing among them must come from withheld samples and the decision-relevant loss function, rather than visual closeness to one dataset.
Multivariable functions and model complexity
The case for multivariable modeling is physical and statistical. Water depth interacts with floor height, basement presence, contents placement, structural materials, duration, contamination, warning, protective works, and event type. In a sample, observed associations can also arise from the correlations among these characteristics. A better-fitting multivariable function is not automatically a causal estimate of any one characteristic.
Wagenaar, de Jong, and Bouwer (2017) assessed supervised flood-loss estimation with enriched data from the 1993 Meuse flood. They reported a 20% reduction in mean absolute error relative to a depth-only baseline on their test set, with tree-based approaches outperforming the Bayesian network in that study. The improvement is specific to their data and validation setting. Its generalizable lesson is methodological: additional variables can carry predictive information, yet their utility depends on data quality and the development population.
Wagenaar, Lüdtke, Schröter, Bouwer, and Kreibich (2018) investigated regional and temporal transferability using multivariable models and datasets from Germany and the Netherlands. Their comparison focuses attention on validation outside the training domain. A function that performs well for one event, region, or claims mechanism can change markedly under another; the role of time must include construction changes, mitigation learning, rebuilding-cost changes, and the composition of affected buildings.
Model flexibility imposes an evidence cost. Trees, random forests, and other supervised techniques can capture nonlinearities, but their output also depends on preprocessing, missing-value conventions, categorical recoding, hyperparameter selection, and leakage prevention. Where only one major flood has been observed, randomly dividing individual claims into training and test sets risks placing similar event-specific conditions in both samples. Event-level and region-level validation are stronger tests of operational transfer.
Curve estimation versus annual risk integration
A damage function estimates conditional consequences given a hazard realization; it does not by itself establish how frequently that event occurs. Expected annual damage requires integrating event-specific loss with a suitable annual hazard or event distribution. If scenarios represent mutually exclusive annual event states with probabilities p_i, then E[D] = sum(p_i × D_i) across the complete specified states. If the input is instead a set of exceedance probabilities, simply multiplying each exceedance probability by the corresponding loss and summing will double count nested events. An exceedance curve requires coherent integration over nonoverlapping probability increments or an equivalent continuous formulation.
USACE's flood risk management methods explicitly distinguish uncertainty in flood levels and damage consequences and place economic estimates in a risk-based decision setting. For insurers, a property-level expected damage estimate may feed catastrophe event loss distributions and portfolio accumulation. Dependence across properties arises through shared weather and hydrologic events; it cannot be inferred from separate building depth-damage curves alone.
Asset Contents and Elevation
Elevation as a physical boundary condition
The vertical relationship between water and the building's vulnerable components often dominates first-order damage estimation. The same floodwater surface can remain beneath one elevated living floor and submerge another. A first-floor elevation recorded relative to local ground must be combined with a flood depth measured above that same ground, or both must be converted to a common datum. Mixing local ground reference, national elevation datum, and floor-height convention produces a systematic error rather than random model noise.
The Hazus flood methodology uses flood depth and building inventory attributes to estimate structure damage, and its user-supplied inventory approach treats structure location and first-floor-related variables as relevant input. A parcel centroid or coarse grid may miss abrupt topography or local protection around a specific building. Where a basement exists, internal flood exposure may begin below the first finished floor; applying an above-floor-only curve can underestimate early mechanical, storage, and foundation-related losses.
Floor elevation has more than geometric significance. It determines when floodwater reaches construction assemblies, utilities, doors, vents, electrical outlets, cabinets, finishes, and contents. The order of component wetting depends on building design. Below-floor spaces can also experience structural loading, buoyancy, and water ingress before the ground-level living space is affected. A building classification that does not capture these distinctions may mask large differences at the same modeled exterior water depth.
The ASCE/SEI 24-24 standard addresses minimum flood-resistant design and construction requirements for structures within its applicable code and regulatory scope. Its subject includes elevations, materials, and floodproofing. It is an engineering design standard, and applicability to an existing building depends on the jurisdiction, adopted building code, construction date, and alteration or substantial-improvement conditions. Its presence in the bibliography is evidence of a recognized engineering control, not proof that a given international portfolio is legally subject to it.
Buildings and contents as different value populations
Buildings, contents, machinery, inventory, and business interruption should be separated whenever their value bases, fragility, location, and coverage differ. Contents can be moved to upper floors in advance of flooding; built-in equipment is less mobile. Residential personal property and commercial stock have different storage practices and restoration costs. Commercial buildings may contain expensive equipment with high failure sensitivity even at relatively low water depths.
The U.S. Standard Flood Insurance Policy distinguishes building property and personal property, with special restrictions concerning certain items in basements and below the lowest elevated floor. In the science of loss modeling, this illustrates why physical contents damage and covered contents damage can diverge. It does not establish an international coverage rule. A model intended for a different country must use the relevant policy wording and applicable local law.
To avoid denominator confusion, every tabulated loss ratio should indicate a numerator and denominator: building repair cost divided by building value; contents replacement cost divided by contents value; contents damage divided by structure value; or total direct loss divided by combined insured value. The USACE memorandum's content convention is especially instructive because its generic content curves were modeled against structure value, unlike conventional contents-value proportions. Reusing the numerical curve without its definition can bias the result by the content-to-building valuation ratio.
Building values and local reconstruction costs
A monetary vulnerability function depends on the value definition. Replacement value reflects costs of restoring or replacing an asset using the specified current price basis. Depreciated value reflects a different economic interest. An assessment based on sales prices can incorporate land and market premiums irrelevant to building repair. An inventory based on floor area and country-average construction costs may be suitable for macroeconomic comparison but require corrections for high-value materials and highly specialized buildings.
The JRC damage database expressly separates fractional curves from country-level maximum damage values and uses construction cost surveys and socioeconomic variables to develop a globally consistent reference. This design makes comparisons possible where direct loss inventories are limited, while also creating sensitivity to assumptions about current building materials, market costs, price levels, depreciation, and the represented asset classes.
Reconstruction pricing after major disasters may differ from ordinary replacement-cost schedules because labor, materials, contractors, permits, drying capacity, and logistical constraints are affected by widespread demand. If a damage function is calibrated in one price year and applied to another, analysts should record the cost-index conversion and any real changes in rebuilding practice. The engineering damage ratio itself may not be stable under a material or labor shortage when the numerator is a cost rather than a count of physically damaged components.
Floodproofing and component placement
Flood damage reduction can be pursued by elevating vulnerable floors and equipment, selecting materials tolerant of wetting, using flood-adapted interiors, relocating contents, managing openings, or designing controlled water entry in suitable spaces. Effects depend on flood intensity, installation quality, maintenance, and the kinds of damage measured.
FEMA's wet floodproofing guidance addresses the use of flood-damage-resistant materials and associated restrictions under applicable U.S. regulatory arrangements. The dry floodproofing technical bulletin describes requirements and certification for applicable nonresidential structures. Dry floodproofing assumes the envelope and systems can resist specified loads and leakage conditions; implementation must be engineered for credible depth, duration, debris, and warning scenarios. A nominally sealed building may remain vulnerable to seepage, utility connections, doors, and failure beyond its design level.
Kreibich, Thieken, Petrow, Müller, and Merz (2005) found from approximately 1,200 households affected by the 2002 Elbe flood that flood-adapted building practices were associated with substantial reductions in building and contents damage ratios. Their paper reported 46% and 53% reductions in building damage ratios for certain measures, and 48% and 53% reductions in contents ratios for the corresponding categories. These are observational findings from the stated study population, with selection and implementation contexts. They cannot be treated as guaranteed percentage discounts for every property.
Transparent hypothetical illustration: building and contents
Consider a hypothetical single residential building with a defined building replacement value of $350,000 and separately measured contents replacement value of $90,000. At a specified internal water depth of 1.00 meter, suppose illustrative, invented functions assign a 0.16 building damage fraction and a 0.40 contents damage fraction. The direct modeled physical loss would be $350,000 × 0.16 + $90,000 × 0.40 = $56,000 + $36,000 = $92,000.
Now suppose a hypothetical physical intervention changes the relevant internal depth and component exposure, and an expressly assumed alternative evaluation produces damage fractions of 0.09 for the building and 0.19 for contents. The illustrative loss becomes $350,000 × 0.09 + $90,000 × 0.19 = $31,500 + $17,100 = $48,600. The difference is $43,400, or approximately 47.17% of the initial modeled loss. Both sets of fractions were invented solely to demonstrate accounting and are not taken from Hazus, USACE, JRC, or an observed mitigation study. This is a mathematical example, not a causal claim about elevating a particular type of property.
For an alternative historical accounting convention, suppose a USACE-style contents function expressed contents damage as 0.12 of structure value. Its contents damage would be $42,000, calculated as $350,000 × 0.12. Multiplying 0.12 by $90,000 would instead yield $10,800 and use the wrong denominator for that assumed definition. The illustration shows why metadata about a function may matter as much as its tabulated coefficients.
Model Uncertainty and Mitigation
Sources of uncertainty
Flood-loss uncertainty has several distinct origins. Hazard uncertainty concerns rainfall and flow inputs, terrain representation, defense performance, water levels, velocity, duration, and the probability of particular events. Exposure uncertainty concerns whether an asset exists and its precise characteristics and value. Vulnerability uncertainty concerns damage conditional on those conditions. Economic uncertainty concerns repair scope and prices. Insurance uncertainty concerns policy coverage and claim handling. These uncertainties need separate description, even when they are ultimately combined in a simulation.
Some uncertainty is reducible by better data. Accurate surveys of first-floor heights and building characteristics may reduce systematic exposure errors. Verified post-event depths can improve calibration. Other uncertainty persists because apparently similar buildings respond differently, especially when contents and precaution differ. Extreme events also create conditions rarely represented in ordinary historical claims samples.
Merz and colleagues (2010) emphasize model validation and uncertainty analysis as fundamental requirements. The USACE risk-management background guidance specifically recognizes economic uncertainty from imprecise stage-damage relationships and property values. A precise hydraulic map combined with an unverified damage function remains an uncertain economic forecast.
Structural and parameter uncertainty
Structural uncertainty concerns which explanatory relationship is appropriate: a single depth curve, multiple grouped curves, a component-engineering function, or a multivariable statistical method. Parameter uncertainty concerns the estimated numerical values inside that chosen form. Model validation must avoid confusing these with the natural dispersion of otherwise unobserved repair outcomes.
Uncertainty bands derived from sampling variability do not automatically cover model misspecification, new construction types, unprecedented flood duration, or local cost escalation. Similarly, an ensemble of several depth-damage functions does not guarantee a statistically calibrated predictive interval if the curves share the same source assumptions. Independence must be examined at the data and methodological level.
The cross-country transferability study by Wagenaar et al. (2018) provides direct motivation for testing models outside the place and period of estimation. A geographical or temporal holdout reveals whether a curve's assumptions survive differences in asset composition, hazard, and measurement. When data are sparse, external validity should be stated with controlled scope rather than asserted from performance on the development event.
Mitigation as a change to vulnerability and exposure
Pre-event risk reduction can alter the frequency with which water reaches a building, the depth experienced at vulnerable components, the expected loss conditional on water exposure, or the value placed at risk. These are different mechanisms. A levee or drainage project may change hazard at the property, raising a floor changes relative depth, moving equipment changes component vulnerability, and relocating the entire building changes exposure. Counting the same mitigation effect in both a hazard layer and a damage curve would double count benefit.
A comparative study of eight paired floods by Kreibich and colleagues (2017) found that adaptation and preparedness changes contributed to smaller consequences in later events, with combinations of hazard, exposure, and vulnerability changes. The analysis supports the importance of behavior and institutional learning. It also demonstrates why a lower realized loss after a later flood is not, without adjustment, an estimate of a single intervention's independent causal effect.
The German work by Thieken et al. (2005) connects household damage with precaution and building characteristics. The more directly mitigation-focused Elbe household study offers evidence for flood-adapted interiors. Insurer product design may encourage such measures, but the effect of insurance on precaution is institutionally contingent. Thieken, Petrow, Kreibich, and Merz (2006) documented mitigation and insurance interactions in Germany, including differences in recovery and incentives.
The coverage transformation
Engineering mitigation can reduce physical loss without producing an identical change in insured paid loss. Coverage limits, deductibles, exclusions, valuation provisions, and distinct building and contents limits reshape the outcome. Where a loss is under the deductible, a physical improvement may have little effect on insurer payment while substantially benefiting the property owner. Where a covered event already exhausts a fixed limit, reducing damage may not change insured payment until the expected damage falls below the limit.
The NFIP's Dwelling Form policy language provides a concrete legal illustration of separately regulated building and personal-property benefits. In insurance models outside the NFIP, the contract may define flood differently, contain sublimits, or cover different categories. A physical flood-vulnerability function should therefore remain separately auditable from the software applying policy terms.
A generic illustrative transformation is P = min(max(C − A, 0), L), where C is the contractually eligible loss amount, A is a hypothetical per-loss deductible, L is a hypothetical payment cap after application of that deductible, and P is the resulting payment under that simplified assumed contract. Real contracts may apply percentage deductibles, multiple limits, reinstatements, salvage, coinsurance, expense provisions, or different calculation orders. This algebra is an illustrative bookkeeping identity under specified assumptions, not a legal statement of general insurance practice.
Empirical Verification, Uncertainty, and Competing Explanations
What constitutes a validation test
A meaningful test begins by declaring the prediction target: conditional building physical repair cost, combined building-and-contents direct loss, property replacement-value fraction, or insured paid claim amount. A model calibrated against one and scored against another answers an equivocal question. Equally important are the evaluation population, event horizon, source uncertainty, and tolerance relevant to the decision.
At the individual-property scale, absolute error can be measured in currency or as a ratio. At the event scale, aggregate bias compares total predicted direct losses to independently observed direct losses over the same affected population and cost definition. For catastrophe decisions, underprediction of severe loss, calibration of conditional loss distributions, and ranking across geographically correlated risks may be more important than an average residential absolute error. The chosen scoring rule should be explained rather than presented as a universal ranking of all models.
Seifert, Kreibich, Merz, and Thieken (2010) examined commercial flood-loss model validation, including leave-one-out validation at the micro scale and comparison with observed loss records across 19 German municipalities at a meso scale. The study reports that the unit of validation and spatial aggregation affect interpretation. Good aggregate performance can coexist with substantial property-level error because over- and underestimates offset when summed.
Data-selection and reporting mechanisms
Insurance claim records capture a selected process. A building must be insured under the relevant product, experience a covered or reportable event, be reported or otherwise registered, and pass the system's administrative rules. Some records contain zero payment; some losses remain below deductibles; some building damage has no insurance; and some costs are allocated to different claim categories. Conditioning on recorded claims can distort the relation between depth and damage in the whole exposed population.
The NFIP claims study offers strong large-sample evidence but also highlights incomplete variables, unit anomalies, and the distinction between total assessed structure value and capped insured value. Its observed population is U.S. NFIP claims rather than the full stock of U.S. flooded buildings. Its authors' conclusion about inadequacy of several commonly used curves is a specific empirical challenge to existing practice, not a claim that all depth-based relationships are useless.
FEMA's NFIP data FAQ clarifies that publicly redacted claims and policy datasets are anonymized, not generally linkable at individual policy level, and unsuitable as direct address-level catastrophe-model inputs. FEMA also notes that some data fields of low confidence were excluded from public products. The official claims dataset catalog identifies the administrative provenance of the redacted data. These first-party statements constrain what can be claimed from open downloads. This monograph does not assert access to the nonpublic depth-enriched historical dataset analyzed by Wing et al.
A validation study based solely on paid-claim amount must separately account for coverage values, deductibles, settlement behavior, claims status, policy vintage, and missing building records. Otherwise the model can confuse insurance conditions with physical fragility. The most persuasive validation pairs engineer-inspected damage and measured water conditions with consistently valued exposure, then separately validates the insurance transformation on matching contracts.
Out-of-event and cross-region validation
Flood events differ in hydrologic formation and duration. Claims from one major flood may share unrecorded contamination, emergency response, reconstruction prices, local buildings, and adjustment practices. Therefore, a random row-level split can overstate generalization if both training and test records represent the same event-specific environment. Holding out entire events, municipalities, building typologies, or years more closely tests real deployment.
The 2018 transferability analysis specifically addresses this problem across regions and periods. The 2017 Dutch multivariable study establishes that added covariates can improve one tested dataset; the paired reading prevents that result from being promoted into a general statement that more elaborate models always outperform locally calibrated simpler curves.
Repeated-event studies further reveal adaptation. When a second flood produces lower losses, a model fitted only on the earlier period may be temporally miscalibrated because warnings, construction, protective works, and contents use have changed. Kreibich et al.'s international paired-event comparison shows how societal and organizational changes interact with physical hazard. This motivates event-vintage documentation, especially in portfolios exposed to repetitive flooding.
Competing interpretations of monotonic damage
One interpretation of depth-damage curves is mechanical: as water reaches additional components, direct physical damage increases for the same unchanged building. Another is statistical: a curve summarizes conditional losses across buildings affected at different depths. A third is administrative: a recorded claim amount changes with insurance, valuation, and reporting. These are different propositions, so disagreement between them need not be an error in the underlying observations.
Wing et al. (2020) document nonmonotonic empirical cross-sectional patterns and substantial conditional dispersion. The result motivates richer descriptions of conditional losses, particularly for individual-property and tail-risk decisions. It does not refute the physical intuition that progressively deeper water can reach more components in an unchanged building. The correct unit and counterfactual must be declared before interpreting a statistical curve as an engineering mechanism.
Other competing explanations include low-value buildings having greater proportional loss because some repair costs are fixed, high-value buildings having expensive but flood-resistant lower levels, protective measures being concentrated in high-hazard zones, and insured-claim sampling excluding unreported low losses. Such explanations should be treated as hypotheses for stratified data or natural experiments; an observed association with property value does not identify a uniquely causal mechanism.
Reproducible minimum audit
For a research-grade function or purchased vendor model, the minimum transparent audit includes provenance and edition of the damage function; the observed population and years; definitions of depth and reference datum; asset classification; denominator and price basis; handling of zero loss and total loss; missing observations and outliers; curve fitting and restrictions; external validation design; residual distributions or error measures; limits of transferability; and documentation of the insurance policy transformation.
Source-to-claim review is separate from computational reproducibility. A paper may be correctly attributed but empirically limited to one region. A curve may be numerically reproduced while relying on a mismatched building class. A software test may pass despite a scientifically unjustified change in the target population. For a defensible claim, bibliographic, conceptual, statistical, and computational conditions must all be met.
Integrative Comparative Scientific Synthesis
Comparison across the principal source traditions
The primary source traditions answer related but distinct needs. USACE EGM 01-03 offers standardized residential functions based on its documented U.S. loss evidence and explicitly delimited application. FEMA Hazus supplies an inventory-sensitive framework for direct physical and economic loss estimation in U.S. planning. The JRC database supports international comparisons using normalized curves and maximum national damages. Household-survey and supervised-learning studies seek richer empirical description. Large-scale NFIP claims research tests the behavior of recorded insurance-derived damage at breadth. Each provides evidence with a defined domain.
The USACE memorandum illustrates the importance of denominators and excluded costs. The Hazus manual illustrates explicit integration of building characteristics with damage and economic calculation. The JRC global report illustrates comparability through common functional categories and national replacement values. The NFIP empirical analysis illustrates why none of these can assume that one depth gives a unique observed result at each real property.
The central conclusion of the comparison is an interpretive one: scale and intended decision determine whether the leading defect is lack of local calibration, insufficient predictor detail, poor representation of conditional dispersion, or ambiguity between physical and insured loss. A regional benefit-cost assessment may tolerate noisy property-level estimates if aggregate bias and risk ranking are adequately tested. A single-building underwriting or claims decision may require direct inspection and contract-specific information. A multinational comparison may need standardized proxy values and explicit uncertainty rather than unverified transfer of one local country's curve.
Comparative evaluation matrix
| Research choice | Evidence needed | Main inferential risk | Appropriate interpretation |
|---|---|---|---|
| Generic residential depth-only function | Defined building class, water-depth convention, loss denominator, original calibration and validation | Hidden heterogeneity and extrapolation | Conditional benchmark for a stated population |
| Component-based engineering function | Building assemblies, equipment elevations, materials, restoration costs | Unverified damage thresholds and component assumptions | Mechanism-based estimate subject to engineering tests |
| Multivariable empirical function | Event and property covariates, adequate sample, independent external evaluation | Leakage, overfitting, proxy variables and unstable transfer | Conditional prediction within a tested domain |
| International normalized function | Comparable damage classes, national values, reference price year | Local construction and price differences | Consistent broad-scale comparison with local uncertainty |
| Insurance-payment function | Exact policy and claim definitions, coverage, paid/eligible loss, selection process | Physical damage confused with covered payment | Insurer-specific financial output for specified terms |
The matrix is an original comparative summary of established approaches, not a newly instituted model, typology, or doctrine. Its rows are analytic aids to source interpretation. Direct comparison of performance requires a common test dataset, aligned targets, and equivalent loss measures.
Worked annual-loss and insurance illustration
Assume an entirely hypothetical, mutually exclusive annual event-state distribution for one property: a moderate scenario occurs with probability 0.10 and generates $15,000 of direct damage; a severe scenario occurs with probability 0.03 and generates $90,000; an extreme scenario occurs with probability 0.01 and generates $350,000; all other years have zero modeled direct damage. These are disjoint event-state probabilities, not nested return-period exceedance probabilities. The expected annual direct damage is 0.10 × $15,000 + 0.03 × $90,000 + 0.01 × $350,000 = $1,500 + $2,700 + $3,500 = $7,700. The remaining probability is 0.86 and contributes zero in this illustration.
Now assume solely for the demonstration that a simplified insurance contract has a $20,000 deductible and a maximum payment of $200,000 after that deductible, with all illustrated physical damage otherwise eligible. Scenario payments are $0, $70,000, and $200,000. Expected annual insurer payment is 0.10 × $0 + 0.03 × $70,000 + 0.01 × $200,000 = $4,100. The direct physical-loss expectation of $7,700 and the assumed insurer-payment expectation of $4,100 answer different questions. Neither figure includes claims expenses, time value, inflation, or stochastic dependence with other buildings. The example is wholly invented and carries no empirical inference.
This distinction is consequential when measuring mitigation. Reducing expected direct damage may improve household welfare, reduce interruption, and decrease recovery requirements even if contractual deductibles or limits cause a different percentage change in expected insurer payment. Likewise, an insurer's limited payment may leave a substantial residual economic burden. The separate NFIP building and contents structure provides a real contractual reason to keep these computations distinguishable.
Transferability across jurisdictions and future building stocks
National building codes, construction techniques, asset valuations, flood-protection systems, repair prices, insurance forms, and public recovery arrangements differ. Consequently, international transfer requires more than currency conversion. It may require changes in occupancy mapping, hazard definition, first-floor conventions, component vulnerability, maximum damage values, and coverage transformations. The JRC's continental and national approach addresses part of this need at broad scale; building-level empirical validation remains necessary when decisions demand local precision.
Where codes have evolved, newly built and older structures may not share the same vulnerability curve. Changes in interior equipment, electrical systems, insulation, and contents can also change damage even when the exterior building form looks similar. The ASCE flood design standard is evidence of evolving engineering requirements. Actual performance depends on adoption and compliance; projected code effects should not be automatically transferred to the existing building stock.
Annual hazard estimation should also be treated as a separate changing input. The European Union's flood-risk planning directive governs assessment and management obligations in its jurisdiction; it does not ensure identical hydrological exposure, property protection, or damages across member states. A plausible future flood-loss scenario must make explicit whether change arises from hazard, exposure, vulnerability, economic valuation, or insurance coverage.
Limits of the evidence and further research
This study draws on accessible authoritative technical documents and original research records. It did not retrieve original raw loss observations, reconstruct USACE regression coefficients, refit JRC curves, or independently reproduce Wing et al.'s restricted historical claims analysis. Its quantitative illustrations are arithmetic demonstrations using declared synthetic assumptions. The scientific contribution is the cross-source reconciliation of definitions, mechanisms, insurance accounting, and validation limits.
The most important empirical gaps concern paired building-level records containing verified water levels, component losses, contents placement, modern price definitions, protective measures, and contract-adjusted payments. Additional gaps include consistent reporting of zero loss, cross-event repeat observations, commercially significant equipment, contamination, multi-peril interactions, and global regional calibration. These questions can be addressed through pre-registered field protocols, secure claim-inspection data linkage, open code, and validation on independent events.
Research that compares damage-function families should report performance by asset value, depth band, hazard type, region, and flood-event vintage, not only pooled error. Distributional quality and systematic error at high losses matter for risk capital and catastrophe protection. Mitigation studies should distinguish physically altered hazard from property-level adaptation and make selection mechanisms explicit. A regulator, insurer, or public authority can then understand what evidence genuinely supports the modeled loss.
Conclusion
Flood depth-damage functions are interpretable scientific and operational tools when their measurement, valuation, and population boundaries are explicit. A relationship between water depth and a percentage loss is incomplete until the relevant floor or ground datum, building class, damaged component, valuation base, and observed source population have been specified. The distinction between contents damage as a fraction of contents value and contents damage as a fraction of structure value is a concrete example of an otherwise easy-to-overlook source of modeling error.
The comparative evidence establishes three findings. First, generic functions support repeatable planning and international comparison but inherit the limitations of their calibration samples and standardized asset categories. Second, building elevation, contents placement, duration, contamination, construction, and mitigation can substantially alter vulnerability; multivariable approaches offer predictive possibilities that must be tested across events and regions. Third, the broad dispersion of observed insured flood losses requires thoughtful treatment of conditional distributions, reporting selection, and contract-dependent outcomes, particularly for individual-property decisions and high-loss risk estimation.
The original analytical synthesis identifies the governing practical requirement: every flood-loss estimate should be traceable through coherent definitions of hazard, exposed asset, damage ratio, monetary value, and insurance obligation, with model validation matched to the actual decision being made. This conclusion integrates established engineering, statistical, economic, and insurance evidence. It introduces no independent theory or proprietary vulnerability framework.
Flood damage science advances when detailed hazard maps are accompanied by equally serious evidence about assets and consequences. For international property insurance, the most defensible result is an estimate whose provenance, uncertainty, and contractual meaning can be independently inspected.
References
American Society of Civil Engineers (ASCE). (2024). Flood Resistant Design and Construction, ASCE/SEI 24-24. Official standard overview. https://www.asce.org/publications-and-news/codes-and-standards/asce-sei-24-24/
de Moel, H., van Alphen, J., & Aerts, J. C. J. H. (2009). Flood maps in Europe: Methods, availability and use. Natural Hazards and Earth System Sciences, 9, 289–301. https://nhess.copernicus.org/articles/9/289/2009/
European Parliament & Council of the European Union. (2007). Directive 2007/60/EC on the assessment and management of flood risks, OJ L 288, 27–34 (November 6, 2007). https://eur-lex.europa.eu/eli/dir/2007/60/oj/eng
Federal Emergency Management Agency (FEMA). (2021). Requirements for the Design and Certification of Dry Floodproofed Non-Residential and Mixed-Use Buildings (NFIP Technical Bulletin 3). https://www.fema.gov/sites/default/files/documents/fema_technical-bulletin-3_1-2021.pdf
Federal Emergency Management Agency (FEMA). (2022, May). Wet Floodproofing Requirements and Limitations (NFIP Technical Bulletin 7). https://www.fema.gov/sites/default/files/documents/fema_nfip-technical-bulletin-7-wet-floodproofing-guidance.pdf
Federal Emergency Management Agency (FEMA). (2025). Hazus Flood Model Technical Manual, Hazus 7.0 (June 2025). https://www.fema.gov/sites/default/files/documents/fema_rsl_hazus-7-fltm_06272025_0.pdf
Federal Emergency Management Agency (FEMA), National Flood Insurance Program. (n.d.; consulted November 16, 2026). FAQs About Policies and NFIP Data. https://agents.floodsmart.gov/flood-maps-and-data/faqs-about-nfip-data
Federal Emergency Management Agency (FEMA). (n.d.; federal dataset record consulted November 16, 2026). FIMA NFIP Redacted Claims (OpenFEMA), Data.gov catalog. https://catalog.data.gov/dataset/fima-nfip-redacted-claims-openfema
Huizinga, J., de Moel, H., & Szewczyk, W. (2017). Global flood depth-damage functions: Methodology and the database with guidelines. EUR 28552 EN, European Commission Joint Research Centre. https://publications.jrc.ec.europa.eu/repository/handle/JRC105688 ; doi:10.2760/16510.
Thieken, A. H., Müller, M., Kreibich, H., & Merz, B. (2005). Flood damage and influencing factors: New insights from the August 2002 flood in Germany. Water Resources Research, 41(12), W12430. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2005WR004177
Kreibich, H., Thieken, A. H., Petrow, Th., Müller, M., & Merz, B. (2005). Flood loss reduction of private households due to building precautionary measures: Lessons learned from the Elbe flood in August 2002. Natural Hazards and Earth System Sciences, 5, 117–126. https://nhess.copernicus.org/articles/5/117/2005/nhess-5-117-2005.html
Kreibich, H., et al. (2017). Adaptation to flood risk: Results of international paired flood event studies. Earth's Future, 5(10), 953–965. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2017EF000606
Merz, B., Kreibich, H., Schwarze, R., & Thieken, A. (2010). Assessment of economic flood damage. Natural Hazards and Earth System Sciences, 10, 1697–1724. https://nhess.copernicus.org/articles/10/1697/2010/
Seifert, I., Kreibich, H., Merz, B., & Thieken, A. H. (2010). Application and validation of FLEMOcs: A flood-loss estimation model for the commercial sector. Hydrological Sciences Journal, 55(8), 1315–1324. https://www.tandfonline.com/doi/full/10.1080/02626667.2010.536440
Thieken, A. H., Petrow, T., Kreibich, H., & Merz, B. (2006). Insurability and mitigation of flood losses in private households in Germany. Risk Analysis, 26(2), 383–395. https://onlinelibrary.wiley.com/doi/10.1111/j.1539-6924.2006.00741.x
United States Army Corps of Engineers (USACE). (2000, December 4). Economic Guidance Memorandum 01-03: Generic Depth-Damage Relationships. https://planning.erdc.dren.mil/toolbox/library/EGMs/egm01-03.pdf
United States Army Corps of Engineers (USACE), Hydrologic Engineering Center. (n.d.; consulted November 16, 2026). Background for Flood Risk Management Studies. https://www.hec.usace.army.mil/confluence/hmsdocs/hmsag/flood-risk-management-frm-studies/background-for-flood-risk-management-studies
United States, Code of Federal Regulations. (Current online text consulted November 16, 2026). 44 CFR Part 61, Appendix A(1): Standard Flood Insurance Policy, Dwelling Form. https://www.law.cornell.edu/cfr/text/44/appendix-A%281%29_to_part_61
United States, Code of Federal Regulations. (Current online text consulted November 16, 2026). 44 CFR § 61.3: Coverage and benefits provided under the Standard Flood Insurance Policy. https://www.law.cornell.edu/cfr/text/44/61.3
Wagenaar, D., de Jong, J., & Bouwer, L. M. (2017). Multi-variable flood damage modelling with limited data using supervised learning approaches. Natural Hazards and Earth System Sciences, 17, 1683–1696. https://nhess.copernicus.org/articles/17/1683/2017/
Wagenaar, D., Lüdtke, S., Schröter, K., Bouwer, L. M., & Kreibich, H. (2018). Regional and temporal transferability of multivariable flood damage models. Water Resources Research, 54(5), 3688–3703. https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2017WR022233
Wing, O. E. J., Pinter, N., Bates, P. D., & Kousky, C. (2020). New insights into US flood vulnerability revealed from flood insurance big data. Nature Communications, 11, Article 1444. https://www.nature.com/articles/s41467-020-15264-2