RESEARCH DOSSIER
Long-Term Care Episode Duration and Competing Exit Risks
Abstract. The duration of long-term care (LTC) need, the duration of an insured claim, and the duration of benefits paid by a contract are three distinct observations of a complex process. Recovery and death are competing exits from an active-claim state, but an administrative record may also end through benefit exhaustion, loss of eligibility, change of care setting, discontinuation of services, transfer of administration, or data closure. Each definition changes the population at risk, the denominator of a termination rate, the interpretation of censoring, and the cash flows derived from an episode model. This monograph examines the empirical and mathematical foundations of episode duration, develops a transparent application of established competing-risks and multi-state methods, and integrates actuarial experience studies, functional classification, contract rules, and cross-jurisdiction evidence. The principal analytical finding is that cause-specific transitions and insured payment states must be reconciled at the level of their definitions before an aggregate duration estimate can inform pricing or reserves. Cumulative incidence quantifies the probability of a particular exit in the presence of other exits; event-free survival and expected time in eligible payment states are different actuarial quantities. Elimination periods, benefit caps, service utilization, recurrence, and the distinction between medical recovery and administrative closure transform the relationship between care need and insurer liability. The research uses source-verified institutional evidence and explicitly hypothetical, reproducible calculations. It does not estimate new population parameters, claim access to proprietary microdata, or introduce a new actuarial theory.
Keywords: Long-term care insurance; episode duration; competing risks; recovery; mortality; claim termination; functional dependency; cumulative incidence; cause-specific hazard; multi-state models; censoring; elimination period; benefit exhaustion; payment duration; experience studies; actuarial reserving.
Introduction
Insurance science requires a precisely identified event before the duration of that event can be measured. A person may require assistance at home for several months, enter an institutional setting, regain one activity of daily living, remain eligible for a public care assessment, submit a private claim, and receive a cash benefit at a later date. None of those dates is automatically the beginning or end of the same statistical episode. A claim system can register a closure while a person's underlying dependency persists; a care episode can conclude without terminating the insured relationship; and death can end an episode while creating a separate obligation under a different component of a combined insurance contract.
The World Health Organization's International Classification of Functioning (WHO, 2001) situates functioning and disability within activity, participation, health, and environmental context. It supplies a classification framework, whereas insurance eligibility is determined by legal and contractual triggers. The difference is consequential: care need in a population survey is wider than some policy definitions, while paid LTC benefits may be narrower than clinical care utilization. The U.S. statutory qualification rule (United States, 2026) includes functional and cognitive criteria for tax-qualified insurance, and the Singapore CareShield Life conditions (Central Provident Fund Board, 2026) use a separate threshold of severe disability. The same individual trajectory can therefore generate distinct entry and exit dates across schemes.
The available actuarial record is unusually informative about the importance of the distinction. The Society of Actuaries' 2000–2016 LTC aggregated databases (SOA, 2020) document claim incidence and claim terminations from participating carriers. The associated technical database report describes how multiple claims separated by six months or less were combined and how claim start was assigned from service and elimination-period information. Those are research design decisions that determine the distribution of observed durations. An otherwise exact survival estimator cannot repair an inconsistent episode definition.
This dossier examines six connected questions. What counts as an episode? Which persons and intervals belong to the exposure denominator? How should recovery and death be estimated as competing exits? How do contractual rules convert duration into money? What evidence can be transported across jurisdictions or products? Which diagnostic results would justify reliance on a model? Throughout, LTC is the principal insurance subject; long-term disability (LTD) studies are used selectively to illuminate methodology, with the distinction between inability to perform work and inability to undertake activities of daily living retained.
The methods combine source criticism, analysis of actuarial experience documentation, established survival-analysis mathematics, comparisons of official public benefit conditions, and numerical scenarios whose parameters are declared hypothetical. For statistical estimands, duration begins at a specified entry time and continues until a defined first exit or censoring. For financial valuation, entitlement and payments are represented separately from underlying need. The resulting contribution is an analytical synthesis and a reproducible specification for source-to-estimand coherence, rather than a new proprietary statistical model.
Conceptual Foundations and Historical Development
Care need, eligibility, claim, and payment
Long-term care denotes assistance associated with sustained functional or cognitive limitations, delivered through formal or informal arrangements in households, communities, or institutions. At minimum, four events must be distinguished. A need onset is the first time an individual satisfies a research definition of dependency. An eligibility onset is the first time that a public program or insurance contract recognizes the relevant qualifying condition. A claim onset is the event date assigned within the insurer's reporting rules. A payment onset is the date or service period from which a covered benefit is payable, often following elimination or certification requirements.
The WHO functioning framework measures health-related functioning and environmental factors. By comparison, section 7702B of the Internal Revenue Code describes certification of inability to perform at least two activities of daily living for a period expected to last at least 90 days, or severe cognitive impairment requiring substantial supervision, among the conditions for tax-qualified treatment. The actual legal text matters: the 90-day functional-certification criterion must not be substituted for a policy's elimination-period length. A care episode can have a qualifying assessment without an immediate payment, and a contractual waiting period can operate even when the statutory criterion has been met.
The NAIC shopper's guide (NAIC, 2022 edition) distinguishes calendar-day and service-day elimination periods and explains that some contracts apply an elimination period again after a later episode. A service-day calculation depends on utilization frequency, making the benefit-onset date partly endogenous to care intensity. Consequently, the financial duration of an episode depends on at least three clocks: biological or functional time, assessment and claim-administration time, and the contractual benefit clock.
Episodes, spells, and recurrence
An episode may mean one uninterrupted clinical state, one connected set of formal care services, one filed claim, or one benefit period under a policy. An insurer may merge separated claims into a single episode for experience reporting, whereas a health system can count a new admission each time a person changes facility. These conventions determine the mass of short spells and the observed recurrence rate. When recovery is followed by deterioration, a first-event competing-risks model records the first recovery and discards the later return unless an additional state is introduced.
The SOA's 2020 aggregate methodology provides an especially relevant operational example: claims separated by no more than six months were combined for study purposes. This choice can transform two short clinical episodes into one longer insured episode. It may be appropriate for comparability with a particular incidence study, but it does not establish a universal clinical definition of recovery. An investigator should publish the gap rule and conduct a sensitivity analysis with alternative recurrence windows.
A useful record distinguishes a person identifier, policy identifier, claim identifier, episode identifier, care-setting spells, assessment history, and payment transactions. Requiring these entities to be identical creates false certainty. The data structure should permit a claimant to leave and return to paid care under a single benefit pool, as well as to receive multiple claims under separate policies. The episode boundary is therefore part of the estimand, not a clerical afterthought.
Exit taxonomy
Recovery, mortality, and other exits should be defined against an explicit initial state. Recovery means the person no longer satisfies the specified dependency or claim eligibility criterion, according to the observed assessment convention. Death is an absorbing biological event. Benefit exhaustion ends payment under a finite benefit pool without necessarily ending care need. A move between facilities is an internal transition unless the policy or analysis explicitly treats the move as terminal. Policy termination and cessation of premiums are materially different from a claimant's recovery.
The SOA's caveats on LTC experience tables (SOA/Towers Watson, 2017) explain that death and recovery are the terminations emphasized in the aggregate experience tables; exhaustion was not counted as termination, while exposure continued until exhaustion. The memo also warns that the claim termination table is based on the original care setting, rather than being designed to estimate transitions among care settings. These details prohibit an unqualified interpretation of published total terminations as every reason an insurer stopped paying.
Development of the actuarial and statistical literature
The emergence of extensive long-term care claims data, particularly after Germany introduced compulsory social LTC coverage in 1995, made duration-dependent transition estimation practicable. Czado and Rudolph (2002), in their survival-analysis application to LTC, demonstrated the use of Cox transition models with censored observations and time-dependent covariates, and linked estimated intensities to multi-state premium calculations. Their paper supplies a specifically actuarial basis for representing care history as transitions rather than as a single average length of stay.
The broader statistical literature supplies complementary estimands. Fine and Gray (1999), in the subdistribution hazards paper, developed regression for the cumulative incidence of a cause in the presence of competing events. Andersen, Geskus, de Witte, and Putter (2012), in competing risks in epidemiology, explain the distinction between cause-specific rates and cause-specific event probabilities. These methods are established statistical foundations. Their application to LTC requires a separate mapping from clinical and administrative transitions to insured contracts.
The ASOP 18 Long-Term Care standard (Actuarial Standards Board, 2022) requires practitioners within its scope to consider pertinent eligibility rules, covered services, benefit amounts, duration limits, and materially relevant assumptions. It also calls for consideration of evolving experience and sensitivity. The present research applies these principles to the measurement of episode duration while retaining the distinction between U.S. professional standards and rules governing other legal systems.
Episode Duration Estimates
Population, time origin, and estimand
For an inception cohort of insured claims, define time zero as the specified first qualifying event. The event-free time T is the duration until the first event among explicitly designated exits. With recovery R and death D as the two biological or functional terminal events, T is the minimum of their latent event times when such latent times are conceptually meaningful. The observed record is usually (Y, delta, J), where Y is the minimum of event time and administrative censoring, delta indicates observed termination, and J identifies the recorded cause. A claim with no termination before study end contributes partial exposure rather than a zero-duration claim.
A cohort of prevalent open claims at a census date answers another question. It overrepresents long survivors and can overstate the typical inception-cohort duration. Incident and prevalent cohorts require different entry mechanisms; left truncation is needed when claims enter observation after onset and selection into observation is conditioned on survival to that entry time. Similarly, starting the clock only after an elimination period selects claims that survive and satisfy administrative criteria. An observed post-elimination survival distribution cannot directly estimate the distribution from initial need onset.
Insured cohorts are also selected by underwriting, plan design, lapse behavior, and claim filing. The evidence from Friedberg, Hou, Sun, and Webb (2023) on lapses in LTC insurance indicates that cognitive impairment and lapse behavior can interact in ways relevant to the composition of policies that remain in force. This finding concerns policy persistency rather than termination of an already active care episode. Its appropriate use here is to identify a pathway of selection before a claim enters the duration analysis.
Measurement of exposure
Continuous-time hazard estimation requires the amount of observed time in the state from which the exit occurs. For cause k, the crude rate over an interval is the number of observed k-events divided by accumulated time at risk, subject to consistent person-time accounting and delayed entry. In a discrete monthly model, event counts must be aligned with monthly risk sets and treatment of entrants, exits, and partial months. Death counts divided by the original cohort size is a cumulative proportion; deaths divided by claimant-months is a rate. They are not interchangeable.
The SOA's LTC aggregate database report (SOA, 2020) describes total reported claim-termination exposure of 10,948,710 in the study's exposure units, 349,110 terminations, and 233,366 deaths, with reported aggregate termination and mortality rates of 3.189% and 2.131% respectively. The study derived a consistent claim-incurred date where source carrier definitions varied, and measured claim exposure from minimum to maximum service date rather than during elimination periods. Those aggregate figures must be interpreted under the report's units, coverage filters, data construction, and excluded-carrier rules. They do not imply that a typical insured claim terminates after the reciprocal of 3.189% reporting periods.
The report notes that fifteen of eighteen participating companies entered the aggregated claim termination database, with three excluded for missing information necessary to compute exposure or reason for termination. This is a nontrivial representativeness issue. Weighting companies by observed exposure cannot eliminate systematic differences between included and omitted carriers unless selection mechanisms are understood. The raw numerator and denominator describe the study population under its cleaning rules; they are not an automatically transportable national incidence or mortality surface.
Censoring and truncation
Administrative study closure, policy records unavailable after a merger, and loss of traceability are forms of censoring when the event of interest is unobserved. Standard nonparametric and semi-parametric approaches require appropriate assumptions about the relationship between censoring and subsequent outcome, conditional on included predictors. A claimant who exits for a separately observed competing cause is not administratively censored for estimation of cumulative incidence; the competing event changes the probability of reaching the event of interest. Artificially treating deaths as ordinary censoring when estimating recovery probability generally targets a different hypothetical quantity.
An observed maximum benefit duration is a contractual cap, not proof of physiological recovery. If exhaustive records identify that exhaustion occurred, it can be represented as an administrative competing transition out of the payment state. If the scientific target is underlying care need, follow-up after exhaustion is necessary to learn when dependency actually ends. Without it, care-duration inference may be nonidentifiable beyond the benefit limit even if the liability duration is fully observed.
The 2017 SOA caveats also highlight limited observations in certain longer policy durations and claim periods. Tail behavior must therefore receive explicit sensitivity bounds and should not be presented as empirically identified merely because a parametric curve can be extended beyond the observed follow-up.
Descriptive survival and conditional remaining duration
Let S(t) denote the probability that a claim remains in its initial active state beyond elapsed duration t under the selected exit definition. The restricted mean episode duration through horizon tau is:
The restriction is important when the mean beyond the observed horizon is sensitive to unobserved tail assumptions. The expected remaining time, conditional on an episode already lasting s, can be written for a fixed horizon h as:
These quantities describe selected duration definitions rather than universal facts about care. A payment-state restricted mean substitutes the probability of being in an eligible and payable state for the initial claim-state survival when recurrent eligibility, temporary suspension, or exhaustion is possible.
Heterogeneity, calendar effects, and long-duration tails
Duration depends on attained age, sex, cognitive status, care need severity, diagnosis, setting, policy type, calendar year, and elapsed duration. Age at onset and attained age should not be interchanged without an explicit time scale. The earliest claim month may exhibit unusually high recovery following an acute event; later months may become dominated by severe chronic impairment and death. A constant aggregate hazard therefore averages over evolving case mix. Apparent duration dependence can reflect both within-person changes and the selective depletion of higher-risk subgroups.
The Japanese administrative cohort studied by Mikoshiba and colleagues, in care-demanding status trajectories (published online 2025; volume 2026), found high persistence in specified need states and extremely low transitions to a no-disability category in its older insured population. The result concerns that study's government assessment categories, ages, dates, and one-year transition intervals. It must be reconciled with the more frequent recorded recoveries in a product such as group LTD, which uses occupational disability criteria and very different entry and exit rules. The contrast demonstrates the methodological importance of phenotype and target population; it does not establish a contradiction in human recovery.
Recovery and Mortality Competing Risks
Cause-specific hazards and cumulative incidence
For a single active-care state with two mutually exclusive observed first exits, let lambda_R(t) be the instantaneous cause-specific hazard of recovery at elapsed time t, conditional on still being active immediately before t. Define lambda_D(t) analogously for death. In a continuous-time setting without additional exits, the all-exit survival probability is
The cumulative incidence of exit k by time t is
For continuous hazards, the left limit is immaterial. At each horizon, S(t)+F_R(t)+F_D(t)=1. In practice, additional observed exits require further hazards and terms in this identity. The expression clarifies why an increase in death hazard can reduce the observed cumulative probability of recovery even if the recovery hazard itself remains unchanged: more individuals die before they can recover.
The distinction is described in Andersen and colleagues' cause-specific risk analysis (2012). A standard Kaplan–Meier analysis that censors competing deaths while estimating recovery describes a counterfactual net failure distribution under additional assumptions, rather than the observed-world recovery cumulative incidence. Aalen–Johansen estimation of competing transition probabilities or direct integration of estimated cause-specific hazards is appropriate when the target is absolute exit probabilities. This is an application of established methodology.
Cause-specific versus subdistribution regression
A cause-specific proportional hazards specification may use
where the baseline and coefficients can differ between recovery and death. Each coefficient compares instantaneous exit rates among individuals who remain in the state. It does not, by itself, give a proportional ratio of cumulative incidence probabilities. When the objective is direct covariate association with the cumulative incidence of a designated cause, the Fine–Gray subdistribution model (Fine and Gray, 1999) provides a different regression formulation.
For LTC valuation, an implementer must choose the model according to the required output. Cause-specific intensities can be combined into transition and payment-state probabilities, and provide a natural basis for scenarios where mortality and recovery change separately. Subdistribution regression is useful for predicting a particular cumulative exit probability under a specified competition structure. Its risk set is constructed for that estimand and should not be interpreted as a direct physical transition intensity to insert mechanically into a multiple-state claims cash-flow model.
Illustrative constant-hazard calculation
Consider a hypothetical newly eligible active claim, with recovery hazard 0.08 per year and death hazard 0.12 per year, constant over time. Assume the events are mutually exclusive first exits, there are no other exits, no recurrence, and no covariates. The total exit hazard is 0.20 per year. Therefore,
After four years, active survival is 0.449329, cumulative recovery incidence is 0.220268, and cumulative death incidence is 0.330403. The sum is one within rounding. The unrestricted expected time in the active state is five years, while the four-year restricted mean is (1-e^{-0.8})/0.2 = 2.753355 years. These are synthetic mathematical outputs, not estimates of LTC population behavior.
If recovery rises to 0.12 while death stays 0.12, the total exit hazard rises to 0.24 and the four-year survival falls to 0.382893. The four-year recovery incidence becomes 0.308554 and death incidence 0.308554. A rise in the recovery hazard reduces death's cumulative incidence because some claimants recover first, even though the death hazard is unchanged. Conversely, if death rises to 0.16 while recovery remains 0.08, four-year active survival is again 0.382893 but recovery cumulative incidence is only 0.205702 and death incidence 0.411405. Identical all-cause survival can therefore conceal materially different exit causes.
This algebra is central to actuarial interpretation. For an immediate payment stream that stops upon either recovery or death, the two 0.24-total-hazard scenarios have the same expected duration. For a combined product with death-linked or recovery-linked additional benefits, the distinction becomes material to total contract value. Under a model allowing recurrence after recovery, the hazards also lead to different expected future entries into the payable state. Cause composition matters even when initial-state survival coincides.
Semi-Markov duration and recurrent states
The Markov assumption conditions future transition intensities on the current state, with any age or calendar dependence explicitly represented. LTC states frequently retain memory of episode duration, prior care admissions, and cumulative functional decline. A semi-Markov approach can condition transitions on elapsed time since state entry, making it useful for recovery patterns concentrated soon after claim onset and mortality patterns that vary with time in institutional care. Czado and Rudolph's LTC transition estimation demonstrates established duration-sensitive survival methods in an insurance setting.
A transparent multi-state structure might distinguish independent functioning, community-based care, institutional care, recovery to noneligible functioning, death, benefit-exhausted but still dependent, and payment suspension. Some transitions are physiologically possible while contractually ineligible; death remains absorbing. An analyst may refine care states by severity, but each additional parameter requires enough observed transitions for stable estimation. Very fine states can increase structural realism while lowering statistical credibility.
For recurrent episodes, a person-level transition history records each entry and exit. The probability of being in the paid state at time t becomes a state-occupation probability that can include multiple spells, not simply S(t) from the first episode. Comparing first-episode duration and lifetime time in care requires this distinction. Recurrence models should account for person-level dependence and observation before death, since death removes the opportunity for future episodes.
The meaning of an observed recovery
Recorded recovery is often an administrative conclusion that a claimant no longer meets a specific threshold. Changes in assessment schedules, documentation quality, rehabilitation access, and benefit rules can alter the observed rate without a commensurate shift in underlying physiology. In cognitive impairment, improvement in care setting or available support can reduce the need for paid services while functional limitations remain. The WHO ICF framework is useful here because personal functioning and environment are recognized jointly, while the insurance classification remains a separate normative determination.
The SOA/LIMRA 2015–2024 LTD termination study (2026) reports more than 43 million life-months exposed, approximately 937,000 recoveries, and over 181,000 deaths from eighteen contributing organizations. Its definition of recovery involves the context of group long-term disability; the observations cannot be combined with LTC care-need terminations as though they measure the same risk. They illustrate the feasibility of cause-specific administrative data collection and the need to preserve product-specific definitions.
Payment Duration Forecasts
From episode time to benefit time
The insurer pays according to an entitlement process that can differ from continuous clinical dependency. Let A(t) indicate that a claimant is alive and in a care-eligible state at duration t, G(t) indicate that applicable waiting-period and benefit-plan conditions have been met, U(t) represent eligible utilization or the fraction of a maximum benefit used at t, and B(t) represent the maximum payable amount per unit time. A schematic discounted present value is
with horizon H and discount factor v(t) selected for the valuation basis. This is an accounting identity under its definitions, not a proposed new model. For an expense-incurred policy, U(t) and B(t) must reflect covered incurred expenses and limits; for qualifying cash indemnity products, utilization may not play the same role. Benefit exhaustion can make G(t) zero while A(t) remains one.
The NAIC guide to elimination periods explains how calendar-day and service-day waiting rules operate and why later episodes may or may not trigger a new waiting period. Two claims with the same survival duration can produce different payments when one receives paid services intermittently. Likewise, a policy's benefit duration can refer to a stated number of months, a daily benefit times a pool of money, or other contractual restrictions. The exact policy wording determines the limiting process.
Worked expected-payment calculation
Extend the preceding hypothetical with a fixed benefit of $20,000 per year, paid continuously while the initial claim remains active, beginning 90 calendar days after time zero, ending at four years from onset, and without discounting. Assume the waiting time is exactly 90/365 years, benefit utilization is one, no benefit exhaustion occurs before the four-year cap, and death or recovery stops payments immediately. The payment expectation is
With recovery hazard raised from 0.08 to 0.12 and death fixed at 0.12, the corresponding expected payment is approximately $46,637.17. Raising death to 0.16 instead and leaving recovery at 0.08 gives the same expected payment because both scenarios have total hazard 0.24 and the toy contract stops on either event. Lowering recovery to 0.04 while keeping death at 0.12 increases the result to approximately $54,252.95. These examples isolate one assumption and are not insurer pricing recommendations. The values omit expenses, inflation, escalation, discounting, lapse, capital, tax, repeated care, and adverse selection.
For a more realistic policy with an explicit cash-flow calendar, the integral becomes a month-by-month expected present value, weighted by the probability of payment eligibility in each month and by the conditional expected cost of services. A 90-day statutory qualifying expectation should not be imported as a universal 90-day elimination period; the above toy contract assumes its waiting rule only for demonstration.
Benefit exhaustion and tails
An unlimited biological care duration can coexist with finite insurer liability. If a claimant reaches a benefit pool limit while remaining severely dependent, a claims system may close the benefit record as exhausted. The correct treatment depends on the estimand. For risk of death or recovery while physically dependent, this is a loss of outcome follow-up or an intermediate state, not a physiological competing death or recovery. For duration of payable liability, exhaustion is an actual termination of payment and must be represented as such. The SOA's 2017 basic-table caveats demonstrate why inherited industry table definitions need exact reconciliation with the intended valuation endpoint.
Long-term care projections are sensitive to infrequent but prolonged spells. Long spells may be underrepresented if studies have short follow-up or exclude closed books with divergent experience. Extrapolating a parametric tail after the last credible duration should produce explicit alternatives, including survival curves with slower and faster late exits. Restricted means at observed horizons remain estimable under less tail dependence than unrestricted lifetime means. A tail scenario should be paired with cash-flow caps and care-severity trajectories to determine whether financial exposure is similarly long.
Jurisdictional and product comparisons
Singapore's CareShield Life benefit conditions specify that eligible monthly cash benefits continue while severe disability involving at least three activities of daily living persists. The official program explanation records the qualifying threshold and lifetime-payment design; “lifetime” here does not mean that a person continues to receive payments after recovery from the qualifying severity. For comparison, U.S. tax-qualified private LTC insurance uses a distinct legal threshold, and specific policies impose their own elimination periods and benefit pools.
These differences create four comparative dimensions: the clinical or functional test, the rule for sustained satisfaction of the test, the actual benefit trigger and payment frequency, and the upper bound on payment. A pooled analysis should first standardize—or explicitly stratify by—these dimensions. Apparent national differences in mean paid duration may arise from case severity, aging, economic access to care, assessment rules, revalidation frequency, or benefit design. Pooling payment durations without observing those variables cannot identify differences in underlying health-state persistence.
The U.S. SOA/LIMRA 2025 LTC study announcement describes an updated initiative for policy years 2000–2023 covering mortality, persistency, incidence, and claim termination. The corresponding data solicitation excludes hybrid and combination products from that stand-alone LTC cohort, an essential scope limit for any subsequent benchmarking. The initiative's existence does not authorize attributing unpublished study results to it. As of the sources reviewed, its announced study program is evidence about expected future data availability, not new finalized numerical termination rates.
Reserving and accounting consequences
Prospective reserves project future net cash flows conditional on policies and claims in force. For an active LTC claimant, the distribution of remaining payment duration is often more relevant than the unconditional episode length from incidence. A change in recovery or mortality assumptions can affect both the total expected payments and their timing; an offsetting effect may arise in combined life/LTC designs when an accelerated LTC benefit reduces or changes a death benefit. The ASOP 18 valuation guidance stresses product features, historical and emerging experience, long-horizon assumptions, and sensitivity testing within its application.
Under financial-reporting regimes applying IFRS 17 Insurance Contracts, cash-flow measurement must reflect the governing accounting framework and contract boundary; no single illustrative survival formula substitutes for an IFRS 17 measurement. A pricing or reserve model should therefore preserve separate layers for observed transitions, projected benefit use, financial discounting, and accounting treatment. This separation permits independently reviewing whether a revision in projected payments came from updated health-state experience, contract design, economic assumptions, or actuarial judgment.
Data, Legal, and Model Validation
Data provenance and linkage
A defensible claim-duration dataset begins with a field-level lineage inventory. Dates should be marked as observed, inferred, imputed, or administratively assigned. Variables should include policy issue and coverage period, functional assessment history, care setting, service start and end, claim receipt and decision, elimination period, payment dates, death date and its independent source, benefit balance, termination reason, and record coverage. The existence of a date field does not prove that it measures an independently witnessed biological event.
The SOA's aggregate LTC data documentation shows how minimum and maximum service dates and a derived claim start can be used to reconstruct exposure. Such choices are suitable for their published design, yet a future multi-state model would need more information about changes in the care setting, time outside services, and death ascertainment. A different inference target must not silently inherit a simplified study file's variable definitions.
The Actuarial Standards Board's ASOP 23 Data Quality (2016 revision, effective 2017) addresses relevance, sufficiency, consistency, limitations, and disclosure in actuarial use of data. Applied to episode research, its emphasis leads to a reconciliation report: counts of claims by source carrier; duplicate person and claim records; overlapping episodes; unknown termination reasons; claims censored at the extraction date; claims with death after administrative closure; and exposure by attained age, elapsed duration, and care setting. This report is more informative than presenting an unexplained percentage of “clean” records.
Informative censoring and missing causes
Censoring assumptions can fail when discontinuation of records depends on health status. A transfer to a new administrator may systematically affect certain policies. An exhausted benefit pool makes continuing claim payments unobservable by design. A missing reason code can correlate with the complexity of terminal cases. Sensitivity analyses should separately consider unknown-cause reassignment, inverse-probability-of-censoring weights where justified, cause imputation with model uncertainty, and bounds that allocate unknown terminations to alternative causes.
Deaths can be undercounted if claims close as suspended before a mortality registry is linked. Recovery rates can be exaggerated if operational closures for missed documentation are treated as improvement. Conversely, a claimant can regain functional independence while the record remains administratively open. No generic adjustment factor resolves both problems. The audit must compare adjudication reason, independently assessed functional change, benefit utilization and subsequent claims histories where accessible.
The SOA caveat memorandum expressly notes reliance on carrier coding of recovery and death, with differences in claim management and plan design as relevant interpretive factors. Its warning supports an empirical validation protocol, not a numerical estimate of coding bias. The necessary reclassification rate remains unidentifiable without independently reviewed records.
Model selection, credibility, and calibration
A model should be as complex as the data and decision require. A nonparametric cause-specific estimator is an appropriate benchmark for first observed exits when there is sufficient follow-up and reliable cause coding. Piecewise constant hazards can be transparent for claims projection. Cox models permit covariate comparisons and time-dependent data under their assumptions. Flexible spline or semi-Markov models may better represent steep early recovery followed by a persistent tail, but need validation over age and duration cells. A multi-state approach is warranted when care-setting transitions, recurrence, or separate payment states affect valuation.
The ASOP 56 Modeling standard requires attention to model purpose, assumptions, controls, validation, and communication within its scope. In a practical LTC application, calibration should compare observed and expected counts by exit cause, policy generation, duration band, attained age, and initial state. For cumulative incidence predictions, compare cause-specific predicted probabilities with empirical curves at the same horizons. Model fit to a combined all-cause termination count can conceal wrong cause mixes and materially misleading projected mortality-related effects.
Validation should include time-split out-of-sample tests where data volume permits, calendar-year backtesting, external study comparison after definition harmonization, parameter stability under different episode gap rules, and a transparent review of unusual tails. A model may fit a historical carrier cohort but fail when benefit design or claims practice changes. Long-term projections should evaluate the impact of shifts in care delivery, medical rehabilitation, morbidity, and survival separately rather than imposing one unexplained trend on all causes.
A quantitative sensitivity matrix
The constant-hazard example supplies a minimal unit test. With lambda_R=0.08 and lambda_D=0.12, it must reproduce S(4)=0.449329, F_R(4)=0.220268 and F_D(4)=0.330403. At every t, the probabilities must be nonnegative and sum to one. Under the 90/365-year waiting period, four-year cap and annual benefit of $20,000, expected payments should be approximately $50,255.22. With total hazard 0.24, expected payments should be approximately $46,637.17, regardless of whether the increase comes from recovery or death under the explicitly narrow toy contract. This invariance is a diagnostic, not a statement about actual combined or recurrent policies.
A second test changes the benefit rule while holding biological hazards constant. Replacing a calendar-day waiting period with service-day accumulation can delay payment for intermittent service users; the amount of delay depends on the utilization schedule and cannot be calculated from an aggregate survival curve alone. A third test allows eligibility to resume after recovery, requiring transition probabilities from the recovered state rather than a one-way survival model. A fourth test imposes exhaustion at a known monetary pool and checks that payment ceases while functional dependency can continue. A fifth test permits death to produce a contractual secondary benefit and verifies that exit-cause composition now matters even when total first-exit hazard is equal.
These tests target structural coherence as well as arithmetic. A spreadsheet or software implementation should expose all time units, rate transformations, time origin conventions, event definitions and payment assumptions. An annual hazard cannot be inserted unchanged as a monthly transition probability; for constant hazard lambda the monthly probability over 1/12 year is 1-exp(-lambda/12), before competing-event allocation. Likewise, a cause-specific monthly first-exit probability must be derived consistently from the total hazard and the cause share under the assumed within-month model.
External validity and legal interpretation
Data from private U.S. individual LTC insurance, public Japanese care assessments, Singapore national disability coverage, and group LTD insurance have different populations, legal claims, administrative routines and benefit designs. The Japanese study of longitudinal care-demand transitions is valuable evidence about persistence within its defined public program. The SOA/LIMRA LTD termination report is valuable evidence about occupational disability claims. Neither establishes directly transferable quantitative recovery parameters for a new private LTC cohort.
Comparative validity requires identifying common health-state definitions, mapping legal eligibility, comparing data collection frequency, reconciling competing exit codes, and testing whether the post-entry populations are comparable. Standardization may reduce case-mix differences but cannot erase unmeasured selection into insurance, care provision, or claims. The distinction between eligibility and payment is particularly salient for cash disability programs and reimbursement-based LTC insurance. Any cross-jurisdiction pooled estimate should present jurisdiction-specific estimates beside the pooled one.
Reproducibility and boundaries of available evidence
This manuscript's original calculations are fully reproducible from stated parameters and equations. The document uses institutional study reports, public standards, legislation, and scientific publications available through public summaries or open pages. It does not claim access to insurer-level data, paid experience-study packages, full closed-access articles, or confidential statutory valuation workpapers. The SOA's LTC study inventory identifies reports and associated data resources, but availability of a study's overview does not establish independent access to its complete participant-level database.
A credible release record should preserve the exact scope of each checked source, including whether a source was consulted as a full institutional page, an abstract, a text-rendered PDF, or an official summary. This affects the strength of conclusions: the model identities and toy outputs can be checked internally, while empirical representativeness and claims-code error rates require access to actual source records. These are scientific boundaries, not tasks disguised as already completed validation.
Integrated Findings and Research Agenda
Evidence-based synthesis
The evidence supports five mutually reinforcing conclusions. First, the event named “long-term care episode” requires an operational time origin and a fixed event taxonomy. Second, death and recovery are mutually exclusive first exits from an active state, whereas benefit exhaustion and other administrative closures may be exits only from a payment state. Third, absolute probabilities of cause-specific exit require a competing-risks calculation; modeling one exit while treating the other as ordinary censoring cannot be read as the observed probability of that exit. Fourth, paid duration depends on policy rules and utilization as well as functional persistence. Fifth, empirical transportability depends on product, jurisdiction, recording practices and case mix.
A direct implication is that one aggregate termination rate supplies insufficient information for a complete insurance liability projection whenever additional benefits, recurrence, or cause-dependent payment rules matter. The synthetic calculation demonstrates the identification problem: the two scenarios with a total hazard of 0.24 generate identical active-state survival and identical restricted payment duration under a simple cessation-at-first-exit contract, while their recovery and death probabilities differ substantially. An actuarial valuation or social-care forecast concerned with what happens after the first exit needs the cause composition and subsequent transition structure.
The 2020 SOA LTC data study provides an institutionally credible example of operational claim construction, but its published aggregates are subject to defined exposure windows, source-carrier exclusions and recording conventions. The subsequent 2000–2023 LTC data request illustrates the next potential update cycle and reinforces the need to distinguish prospective data collection from already published experience. The 2026 LTD study confirms that cause-specific termination information can be collected at large scale in another insurance domain, while its product boundary must remain explicit.
Research priorities grounded in the available sources
The first priority is an independently adjudicated linked dataset combining functional assessment, service utilization, insured entitlement, payment history, recovery decisions and death registration. Such linkage could estimate misclassification of recoveries and distinguish biological cessation from administrative closure. The second priority is a harmonized episode construction experiment that compares alternative recurrence gaps and assesses their impact on mean paid duration, cause-specific cumulative incidence and total lifetime care time. The third priority is long-follow-up information about claimants after policy benefit exhaustion, where observing only payments may conceal persistent dependency.
A fourth priority is multi-state model validation under changing benefit definitions, including cash versus expense-reimbursement designs and different ADL thresholds. A fifth is the measurement of calendar-time changes in recovery and mortality using consistent risk sets rather than crude rates distorted by evolving age and duration composition. A sixth is explicit distributional analysis: whose episode is lost from the dataset when policy lapse, inability to claim, administrative appeal, or benefit exhaustion occurs? These questions connect duration science to access and measurement without presuming new empirical results.
The applied research program should report both statistical and financial outputs: cause-specific cumulative incidence, active-state survival, restricted duration, expected paid months, benefit-pool exhaustion probability and discounted payment projections. Each must be labeled with a time origin, conditioning population, data period, policy specification and cause taxonomy. Reports can then be compared on their declared estimands rather than on labels such as “average claim length.”
Final interpretation
Long-term care episode duration is a joint product of health-state persistence, observed transitions, contract eligibility, and the researcher's definition of an episode. Recovery and mortality compete in determining first exit; care-setting changes and recurrence create histories that extend beyond a single first-exit observation. Payment ends when the applicable insured obligation ends, which may occur before or after the relevant clinical transition is recorded. The most defensible actuarial practice therefore begins with a documented state-and-event definition, estimates the hazards supported by the available data, and translates state occupation into cash flows under the actual contract.
This synthesis follows established actuarial and statistical methods. It makes an original cross-source interpretation of their application to LTC episode and payment definitions, while leaving empirical claims at the strength justified by the cited evidence. It proposes neither a new doctrine nor an unsupported priority claim.
Conclusion
An insured long-term care episode acquires a meaningful measured duration only when its entry criterion, exit events, observation process and payment rules have been specified together. The analytical separation of care need, claim status and payment entitlement prevents recovery, death, exhaustion and administrative closure from being mistaken for equivalent endpoints. Cause-specific hazards and cumulative incidence describe distinct statistical objects; multi-state transition probabilities become necessary when care settings and benefit eligibility change over time or episodes recur.
The examined actuarial studies, official standards, legal rules and scientific literature establish a coherent framework for estimating and interpreting competing exits. The original worked example proves, within declared assumptions, that identical total exit hazards can conceal different recovery and death compositions while yielding the same payment expectation for a contract that stops payment on either first exit. It also demonstrates how a waiting period and benefit cap turn survival into a restricted expected payment. Empirical estimates for an insurer or public program require its actual records, eligibility definitions, mortality ascertainment and validation.
The principal practical result is a rigorous specification for claim-duration research: establish reproducible episodes; preserve person and policy histories; identify causes separately; estimate and validate transition probabilities; calculate payment-state occupancy under documented contracts; and disclose tail uncertainty, selection and record limitations. Such work can support credible pricing, reserving, care planning and comparative research when its evidence and its intended uses are held to the same explicit definitions.
References
Actuarial Standards Board. (2016; effective 2017). Actuarial Standard of Practice No. 23, Data Quality. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2019; current online standard). Actuarial Standard of Practice No. 56, Modeling. https://www.actuarialstandardsboard.org/asops/modeling/
Actuarial Standards Board. (2022). Actuarial Standard of Practice No. 18, Long-Term Care (revised March 2022; effective September 1, 2022). https://www.actuarialstandardsboard.org/asops/asop-18-long-term-care/
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. https://pubmed.ncbi.nlm.nih.gov/22253319/
Central Provident Fund Board, Singapore. (2026 access). CareShield Life: Benefits, eligibility, and payment conditions. https://www.cpf.gov.sg/member/healthcare-financing/careshield-life
Central Provident Fund Board, Singapore. (2026 access). What is CareShield Life? https://www.cpf.gov.sg/service/article/what-is-careshield-life
Czado, C., & Rudolph, F. (2002). Application of survival analysis methods to long-term care insurance. Insurance: Mathematics and Economics, 31(3), 395–413. Open institutional research record and working-paper version. https://epub.ub.uni-muenchen.de/1649/
Fine, J. P., & Gray, R. J. (1999). A proportional hazards model for the subdistribution of a competing risk. Journal of the American Statistical Association, 94(446), 496–509. https://doi.org/10.1080/01621459.1999.10474144
Friedberg, L., Hou, W., Sun, W., & Webb, A. (2023). Lapses in long-term care insurance. Journal of Risk and Insurance, 90(3), 569–595. https://doi.org/10.1111/jori.12425
IFRS Foundation. (2026 access). IFRS 17 Insurance Contracts: Issued standard overview and effective-date information. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
Mikoshiba, M., Kawamura, A., Awatani, T., & Noguchi, H. (2025 online; 2026 volume). Persistence and irreversibility of care-demanding status: Insights from long-term care claims data in Japan. BMC Geriatrics, 26, Article 139. https://link.springer.com/article/10.1186/s12877-025-06854-0
National Association of Insurance Commissioners. (2022 edition; guide text revised 2019). A Shopper's Guide to Long-Term Care Insurance. https://content.naic.org/sites/default/files/publication-ltc-lp-shoppers-guide-long-term.pdf
Society of Actuaries and Towers Watson Delaware, Inc. (2017). Caveats for Use of Long Term Care Experience Basic Tables (May 2017). https://www.soa.org/globalassets/assets/files/research/exp-study/2017-caveats-ltc-basic-tables.pdf
Society of Actuaries, Long-Term Care Experience Committee. (2020). Long-Term Care Intercompany Experience Study: Aggregate Database 2000–2016 Report (August 12, 2020). https://www.soa.org/4a31a6/globalassets/assets/files/resources/experience-studies/2020/2000-2016-ltc-aggregate-database-report.pdf
Society of Actuaries Research Institute. (2020). 2000–2016 Long-Term Care Aggregated Databases: Study overview and source files. https://www.soa.org/resources/experience-studies/2020/2000-2016-ltc-aggregate-database/
Society of Actuaries Research Institute. (2026 access). Long-Term Care Experience Studies: Research inventory. https://www.soa.org/research/topics/ltc-exp-study-list/
Society of Actuaries Research Institute. (2025). 2000–2023 Long-Term Care Data Request: Claim incidence, termination, and utilization. https://www.soa.org/research/opportunities/2025/2000-23-ltc-datarequest/
Society of Actuaries Research Institute, LIMRA, & National Association of Insurance Commissioners. (2025, August 4). Industry leaders launch new long-term care insurance industry experience study. https://www.soa.org/resources/announcements/press-releases/2025/industry-leaders-launch-ltc-study/
Society of Actuaries Research Institute & LIMRA. (2026, September 30). 2015–2024 Group Long-Term Disability Termination Experience Study. https://www.soa.org/resources/experience-studies/2026/2015-2024-group-long-term-disability-term/
United States Congress. (2026 access; laws in effect through Public Law 119-120). Internal Revenue Code, 26 U.S.C. § 7702B, Treatment of qualified long-term care insurance. https://uscode.house.gov/us/usc/119-120/t26/s7702B/
World Health Organization. (2001; current online classification). International Classification of Functioning, Disability and Health (ICF). https://www.who.int/classifications/international-classification-of-functioning-disability-and-health