Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Claims Settlement Data Selection and Closed-File Bias

Abstract. Closed-file insurance claims data condition observation on administrative completion, payment, or other outcome-dependent criteria. This monograph investigates how such selection affects settlement severity, duration, payment probability, claimant outcomes, and reserve evaluation. It distinguishes reported cohorts, accident cohorts, calendar closure flows, open inventories, and payment records; develops probability identities and constructed counterexamples; examines right censoring, reporting delays, competing events, and reopening; and assesses weighting, imputation, and sensitivity bounds. Actuarial research and selected official reporting and professional standards ground the analysis. The central conclusion is that population inference requires explicit estimands, complete eligible inventories or justified observation models, and uncertainty disclosure. Closed-file evidence remains scientifically valuable within its defined frame. No proprietary dataset, new theoretical system, or empirical market-wide magnitude is asserted.

Keywords: insurance claims; closed-file bias; claim settlement; selection bias; informative censoring; delayed reporting; inverse probability weighting; missing data; actuarial reserving; claimant outcomes; claims analytics

Introduction

An insurance claim is a time-dependent contractual and administrative process. An extract of closed files is a conditional observation of that process: it records cases that reached a specified administrative state before a specified date. Research on the severity, duration, or quality of settlement must therefore identify the population represented by the extract and the population about which it intends to draw conclusions. The distinction matters even when every record in the extract is accurate.

This dossier examines closed-file bias as the discrepancy attributable to selecting claims according to settlement, closure, payment, or closely related observation requirements when the scientific target includes other claims. The object is the claim-level observation mechanism, rather than the representativeness of underwriting exposures in general. A closed-file study may answer a legitimate question about completed files. Bias arises when the resulting conditional distribution is treated as the distribution of a different population or when selection changes an association of substantive interest.

Three questions organize the investigation. Which operational and statistical mechanisms determine entry into closed-file samples? Under what assumptions can incomplete histories identify settlement costs, durations, and claimant outcomes? How should uncertainty in unobserved claims constrain comparisons, predictions, and audits?

The method is a targeted narrative synthesis of statistical literature, actuarial research, empirical malpractice studies, and selected official United States and United Kingdom instruments, combined with elementary probability derivations and constructed counterexamples. These examples are fully hypothetical, with values chosen to isolate mechanisms; the study does not estimate any insurer's actual bias or replicate proprietary datasets. The legal examples illustrate specific reporting and conduct requirements, not universal national law or a global regulatory census. Some publications are available only through abstracts, which are used only for the limited propositions they explicitly support. The primary scientific contribution is the integration of observation timing, administrative state, outcome definitions, and decision evaluation within the approved research scope.

Open and Closed Claim Populations

Units and time origins

A loss event can produce several claims, a claimant can have several coverage components, and a claim can have multiple demands, payments, or closure episodes. At least five units must be distinguished: incident, policy claim, claimant, coverage item, and financial transaction. A closure-rate denominator based on claim identifiers cannot be compared directly with a numerator counting payment transactions. Similarly, a bodily injury claim with three injured persons may count once administratively but three times in an individual compensation study. The unit of observation must follow the estimand; separate linkage keys may be needed for all five units.

Let A denote accident or loss date, R first notification or registration, C first administrative closure, and V the valuation cutoff. The report delay is R−A; the first closure duration measured from registration is T=C−R. Accident-to-closure duration includes both components. For a claim registered before V, the observable age is V−R. A requirement that C≤V filters claims precisely according to their observed age and duration. A historical extract whose files were closed during a calendar period differs from an inception cohort of all claims registered during a period and followed to a common age.

A current closure flag is different from first closure, last observed closure, and ultimate resolution. A file marked closed can later reopen, for example after supplemental documentation, additional indemnity, an appeal, or a correction. There is generally no observable calendar date at which eventual permanence can be established without an assumption or a defined follow-up horizon. A research design should use a finite endpoint such as first closure, closure without reopening for twelve months, or status at a fixed horizon. Each endpoint answers its own question. Lindholm and Zakrisson's (2022) model of claim openness explicitly accommodates closing and reopening, illustrating why a state history is richer than a single claim flag.

Payment amounts and contracts

Define Y as the ultimate gross indemnity for the specified unit on a stated nominal or real currency basis, including zero-indemnity outcomes. Let P(V) denote cumulative gross indemnity by V, and Q(V) the recorded estimate of remaining indemnity. In that convention P(V)+Q(V) is case-incurred indemnity, not a realized ultimate outcome. A separate expense process may include allocated adjustment expenses, litigation costs, and other defined items. Recoveries, subrogation, reinsurance, salvage and exchange-rate effects require explicit gross/net conventions. If payments can be reversed, even cumulative paid amounts need not provide strict lower bounds on ultimate net cost.

Contractual entitlement, agreed settlement, payment completion, and file closure can have different dates. A denied claim may have zero indemnity and material investigation expense. A file may be closed because a claimant stopped pursuing it without any adjudication of underlying merit. A paid-only sample excludes zeros and changes the estimand from expected indemnity per claim to expected indemnity conditional on payment. In a fixed cohort, E(Y)=Pr(Y>0)E(Y|Y>0) for nonnegative indemnity. This simple decomposition shows why even a perfectly measured positive-payment mean cannot supply the overall mean without an appropriate payment probability.

Stocks, flows, and denominators

For a constant population of unique claim units over a reporting interval, an open-inventory balance is O_end=O_start+N+U−F, where N is first registrations that enter open status, U is transitions from closed to open, and F is transitions from open to closed. Instantaneous registrations and closures must be included consistently as both events. Portfolio transfers, deleted duplicates, migrations, or classification corrections require separately reconciled signed adjustments. Multiple reopenings and closures count as transitions, while cohort denominators count unique units. Confusing these creates apparent backlog improvement even when files merely change administrative status.

An operational report can combine a calendar-period flow of completed files with an opening-stock count, but these quantities have different sampling frames. Closure cohorts mix old and new reported claims, and their mix can change with backlog clearance or staffing. A reported-cohort rate at a fixed age answers how an inception group progressed. An accident-cohort rate additionally requires accounting for claims that have not been reported. An inventory of currently open cases tends to contain more long-lived cases than a flow of newly reported claims. Neither the stock nor the flow automatically substitutes for the other.

Scope of an observed claim

A reported-claim inventory excludes losses that have not yet been reported and losses that are never submitted. In notation, the target might be all covered incidents, all submissions, all registered claims, all accepted claims, or all claims with completed payments. Movement between these sets can depend on eligibility, awareness, contractual thresholds, insurer registration rules, and policyholder decisions. These different selection stages must be separated to prevent the administrative universe from being mistaken for the universe of insured harm.

The empirical study by Schiff et al. (2013), Massachusetts primary care closed claims, explicitly addresses missing unfiled incidents and delays between occurrence, filing, and closure. Its closed-claim design remains informative for the selected cases, while the acknowledged limitations constrain generalization to patient safety events or contemporary practice. That distinction is transferable to other classes of insurance without assuming equal reporting behavior across them.

Selection Mechanisms and Missing-Not-at-Random Data

Selection and the outcome distribution

Consider a fixed registered cohort, an ultimate outcome Y, covariates X known for every claim in the cohort, and S=1 when the extraction rule supplies a finalized outcome. The extraction can require closure at V, positive payment, documentation completeness, or a union of conditions. Let p=Pr(S=1), assumed positive. For finite expectations,

E(Y|S=1)−E(Y)=Cov(Y,S)/p.

This follows from E(YS)=pE(Y|S=1). It states the sign and magnitude of the discrepancy in terms of the covariance with inclusion; the direction is therefore an empirical and structural question rather than a definition. If expensive claims tend to take longer, early closure can reduce the observed mean. If a rapid settlement initiative targets expensive files while small disputed claims remain pending, early closure can increase it. A positive-payment filter can raise the mean while a premature-closure filter reduces it. Their combined effect depends on joint selection.

The observed density is proportional to the target density multiplied by the inclusion probability at each outcome and covariate value. In symbols, for a density conditional on X=x,

f(y|x,S=1)=f(y|x)Pr(S=1|Y=y,X=x)/Pr(S=1|X=x).

This relation also holds for probability masses. The observed conditional distribution combines the substantive outcome process with the observation process. A large sample size reduces random error in estimating the selected distribution but does not remove the multiplication by unequal selection probabilities.

Collider selection and process comparisons

Suppose latent claim complexity U affects eventual indemnity Y and closure speed S, while claims-handling intervention Z also affects closure speed. One plausible causal diagram is Z → S ← U → Y, with a possible separate arrow Z → Y. Restricting analysis to S=1 can associate Z with U even if Z was assigned independently of U. This is the structural mechanism discussed by Hernán, Hernández-Díaz, and Robins (2004) in their analysis of selection bias. It is distinct from confounding, in which a common antecedent influences exposure and outcome. Both can coexist in administrative claims research.

An example illustrates the effect without confounding. Suppose two independent binary variables, intervention Z and complexity U, each take values 0 and 1 with equal probability. Let indemnity be 100 for low complexity and 500 for high complexity, regardless of intervention. Let closure by V occur whenever Z=1 or U=0. Among closed untreated claims, complexity is always low, and mean indemnity is 100. Among closed treated claims, both complexity classes occur equally, and mean indemnity is 300. The true mean under either intervention is 300. The intervention has no effect on indemnity in this constructed example, yet conditioning on closure produces a difference of 200. Randomization of assignment alone cannot rescue an analysis restricted to a post-assignment selected outcome.

The same mechanism complicates comparing adjusters, insurers, regions, legal representation groups, and automation programs. Assignment may depend on initial severity; closure may depend on both assignment and severity; payment rules may change during the observation window. A causal impact design must establish the time of intervention, the eligible cohort, the treatment strategy, sufficient pre-intervention confounders and appropriate handling of post-treatment missingness, and allow for outcome-altering process changes. A descriptive case-mix adjustment is not automatically a causal estimator.

Missingness mechanisms

Missing completely at random would require inclusion to be independent of the variables relevant to the analysis. Missing at random (MAR), in a simple outcome-missingness setting, permits S to depend on observed X but requires conditional independence from the missing Y given X. Missing not at random (MNAR) permits dependence on Y even after conditioning on available observed information. Actual selection may also depend on partially observed histories, such as negotiations and interim payments; the definitions must then be applied to the available data pattern rather than just baseline variables.

Rubin's (1976) inference and missing data distinguishes conditions under which the missingness model can be omitted from likelihood or Bayesian inference. It is not sufficient to label data MAR if relevant predictors were never collected or if missingness parameters are entangled with outcome parameters. Administrative censoring at a fixed calendar date does not alone establish MAR for ultimate indemnity: completion before that date is a function of duration, and duration may share latent determinants with indemnity.

Missingness is also different from structural zeros. A rejected claim with known zero indemnity has an observed outcome under the specified measure, whereas a pending claim with no payment yet has an incomplete outcome. Replacing missing ultimate amounts with zero confounds an unresolved claim with a financially resolved zero. Conversely, using a case reserve as if it were realized indemnity combines a judgment at valuation with a later outcome.

MAR cannot generally be established solely from observed final amounts because the relevant counterfactual information concerns unobserved outcomes. An audit can test conditional differences on observed characteristics, evaluate predictions in mature cohorts, and assess the plausibility of mechanisms. A sensitivity analysis remains necessary when closure may depend on unobserved determinants.

Censoring and Delayed Reporting

Right censoring versus truncation

For registered claim i, let T_i be time from registration to a specified first event and L_i be V−R_i, the available follow-up time. Observe t_i=min(T_i,L_i) and event indicator d_i=1 when T_i≤L_i. An open file contributes known event-free time through t_i. Ordinary right censoring retains that record and its risk time; a closed-only sample discards it. The latter is a duration-dependent selection or truncation of event times, not an ordinary censored dataset.

A claimant also may enter a system after loss occurrence. If the time scale starts at occurrence and only claims with R≤V are present, late notifications with R>V have absent records, not merely incomplete event times. Delayed entry into an observation system can be handled by left-truncated risk sets where entry is observed and the necessary conditional independence holds. But a reported-only database cannot enumerate never-reported losses by applying a left-truncation formula alone. Policies, exposures, external incident systems, or historical reporting-delay models are required to define or estimate the absent population.

Changes in notification technology, reporting thresholds or policy wording may alter report-delay distributions. A reporting-delay model estimated on old accidents is transportable to recent accidents only to the extent that relevant mechanisms remain comparable, or are adequately modeled. Calendar-time adjustment and policy-version stratification are consequently substantive choices.

Product-limit estimation

Kaplan and Meier's (1958) product-limit estimation addresses event times with incomplete observation under conditions on the loss-to-observation mechanism. In a registered claim cohort, an empirical survival estimate of time to first closure is the product over event ages j of (1−d_j/n_j), where n_j is the number at risk just before age j and d_j the number of first closures at that age. Censored claims contribute to the risk set until their last observed age; they do not count as closures. Independent censoring, possibly within correctly specified strata, is required for the corresponding marginal or conditional inference. A common valuation cutoff can satisfy suitable conditions within an inception cohort, but staggered entry combined with calendar-driven outcome change can violate unconditional independence.

Consider ten claims with first-closure events for two at month one, three at month two, and one at month three, and the remaining four open at a common month-three cutoff. With no earlier censoring, estimated survival at month one is 8/10=0.8, at month two 5/10=0.5, and at month three 4/10=0.4. The closed-only empirical proportion completing by month two is 5/6=83.33%, which answers a selected-sample question. The cohort proportion is 50%. A Kaplan–Meier median is the first age at which estimated survival is at most one-half: month two. Because 40% remain unresolved, unrestricted mean duration requires extrapolating the unobserved tail, while the restricted mean through month three is 1+0.8+0.5=2.3 months in this discrete example.

Competing events and reopening

For a first-exit endpoint, paid settlement, denial, withdrawal and other closure categories can be treated as mutually exclusive competing events when the classification is exhaustive and the first occurrence ends the designated risk period. Cause-specific hazards describe event rates among claims still at risk. The cumulative incidence of category k through age t sums or integrates overall event-free survival times the cause-specific hazard for that category. Treating every other closure type as ordinary independent censoring and interpreting one minus the resulting product-limit survival as the observed probability of paid settlement generally targets a different hypothetical quantity and can overstate actual cumulative incidence.

The preceding distinction follows from elementary risk-set probability: a share of claims removed by denials can no longer experience paid settlement as their first terminal event. If paid settlement can occur after an initial denial or reopening, the endpoint changes to a multistate process. States may include pending investigation, accepted but unpaid, disputed, paid and closed, denied and closed, and reopened. A transition model estimates movements between defined states; a closure event alone does not establish an absorbing state. If the time already spent in a state matters, a semi-Markov representation or elapsed-duration covariates may be appropriate.

Lindholm and Zakrisson (2022) model reopening in an aggregate reserving setting and illustrate performance using simulated claims. Their work provides an actuarial counterpart to the operational distinction between inventory and transitions. It does not establish that one parametric specification fits all insurers or lines of business.

Partial amounts and temporal leakage

A payment history provides information even before closure. An open claim with payments of 20, 15 and 10 has a cumulative amount of 45 in that currency unit, and its eventual gross amount may exceed 45 under nonnegative, nonreversible payments. Each payment is an event, and the remainder is an uncertain quantity. Replacing this history with missingness discards observed evidence. The dates at which information became available must be separated from the dates to which it refers: a medical report added after validation cutoff, even if describing an earlier accident, is unavailable for a prediction at that cutoff. A retrospective data warehouse therefore needs recorded-at timestamps as well as effective event dates.

Reserve estimation has a different time perspective from settlement-outcome evaluation. The remaining cost at V includes additional payments on reported claims and relevant incurred-but-not-reported obligations. Depending on accounting definitions, payments on previously closed claims that reopen may be included in reported-claim development. Historical settlement averages are informative only when linked to these future-payment components, claim mix, and the valuation basis.

Weighting, Imputation and Sensitivity Bounds

Inverse probability weighting

Suppose X is observed for every registered claim, S indicates whether ultimate Y is observed, and pi(X)=Pr(S=1|X)>0 in all relevant strata. Assume conditional exchangeability of observed and missing outcomes given X, adequate probability estimation, and finite moments. Then

E[S Y/pi(X)]=E[E(SY|X)/pi(X)]=E[E(Y|X)]=E(Y).

In a finite cohort of size N with known design probabilities the Horvitz–Thompson-style estimate of the mean is the sum of S_iY_i/pi_i divided by N. A normalized weighted mean divides the weighted outcome sum by the sum of S_i/pi_i. These forms can have different finite-sample properties but converge to the same target under suitable conditions. The weights must represent the actual extraction rule and the target frame. A payment-inclusion weight at the transaction level cannot automatically be used as a claim-closure weight.

Seaman and White's (2013) review of inverse probability weighting discusses why complete-case restriction can bias results and considers imputation, weight stabilization, trimming and augmentation. At low inclusion probabilities, large weights concentrate influence on a few cases. An appropriate diagnostic set includes overlap by relevant strata, probability calibration, weight quantiles, maximum weights and effective sample size, defined descriptively as (sum w)^2/sum(w^2). This quantity diagnoses concentration; it is not a general substitute for a design-appropriate variance estimator. Trimming reduces variance at the potential cost of residual bias or a changed target. Both the threshold and the resulting population coverage must be reported.

In a registered cohort containing 100 simple and 100 complex claims, suppose 90 simple claims close with mean eventual gross indemnity of 100 and 30 complex claims close with mean of 500. Assume the remaining outcomes have these same means within complexity strata. The unweighted closed mean is (90×100+30×500)/120=200, whereas the cohort mean is 300. With closure probabilities 0.9 and 0.3, inverse-probability contributions total 10,000 and 50,000, and the estimated cohort mean is 60,000/200=300. This is a constructed example in which observed complexity is sufficient for the correction.

Now let the 70 still-open complex claims instead have mean 900, while observed complex claims remain at 500; assume the 10 open simple claims remain at 100. The observed inputs, weights and corrected estimate are unchanged, but the actual cohort mean is (10,000+15,000+63,000)/200=440. A model using only complexity cannot identify the missing outcome difference. The gap of 140 is not cured by more precise estimates of the same probabilities. This counterexample locates the substantive limit of selection weighting.

Outcome regression and imputation

Regression predicts missing outcome values by modeling the conditional outcome distribution given covariates and, where available, partial claim histories. If correct and transportable to open files, the model supports standardized means over the complete inventory. For a nonnegative gross loss with many zeros, a two-part model can distinguish probability of positive indemnity and positive severity. A separate component may predict adjustment expenses. Model predictions must respect contract limits and the chosen financial scope; cost totals can exceed indemnity limits if expenses outside the limit are included.

Multiple imputation propagates uncertainty by drawing plausible missing outcomes under an explicit predictive distribution and combining estimates across completed datasets. Deterministic single imputation of conditional means understates the dispersion of completed outcomes and can distort downstream nonlinear metrics. Imputation must preserve dependencies among ultimate cost, observed partial amounts, payment indicator, resolution type and duration. For example, a simulated future payment should be compatible with the observed ledger, and a sampled positive amount should not be paired with a terminal zero-payment state under a definition that rules that out.

A doubly robust mean estimator combines an estimated outcome regression m(X) with an inverse-weighted residual correction:

mean over i of {m(X_i)+S_i[Y_i−m(X_i)]/pi(X_i)}.

Under MAR, positivity, suitable sampling and regularity conditions, consistency may hold when either the outcome regression or the selection probability model is correctly specified. If both are wrong, or the observation depends on the unseen outcome given X, double robustness supplies no general guarantee. Flexible fitting and cross-fitting can reduce some estimation biases, but do not create support where no comparable outcome has been observed.

Sensitivity analysis and partial identification

Let p be the closed proportion, mu_c the mean known ultimate outcome for those closed, and mu_o the mean for the remaining claims. Then the overall mean equals p mu_c+(1−p)mu_o. With no restriction on mu_o other than nonnegative gross indemnity, the closed files supply a lower bound p mu_c but no finite upper bound. When known contractual caps apply to the specified indemnity and a claim's lower bound L_i is valid under its payment convention, the cohort mean lies between (sum of observed Y plus sum of open-claim L_i)/N and (sum of observed Y plus sum of appropriate open-claim U_i)/N. Caps must account for coverage stacking, aggregation and the unit. Uncapped legal expenses, interest or future net recoveries may invalidate the proposed endpoints.

A pattern-mixture sensitivity parameter delta=mu_o−mu_c gives mu=mu_c+(1−p)delta. In a cohort with 60% closed and closed mean 200, open mean 100 implies overall mean 160; open mean 500 gives 320; open mean 900 gives 480. These are conditional scenarios, not sampling confidence intervals. The construction can be stratified on X to compare open cases with comparable closed ones, allowing domain experts to specify plausible departures from MAR. Tail-sensitive quantities, such as reserve quantiles or large-loss proportions, need distributional sensitivity analysis rather than a mean shift alone.

For binary outcomes, bounds are particularly transparent. If 120 of 200 eligible claims are resolved and 96 of those have positive indemnity, then the completed-file fraction is 80%, but the cohort fraction of eventual positive outcomes is between 96/200=48% and (96+80)/200=88%, absent further information. If some open files already have irreversible positive payments, the lower bound can improve. Model-based intervals cannot replace identification bounds unless their assumptions are accepted and their uncertainty interpretation is clear.

Choosing a correction

The choice depends on the available observation frame. With an inventory of all registered files and credible conditional independence, weighting, regression, or hybrid estimation can target registered-cohort outcomes. With reliable event histories, survival and multistate techniques can address durations and competing first events. With only selected closed outcomes and no denominator, representativeness is generally not identified; a new inventory, external sampling frame, independent evidence, or explicitly bounded inference is needed. The method should be chosen after establishing what information exists, not from the sophistication of an available algorithm.

Comparative Claims Model Evaluation

Actuarial evidence and its scope

Antonio and Plat's (2014) micro-level stochastic loss reserving describes an empirical European liability application using occurrence, reporting, payments, and final settlement to project development of unresolved files. The published abstract reports an advantage over aggregate methods in that case study. It supports the relevance of detailed histories without establishing a universal ranking among reserving approaches.

Lindholm and Zakrisson's (2022) collective reserving model retains aggregate count and payment information while incorporating the dynamics of open files. Its numerical comparison relies on simulated histories. A granular model can exploit individual features but demands greater data availability and careful validation; an aggregate model can retain valuable stock and transition information without storing every individual predictor. These approaches occupy different points on a data and complexity continuum.

Calcetero Vanegas, Badescu, and Lin's (2025) micro-level Chain-Ladder method provides a direct actuarial treatment of inclusion weighting in reserving and connects a traditional aggregate method to a sampling interpretation. The authors' revised preprint describes a temporal evaluation on an insurance dataset and discusses instability under some weight configurations. The present dossier does not reproduce that private-data exercise or identify the version-of-record numerical results with the preprint without comparison. It uses the documented connection between observation probabilities and reserve estimation to explain why sampling assumptions belong inside claims model design.

Validation cohorts and leakage controls

A fair test begins with a frozen registered or accident cohort and a series of historical valuation snapshots. Define the prediction date, information actually available then, target horizon, and outcome adjudication rule. Every eligible claim at the prediction date belongs in the evaluation frame, whether open, closed, denied, partially paid, or later reopened, as appropriate to the endpoint. Mature follow-up or a specified censoring method determines which outcomes can be scored without assuming future information. Recent incomplete cohorts can be evaluated for observed interim targets separately.

Three comparisons are useful. A closed-only baseline estimates outcomes from completed files and documents its selection rule. An inventory-based adjusted baseline uses common pre-cutoff features and observed missingness mechanisms. A process-aware model also uses partial payment and status histories. All are trained and tested on comparable cohorts and targets. An aggregate development model provides an additional benchmark for portfolio reserves. A model with accurate aggregate reserves can still allocate them poorly across individual claims; a model with good claim-level average error can still produce an inaccurate tail reserve.

Temporal splitting is essential for operational forecasting. A random split of episodes can place the same reopened claim, claimant, accident, catastrophe or litigation group in both training and test sets. Such splits may exaggerate performance. Stable grouping by claim and relevant common loss event, followed by chronological separation, helps address correlated records and repeated outcomes. Hyperparameters and missingness models must be fitted inside training partitions; outcome-based feature selection, payment updates, later reserve revisions, and retrospective fraud labels must not enter information available at the earlier prediction date.

Scoring cost and duration

For mean ultimate indemnity, cohort-level signed error reveals systematic over- or underestimation; mean absolute error and root mean squared error reveal different sensitivities to large losses. Calibration by case-mix and age group tests whether predicted expected totals align with observed totals. For probabilistic costs, proper scoring rules or quantile losses can assess the distribution conditional on mature outcomes or defensible observation adjustments. Prediction intervals concern future outcomes; confidence intervals concern an estimated population quantity; sensitivity intervals concern assumptions or identification. A report should avoid conflating them.

For time to first closure, survival calibration at common horizons and appropriate censoring-adjusted scores are relevant where their assumptions hold. Cause-specific event probabilities should be compared with cumulative incidence for the defined first-event categories. For reopening, predictions of being open at a future time differ from time until first reopening or total remaining open days. Metrics must reflect the actual operational question and risk set.

For reserve forecasts, reconstruct paid-to-date and later development at each historical valuation, isolate reported and unreported components where possible, and compare predictions of subsequent cash flows with amounts observed over the chosen maturity horizon. Heavy-tail and catastrophe dependence require attention to clustering and exceptional losses. A single backtest in a stable period does not establish robustness to inflation, litigation changes or workflow shifts.

A rank reversal from administrative selection

Consider two hypothetical claims teams with 100 eligible claims each. Team A has 80 closed at cutoff, 72 with positive payment, and 20 unresolved; team B has 40 closed, 38 with positive payment, and 60 unresolved. Among resolved claims, A shows 90% positive payment and B 95%. The observed complete-file ranking favors B. If among open claims A later pays 18 and B later pays 30, final cohort payment proportions are A 90% and B 68%. Both scenarios are compatible with the cutoff table. Before observing the remaining outcomes, the bounds for A are 72–92% and for B 38–98%. Neither team can be ranked unconditionally at that cutoff from these figures.

The counterexample concerns the statistical definition of a positive payment, not whether any claim deserved payment. Entitlement and fair treatment require additional information. Even mature cohort rates can differ because of coverage, claim merit, or claimant needs. A scientific evaluation should stratify or standardize meaningful case mix, examine rejected and withdrawn cases, and avoid interpreting every positive-payment rate as a causal or normative performance score.

Robustness to institutional change

Different insurers can use administrative closure for different purposes. One system may close after a payment instruction, another after reconciliation, and another after litigation or appeal exhaustion. Acquisition, outsource transitions, digitization or a change in closure targets can break the history. A comparison should map status definitions, check mass shifts, inspect unexpected spikes in closure time, and reconcile transition counts against financial ledgers before attributing performance changes to substantive mechanisms. Domain shifts must be treated as changes in both the outcome and the observation process.

A useful stress exercise re-estimates results under alternative reporting-delay assumptions, closure definitions, weight truncation, and open-outcome shifts. Evidence that a result remains stable over plausible scenarios supports robustness within those scenarios. If rankings reverse, portfolio totals exceed tolerances, or vulnerable claimant groups vanish from the effective sample, the appropriate conclusion is conditional or bounded rather than a single decisive ranking.

Decision Implications and Audit Trails

Claims conduct and measurement

The United Kingdom Financial Conduct Authority's ICOBS 8 claims handling rule 8.1.1R specifies prompt and fair handling, progress information, avoidance of unreasonable rejection and prompt settlement after terms are agreed (FCA, 2008, as amended). These are obligations for relevant insurers within the rule's scope; no statistical proxy alone proves compliance. A closed-file turnaround metric may omit precisely the pending claims whose treatment requires scrutiny. For oversight, open duration, backlog age, reasons for delay, complaint handling, repeated requests and reopened outcomes need visibility alongside completed-file statistics.

The FCA's general insurance value measures reporting rules (2021, as amended) provide a particularly useful example of explicitly defined denominators. Average payout uses accepted claims with payments or benefits and closure at period end. The acceptance-rate definition uses registered and rejected counts, while registration and rejection have stated exclusions and period conventions. The prescribed figures answer regulatory-reporting questions and should be implemented as specified. They should not be silently interpreted as inception-cohort probabilities of eventual acceptance or mean ultimate indemnity for all submitted claims. Their cost measure also includes specified nonmonetary benefits and claim-specific handling costs, so comparisons with an indemnity-only severity model require adjustment or explicit separation.

In Washington State, RCW 48.140.020 defines a closed-claim reporting obligation for medical malpractice insurance within its statutory scope. A database collected pursuant to such an obligation is a valuable reporting frame whose coverage depends on the statutory entities, claim definitions and timing. It is not an outcome inventory of every medical incident or all unresolved demands. Legal mechanisms for closed-claim reporting should be recorded as part of dataset provenance. The two examples illustrate different purposes—consumer value measures and malpractice closed-claim reporting—and cannot be homogenized into a single global denominator.

Actuarial professional standards

For U.S. actuarial practice within its scope, the Actuarial Standards Board's (2016) ASOP No. 23 addresses data quality, including appropriate selection and review. The board's (2007) ASOP No. 43 addresses unpaid property/casualty claim estimates, including the intended measure, assumptions, data, reasonable alternatives, validation and disclosure. Its distinctions among accounting, valuation and review dates are particularly important when reconstructing historical datasets. Their professional and jurisdictional scopes must be preserved rather than treated as universal legislation. A scientific study can adopt the methodological disciplines of data provenance, defined measures and sensitivity while maintaining a separate account of actual regulatory obligations.

Decision support and claimant outcomes

Closed-file selection can affect resource allocation if a model understates the expected effort for complex active files. It can also affect triage if labels for disputes or fraud appear primarily in mature cases. A predictor trained on late-stage documentation may seem accurate retrospectively but be unusable at first notification. Its predictions may change actual observation probabilities when the organization uses it to prioritize closure. This feedback means the training distribution is partly produced by previous decision policies.

An evaluation of a new claims process should record the decision policy, assignment time, intervention, follow-up, reasons for noncompletion, and subsequent changes in policy. It should monitor the entire eligible population as well as subgroups affected by differences in documentation access, geography, disability, representation or communication barriers where lawful and justified. Use protected or sensitive characteristics only with appropriate authorization, minimization, privacy safeguards and a specified fairness purpose. Disparate completion rates or cost estimates warrant investigation, but cannot by themselves establish unlawful discrimination. Absence of measured disparity in selected closed files likewise cannot establish fairness in the whole inventory.

Fair treatment evaluation has outcome and process dimensions. A claimant can eventually receive appropriate indemnity after an excessive delay, or a timely denial can be contractually appropriate or inappropriate. A meaningful audit distinguishes time to first response, time to decision, payment completion, reasons for rejection or withdrawal, complaint and appeal outcomes, and claimant communication. File status serves as one administrative measurement within this wider evidence chain.

Reproducible data lineage

The following elements belong in an analytical data specification: the source system and version; unit and linkage keys; event and recorded-at timestamps; extraction cutoff and query; claim state definitions; financial and currency basis; population boundary; exclusion counts and reasons; reporting and closure cohorts; missingness indicators; reconciliation to inventories and ledgers; model fitting and validation windows; assumptions and uncertainty analysis. An auditable extract should permit the cohort to be reconstructed from a broader inventory, not only from the final analytical table.

The audit should preserve both first and subsequent closures, payments and reversals, case-estimate changes, and data corrections. Record deletion or merging requires provenance. A closed flag should be traceable to the event that established it. Historical snapshot reconstruction requires transaction or version histories; current overwritten records cannot always reproduce what was known at an earlier valuation. The organization should distinguish immutable evaluation snapshots from personally identifiable operational records and apply appropriate retention, access and security controls.

A reproducible selection report begins with counts of unique registered units, then separates those excluded by coverage, cutoff, missing fields, closure, payment and other rules. Exclusions should be presented as mutually exclusive sequence steps or as explicitly overlapping sets. For each critical stratum document the eligible count, closed count, paid count, still-open count, and distribution of observed age. Reconcile monetary totals and missingness patterns. Report the original unadjusted estimand and any adjusted or bounded result with its own target definition.

Interpreting evidence strength

A defensible completed-case descriptive analysis can have high internal measurement quality but limited population coverage. An inventory-based estimator can reduce selection discrepancies yet depend on unverified assumptions about future amounts. A process-aware model can represent reopening but be sensitive to sparse transitions and structural changes. No hierarchy of algorithms removes the need for a well-defined sampling frame and observation process.

An internal audit should ask whether an inference would remain valid if closures accelerated without any change to ultimate costs, whether the denominator would change after a data migration, whether model labels include post-cutoff information, and whether open-file outcomes compatible with the evidence could reverse the conclusion. These checks expose selection-induced apparent improvements and distinguish exploratory analytics from decision-grade evidence.

The external literature used here spans abstract-only and inspected full-text material. The dossier confines factual descriptions of restricted works to their accessible summaries; mathematical counterexamples and derivations are transparent authorial analysis. The examples have no empirical insurer identification. Their purpose is to establish logical possibilities, clarify required assumptions and guide subsequent properly authorized inventory-based validation.

Conclusion

Closed-file bias arises when administrative completion or payment selection changes the distribution represented by claims data relative to its intended target. The phenomenon encompasses outcome-dependent inclusion, risk-set depletion, delayed reporting, lost information from partial histories, and causal distortion from conditioning on a common consequence. Its magnitude and sign are contingent on the underlying claim and observation processes. The probability identity relating selected and full-cohort means establishes this directly.

The scientific remedy begins with the registered or accident population, a precisely defined observation date and financial measure, and meaningful claim-state histories. Survival analysis retains time at risk; multistate methods distinguish closure from permanent resolution; weighting and imputation can recover population quantities under support and missingness assumptions; sensitivity analysis and bounds communicate what remains unidentified. Statistical adjustments cannot recover information from an absent population without additional evidence or assumptions.

Actuarial research demonstrates the relevance of event histories, claim openness and inclusion probabilities to reserving, while regulatory reporting examples show why official counts and denominators must be interpreted according to their purposes. A closed-file extract can support a valid completed-file study or a specified compliance report, provided the resulting claims remain within that frame. A fair settlement evaluation requires an eligible-cohort perspective together with the substantive contract and conduct evidence appropriate to the question.

For insurance decision-making, the central implication is that an administrative outcome is also a measurement mechanism. Settlement policy, closure speed and data availability can change simultaneously. Evaluation must therefore be anchored to the information available at each decision time, tested against independent historical inventory cohorts, and accompanied by explicit identification and uncertainty limits. This synthesis establishes no new theory or proprietary empirical estimate. It provides a coherent scientific basis for identifying, diagnosing and controlling the distortions introduced by incomplete claim histories.

References

  • Actuarial Standards Board. (2007). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. Source

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

  • Antonio, K., & Plat, R. (2014). Micro-level stochastic loss reserving for general insurance. Scandinavian Actuarial Journal, 2014(7), 649–669. DOI: 10.1080/03461238.2012.755938. Source

  • Calcetero Vanegas, S., Badescu, A. L., & Lin, X. S. (2024). Claim reserving via inverse probability weighting: A micro-level Chain-Ladder method. arXiv:2307.10808, revised preprint. Source

  • Calcetero Vanegas, S., Badescu, A. L., & Lin, X. S. (2025). Claim reserving via inverse probability weighting: a micro-level Chain-Ladder method. European Actuarial Journal, 15, 89–135. DOI: 10.1007/s13385-024-00395-3. First online August 28, 2024. Source

  • Financial Conduct Authority. (2008, as amended). Insurance: Conduct of Business Sourcebook, Chapter 8: Claims handling, rule ICOBS 8.1.1R. FCA Handbook. Source

  • Financial Conduct Authority. (2021, as amended). Supervision Manual, SUP 16.27: General insurance value measures reporting, rules 16.27.3–16.27.11. FCA Handbook. Source

  • Hernán, M. A., Hernández-Díaz, S., & Robins, J. M. (2004). A structural approach to selection bias. Epidemiology, 15(5), 615–625. Source

  • Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. DOI: 10.1080/01621459.1958.10501452. Reprinted in Breakthroughs in Statistics, pp. 319–337; linked Springer reprint. Source

  • Lindholm, M., & Zakrisson, H. (2022). A collective reserving model with claim openness. ASTIN Bulletin, 52(1), 117–143. DOI: 10.1017/asb.2021.33. Source

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

  • Schiff, G. D., Puopolo, A. L., Huben-Kearney, A., Yu, W., Keohane, C., McDonough, P., Ellis, B. R., Bates, D. W., & Biondolillo, M. (2013). Primary care closed claims experience of Massachusetts malpractice insurers. JAMA Internal Medicine, 173(22), 2063–2068. DOI: 10.1001/jamainternmed.2013.11070. Source

  • Seaman, S. R., & White, I. R. (2013). Review of inverse probability weighting for dealing with missing data. Statistical Methods in Medical Research, 22(3), 278–295. DOI: 10.1177/0962280210395740. Source

  • Washington State Legislature. (2006, as amended). Revised Code of Washington, § 48.140.020: Closed claim reporting requirements. Source