Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Claim Loss Quantification from Partial Records and Uncertain Evidence

Abstract. Individual insurance claims often require monetary quantification after records of the insured interest, event, quantity, cost, or pre-loss economic performance have become incomplete. This monograph investigates the epistemic and operational consequences of partial observation. It distinguishes economic loss, covered loss, contractual benefits, settlement and payment; examines missingness processes, source provenance, entity matching, inventory reconstruction, valuation bases, constrained ranges, imputation, sensitivity, and nonlinear contractual effects; compares selected international, UK, California, US federal flood, and Australian claims conduct requirements; and develops reproducible synthetic cases and an independent case-audit design. The research combines source-critical selective review with transparent numerical derivations and institutional comparison. It finds that defensible quantification depends on the structure and provenance of missing evidence, explicitly conditional valuation assumptions, and the separation of factual uncertainty from coverage and calculation. Evidence-linked bounds, carefully qualified predictive estimates, and traceable revisions make uncertainty intelligible and independently examinable. The contribution is an original analytical synthesis of existing methods and institutional evidence; no new theory, empirical dataset, or universal legal rule is claimed.

Keywords: Claims science; claim loss quantification; partial records; missing data; loss adjustment; evidence provenance; valuation; partial identification; sensitivity analysis; uncertainty bounds; claims handling; case-level audit; insurance indemnity; business interruption.

Introduction

An insurance claim becomes a quantification problem when the event, the affected interest, or its monetary consequences can be established only incompletely. A building may be inspected after emergency demolition; stock records may have been destroyed by the same fire that destroyed inventory; a vehicle may be appraised from dated photographs after disposal; an interrupted business may have lost the ledger needed to establish its ordinary performance. The object of this research is the amount of an individual loss under those conditions. It examines how evidence can delimit a defensible value, which assumptions narrow that value, and what an institution should preserve so that an independent reviewer can reproduce the reasoning.

The distinction between economic loss, covered loss, contractual benefit, agreed settlement, payment, and accounting liability organizes the investigation. A missing invoice can make a cost uncertain without making the insured event uncertain. Conversely, an excellent repair invoice can establish exactly what a contractor charged while leaving both causation and policy eligibility unresolved. A paid amount records financial execution; it can reflect compromise, a deductible, an advance, a limit, or a disputed item. The observed payment is therefore not a neutral label for underlying loss. These distinctions keep case-level estimation within its scientific boundary and prevent the use of portfolio reserves as substitutes for evidence of individual entitlement.

The investigation asks three questions. First, how do missingness, provenance, and timing affect the amount recoverable through evidence? Second, which arithmetic, statistical, and valuation procedures produce results that remain intelligible under independent scrutiny? Third, how do institutional requirements for proof and fair treatment alter the way those procedures are used? The answers are developed through a selective source-critical review of foundational statistics, actuarial standards, official insurance conduct requirements, and regulator studies, combined with transparent hypothetical calculations and contrary cases. The research does not use confidential claims files, estimate real compensation, or claim a new statistical theory. Its contribution is an original synthesis of established analytical results applied to the structure of individual claims.

The sources were selected for direct relevance to missing observations, valuation, provenance, fair investigation, and reproducibility. Searches used the expressions inference and missing data, partial identification, insurance claim proof of loss, property settlement estimates, claims handling regulation, digital evidence preservation, and actuarial data quality. Official regulatory texts and published standards were prioritized over commentary. Publisher abstracts or overviews were used only for the propositions they actually establish. The comparative treatment covers international supervisory guidance, United Kingdom conduct rules and reviews, California property rules, the US federal flood dwelling form, and Australian supervisory guidance and review evidence. These examples illuminate different institutional settings; they do not form an exhaustive survey of national law. Numerical cases are constructed for this study, not sampled observations. All amounts in numerical examples are hypothetical US dollars unless explicitly specified.

Loss Units and Missing Evidence

The quantity to be estimated

A loss estimate requires an explicit object, unit, valuation basis, reference time, and boundary. A unit can be a damaged asset, repair operation, stock category, period of interruption, necessary additional expense, or medically relevant service. The scope must establish which units are being counted, which have independent value, which are components of other units, and which are alternative ways of restoring the same function. A replacement cost and a repair cost for the same component are competing measures rather than two additive losses. A claimant's repair invoice and an insurer's estimate can describe the same work using different line items; summing them duplicates one economic event.

Let the factual loss vector be x, with entries corresponding to separately identified units. Let E denote the available information, A the valuation assumptions, and W the applicable policy wording and legally determined coverage facts. For an additive example, gross economic loss G equals the sum of each damaged quantity q times its relevant unit price p, plus separately justified common costs C. A covered measure H is a function of G and the contractual classifications. An illustrative net amount P = min(M, max(0, H - D)) applies only to a simple single-limit contract where the deductible D precedes the limit M. Other contracts may apply limits, deductibles, coinsurance, salvage, depreciation, or reinstatement conditions in different sequences. In particular, P is not generally a linear function of G.

This notation makes the questions visible: quantity, price, scope, and entitlement are different unknowns. An observation of the total payment cannot ordinarily be inverted to find gross loss when a cap binds. With P = 10,000 and M = 10,000, many covered values can yield the same payment. In an uncapped example, a deductible of 500 could be added back only if the payment is known to be the complete application of the stated contract and contains no other adjustments. Even elementary arithmetic requires institutional context.

Completeness, accuracy, and relevance

A document-rich file may remain incomplete when all documents concern a single category of expenditure and another category is absent. Completeness concerns the correspondence between the records and the defined population of loss units. Accuracy concerns whether the recorded quantities and attributes are correct. Relevance concerns whether those quantities and attributes answer the valuation question. A genuine receipt for an unrelated asset satisfies authenticity but not relevance. An accurate record of current retail prices may lack relevance to a contractual valuation date six months earlier.

The record should distinguish whether a unit was observed, reported but unverified, inferred from other records, bounded without a point estimate, disputed, explicitly absent, or not applicable. An empty numeric field is an administrative state, not a measured zero. A zero requires a basis: no damaged unit, no covered expense under a specified interpretation, or verified zero quantity. In a spreadsheet, both zero and an empty cell can disappear into an identical sum, creating a particularly dangerous form of loss suppression.

Actuarial Standard of Practice No. 23 distinguishes the review of data from a more extensive audit. Its data quality guidance (Actuarial Standards Board, 2016, sections 2 and 3) is directly relevant when an actuary uses claims records for an applicable assignment; the distinction also clarifies the epistemic standard of a reconstructed case. An analyst who reads the inventory for internal consistency has performed a review, not independently confirmed all items with original suppliers. The standard also distinguishes data from assumptions and excludes generation of a wholly hypothetical dataset from its scope. The examples below consequently illustrate methods rather than assert compliance with a professional audit standard.

Mechanisms of missingness

Denote by R an indicator of whether an evidentiary item is observed. A missingness process may depend on the observed characteristics of a loss, on an unobserved characteristic, or on neither. In the conventional statistical vocabulary, missing completely at random describes independence from the relevant data; missing at random allows dependence on observed data subject to the relevant conditioning definition; missing not at random allows dependence on the missing values even after observed information is accounted for. These categories concern a specified probabilistic model and pattern of observation. They are not labels that establish themselves from the physical appearance of a file.

Rubin's inference and missing data (1976) establishes in its publisher abstract the conditions under which direct-likelihood or Bayesian inference may disregard the missingness process: missing at random and distinct parameters. The paper's statement is a result about inference under specified conditions. Applying it to insurance requires an additional inquiry into the information-generating process. If high-temperature damage destroys labels and documents in the most severely damaged rooms, missing asset details may be associated with loss magnitude. If the claims portal randomly loses an attachment because of a system failure unrelated to its content, the credibility of an independence assumption may be greater, although the system's behavior would still need evidence.

A further distinction is between a missing attribute and an unknown population. In an inventory with 100 known items, 12 absent purchase prices leave a fixed denominator. After total destruction of a warehouse, the number of distinct items may itself be uncertain. A method that imputes prices for known rows cannot recover an unknown number of missing rows. Inventory identity must be reconstructed before price imputation can be evaluated. With no credible limit on quantities or prices, finite upper bounds may be unavailable. A policy limit bounds an applicable payment, not the underlying economic loss.

Selection can be generated by human behavior as well as the insured event. Small cash purchases may have fewer receipts; expensive items may retain serial numbers; repeated transactions may be recorded more consistently than gifts; a business under financial stress may have delayed accounting updates. These are plausible mechanisms, not empirical claims of prevalence. Their consequence is that routine substitution of an observed average assumes something substantive about the missing units. The severity and direction of the resulting error must be tested rather than inferred from the completeness percentage alone.

Three independent uncertainties

The quantity of physical damage can be uncertain while the contractual valuation basis is clear. The valuation basis can be contested while physical damage is agreed. The underlying facts and contract can both be agreed while an implementation error miscalculates depreciation. These three domains correspond to evidentiary uncertainty, interpretive uncertainty, and computational uncertainty. Their remedies differ. More photographs may resolve physical scope, legal interpretation may resolve eligibility, and a reproducible calculation may eliminate arithmetic mistakes. Treating all three as a single confidence score obscures where additional work can improve the result.

A defensible case record therefore identifies each uncertainty with an affected amount, the possible directions of change, and the specific evidence or decision that could resolve it. Where the missing item cannot affect the chosen payment because a contractual limit already binds under all feasible scenarios, the remaining economic uncertainty can still matter for other purposes. The significance of information depends on the decision under consideration.

Record Authenticity and Reconstruction

Establishing provenance

The original record is evidence of an assertion made through a particular system or person at a particular time. It may describe a purchase, a physical observation, a payment, or a retrospective estimate. These categories should remain separate throughout handling. An invoice may be genuine but provisional; a photograph may show a real object but not establish its location or ownership; a bank transfer may combine several transactions and disclose no physical quantities. Evidence should be assessed at the proposition level.

Every evidence entry should preserve the source custodian, date of acquisition, document date, purported event date, reference to the original, transformations performed, and the person or system responsible for interpretation. The insured event may precede the production of the document by days or years. A repair quotation produced after a catastrophe is contemporaneous evidence of a supplier's quotation, not a contemporaneous observation of the pre-loss building. That distinction is essential when prices, availability, and scope have changed.

Guttman et al. (2022) discuss the distinctive preservation problems of digital material in Digital Evidence Preservation. In claims work, an appropriate application is to retain original files and separate derivative representations. A photograph converted to a smaller image for review may lose metadata or detail. A spreadsheet exported to PDF may omit formulas and hidden rows. A hash of the original byte stream enables detection of changes to that stream, provided that the recorded digest itself is trustworthy. A matching digest establishes integrity relative to the recorded bytes, not the truth of the depicted scene, the authenticity of a signature, or the accuracy of the document's contents. These practical implications are analytical applications of evidence preservation, not assertions that NIST imposes insurance-specific obligations.

Access controls must reflect the sensitivity of claim records. The record can identify the verification of a medical or financial document without unnecessarily duplicating private details for every reviewer. Redactions must be made in derived copies, retaining protected originals subject to lawful retention, access and deletion requirements. A reproducibility obligation does not authorize indiscriminate acquisition of private information.

Triangulation and dependence

The reconstruction of a missing record relies on independent relationships among observations. A supplier invoice can be cross-checked against payments, stock movement records, purchase orders and physical capacity. An asset photograph can be correlated with a dated maintenance record and a serial number. A building dimension derived from a plan can be checked against surviving foundations and measurements. A corroborating observation is strong when its errors and incentives differ from those of the initial source and when both genuinely refer to the same unit.

A counterexample is the circulation of one adjuster's estimate through an insurer's report, a contractor's scope, and an automated summary. Three documents then reproduce a single method rather than offer three independent estimates. Likewise, two price guides may share the same transaction source. Independence is a property of the production process, not a count of documents. Treating internally copied evidence as independent artificially narrows uncertainty.

Entity matching should precede aggregation. Use stable item identifiers when possible; otherwise document the matching variables, date tolerance, location and quantity, and distinguish uncertain matches. Suppose an invoice records four pumps, a bank payment covers the full invoice, and a warehouse movement shows four pumps received. The three records support one transaction of four pumps, not twelve pumps. A subsequent stock transfer may change ownership or location without changing the total originally acquired. Quantity reconstruction must therefore include both entry and exit.

In inventory cases, a basic accounting identity is opening inventory plus acquisitions minus disposals and other recorded outflows equals closing inventory, subject to measurement adjustments. A physical loss is one possible outflow. When records are incomplete, the identity constrains the missing terms. For example, 90 opening units and 45 documented arrivals with 80 verified departures imply 55 units before an event if there were no other movements. If unrecorded departures are bounded between zero and eight, the pre-event stock lies between 47 and 55; if the bound is unsupported, it cannot be used as a factual constraint. If 12 units survive, the physically affected stock lies between 35 and 43 conditional on those assumptions. Price must then be established on the contractual valuation basis, not substituted by the purchase price without analysis.

Handling contradictory and low-quality sources

Contradictory observations warrant competing explanations. A repair contractor might quote a larger scope than an engineer because the contractor assumes replacement rather than repair. A narrower estimate might omit access work. A pre-loss photograph might be incorrectly dated. The analyst should reconcile shared line items, isolate methodological differences and identify what each source actually observes. An arithmetic average of two numbers obscures these differences.

The treatment of an unverified claimant statement depends on its materiality and on applicable procedures. It can support a hypothesis that should be tested through available independent records. It should be retained and evaluated instead of silently deleted or converted into a confirmed observation. A persuasive explanation of why an item was accepted or rejected should identify the evidence, not merely disclose a numerical confidence grade. A suspicion of manipulation is a separate investigative matter. An inconsistency can arise from transcription, memory, a supplier error, ambiguous units, or intentional conduct; the inconsistency alone does not establish fraud.

Reconstruction is iterative. A later original record may confirm or displace an earlier inference. Versioning must preserve the original estimate and the evidence then available, the reason for the revision, affected lines, decision time, and author. A corrected total without the prior reasoning cannot support an audit of whether the original decision was reasonable at the original date. For scientific analysis, the goal is both the best present estimate and the ability to understand how it changed.

Valuation under Partial Observation

Establishing the valuation basis

Each type of loss requires an explicit measurement rule. A replacement-cost estimate concerns the cost to restore the defined property with suitable materials and labor on the applicable basis. A market-value estimate concerns an exchange price at a specified date in a specified market. Actual cash value has policy- and jurisdiction-dependent meaning. A repair estimate measures restoration operations; a total-loss estimate may instead depend on pre-loss value and residual value. A fixed benefit can depend primarily on satisfaction of a trigger and the stated sum rather than proof of the same dollar amount of economic damage. These distinct objects require different evidence.

The cost of repair is assembled from scope, quantity, unit rate and legitimate common components. Common costs include items such as mobilization or necessary protection where the measure permits them. Taxes, overhead, disposal, access, and code-related costs are included only when relevant to the valuation basis and the separately determined coverage. Separating those components prevents both omission and double counting. An estimate that uses an all-inclusive unit rate and also adds the same labor overhead outside that rate is inconsistent. The gross bill, paid invoice, and hypothetical replacement estimate should have independently labeled fields.

Comparable-market analysis requires asset identity, age, mileage or usage, condition, configuration, geography, date, and transaction type. An asking price is not necessarily an executed transaction price. A completed sale under unusual conditions may fail to represent the relevant market. A short list of comparables can be made more transparent through an adjustment table, where each addition or subtraction has a reason and a source. Missing comparable attributes should be carried as uncertainty in the adjustment, not assigned arbitrary corrections with false precision.

The FCA's (2024) vehicle valuation review identifies the importance of explaining the pre-accident valuation and deductions to customers and maintaining effective processes for disputes. This institutional finding reinforces the analytical requirement that valuation evidence remain understandable to another party. A single quoted price without a record of comparables, deductions or condition cannot reveal whether an error arose from a mismatched vehicle or an unreasonable adjustment. It is possible to arrive at a plausible total through an inadequate procedure.

Temporal reconstruction and external prices

An insured event, valuation date, quote date, expenditure date, decision date and payment date are distinct. Prices may change between them; local market availability can be disrupted; the contract may specify a particular date or incurred-cost condition. An observed recent price should be converted to the relevant basis only when a defensible conversion is available. If an appropriate price index I exists, a simple rebasing calculation is P(t0) = P(t1) times I(t0)/I(t1). The index must measure a comparable category and geography. A broad consumer price index may be unsuitable for specialized machinery or disaster-affected structural components.

Assume a later quote of 12,600 at index 105 and a target date index of 100. The rebased price is 12,000. This is a mathematical illustration, not evidence that the index is suitable. If the later quote also includes a distinct temporary local scarcity premium, the index correction may not remove it. An alternative market quotation or a component analysis may be needed. Where the contract uses the actual reasonable cost of reinstatement rather than a date-rebased market price, the rebased figure may answer the wrong question. The analyst must record the contractual basis before transporting values across time.

Foreign currency requires an analogous separation. A supplier quote may be denominated in euros and claim settlement in dollars. The correct conversion convention depends on the policy and applicable rules, and the exchange date must be identified. A converted amount and an original amount both need to be retained with their units. Costs inclusive and exclusive of recoverable taxes cannot be mixed without adjustment. Inflation and exchange changes should not be folded invisibly into a single depreciation coefficient.

Deterministic reconstruction and bounded quantities

When available data support an exact accounting relationship, deterministic reconstruction can be stronger than a statistical estimate. When they support only upper and lower limits, interval reconstruction is more credible than an unsupported point. The problem may be represented as a feasible set F(E,A) of vectors consistent with evidence E and assumptions A. For a chosen valuation functional g, the lower and upper values are infimum and supremum of g(x) over that set. These are conditional identification bounds, not confidence intervals. They may coincide if constraints determine the relevant total.

Manski (2003), in the publisher overview of Partial Identification of Probability Distributions, develops the general perspective that incomplete observations can support useful inference without point identification. The present case-level formulation is a direct application of feasible-set reasoning, not a theorem attributed to an unread chapter. A claim's unknown physical quantities are not identical to a population parameter. Nevertheless, both contexts require that observation and identifying assumptions be kept separate.

Consider three stock categories. Verified items contribute 7,200. For the second category, available evidence bounds damaged quantity between 6 and 10, unit value between 300 and 450. A third category contributes between 500 and 900. If all combinations are feasible and all amounts nonnegative, the total interval is 7,200 + 6 times 300 + 500 = 9,500 to 7,200 + 10 times 450 + 900 = 12,600. A tighter bound may follow from a shared capacity constraint or correlation between grades and unit values. Multiplying marginal extremes is valid as an outer bound under the stated nonnegative ranges; equality with the actual feasible bounds requires simultaneous feasibility.

For a simple linear program with known unit rates and linear stock constraints, a min/max optimization identifies feasible endpoints. With uncertain unit rates multiplied by quantities, the objective may be bilinear; a linear procedure may require a suitable transformation, relaxation, or scenario representation. Mathematical feasibility is distinct from evidentiary support. A solver can compute a precise optimum over assumptions that are scientifically unjustified. The evidence ledger must therefore identify each constraint and why it is accepted.

Statistical prediction and imputation

Imputation supplies values for missing data using a model or analogous observations. A conditional mean can be useful when the target is expected loss, a representative dataset supports the conditional model, and missingness assumptions are defensible. A conditional median may reduce expected absolute error. These objectives should be explicit. A point estimate optimal under squared loss is not automatically optimal for a contractual decision, and a statistically likely value is not proof that a particular asset existed.

Multiple imputation creates repeated draws under a model to represent uncertainty associated with missing data. It can help avoid treating all filled values as if they were observed. For a single claim, however, the relevance of the imputations depends on the adequacy of the model for the particular case and the quantity being estimated. A predictive interval for a missing item is different from an interval for the estimated conditional mean. Imputed assets should never silently acquire the evidentiary status of documented assets.

A synthetic inventory has 12 known items with eight observed values averaging 200. If four missing values share the same conditional mean, their predicted aggregate is 800 and the estimated total is 2,400. Suppose instead that the missing category contains specialist items and credible external evidence supports means between 300 and 500. Then the same eight values yield expected totals between 2,800 and 3,600 conditional on those alternative means. This sensitivity analysis does not establish a probability distribution for any specific claimant. A related physical check might show that the four missing items cannot all have been specialist units. Such a constraint would change the feasible scenarios.

Bayesian analysis can combine prior information with a likelihood when both are explicit and defensible. It is particularly important to inspect prior influence when the file offers little information. If a model starts with a narrow prior based on routine small claims, a catastrophe loss may be pulled toward the routine distribution even when its missingness mechanism and asset mix differ. Posterior precision is conditional on the likelihood and prior, and does not establish legal proof. A broad prior cannot repair a misspecified likelihood either. The scientific question is how the resulting inference changes under well-motivated alternative assumptions.

Business interruption and counterfactual valuation

Business interruption introduces a fundamentally unobserved quantity: the economic result that would have occurred in the relevant alternative history. The available evidence includes pre-event accounts, orders, contracts, production capacity, industry conditions, actual sales during interruption, saved expenses and additional mitigation expenses. The estimate depends on how the policy defines the reference and indemnity period, the insured peril and allowable trends, and the relevant measure of earnings. These elements must be separately documented.

A simplified hypothetical loss has counterfactual sales 100,000, actual sales 60,000, an applicable contribution margin of 40 percent, and eligible additional expense of 5,000, with no overlap or other adjustment. The resulting amount is 21,000. If independently supportable counterfactual sales range from 90,000 to 115,000 with actual sales, margin and expense fixed, the amount ranges from 17,000 to 27,000. The example assumes the margin already reflects the identified variable-cost savings; subtracting those savings again would double count. It is an economic illustration, not a generic policy formula.

The FCA's account of the business interruption test case records the January 15, 2021 UK Supreme Court decision on sampled policy wordings and explains that valid claims under the relevant analysis should not be reduced on the basis that loss would have occurred from the pandemic anyway. The case illustrates how contractual counterfactual definition can be decisive even when revenue data are complete. This study relies on the regulator's official summary rather than claiming to interpret the full judgment or later case law. Conclusions should be confined to the described policy issues.

A forecast trained on realized performance from an already affected period risks making a loss-induced downturn part of the baseline. Conversely, assuming an unusually favorable pre-event month persists may exaggerate loss. Seasonal comparisons, contemporaneous bookings, capacity and external trends can be helpful, but the counterfactual should exclude only circumstances required by the relevant contractual and legal interpretation. Statistical adequacy and contractual adequacy are separate tests of the same valuation.

Sensitivity and Error Bounds

Sources of valuation variance

A loss estimate can vary because an asset count is uncertain, a quote is provisional, an allocation is contested, a future expense is unresolved, or a model is sensitive to specification. Uncertainty should be classified before it is aggregated. Sampling error, measurement error, uncertainty arising from missing observations, valuation-date uncertainty, and disagreement on applicable scope do not have the same interpretation. An interval produced by plausible extremes is a scenario range. A confidence interval for a parameter requires a sampling and inference procedure. A credible interval requires a posterior model. These should be labeled distinctly.

For the simple value V = qp, an incremental change is p delta-q + q delta-p + delta-q times delta-p. Around a baseline q = 20 and p = 500, a two-unit increase and a 50-dollar price increase yield V = 22 times 550 = 12,100, compared with 10,000. The interaction term is 100. If both inputs are treated as independent small variations, the interaction may be neglected as a first-order approximation, but the exact change must be used for material movements. Where positive correlation reflects a common disaster effect, the joint range may be wider than one-at-a-time sensitivity suggests.

When a justified probability model exists, for additive loss Y = X1 + X2, Var(Y) = Var(X1) + Var(X2) + 2 Cov(X1,X2). For hypothetical standard deviations 300 and 400 with correlation 0.5, the variance is 90,000 + 160,000 + 120,000 = 370,000, and the standard deviation approximately 608.28. With independence it is 500. This calculation illustrates why shared contractors, shared indices and shared evidence-generation mechanisms can invalidate independence assumptions. It does not assert the existence of those probabilities in an actual claim.

A scenario table should report the selected inputs and their consequences, and explain which combinations are supported. Assigning each uncertain input its high value may describe a stress scenario rather than a jointly likely state. Conversely, a base case that assumes all missing data are typical may fail to represent a plausible adverse joint scenario. A probability distribution should be introduced only with a model and evidence sufficient to defend it. The fact that a range has two endpoints does not mean their midpoint is the expected value.

Nonlinear contractual transformations

For illustration, consider a single deductible of 1,000 and a maximum payment of 9,000. A covered amount between 9,500 and 12,600, derived from the earlier stock scenario, yields P between 8,500 and 9,000 when P = min(9,000, max(0, H - 1,000)). The top of the economic range is truncated at the contractual limit. The range width contracts from 3,100 to 500. That does not imply the event is understood more precisely; it means the contract limits the monetary consequence of the unresolved facts under this stipulated formula.

In another case, a deductible of 1,000 and covered loss between 500 and 1,500 give payment between zero and 500. Here the uncertainty crosses the threshold at which a payment becomes positive. A small change in scope can be decision-critical. The contrast shows why magnitude alone cannot determine an investigation's priority. The marginal value of information is high when it determines whether the contractual threshold is crossed, low for a specific payment when all scenarios yield the same amount, and potentially high again for another task such as subrogation.

Nonlinearity also matters for estimated distributions. For two hypothetical equally likely covered losses, 500 and 1,500, a deductible of 1,000 produces payouts zero and 500 with expected payout 250. Applying the deductible to the mean loss of 1,000 produces zero. Consequently, the expected contractual payment generally differs from the contractual payment evaluated at expected gross loss. In statistical models of uncertain claims, applying the contract within each scenario preserves this distinction. The illustration uses explicitly assigned equal scenario weights for mathematics only.

Independent validation and error assessment

The modeling standard (Actuarial Standards Board, 2019, sections 3.1.5 and 3.6.2) connects data suitability to the intended use of an actuarial model and requires relevant validation of outputs within its scope. An individual claim reconstruction can apply the underlying discipline by verifying input lineage, recomputing amounts and checking whether predicted results agree with known independent constraints. It must retain the distinction between mathematical validation and empirical performance validation.

Retrospective validation can be designed with a set of well-documented claims. Select the reference amount appropriate to the study: independent replacement estimate, verified eligible repair invoice, or other validated economic measure. Artificially mask selected records using mechanisms that represent plausible real-world absence. Estimates obtained from the masked file can be compared with the known reference under each mechanism. The process should distinguish tests of numeric prediction from tests of item existence and from legal eligibility. A paid settlement is not an adequate gold standard when it may incorporate compromise or systematic bias.

Mean signed error measures directional tendency: average of estimated minus reference loss. Mean absolute error measures average error magnitude; squared error emphasizes larger departures. If a method produces predictive intervals with known targets, empirical coverage and interval width should both be evaluated. Wider intervals can cover more cases without delivering useful precision. A narrow interval with poor calibration is equally inadequate. For true identification bounds, empirical inclusion tests data and assumptions on observed cases, but coverage cannot be given a nominal confidence level simply by convention.

Leakage is a risk when a price estimate uses post-settlement repair information that would not have been available at the decision date. Claims from the same catastrophe may share sources and errors. A validation split by individual rows at random can therefore place nearly identical cases in training and testing. Splitting by claim event, provider or time, where appropriate, tests transportability more meaningfully. An evaluation should also retain unresolved and denied cases where lawful and analytically relevant; complete-case testing alone may remove the population for which reconstruction is most needed.

A small synthetic illustration makes the distinction between bias and spread concrete. Reference values of 1,000, 2,000 and 3,000 and estimates of 1,100, 1,700 and 3,200 produce errors 100, -300 and 200. Mean signed error is zero, mean absolute error is 200, and root mean squared error is about 216.02. Zero average bias in these three points does not establish accuracy or fairness. The method might have systematically underpredicted a particular loss category and overpredicted another. A useful evaluation stratifies by the mechanism of missingness, line of insurance, severity and access to evidence, subject to privacy and statistical adequacy.

Information acquisition and proportionality

The value of a new observation depends on what it can resolve and on the institutional cost of obtaining it. If obtaining a second appraisal reduces the maximum disputed amount by little, it may be disproportionate. If a particular supplier reconciliation resolves a high-value missing category, it can be decisive. The comparison should include delay, privacy intrusion, claimant ability to obtain evidence, financial distress and the possibility of a mistaken adverse decision. A simple expected-error reduction rule does not itself determine the legally required investigation.

In a bounded numerical case, information can be prioritized without inventing probabilities. Calculate how much the lower and upper limits change if each uncertain input is resolved. A known quantity may leave wide rate uncertainty, while a rate quote may do little when the number of units is unknown. The results support focused questions: which record establishes damaged quantity, which identifies the appropriate unit price, which confirms the relevant date? Any evidentiary threshold imposed by applicable law remains an independent constraint.

Comparative Claims Evidence Rules

International supervisory expectations

An estimate is used inside a claims-handling process, and the process influences which data become available. The IAIS ICP 19 conduct benchmark (online consolidated text, standard 19.10) establishes a supervisory expectation for transparent, fair and timely claims work. Its guidance addresses documented procedures, explanations of decision factors, competent assessment, dispute resolution and outsourced processes. It is a supervisory standard for implementation through relevant frameworks, not a universal cause of action or directly binding contractual rule.

The analytical implication is that a valuation's intelligibility is part of the process through which evidence can be contested and improved. When an insurer discloses the basis for depreciation and the identity of comparable assets, the claimant can respond to an incorrect condition assumption or missing item. The possibility of correction can improve evidentiary quality. A generic rejection with no explanation removes that opportunity even where the original estimate was mathematically careful.

United Kingdom

The FCA's ICOBS 8.1.1R sets conduct obligations for insurers within the rule's scope, including fair and prompt claims processing, reasonable assistance and progress information, avoidance of unreasonable rejection and prompt settlement following agreement. It does not prescribe a single statistical imputation method or replace specific contractual proof requirements. In a partial-record case, reasonable guidance can help identify what evidence is required and what alternative evidence might bear on the amount.

The regulator's business interruption claims review (FCA, 2022) identified a shortcoming in quality reviews that concentrated on quantum rather than broader experience. An accurately calculated final amount can coexist with poor requests, unnecessary delay, or an inadequate explanation. The review is evidence of issues observed in that specific supervisory work, not an estimate of the prevalence of such deficiencies throughout the market.

California

The California Department of Insurance's claims regulations index identifies separate general settlement rules and line-specific provisions. This structure matters for comparison because a property-specific requirement cannot be applied indiscriminately to every insurance claim. Under California title 10 section 2695.7(d), insurers are required to investigate fairly, objectively and thoroughly, and persistent requests for immaterial or unnecessary information are restricted. The provision concerns investigation and conduct in its applicable context.

For applicable first-party property settlement based on a written scope or estimate prepared by or for the insurer, section 2695.9(d) requires the insurer to provide the claimant with the documents underlying settlement. The evidence record should preserve the exact version communicated, because a later internal revision may not reflect the basis the claimant received. This requirement supports a comparison of scope transparency; it does not imply that every difference between contractor and insurer estimates is resolved in favor of one source.

United States federal flood insurance

A contrasting example is the federal National Flood Insurance Program dwelling form. Section VII.G of the Standard Flood Insurance Policy requires a damaged-property inventory and supporting documentation and specifies a sworn proof of loss within the ordinary 60-day policy text. Section VII.G.9 also allows the insurer, at its option, to accept the adjuster's report instead of a separate proof of loss, subject to the form's signature and possible sworn-report requirements. This is a form-specific requirement. Authorized event-specific waivers or extensions and the policy edition in force must be verified separately; the ordinary text alone does not determine an actual claimant's deadline.

The federal form shows that strong reconstruction of physical quantities and rates can coexist with formal submission requirements. A mathematically valid estimate does not itself fulfill a sworn proof obligation. Conversely, the submission of a sworn statement is not an independent statistical verification of every underlying amount. The researcher should record both procedural status and evidentiary status. No general inference about all US private insurance contracts follows from the NFIP form.

Australia

ASIC's (2025) claims handling licensing guidance explains the application of licensing obligations to relevant claims handling and settling services. Table 4 addresses timeliness, the burden and relevance of information requests, transparent communication, responsive treatment of vulnerability, and opportunity to respond to adverse findings. The guidance's legal authority and audience differ from those of a statute or a policy form; its relevance lies in explaining regulator expectations for applicable licensees.

ASIC's (2023) home insurance claims review examined administrative data on 218,256 home claims lodged with six participating insurers from January through March 2022, with subsequent progress observed through September. The work also considered 150 targeted files, 40 interviews and 25 case studies. Its Figure 1 identifies the targeted qualitative work as nonrepresentative. The report's combination of administrative and qualitative information illustrates why evaluation of claims requires more than an average settlement amount; it documents observed issues rather than estimating causal effects of a specific missing-data procedure.

Comparative interpretation

These examples differ in legal form: international supervision, domestic conduct regulation, federal contract form, licensing guidance and regulator observations. They nevertheless expose a shared operational question: who may demand which records, for what purpose, at whose cost, and with what explanation? A stricter documentary rule may make a claim procedurally incomplete despite a plausible economic range. A less burdensome procedure may make alternative evidence possible without lowering the validity of the resulting estimate. The actual influence on outcomes depends on contract, applicable law, administrative implementation and claimant circumstances.

For a hypothetical loss with an accepted unit inventory except for one damaged asset, suppose a missing receipt leaves that asset valued anywhere from 500 to 1,000 on the evidence. Under a procedure that accepts independent supplier confirmation, the estimate may be refined to 800. Under an assumed strict documentary rule the asset remains disputed pending required proof. Under an arbitrary practice of filling missing entries with zero, the amount becomes zero despite the identified uncertainty. The third procedure is a statistical convention, not a legal conclusion. No jurisdiction above is asserted to require that zero-fill practice. The example demonstrates that institutional evidence admission changes the information available to the estimator and that numerical conventions can have distributive consequences.

The comparison does not establish a universal burden of proof, proof standard, entitlement to an advance, or waiver of contract terms. Those matters require the governing legal provisions, policy version, and circumstances of the specific case. What can be generalized at the level of claims science is the need to separate substantiated valuation from procedural sufficiency and to explain the connection between evidence and a decision.

Replicable Case-Level Audit

A complete synthetic reconstruction

Consider a hypothetical commercial contents insurance claim after a localized fire. Coverage for the relevant event and identified contents is assumed for the main calculation. The valuation basis is stipulated to be replacement cost at the valuation date, without depreciation, taxes or other coverage adjustments. An insured can recover only one restoration measure for each item. No actual supplier, claimant or policy is represented.

The file contains an independent inventory of equipment group A, a partially recovered supplier record for group B, photographs for group C and a separate cleanup quotation. The inspection and warehouse records for group A agree on 12 damaged units, valued at 400 each. Group B stock reconciliations support 5 to 8 damaged units and comparable supplier quotations support 700 to 900 per unit. Group C contains ten less expensive items at 120 each; only six are initially supported by documentation and later independent photographs establish four additional items. Assume the photographs can be matched reliably to unique damaged items. Cleanup is necessary and eligible, with a range of 600 to 900 after duplicate work has been removed.

Loss unit Quantity or scope Value basis Conditional range
Group A 12 units 400 per unit 4,800 to 4,800
Group B 5 to 8 units 700 to 900 per unit 3,500 to 7,200
Group C 10 units 120 per unit 1,200 to 1,200
Cleanup Independent eligible cost Quotation range 600 to 900
Gross covered amount Additive; no overlap Stipulated basis 10,100 to 14,100

Without further joint restrictions the interval endpoints are obtained by summation. Assuming all combinations are feasible, its endpoints are attainable in the simplified problem. The initial documentary-only group C contribution would have been 720. The additional corroborated items add 480, and the traceable record distinguishes this revision from a price change. If the photographs were instead repeated views of the same items, the added 480 would be unsupported. The presence of more images does not by itself increase the unit count.

The hypothetical contract applies a 1,000 deductible followed by a 12,000 maximum payment. The corresponding payout interval is 9,100 to 12,000. A central planning scenario with group B quantity 6, rate 800, cleanup 750, and all other stipulated values yields gross 11,550 and payment 10,550. This scenario is not an expected loss and does not supersede the range. A chosen final settlement would require the applicable contractual and evidentiary decision, not merely the midpoint or the planning scenario.

Now suppose an independently recovered group B stock movement supports exactly seven units. A dated supplier quotation establishes the applicable unit cost at 820. Group B then contributes 5,740. Keeping cleanup uncertain, gross covered loss becomes 12,340 to 12,640 and the corresponding payment becomes 11,340 to 11,640. Both lower and upper calculations can be reproduced. The width of gross uncertainty falls from 4,000 to 300. The width of payment uncertainty falls from 2,900 to 300. The evidence was informative about both the factual valuation and the possible amount payable.

A contrary interpretation excludes cleanup under the stipulated policy. With group B established and group C corroborated, gross covered loss then becomes 11,740 and net payment 10,740. The factual invoice remains part of the economic consequences but is outside covered loss under that assumed interpretation. The record should identify this as a coverage classification rather than a change to physical damage. If the eligibility decision remains unresolved, the conditional covered scenarios should both be kept: 10,740 net with excluded cleanup, or 11,340 to 11,640 with included cleanup. Reporting the convex hull 10,740 to 11,640 is a convenient outer range, but it includes unsupported intermediate values. Keeping the disjoint alternatives visible is more informative.

If the insurer had already paid a documented advance of 4,000, outstanding net payment under the main established-quantity scenario would be 7,340 to 7,640, assuming the advance applies fully against the same benefit. This is a payment reconciliation rather than a revised economic loss. Any separate salvage recovery is handled on the appropriate policy and accounting basis; a recovery retained by the insurer must not automatically be subtracted again from a claimant's payment. The payment ledger should identify recipient, date, currency, purpose and allocation.

Capturing a repeatable record

A case audit begins with a frozen information cutoff. Preserve the original records and protected access references, a map of identities, the valuation basis, units and currency, the contract version and separately established eligibility decisions. Every input should identify its evidence or explicit assumption. Every transformation should be reproducible. The evidence inventory should support review of source quality, duplication and contradiction. The decision record should distinguish provisional, disputed, accepted, excluded and paid amounts.

For the worked fire case, a reviewer should be able to reconstruct the table from each stated input, repeat the deductible and limit order, identify the observations that narrowed group B, and determine the effect of alternative cleanup treatment. A second reviewer who agrees with all inputs but obtains a different total has found a computational discrepancy. Reviewers who disagree on group B quantities have identified an evidentiary discrepancy. Reviewers who disagree on cleanup eligibility have identified a contractual discrepancy. Each has a different escalation path.

A minimal change log might state that the group C amount increased from 720 to 1,200 after four independently identified units were corroborated. The group B interval changed from 3,500-7,200 to 5,740 after quantity and price evidence were accepted. The cleanup interval remained 600-900. The gross and payout recomputations follow from those changes. A human-readable explanation and machine-readable line-item record should agree. The text record is not a substitute for the structured evidence, and the structured evidence is not a substitute for an intelligible explanation.

A review checklist should test the decision date, asset identity, unit conversions, matching of replacement and repair alternatives, formula sequence, sign of adjustments, deductible and limit placement, source dates, and treatment of missing amounts. It should examine whether excluded items have been mistaken for zero economic loss. A second reviewer should examine a selected set of records against source material, rather than merely recalculate the spreadsheet from transcribed inputs. The audit should preserve disagreements and their resolution instead of recording only a final pass mark.

Automation and accountability

Automated extraction can reduce repetitive transcription but must preserve a relation to the original evidence. Optical recognition errors can alter decimal separators or units; a language model can create a plausible asset description that does not appear in a source. Inferred or machine-extracted values should therefore remain distinguishable from validated observations. Restrict automation to tasks whose output can be verified at the required level. In a high-value file, an apparently small transcription error may be decisive.

A valuation model should maintain versioned parameters, reference data, geographical scope, intended target, missing-data treatment and overrides. An override needs an explanation and should remain visible after later updates. A model trained on settled amounts is predicting a historical decision process unless an independent economic target has been established. If historical outcomes reflect inconsistent deductions or incomplete evidence, the model can replicate those inconsistencies at scale. Predictive accuracy against payment is not sufficient evidence of appropriate loss measurement.

The availability of alternative records also creates a fairness concern. A claimant who keeps extensive digital records may be able to establish an asset sooner than a claimant whose documentation was destroyed. Different evidence can appropriately lead to different certainty, but an unexplained rule that equates documentation absence with zero loss may convert record-keeping differences into unexplored monetary disparities. Review must distinguish differences justified by the relevant proof rules from errors due to the administrative design of the estimation process.

Research design for independent empirical testing

Future empirical work could assemble de-identified claims with known valuation bases and independent reference measurements. It would define the target before sampling, retain a record of exclusions, and stratify by loss type and cause of record incompleteness. It could then compare deterministic reconstruction, feasible bounds, model-based estimates and hybrid review under specified masking mechanisms. Claim clusters sharing event, adjuster or supplier should be handled explicitly. Operational outcomes should include time, revision frequency and burden of requests as well as estimation error.

A credible study would separately report the performance of the methods under observed random-like omission and under informative omission. The latter might be simulated by preferentially hiding the most expensive units or those from severely damaged areas. An alternative mechanism could preferentially hide low-cost purchases. Results would depend on these constructed mechanisms and the composition of the dataset; a single performance number would conceal the underlying assumptions. In actual observational data, the values underlying genuinely missing observations are usually unavailable, so a complete-case masking experiment cannot alone establish real-world calibration.

For an independent field audit, selected reopened claims could be compared with subsequently obtained documentation, with all subsequent evidence distinguished from material available at the original valuation date. Selection of reopened claims is potentially informative and cannot simply be generalized to all claims. Where no reliable reference can be obtained, the study should report unresolved outcomes instead of treating the latest adjudication as flawless ground truth. These limitations define the research program rather than invalidate the use of carefully scoped comparisons.

Integrated Analytical Synthesis

The material supports three connected conclusions. First, the monetary uncertainty of an individual loss depends on the structure of the missing evidence rather than the count of absent documents. Missing records can conceal unit identity, price, scope, temporal basis, dependency, or the entire population of affected assets. Consequently, a simple completion percentage cannot establish the reliability of the estimate. An inventory with three uncertain high-value units may be more consequential than one with fifteen missing details about verified low-value units.

Second, defensibility arises from the relation among observations, explicit assumptions, valuation rules and calculation. Identification bounds reveal what remains compatible with the stated facts and assumptions. Statistical estimates answer model-dependent predictive questions. Contractual transformations determine how the same economic uncertainty translates into monetary consequences for a particular benefit. The strongest practice is to preserve these levels and communicate the exact meaning of their outputs. A range can be more defensible than a point while still being too broad for a final decision. A point can be useful for operational planning without constituting evidence of a particular historical fact.

Third, institutional procedures influence the obtainable evidence. These procedures can improve accuracy when they focus requests and give an opportunity for correction. They can degrade information when they demand immaterial documents, fail to explain adjustments, or encourage premature closure. Different legal and regulatory forms establish different obligations. The comparative sources support treating procedural sufficiency, economic quantification, and contractual outcome as connected yet independently reviewable components.

A case-level estimate should remain revisable when a material contradiction or new verified record appears. Closed claims may remain closed for legal or administrative reasons; scientific versioning preserves the historical estimate and the context of any new knowledge without asserting a right to reopen. The ability to reproduce every version is more informative than the presentation of one apparently final number. It protects the distinction between decision quality assessed at a historical date and the quality of knowledge obtained later.

The contribution of this dossier is an original integration of missing-data reasoning, valuation arithmetic and constraints, claims process and auditability, including synthetic contrary cases that expose mistaken aggregation, midpoint selection and nonlinear payment conversion. It advances a substantive analytical interpretation using established methods, not a new named doctrine or an empirical discovery. The research questions can therefore be answered at the level of theoretical and applied claims science while leaving specific legal decisions and calibration studies to the evidence required for them.

Research Limitations and Scholarly Controls

The literature review is selective rather than systematic. Rubin's article and Manski's book are represented only through verified publisher abstract or overview material; their proofs and detailed chapters were not independently examined. Their cited propositions are restricted accordingly, and the mathematical examples are derived within this monograph. The NIST report is represented through the official landing-page abstract and bibliographic record. The practical preservation analysis is an application by the author rather than a claim to reproduce the report's full procedural requirements.

Regulatory sources were consulted as retrieved during the research session. An online consolidated text may subsequently change; a rule's application to a real claim requires verification of its effective version, jurisdiction, product and policy. The FCA test-case account is a regulator summary, not a comprehensive analysis of the judgment or later appellate authority. The Australian report's sample and methodological limitations are retained. No claim to have collected its original data, conducted a representative survey, or measured comparative causal effects is made.

The quantitative cases are synthetic and deliberately simplify contract terms, tax, valuation timing and scope. Their role is to expose mechanisms and support mathematical reproducibility. They neither estimate real-world error distributions nor recommend a settlement value. There is no underlying confidential claim file and no independent external peer review. Individual case investigation, professional advice and statutory interpretation require appropriate authorized expertise and evidence.

The internal scholarly review examines consistency of estimands, observation status, assumptions and contractual transformations; recalculation of illustrative arithmetic; alignment of regulatory propositions with the specified clauses or checked summaries; and scope of the cited research. Editorial review checks exact title, unnumbered organization, American English, coherent source anchors, full bibliography and reader-facing separation of examples from empirical findings. Technical repository checks are recorded separately in the research package. Scientific and technical review establish different kinds of assurance.

Conclusion

Claim loss quantification from partial records is an exercise in reconstructing an economic quantity under identifiable evidentiary, valuation and contractual constraints. Its scientific substance lies in establishing the object of measurement, discovering which facts the available records actually support, calculating the consequences of unresolved information and preserving the reasoning through which a result is obtained.

The investigation establishes that missingness can be informative, that observation count alone does not establish evidence quality, and that interval estimates can represent uncertainty without assigning unjustified probability. It explains how deterministic identities can narrow quantities, how market or cost evidence supplies valuation inputs, and how predictive methods require explicit assumptions and validation. Numerical illustrations show that nonlinear policy conditions can materially alter the relation between uncertainty in gross loss and uncertainty in payments, even when the policy calculation itself is exact.

Comparative conduct sources establish the relevance of clear reasons, proportionate information acquisition, documented investigation, competence and opportunities for contestation within their respective scopes. The resulting practical conclusion is that a case estimate should disclose the amount established, the amount conditionally inferred, the unresolved range, the decision basis, and the evidence that could alter the result. A reproducible audit trail makes these distinctions operational. It also enables later correction without confusing a revised understanding of loss with an unexplained revision of payment.

The central achievement of the inquiry is a precise account of how different forms of incomplete knowledge affect individual claim quantum and how those effects can be made visible within an accountable insurance process. The methods and proposed empirical tests remain within established statistics, valuation and claims science. Their validity in particular insurance settings depends on the supporting evidence, applicable contract and law, and independent examination of the case.

References

  • Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581-592. DOI: 10.1093/biomet/63.3.581. Source

  • Manski, C. F. (2003). Partial Identification of Probability Distributions. Springer. DOI: 10.1007/b97478. Source

  • Guttman, B., White, D. R., Williams, S., & Walraven, T. (2022). Digital Evidence Preservation: Considerations for Evidence Handlers. NISTIR 8387. National Institute of Standards and Technology. DOI: 10.6028/NIST.IR.8387. Source

  • Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality. Final revision, December 2016. Source

  • Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Source

  • International Association of Insurance Supervisors. (2024 edition, online consolidated text retrieved November 19, 2026). Insurance Core Principles and ComFrame, ICP 19, Conduct of Business, standard 19.10 and guidance 19.10.1-19.10.12. Source

  • Financial Conduct Authority. (2008, as subsequently consolidated; accessed November 19, 2026). Insurance: Conduct of Business Sourcebook, ICOBS 8, Claims Handling, rule 8.1.1R. Source

  • California. (Current text consulted November 19, 2026). California Code of Regulations, title 10, section 2695.7, Standards for Prompt, Fair and Equitable Settlements. Legal Information Institute reproduction. Source

  • California. (Current text consulted November 19, 2026). California Code of Regulations, title 10, section 2695.9, Additional Standards Applicable to First Party Residential and Commercial Property Insurance Policies. Legal Information Institute reproduction. Source

  • California Department of Insurance. (Current index consulted November 19, 2026). Fair Claims Settlement Practices Regulations, California Code of Regulations, title 10, chapter 5, subchapter 7.5. Source

  • United States. (Current text consulted November 19, 2026). 44 CFR Part 61, Appendix A(1), Standard Flood Insurance Policy, Dwelling Form, section VII.G, Requirements in Case of Loss. Legal Information Institute reproduction. Source

  • Financial Conduct Authority. (2024). Findings of multi-firm review into insurers' valuation of vehicles. Source

  • Financial Conduct Authority. (2022). Claims handling lessons learned from business interruption insurance. Source

  • Financial Conduct Authority. (2021 judgment and subsequent updates; consulted November 19, 2026). Business interruption insurance, Supreme Court judgment section. Source

  • Australian Securities and Investments Commission. (2025). Claims handling and settling: How to comply with your AFS licence obligations. Information Sheet INFO 253, June 2025 file; table 4, pp. 13-15. Source

  • Australian Securities and Investments Commission. (2023). Navigating the storm: ASIC's review of home insurance claims. Report REP 768, August 2023, pp. 3-5 and Figure 1. Source