RESEARCH DOSSIER
Survival Analysis and Kaplan-Meier Estimation in Insurance
Abstract. Insurance generates a distinctive class of time-to-event observations: death during cover, the first insured loss, progression into disability, recovery from a claim, policy lapse, settlement of an open claim, and termination of a benefit stream. The Kaplan–Meier product-limit estimator permits nonparametric estimation of a survivor function when durations are incompletely observed through right censoring. Its validity rests on the definition of the event and time origin, the construction of risk sets, the censoring and entry mechanisms, and the population about which an inference is intended. This monograph derives the product-limit estimator from conditional event probabilities and counting-process risk sets; distinguishes survival, cumulative hazard, cumulative incidence, and contractual cash-flow quantities; examines Greenwood uncertainty, transformed intervals, simultaneous bands, and logrank comparisons; and develops reproducible illustrative calculations. It identifies how delayed entry, recurrent claims, policy termination, calendar-time changes, competing risks, interval observation, informative censoring, and selection into claims systems change the estimand or undermine identification. Applications connect nonparametric event-time description to mortality experience, disability termination, claim settlement, retention, and limited-horizon benefit costs. The central analytical finding is that a survival curve becomes actuarially interpretable only through an explicit observation-to-contract map: the statistic must correspond to a well-defined covered population, exposure clock, endpoint, benefit eligibility rule, and observation process. Accordingly, the manuscript supplies an implementation and validation protocol, including risk-set reconciliation, sensitivity analyses, numerical examples, and boundaries on extrapolation. The contribution is an insurance-specific synthesis and worked methodology; it introduces no new statistical theory or empirical population estimate.
Keywords: Kaplan–Meier estimator; survival analysis; product-limit estimator; insurance; claim duration; life insurance; health insurance; disability insurance; right censoring; delayed entry; left truncation; risk sets; Greenwood variance; logrank test; competing risks; cumulative incidence; restricted mean survival time; actuarial model risk; claims data quality.
Introduction
A time-to-event observation has a start, an endpoint, and an observation horizon. These three elements carry special weight in insurance because a contract changes the identity of the event that matters. Death may be a mortality event, a benefit trigger, or an endpoint preceding a different benefit; disability recovery may terminate payments while a death or retirement provision also terminates the same claim. Claim reporting occurs after an underlying loss and is measured on a different clock from claim settlement. A policy that disappears from a company dataset might have lapsed, transferred, expired, or become administratively invisible. A survival estimator can process the recorded dates correctly while yielding an economically misleading interpretation if those distinctions are lost.
Kaplan and Meier's (1958) nonparametric product-limit paper derives estimation of survival from observations interrupted by losses to follow-up or administrative termination. The method is a foundation for event-history inference, including insured populations, but insurance applications require a further contractual specification. The analytical question is therefore twofold: what survival function do the data identify, and what insurance decision or description does that function actually support?
This research examines the standard nonparametric method and its boundaries. Its evidence base comprises primary statistical research, publisher-verified books and actuarial standards. Inference is developed from probability identities and explicit constructed examples rather than a proprietary insurer portfolio. All example observations and amounts below are hypothetical and are included for arithmetic verification; no numerical figure describes the experience of a real carrier or country. Regulatory discussion of U.S. actuarial standards identifies their professional scope and does not universalize them as international law.
Study design and limits of inference
The unit is specified as a policy, individual life, distinct claim, or episode, depending on the question. The event time is nonnegative; only the event definition, not the term “survival,” determines whether persistence is desirable. The statistical study estimates survival conditional on an admissible sampling and observation scheme. Insurance valuation may require additional ingredients: benefits, payment schedules, discount curves, expense assumptions, dependence, selection, and forward-looking adjustment. This manuscript explores those interfaces but confines detailed financial modeling and multistate valuation to their own disciplines.
Time-to-Event Insurance Outcomes
The event-time random variable
Let denote the elapsed time from a specified origin to a specified event. Define the survival function, distribution function, and, when a density exists, the instantaneous hazard by
For an absolutely continuous event time, the cumulative hazard is and . The hazard is an event rate conditional on having remained event-free; it is not itself the probability of an event during a year. A discrete-year actuarial death probability is . Its relationship to a continuously varying hazard is when the hazard is well defined. A policy-duration hazard and an age-specific mortality hazard refer to different time axes and conditioning populations.
A useful classification asks which exact transition generates the observed duration. For life insurance, duration from policy inception to death is distinct from attained age at death, and death after lapse is absent from an active-coverage death-benefit record unless independent follow-up exists. For disability, the claim duration until recovery answers a different question from the time until any benefit termination. In health coverage, time to hospital admission may restart after each discharge; treating every recurrent admission as if it were one first event changes the target. In property and casualty claims, time from occurrence to report measures reporting delay, while time from report to final closure measures handling duration. The latter is conditional on reporting and on a definition of permanent closure; reopened claims require a documented rule.
Insurance time origins and estimands
An estimand is the precisely defined population quantity that an analysis seeks to determine. Some common contrasts reveal why a single generic “survival rate” is insufficient:
| Insurance question | Time origin | Event | Event-free interpretation | Important selection condition |
|---|---|---|---|---|
| Life cover mortality | Policy issue or exposure origin | Death | Survival among the specified insured cohort | Active cover and subsequent lapse/death ascertainment |
| Disability benefit | Claim onset or eligibility start | Recovery or benefit termination | Continued disability or continued payments, respectively | Accepted claim and waiting-period qualification |
| Claims operations | Report receipt | Final settlement | Claim remains open | Reported and registered claims only |
| Retention | Policy commencement or renewal | Lapse or cancellation | Policy remains in force | Transfer, expiry and death definitions |
| First insured loss | Attachment of cover | First covered occurrence | No qualifying event yet | Exposure periods and eligibility exclusions |
Different event definitions can share dates but carry different financial consequences. A death that terminates an annuity may reduce future liability, whereas death that triggers term-life insurance may create an immediate claim. Recovery from disability can shorten payment duration. Cancellation often stops exposure and may itself be a cause-specific endpoint, rather than an innocuous observation loss. The survival function always carries the chosen interpretation within its conditioning population.
The 2015 Valuation Basic Report and Tables, issued by the Society of Actuaries, illustrate that institutional mortality applications classify experience by variables including sex, smoking status, and age conventions. A raw Kaplan–Meier curve is a statistic of its observed cohort; it becomes a pricing assumption only after an additional process evaluates representativeness, contract and coverage conditions, trend, selection, and the intended use.
Calendar time, duration, and attained age
Insurance event history commonly has three clocks: calendar date , elapsed contract duration , and attained age . For a policy issued at age on date , and only in idealized exact-time units. Administrative systems introduce day-count and age-nearest/age-last conventions, endorsements, interruptions, and renewals. Mortality may depend on attained age and calendar year, while lapse is strongly structured by policy duration. Pooling cohorts from different issue years on duration since issue can confound secular change with duration, and pooling lives by age can obscure duration-dependent underwriting selection.
For every curve, the manuscript therefore recommends recording the population inclusion date, time scale, entry rule, event definition, coverage status at origin, censoring rule, and end-of-observation date. These are part of the scientific identity of an estimate rather than optional metadata.
Survival probability and economic quantity
A survival probability has units of probability; expected event-free time has units of time; projected claim cost has units of currency. A nonparametric estimate of the first can support calculations of the latter only when the translation is explicit. For example, a hypothetical benefit of per unit time while a claimant remains eligible yields the undiscounted restricted expected benefit integral . Under deterministic discount factor and specified benefit conditions, the corresponding time-zero expected present value is . If multiple termination modes or partial benefits exist, a state occupancy model or mode-specific construction is necessary. The equations are accounting identities conditional on their assumptions, not estimates of real liabilities.
Risk Sets and Censoring
The observation scheme
Let be the right-censoring time and . Let indicate that the specified event was observed. The observed record says either that the event occurred at or that the subject remained event-free at least until the censoring time. The censoring indicator must be generated by event classification, rather than merely by whether a database row has a blank settlement date.
Administrative right censoring occurs when observation ends at a prespecified valuation or extraction date. Other examples include loss to follow-up, record transfer, and withdrawal. Kaplan–Meier inference requires a mechanism making observed future event experience representative of the people still under observation, given the variables or strata used for adjustment. The classical simple formulation uses independent event and censoring times. Kaplan and Meier's original independent-censoring discussion explicitly treats this assumption as something requiring scrutiny.
In an insurance dataset, censoring may be substantively informative. A claimant whose case is moved to litigation may disappear from one operational extract; a member who switches health plans may differ in morbidity; a disability file closed to reporting may remain clinically disabled. If transfer, withdrawal, or administrative loss predicts event times, an unadjusted Kaplan–Meier estimate need not identify the target marginal survivor function. Stratification can help when independent censoring holds conditional on measured variables, but marginal curves assembled from differently selected conditional groups still demand a defined standardization population.
Risk sets and ties
For an event time , let be the number observable, eligible, and event-free immediately before , and let be the number experiencing the event at . Under the common convention that events at a tied recorded time precede censorings at that time, individuals censored at are included in and removed after the event count. This convention should be specified and consistently implemented. A different ordering for day-aggregated records can change estimates.
The counting-process representation makes the eligibility rule exact. Define , and at-risk indicator for a simple cohort observed from zero. The pooled processes are and ; the change equals and . Aalen's (1978) counting-process formulation establishes the martingale and integrated-hazard framework that supports modern event-history inference. The definitions must be refined when observation begins after time zero.
Delayed entry and left truncation
Suppose an individual can enter the observable dataset at time , having survived the event-free state until entry. The observed interval is and the counting-process risk indicator is
Only subjects satisfying are in the event-time risk set. A subject who joins a portfolio transfer at year five cannot be counted as under observation during years one through four; including that subject retrospectively introduces guaranteed survival time. Conversely, discarding the first five years of all contracts changes the inferential population. The censoring and truncation treatment of Klein and Moeschberger (2003) develops the likelihood and risk-set changes for such samples.
Delayed-entry product limits can estimate conditional hazards over supported durations when the entry mechanism is independent of the event process in the relevant conditional sense. They cannot reconstruct survival before the earliest observed entry or unobserved early events solely from a late-entry sample. Transfer portfolios and legacy administrative migrations therefore need a documented history-reconstruction policy. Policyholders with unknown inception dates should not be placed at time zero by convenience.
Left and interval censoring
A disability event known to have occurred between two medical examinations is interval censored; an onset known only to predate first observation is left censored. Ordinary right-censored Kaplan–Meier software does not automatically produce a valid estimate for these schemes. Turnbull's (1976) self-consistency estimation for incomplete data generalizes nonparametric estimation to grouped, censored, and truncated observations. A conversion of unknown interval events into recorded visit dates creates measurement error and event-time artifacts; whether interval-censoring analysis is required follows from the actual observation mechanism.
Episodes, recurrence, and dependence
One person can contribute several policies, and one policy can contribute multiple claims. Repeated episodes are correlated, share underlying characteristics, and may have altered hazards after prior events. A Kaplan–Meier estimator of time to first event can be applied to an independent subject-level cohort, but treating all recurrent episodes as independent lives will typically overstate precision. Claim-level durations may legitimately be estimated descriptively, provided the estimand is an episode-weighted rather than person-weighted distribution and uncertainty accounts for clustering. Dependence also arises within employer groups, geographic catastrophe clusters, households, and repeated underwriting decisions. Standard Greenwood variance assumes a sampling structure that may not encompass these dependencies.
Identification and observation-support audit
An estimate is identified only within the range of times for which an adequate risk set is available under the assumed censoring and entry mechanisms. At long durations, a high-looking survival plateau may reflect administrative censoring of almost everyone. A median is unobserved if the estimated curve does not reach 0.5. It is invalid to report the final observed censor time as the time of the median event. A credible insurance analysis must publish event counts, censor counts, risk-set sizes, entry distributions, observation horizon, and sufficient stratification to reveal how much information remains at each duration.
Kaplan–Meier Product Limits
Derivation from conditional survival
Let the distinct observed event times be . Conditional on the population at risk just before , the fraction with no event at that time is estimated by . Survival to time factors into the probability of surviving each earlier event time. The empirical product-limit estimator is therefore
Between event times the estimate is constant and it drops at each observed event. Pure censoring times do not cause drops, though they reduce later risk sets. For a cohort with complete observations and no censoring, the product telescope reproduces the empirical survivor fraction. Under independent censoring and regular sampling assumptions, the estimator is the nonparametric maximum-likelihood solution for the right-censored failure distribution over its identifiable support, as developed by Kaplan and Meier (1958). The method's apparent simplicity reflects precise accounting of conditional event exposure.
There are two interpretations worth keeping separate. The finite-sample estimate is a product of event-free proportions among the observed risk sets. Its population interpretation relies on assumptions that allow those observed proportions to stand for conditional event survival in the target population. The same arithmetic can be computed under informative censoring, but the population claim then changes or becomes unsupported.
Likelihood interpretation
For an idealized independent right-censored sample from a continuous distribution, an event at contributes and a right censoring at contributes to the part of the likelihood involving the event-time distribution, conditional on the censoring mechanism. Allocating discrete mass at observed event times gives the event-time likelihood components proportional to , where is a conditional event probability at . Maximization gives , yielding the product limit. This intuitive factorization requires attention to ties and to the observation model; it is not an argument that an arbitrarily selected administrative extract is random.
Cumulative hazards
The Nelson–Aalen cumulative-hazard estimator is
Nelson's (1972) hazard plotting treatment and Aalen's counting-process work connect this estimator to event processes. The exponential transform is close to, but is generally not exactly equal to, the finite-sample Kaplan–Meier curve. At one event time, the former multiplies by , while the latter multiplies by . The approximation becomes closer for small conditional event fractions. Choosing one of these quantities is a methodological decision; their estimated targets coincide under the usual continuous-time large-sample setting.
Worked synthetic portfolio: claim closure
Consider twelve hypothetical claims registered at the same logical origin. The event is first final closure; a claim observed without final closure by its administrative follow-up date is right censored. The following risk sets describe a deliberately constructed dataset in months. Where censoring occurs between event times, it reduces the next risk set. None of the observations represents a real insurer.
| Month | Risk set just before event | Closures | Censorings after event/before next event | Kaplan–Meier survival after event |
|---|---|---|---|---|
| 2 | 12 | 2 | 1 at month 3 | 0.833333 |
| 4 | 9 | 1 | 2 at month 5 | 0.740741 |
| 6 | 6 | 2 | 1 at month 7 | 0.493827 |
| 8 | 3 | 1 | 2 at month 9 | 0.329218 |
For example, ; ; and . The censorings at months three, five, and seven do not directly change the curve. Nine observations have ended by closure or censoring before month eight, leaving three people at risk at the last listed event time. The final two claims are censored at month nine, so the curve does not fall to zero by that date. Counting these censored claims as settlements would be a data-classification error.
The median first-final-closure time in this synthetic example is six months, the first time . The statement is confined to the estimated survivor function for the hypothetical cohort under the analysis assumptions. An insurer wanting a forecast for a new portfolio would need evidence on report mix, calendar effects, litigation, case reopening, and the adequacy of censoring assumptions.
The restricted mean and its actuarial reading
For a finite horizon , define restricted mean survival time by
The natural plug-in estimate is . The synthetic claim portfolio gives, for months,
The integral treats all claims as active from their common time origin until first final closure or the horizon, not until every subsequent payment ceases. Were the hypothetical per-claim servicing cost a constant 100 monetary units per month while open, the truncated expected servicing cost under this simplified rule would be about 613.58 units. If administrative expenses respond to claim complexity or litigation state, the product of a common cost rate and survival no longer captures the expense mechanism. Uno and colleagues (2014) discuss restricted-mean comparisons as a useful time-scale-based alternative to an exclusive focus on hazard ratios.
What nonparametric estimation leaves unknown
The estimator gives a step-function description on the observed support; it neither estimates a smooth instantaneous hazard without additional methods nor extrapolates a tail that has not been observed. A lifetime annuity or long-duration disability reserve typically needs a credible tail beyond the available support. That step requires a separately documented model or an external mortality or termination assumption. A Kaplan–Meier curve should not be extended horizontally forever as though the absence of observed late events established zero future hazard. Where the last observation is a censoring, the likelihood carries no point identification of the ultimate event-time tail without added assumptions.
Reproducible calculation specification
An implementation should first sort event and censoring records by elapsed event-time scale; group truly tied event times; compute the eligible risk set immediately before each event time; remove events and same-time censorings in the prescribed order; and calculate the cumulative product in stable floating-point precision. The software output should reproduce a hand-checked small example and reconcile counts against the transaction-level data. For left-truncated data, risk sets are reconstructed at every event time using the entry and exit intervals, rather than simply by subtracting prior departures from a fixed initial sample.
Confidence Bands and Group Comparisons
Sampling uncertainty and Greenwood's estimator
For distinct event times with , the commonly used Greenwood large-sample variance estimate is
The summands represent estimated uncertainty from sequential conditional-event fractions. Small terminal risk sets make the variance sensitive, even when the step curve itself looks smooth. In the synthetic portfolio, the sum at month eight is , yielding a standard error of about at . These figures describe model-based sampling uncertainty under the simple independent-cohort design; they do not include uncertain data linkage, unmeasured informative censoring, or future claim mix.
A plain pointwise Wald interval can leave the allowable probability range near zero or one. Probability-preserving transformations are often preferable. With , the delta-method standard error is approximately , where . Transforming the interval on the -scale back to keeps endpoints between zero and one, though endpoints at or require special handling. Different interval constructions and bootstrap procedures can differ materially in short or highly censored samples.
Pointwise intervals and simultaneous bands
A 95% pointwise interval is designed to cover the survival probability at one prespecified time under repeated sampling. A simultaneous 95% band targets coverage across a specified time interval. Showing pointwise intervals at every month and describing the entire display as a 95% simultaneous band misstates inference. Hall and Wellner's (1980) simultaneous confidence-band construction develops asymptotic bands for right-censored survival functions. This matters in claims operations because a duration curve is regularly used to support conclusions over a range of settlement ages rather than one selected month.
Before presenting intervals, an analyst should disclose the confidence method, sampling unit, treatment of clustering, censoring assumptions, and time range. The survival-plot guidance of Pocock, Clayton, and Altman (2002) explains how truncation of plots and missing information on risk sets can distort interpretation. In insurance documents, display both event counts and numbers at risk under key durations, especially after the tail begins to thin.
Comparing insurance cohorts
Suppose groups A and B represent policy issue vintages, underwriting classes, product versions, or operational claims teams. Their Kaplan–Meier curves estimate separate event-free distributions conditional on group selection and observation rules. Group differences are descriptive unless the design warrants a causal interpretation. A product-vintage comparison can mix duration effects with secular changes in inflation, settlement technology, policy wording, and claimant characteristics.
The logrank test compares observed events to expected events over pooled risk sets. At ordered event times , write for group A's at-risk count, for the total event count, and for A's events. Under a common event hazard, the expected A count is . A two-group statistic can be written , where for
A large-sample standardized statistic is assessed against a standard normal reference, or its square against a one-degree-of-freedom chi-square reference, subject to design and censoring assumptions. Mantel's (1966) survival-data rank comparison and Peto and Peto's (1972) logrank derivation provide foundational treatments. Weighted versions may be chosen for scientifically justified alternatives, but choosing weights after inspecting the curve increases selective-reporting risk.
When comparisons change over time
If curves cross, a single logrank result may obscure opposing early and late differences. A comparison of restricted means to a fixed common horizon describes the average event-free-time difference, which can be more actionable for claim handling capacity than an average hazard ratio. Where covariate adjustment is required, Cox's (1972) proportional-hazards regression is a separate semiparametric tool, and proportionality requires investigation. Grambsch and Therneau's (1994) weighted-residual diagnostics provide formal methods for assessing time-varying effects in the Cox setting. None of these methods alone repairs a misdefined population or an unobserved event.
Multiple groups and development samples
Comparing dozens of underwriting segments or claim subtypes creates multiple-testing and sparse-event concerns. A visible ranking of group curves after selecting the most favorable inspection horizon is an exploratory finding, not a prospectively validated difference. Independent validation by issue vintage or calendar period, uncertainty intervals on meaningful effect sizes, and documentation of selection criteria support better decisions. For clusters such as several claims per person, either estimate a person-level first-event outcome or use inference that respects group dependence; naively treating every claim as an independent life exaggerates the amount of information.
Competing Observation Mechanisms
Censoring, policy termination, and competing events
Censoring refers to incomplete observation of an event time under a specific observation process. A competing event changes the possible event history: after it occurs, the event of interest is either impossible or has a different meaning. These two conditions should be distinguished before any software is selected. Death competes with recovery as a first termination event for some disability benefit definitions; cancellation competes with death as an active-policy termination; and withdrawal, litigation, and administrative transfer may be censoring mechanisms or separate operational events depending on the analytic question.
For mutually exclusive first-event causes , define the cause-specific hazard , all-cause survival , and cause-specific cumulative incidence
Under a continuous-time model, the total cause-specific hazard is , while a single cause's cumulative incidence depends on all hazards through the all-cause survival. The complement of a Kaplan–Meier curve obtained by incorrectly treating other observed causes as independent censoring estimates a hypothetical net risk under a cause-specific-hazard construction, not the actual cumulative probability of the cause in the presence of competitors. Andersen and colleagues (2012) explain the rate–risk distinction under competing risks, and Gooley and colleagues (1999) establish the cumulative-incidence interpretation.
Worked synthetic competing-risk contrast
Take ten independent hypothetical covered lives under complete observation to a common horizon. At month one, two die while covered. At month two, three living policies lapse permanently. At month three, one of the five remaining active lives dies while covered. For the first active-coverage termination event, the all-cause survivor proportions after months one, two, and three are , , and . The empirical cumulative incidence of covered death by month three is the two deaths among ten at month one plus at month three, or . The incidence of lapse is , and the remaining active fraction is , so the components sum to one.
If the three lapses at month two were instead recorded as ordinary right censorings for the covered-death endpoint, an otherwise naive cause-specific Kaplan–Meier calculation would give survival , and a complement . The difference between 0.36 and 0.30 illustrates a change of estimand, not a software defect. Whether post-lapse death remains an analyzable event also depends on the desired target: mortality in all original policyholders requires continued ascertainment after lapse, whereas covered death under the original contract is a first-event system with lapse as a competing mode. Gray's (1988) cumulative-incidence group tests support comparisons where a particular competing-event probability is the target.
Multistate relationships
Some insurance histories contain more than one episode or pathway: healthy to disabled, disabled to recovered, recovered to disabled again, or a claim reported, under investigation, settled, reopened, and settled finally. A first-event Kaplan–Meier curve describes only one chosen transition out of a state. When the question concerns the probability of occupying a state at time , transition-probability methods are required. Aalen and Johansen's (1978) empirical transition-matrix estimator extends product-limit estimation through a matrix product integral for censored multistate histories. The assumptions and interpretation of state occupancy depend on the modeled transition process and observation scheme. This monograph uses that literature to establish the boundary between a single-event duration curve and multistate insurance cash-flow valuation; it does not replace a detailed multistate actuarial model.
Informative censoring and sensitivity
The independent-censoring assumption may be evaluated empirically through available proxies but cannot be established solely by finding balanced observed variables. An extract should record reason codes for censoring and summarize the event history preceding record loss, stratified by calendar period, claim severity, product, intermediary, and known state. Censoring that increases abruptly at a system migration date is a concrete warning that technical data processes might drive the apparent duration curve. In such cases, a re-linkage to prior and successor systems is often more valuable than another estimator.
A basic sensitivity exercise compares (a) administrative censoring only, (b) additional censoring of transfers and withdrawals, and (c) plausible alternative classification of uncertain closures. These are distinctly labeled estimands and scenarios; the analysis should not call the most favorable scenario an identified population truth. Conditional inverse-probability weighting may address censoring dependent on recorded history if the censoring model is adequate and probabilities of remaining observed are positive. Satten and Datta (2001) show how the Kaplan–Meier estimator admits a weighting representation. In a richer insurance analysis, time-varying weights would require reliable observation-history data, diagnostics for extreme weights, and uncertainty reflecting estimation of the weights. No such model is estimated here.
Periodic observation and interval uncertainty
Claims eligibility and recovery may be assessed only at scheduled examinations. If a disability ceases between two assessments, the actual event lies within an interval; taking the later visit as exact recovery introduces a systematic delay tied to examination schedules. Turnbull's incomplete-data estimator is a defensible starting point where event-time intervals are available. If the target is instead the administrative payment stop date, that later date can be an exact operational event. These are distinct estimands, so a precise choice of endpoint resolves many apparent disagreements between claims and clinical systems.
Selection into the observed claims universe
Observing only reported, accepted, or paid claims conditions on processes preceding the measured settlement time. An estimate of time from report to final closure for reported claims cannot, without additional information, establish time from loss occurrence to settlement for all losses. Likewise, a duration curve among approved disability claims excludes early rejections and claimants who fail a waiting period. Selection can alter the level and shape of a curve even if subsequent censoring is independent. In portfolio mergers, only contracts surviving until system conversion may be represented; delayed-entry methods recover risk-set consistency within identifiable support but do not supply missing event histories for contracts that disappeared earlier.
Model risk and actuarial evidence
The Actuarial Standards Board's U.S. ASOP No. 23 on Data Quality addresses selecting, reviewing, using, and relying on data in actuarial services. Its ASOP No. 56 on Modeling addresses model design, validation, use, and limitations. These standards apply within their stated U.S. professional scope and are cited here as examples of disciplined actuarial practice, not as worldwide regulatory obligations. For an event-history estimator, an appropriate review considers event eligibility, completeness of reporting and exit dates, consistency of source systems, model assumptions, reproducibility, and the intended use of estimated duration curves. The strongest calculation cannot compensate for an unreconciled time origin or unreliable administrative event codes.
Insurance Duration Applications
Mortality under insurance selection
For a life portfolio, an issue-cohort survival curve has a precise target only after defining whether follow-up continues after surrender, whether death data are linkable to external registries, and whether entrants are all insured lives or a selected exposure subset. An analysis of death during an active term contract is a different event-history system from population survival after the contract ends. Underwriting selection, age and duration effects, exposure mix, country-specific death registration, and policyholder attrition may all matter. Kaplan–Meier estimates can describe an insured sample over observed durations; durable mortality bases for future valuation require age and calendar-time structuring, selection adjustment, mortality improvement judgment, and often considerably longer histories.
When following an issue cohort without loss to follow-up, the probability of survival from age to is , and a one-year conditional death probability is when . Applying this ratio to an insurance portfolio whose composition changes through lapse and underwriting class migration presupposes that the survival curve refers to the intended mortality population. The Society of Actuaries valuation tables are a documented example of mortality experience structured by conventional underwriting characteristics; they should not be equated with the uncensored empirical survival of a single arbitrary portfolio.
Disability termination and benefit continuation
Disability analysis may define as time from accepted disability onset to the earliest termination of the payment condition. If benefit termination can occur by recovery, death, retirement age, expiration of policy benefits, and adjudication, the associated event-state architecture needs to reflect the purpose. For “any benefit termination,” all those modes may form a composite endpoint. For “recovery while disabled,” death and other termination modes compete. Eligibility may begin only after a waiting period, creating selection and possibly delayed entry. A simple KM estimate of any termination may correctly answer the persistence-of-payment-state question while giving a misleading picture of biological recovery if causes are mixed together.
A bounded expected benefit-duration measure is , where the survival state represents continued contractual payment eligibility. Payment frequency, amounts, offsets, wage inflation, and partial return-to-work can make expected dollars differ from the product of expected duration and a flat monthly amount. A decision-grade valuation should use a clearly defined benefit state and incorporate the necessary economic assumptions rather than multiply a curve by an unexplained average claim cost.
Claim-reporting and settlement processes
For property or casualty claims, reporting lag and settlement lag use loss-occurrence date , report date , and final closure date . The identity holds for complete records using consistent date definitions, but the distribution of the total delay cannot generally be recovered by adding separately estimated mean survival curves. Dependence between reporting and settlement lags matters, especially for large claims and litigation. A closed-claims-only sample shortens observed settlement durations because long-lived open claims are excluded; right-censored open claims must enter the risk sets if the question is the distribution of first final closure among reported claims.
Reopened claims present a particular identification choice. If the target is first administrative closure, a reopened claim has had its first event already. If the target is durable final settlement, temporary closures should not be treated as event times; a follow-up horizon or a multistate description is needed. Operational status definitions should be stable across technology releases and claims teams. A mechanical increase in “closure rate” after a code change may reflect recoding rather than improved handling.
A practical claims display provides event-free/open proportions at reporting ages of 30, 90, 180, and 365 days, along with claim counts at risk and a statement of the extraction date. Because the last months may contain few claims and many censorings, it also reports a common restricted horizon. Comparing closed-only average settlement days with the restricted mean among all reported claims mixes two different denominators and should be avoided.
Retention, surrender, and coverage exposure
For retention analysis, lapse may be the event and policies retained at a fixed horizon are treated as still in force. Death, contractual maturity, transfer, and customer-initiated cancellation require their own classifications. A curve of time to any exit addresses persistence of the contract, while a cause-specific lapse probability addresses one exit reason. Product design, distribution channel, renewal features, and sales vintage may alter the trajectory. A lapse curve without allowance for scheduled policy maturities can make a product with shorter contractual terms look artificially less persistent.
An actuarial exposure denominator is not automatically the raw number of policies. Multiple lives per policy, endorsements, seasonal cover, suspension periods, and geographical location of insured risks may change exposure. A person-day or policy-day ledger reconciled to the event-history start and end rules offers a transparent basis for comparison. Where entry occurs through staggered issue dates and a common end-of-study date, duration analysis must explicitly align each subject to policy age rather than to calendar extraction age.
Pricing, reserving, and operations: three different uses
An observed settlement curve can support a description of how long current claims remain open; staffing plans can use restricted expected open-claim time when the future mix is comparable. Claim reserve estimation requires amounts, development patterns, recoveries, discounting where applicable, dependencies, and often a valuation-specific conditional distribution. Life insurance pricing likewise requires future mortality and financial terms; a historical Kaplan–Meier estimate alone is insufficient to set premiums. Scientific integrity therefore requires labeling a curve as one of three things: a descriptive empirical distribution, an input to a predictive procedure, or a direct component of a defined valuation equation. These labels correspond to increasing evidentiary requirements.
Practical reproducibility protocol
A reusable event-history analysis should preserve the following items together with the results:
- The declared study population and claim/policy/person-level sampling unit; start and stop rules; handling of multiple records and reinstatements.
- The endpoint definition, time scale, calendar cut, exact event status mapping, reason codes for exits, and treatment of ties.
- Source-system provenance, deduplication keys, linkage quality, missing dates, logical date-order tests, and the population excluded by missingness.
- Administrative censoring versus withdrawals, transfers, competing events, and interval observation; the assumed independence or conditional independence grounds.
- Delayed-entry intervals, minimum and maximum supported durations, observed event and censor counts, numbers at risk at prespecified horizons, and the terminal risk set.
- Pointwise intervals and/or simultaneous bands with their actual constructions; whether dependence or clustering was taken into account.
- Predeclared group comparisons, multiplicity decisions, validation period, and sensitivity scenarios for uncertain event/censor classifications.
- Restricted-horizon metrics, any benefit/cost translation, model limitations, and a separate authorization trail for pricing or reserving use.
The protocol operationalizes the distinction between a statistically computed curve and an insurance inference with evidentiary support. Reproducibility means that the same declared event logic, population extract, tie convention, and software produce the same risk sets and estimates; it does not mean that a different portfolio or a later calendar period must exhibit the same survival function.
Independent analytic synthesis
The conceptual importance of Kaplan–Meier estimation in insurance follows from the conjunction of three facts. First, contracts and administrative data systems produce staggered observation and right-censored durations, precisely the setting in which an empirical event fraction can fail. Second, contractual states determine whether the event is a cost trigger, a cessation of exposure, a claim-handling milestone, or a benefit-termination mode. Third, inference from that state is constrained by who became observable and why observation stopped. These factors form a sequence of decisions: define the contractual event; identify the population and observation clock; establish the sampling and censoring mechanism; estimate only on the supported domain; communicate uncertainty; then translate the result into a specific actuarial or operational quantity.
The sequence yields several practical deductions. Risk-set reconstruction should precede any statistical model comparison because an incorrect denominator changes all subsequent estimators. A cause-specific business question should precede the decision to treat an exit as censoring because that decision defines which risk is actually reported. Follow-up quality may be more valuable than numerical sophistication when censoring and data migrations dominate the tail. And restricted-horizon quantities often provide a more defensible operational basis than extrapolated medians or indefinite tail assumptions. These are interpretations and methodological recommendations grounded in the cited literature and the worked examples, not newly named principles or a proposed statistical theory.
Research needs and feasible extensions
Several inquiries are especially useful for future independent empirical studies: how administrative migrations affect censoring by claim status; how strongly settlement times depend on claim severity after reporting; whether insured-mortality follow-up after surrender is systematically missing; how benefit waiting periods alter disability termination estimates; and whether calendar-vintage stratification changes observed lapse curves. Each inquiry requires a defined dataset, record-linkage provenance, and a preregistered estimand. Further modeling may employ competing-risk methods, multistate occupancy, inverse-probability weighting, hierarchical clustering, or externally validated tail extrapolation. Selection among these methods follows the evidence-generating process and actuarial decision, rather than an a priori preference for complexity.
Conclusion
Kaplan–Meier estimation provides a coherent nonparametric representation of event-free survival under right-censored observation. Its product-limit form is derived from conditional event survival in correctly assembled risk sets; Greenwood uncertainty and formal group comparisons quantify aspects of sampling variation; left truncation and interval observation require design-specific treatment; and competing risks require cumulative incidence when the question concerns an actual cause-specific event probability. These conclusions follow directly from the established methodological literature and the fully specified worked examples.
The actuarial meaning of the curve arises from the relation between statistical event, insurance contract, exposure population, observation process, and economic target. A curve for claim closure cannot be treated as a mortality curve, and a death probability among active policyholders cannot be read as cohort mortality after surrender without additional follow-up. The distinction between observed data and a forward-looking pricing or reserving input is similarly essential. Reporting a clear time origin, event classification, risk sets, censor reasons, supported horizon, uncertainty, and selection mechanism establishes an auditable basis for interpretation.
Two calculations anchor the conclusions. The twelve-claim illustration gives an estimated open-claim fraction of at eight months and a restricted mean of about 6.135802 open months; the competing-risk illustration shows a covered-death cumulative incidence of 0.30 rather than the 0.36 produced by a naive cause-specific Kaplan–Meier complement. The first is an estimation exercise with right censoring, and the second is an estimand demonstration. Both are synthetic and have no inferential claim about market experience.
The research contribution is the systematic alignment of standard survival inference with the heterogeneous clocks, termination rules, and observation constraints of insurance records. The result supports reproducible descriptive research and disciplined use of time-to-event evidence in international insurance and risk science. The manuscript makes no claim of discovering a new estimator, constructing a new doctrine, independently verifying proprietary experience, or establishing a novel general theory.
References
Aalen, O. O. (1978). Nonparametric inference for a family of counting processes. The Annals of Statistics, 6(4), 701–726. DOI: 10.1214/aos/1176344247
Aalen, O. O., & Johansen, S. (1978). An empirical transition matrix for non-homogeneous Markov chains based on censored observations. Scandinavian Journal of Statistics, 5(3), 141–150. JSTOR stable identifier: 4615704
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality. Actuarial Standards Board.
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Actuarial Standards Board.
Andersen, P. K., Geskus, R. B., de Witte, T., & Putter, H. (2012). Competing risks in epidemiology: Possibilities and pitfalls. International Journal of Epidemiology, 41(3), 861–870. DOI: 10.1093/ije/dyr213
Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202 (article text; accompanying discussion in original issue extends to p. 220). DOI: 10.1111/j.2517-6161.1972.tb00899.x
Gooley, T. A., Leisenring, W., Crowley, J., & Storer, B. E. (1999). Estimation of failure probabilities in the presence of competing risks: New representations of old estimators. Statistics in Medicine, 18(6), 695–706. DOI: 10.1002/(SICI)1097-0258(19990330)18:6%3C695::AID-SIM60%3E3.0.CO;2-O
Grambsch, P. M., & Therneau, T. M. (1994). Proportional hazards tests and diagnostics based on weighted residuals. Biometrika, 81(3), 515–526. DOI: 10.1093/biomet/81.3.515
Gray, R. J. (1988). A class of K-sample tests for comparing the cumulative incidence of a competing risk. The Annals of Statistics, 16(3), 1141–1154. DOI: 10.1214/aos/1176350951
Hall, W. J., & Wellner, J. A. (1980). Confidence bands for a survival curve from censored data. Biometrika, 67(1), 133–143. DOI: 10.1093/biomet/67.1.133
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
Klein, J. P., & Moeschberger, M. L. (2003). Survival Analysis: Techniques for Censored and Truncated Data (2nd ed.). Springer. DOI: 10.1007/b97377
Mantel, N. (1966). Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemotherapy Reports, 50(3), 163–170. PMID:5910392
Nelson, W. (1972). Theory and applications of hazard plotting for censored failure data. Technometrics, 14(4), 945–966. DOI: 10.1080/00401706.1972.10488991
Peto, R., & Peto, J. (1972). Asymptotically efficient rank invariant test procedures. Journal of the Royal Statistical Society: Series A (General), 135(2), 185–198. DOI: 10.2307/2344317
Pocock, S. J., Clayton, T. C., & Altman, D. G. (2002). Survival plots of time-to-event outcomes in clinical trials: Good practice and pitfalls. The Lancet, 359(9318), 1686–1689. DOI: 10.1016/S0140-6736(02)08594-X
Satten, G. A., & Datta, S. (2001). The Kaplan–Meier estimator as an inverse-probability-of-censoring weighted average. The American Statistician, 55(3), 207–210. DOI: 10.1198/000313001317098185
Society of Actuaries. (2015; report updated 2018). 2015 Valuation Basic Report and Tables. Society of Actuaries. Publisher URL in linked title
Turnbull, B. W. (1976). The empirical distribution function with arbitrarily grouped, censored and truncated data. Journal of the Royal Statistical Society: Series B (Methodological), 38(3), 290–295. DOI: 10.1111/j.2517-6161.1976.tb01597.x
Uno, H., Claggett, B., Tian, L., Inoue, E., Gallo, P., Miyata, T., Schrag, D., Takeuchi, M., Uyama, Y., Zhao, L., Skali, H., Solomon, S., Jacobus, S., Hughes, M., Packer, M., & Wei, L.-J. (2014). Moving beyond the hazard ratio in quantifying the between-group difference in survival analysis. Journal of Clinical Oncology, 32(22), 2380–2385. DOI: 10.1200/JCO.2014.55.2208
Editorial provenance: original analytic synthesis by Angela Bogdanova (Artificial Sapiens); research manuscript prepared for Insurance by Bogdanova. Published by Aisentica Development. Public release status: NOT RELEASED. First-publication date: unassigned.