Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Long-Term Care Entry Incidence by Functional State

Abstract. Long-term care (LTC) entry incidence is the frequency with which individuals occupying a defined functional and insurance-exposure state enter a qualifying care or benefit episode over a specified time interval. Its estimation requires an explicit distinction among functional deterioration, care use, certification, claim incurrence, administrative adjudication, and commencement of payment. This monograph develops a source-grounded analysis of the incidence estimand, denominators, competing events, functional measurement, transition processes, experience calibration, and contractual comparability. It integrates the World Health Organization's functioning framework; foundational activities-of-daily-living instruments; original U.S. insurance and tax provisions; Society of Actuaries insurer experience definitions; longitudinal population research; and selected long-term care systems in Europe and Asia. Established multi-state and competing-risks statistics supply the formal methods. A transparent hypothetical calculation demonstrates why a single crude claim rate cannot identify age- and state-specific care-entry hazards. The synthesis locates the most consequential uncertainty in the map between latent functional need and observed payable incidence, the composition of the active risk set, and the timing and contractual meaning of an insured event. No new actuarial theory, proprietary dataset, or causal estimate is claimed.

Keywords: long-term care insurance; incidence; functional state; activities of daily living; instrumental activities of daily living; severe cognitive impairment; claim trigger; multi-state model; competing risks; exposure; active lives; morbidity; selection; claims adjudication; actuarial calibration; international comparison.

Introduction

A person can need help, receive help, qualify under a public statute, incur an insured loss, submit a claim, and begin receiving money at six different times. Treating those observations as a single event distorts long-term care incidence. The correct object of measurement is a transition from a specified origin state to a specified destination state, observed in a defined population under an explicit contractual or program rule. This framing is essential to pricing, projection of long-tail liabilities, benefit design, clinical assessment, fairness, and public-system comparison.

The WHO classification of functioning supplies a common vocabulary for activities, participation, health conditions, and environmental influences (World Health Organization [WHO], 2001). Its multidimensional understanding of disability is especially important when physical limitations coexist with preserved independence because of assistive devices, supportive environments, or family care. The OECD care affordability study likewise uses activities of daily living (ADL) and instrumental activities of daily living (IADL) to classify care needs, but reports their population distribution rather than the incidence of privately insured payable events (OECD, 2024).

A long-term care insurance benefit can be triggered by performance limitations, cognitive impairment, a certified care plan, eligible service use, or combinations of these terms. In the United States, the federal qualified-LTC definition contains a specified functional or cognitive threshold and practitioner certification (United States, 2026). The Singapore CareShield Life assessment applies a three-ADL threshold within a different statutory cash-benefit design (Central Provident Fund Board [CPF], n.d.). The German and Dutch public systems use other entitlement structures, discussed below. Such differences alter the numerator and eligibility time at which an entry is recorded.

Three questions organize this study. First, whose time is at risk, and which transition defines a claim? Second, how do functional trajectories, certification, and death generate and obscure observable entries? Third, how can heterogeneous insurer, survey, and public administrative data be used without claiming a comparability or precision they do not possess? The analysis applies established incidence accounting, event-history reasoning, and statutory comparison, accompanied by original derivations and explicit hypothetical stress tests. It does not replicate restricted insurer microdata or estimate a new population incidence table.

Conceptual Foundations and Historical Development

The statistical event and the insured event

Incidence is a flow: new events during a period relative to the population-time capable of experiencing them. Prevalence is a stock: individuals already occupying a state at a particular time or during a period. A claim count is an administrative observation; its relationship to underlying disability onset depends on the definition and data collection process. A payment count may include subsequent payments for an earlier claim and cannot serve as an onset numerator without episode reconstruction.

For an insured population, define the sequence conceptually as health and functioning; material decline in functioning; qualifying persistence or severity; practitioner or approved-assessor certification; claim notice; verification and decision; satisfaction of any applicable elimination condition; and payable episode. The sequence may contain overlapping events and may not be monotonic. A cognitive-impairment qualification can exist without the same number of physical ADL limitations; assistance may be needed intermittently; and certification may occur after the date attributed to claim incurrence. An individual receiving informal care is not thereby necessarily on an insured claim.

The distinction has direct actuarial significance. An insurer valuing the probability of a first payable LTC episode requires an incidence basis tied to contract terms and benefit inception. A public-health analyst estimating onset of a new ADL limitation requires a health-state transition basis, including persons who have no qualifying policy. An administrator estimating initial eligibility decisions requires an applications-and-certifications basis. These rates can answer complementary questions, but their denominators and observation rules are different.

Functional ability, ADL, IADL, and cognition

The historical basis for functional measurement includes Katz and colleagues' standardized ADL index (Katz et al., 1963) and Lawton and Brody's instrumental daily-living assessment (Lawton & Brody, 1969). The former emphasizes essential self-care activities; the latter extends assessment to tasks such as household management and independent community living. Their instruments identify clinically and socially meaningful dimensions, yet a research measurement instrument is not itself an insurance-policy trigger.

The ICF describes functioning in context. A limitation in an activity can be affected by capacity, environmental barriers, assistive technology, social support, and the conventions of assessment. The WHO's functional ability account describes ability as an interaction of intrinsic capacity and environmental factors (WHO, 2020). Thus, two individuals with similar diagnoses can occupy different observed functional states, and an individual can cross an insurance threshold after a change in circumstances even without a new pathological event.

A useful measurement record preserves the individual ADL items rather than only an aggregated score; the degree and kind of assistance; whether ability was assessed with customary aids; the assessment date; the historical reference period; and the source of the information. It separately identifies cognition, supervision and safety risk. Clinical classifications often allow gradations from independent performance to cueing, supervision, hands-on help, or total assistance. Policy language may group these gradations differently. Collapsing them into a single unqualified label such as dependent creates uncontrolled measurement error.

IADL deficits can precede basic ADL limitations. Consequently, an IADL-only state can carry elevated future entry risk without satisfying a two- or three-ADL insurance trigger. The distinction is relevant both to anticipatory pricing and to equity: low availability of unpaid care can increase observed service use among persons with similar measured impairment, while extensive family care can delay presentation for a payable service. Those are mechanisms to be evaluated, not numerical effects that can be assumed.

Why an episode can end and begin again

Long-term care trajectories include improvement, deterioration, institutional transfer, changes in caregiving arrangements, hospitalization, and death. Ankuda and colleagues' longitudinal incident-disability study followed 8,713 U.S. adults age 65 or older after incident ADL or IADL disability. At two years, 22.1% had recovered and 46.8% had died; care arrangements and functional states shifted substantially (Ankuda et al., 2020). These are outcomes in a particular Health and Retirement Study cohort, measured biennially; they are not prospective private-LTC insurance claim termination probabilities.

Recurrent episodes require a rule for separating new entries from interruptions of a continuing episode. If a claim is temporarily suspended and resumes, one administration may retain one incurred claim while another counts a renewed claim. The SOA's 2000–2016 LTC aggregation explicitly grouped service dates that overlapped or were separated by six months or less, producing a common definition across contributors. The choice is consequential for incidence and recurrence analysis, even if cumulative dollars paid remain identical.

The institutional history of thresholds

Functional assessment has become a gateway to insurance entitlement. The U.S. federal tax-qualification framework in 26 U.S.C. § 7702B refers to inability to perform at least two listed ADLs for an expected period of at least 90 days, a legally defined alternative level of disability, or substantial supervision due to severe cognitive impairment; certification by a specified practitioner is required. The listed activities are eating, toileting, transferring, bathing, dressing, and continence (United States, 2026). The anticipated-duration criterion must be distinguished from a policy's elimination period; neither should silently be reinterpreted as 90 days of mandatory nonpayment.

The NAIC model regulation contains benefit-trigger standards, assessment provisions, and independent review procedures (National Association of Insurance Commissioners [NAIC], 2017). It is model regulatory language, not automatically effective law in every U.S. state or every policy generation. The model text does not itself determine the covered benefits, service conditions, benefit limits, or state adoption dates of a particular contract. Contract-specific comparison must preserve these layers.

The conceptual history therefore spans clinical instruments, functioning classifications, underwriting and product design, formal eligibility rules, and longitudinal actuarial experience. Each layer contributes information; none supplies by itself a universal incidence parameter.

Incidence Population and Denominator

Defining the risk set and observed exposure

Let i index individuals, a denote attained age, t calendar time, s the origin functional state, and r the product or program rule. Let Y_i,s(t) equal one when individual i is observed, legally covered, in state s, and eligible to generate the target entry at time t, and zero otherwise. Define the exposure for an interval [t0,t1] as

E_s,r = sum_i ∫_[t0,t1] Y_i,s(u) du,

expressed in person-years when time is measured in years. Let N_s→P,r count first transitions from state s to the defined payable-entry state P during that same risk interval. A conventional central estimate is

m_s→P,r = N_s→P,r / E_s,r.

This is an events-per-risk-time rate under the specified recording rule. It becomes a direct probability only after a time scale and survival assumptions are specified. Under a constant hazard h over one year with no competing decrement, the corresponding entry probability is 1 − exp(−h). With death, lapse, recovery, and competing entries, the cumulative incidence requires the full transition system.

The denominator must be conditional on the event being possible. A policyholder already in an active benefit episode is generally not at risk of a first incidence event. A policyholder whose contract lapsed is not exposed for a later covered entry under that policy. A person who joins midway through the period contributes only eligible time. An individual living with severe impairment before policy issue may be excluded by underwriting or contract terms. Using the entire population of older adults, the number of insured policies in force, annual premium volume, and active person-years interchangeably produces different and often misleading rates.

A rate based on policy-years also requires an explicit unit for people with multiple policies. One person may have two contracts and one underlying episode; two contracts may generate two administrative claims. Person-based event counts and policy-based liability counts serve different questions. The appropriate aggregation is defined before analysis and recorded in the source metadata.

The SOA's exposed-life accounting

The 2020 SOA aggregate database report describes an insurer experience file for 2000–2016 derived from 18 contributing companies, representing 80% of 2016 U.S. LTC earned premium. The aggregate incidence component comprised 61,323,281 exposure units and 620,591 incurred claims, with a reported overall incidence rate of 1.012% (Society of Actuaries [SOA], 2020). These are data from submitted policies and claims, grouped under the study's definitions; they are not a weighted national survey of all people who need LTC.

Critically, the source differentiates total-lives exposure from active-lives exposure. Total-lives time spans the observation of the in-force policy without removing time on claim. Active-lives exposure reduces that time for intervals already on claim. The study applies Balducci-style exposure conventions, assigning a full year in the policy year of incurrence, while partial years may be assigned for death or lapse. It also removes claims that began before exposure start and combines closely spaced claims using the six-month episode rule. A naive quotient of unique people with disability divided by total insured lives would therefore fail to reproduce the published statistic.

Its information-release constraints matter. Dates including claim incidence are reduced to years, attained ages 90 and above are aggregated, and records with very low exposure are merged into unknown categories for privacy. The publicly aggregated dataset is valuable for descriptive pricing research and historical benchmarking, but cannot identify day-level functional state trajectories or the causal effect of a specific ADL transition on claim entry. The SOA database methodology documents the units, suppression and grouping choices in detail (SOA, 2020).

Ages, durations, cohorts, and time scales

Entry incidence is ordinarily steeply dependent on attained age, but age alone is insufficient for experience reconciliation. Attained age, issue age, policy duration, issue cohort, calendar year, underwriting practice, gender, benefit configuration, tax qualification, care setting, and prior health history may all be correlated. An apparent calendar trend can be caused by a cohort with different underwriting, a change in claims handling, or a shift toward older attained ages. Statistical adjustment must specify which variables are observed before onset and which are determined by prior selection or claims behavior.

Attained age is a natural time scale for progression of dependency; duration since policy issue captures selection and underwriting wear-off; duration since entering a functional state captures disease and care-path dependence; and calendar time captures institutional and environmental change. In individual histories these clocks operate simultaneously. A model that uses only policy duration to represent biological deterioration risks assigning an age effect to select underwriting eras. Conversely, an age-only model may mistake changes in applicant selection for changing morbidity.

Conditioning on an already impaired origin state is analytically different from estimating incidence across all active policies. If s describes one IADL limitation, its exposure comprises only those observed in that state and still at risk for the payable transition. If the historical claims dataset does not contain measured s, the desired conditional rate is not identifiable without a linked functional dataset and a defensible transport or imputation model. This limitation cannot be repaired by finer grouping of age alone.

Delayed entry, censoring, death, and selection

Prospective observation may begin after an individual entered the underlying population state. Left truncation is handled by including a person in the risk set only when first observed and eligible. Right censoring is defensible when follow-up ends for reasons conditionally independent of the event under the fitted model; loss of contact may violate that assumption. Death is a competing event for first entry, rather than a noninformative censoring event, because a deceased person cannot subsequently enter an episode. Policy lapse has a different interpretation: it removes insured exposure, while dependence might still occur in the population.

This distinction generates two targets. The insured-portfolio incidence measure answers what happened while coverage existed. A population-based dependency estimate seeks health transitions regardless of continued insurance. If healthier individuals are more likely to lapse, restricting observation to persistent policies changes the case mix. If newly impaired policyholders are more likely to retain coverage, selection into the active risk set is dynamic. The separation between hazard among active policies and incidence among original policy purchasers must remain explicit.

Finkelstein and McGarry's LTC private-information evidence identifies multiple dimensions of privately known risk and insurance preferences in the LTC market (Finkelstein & McGarry, 2006). Consequently, insured cohorts need not reflect population morbidity distributions, and underwriting effects cannot be inferred from a single crude selection ratio. A model can be internally calibrated on a portfolio yet externally inapplicable to a mandatory social scheme or a newly insured group.

Incidence, prevalence, utilization, and cost

Population need is relevant to projecting demand and public expenditure, but it cannot be substituted for the flow of new insured payable claims. OECD (2024) reports that nearly one-quarter of older people across OECD countries have some LTC needs under its study definitions, with approximately 12% classified as low-need, 8% moderate-need, and 4% severe-need on average. The OECD needs distribution is cross-sectional and definition-dependent. It is evidence about the stock of needs, not an actuarial rate of first insured entry at a known age and functional origin state.

Similarly, Medicaid service utilization describes access to and financing of services. The Medicaid HCBS statistics report that 86.2% of Medicaid long-term services and supports users received home- and community-based services in 2021, accounting for 63.2% of spending (Centers for Medicare & Medicaid Services [CMS], n.d.). These percentages condition on users or spending and do not estimate the chance of first functional impairment among all eligible people.

The economic object extends beyond incidence: expected benefit outgo is a product of entry rates, subsequent survival and exit dynamics, service intensity, covered cost, contractual maxima, and inflation or benefit growth. Separating those components is necessary for interpretation and for credible sensitivity analysis.

Functional Assessment Transitions

From a functional description to an event-history state

A functional state is a defined set of observable conditions at a time. It may be represented by the vector S(t) = [A(t), I(t), C(t), H(t)], where A records the item-specific ADL limitations and assistance levels, I records IADL capacity, C records cognitive impairment and supervision need, and H records the social and physical setting relevant to assessment. The vector can have many combinations; any reduced operational classification must document the rule used to combine them. Age, policy duration, and product are covariates or strata, not elements that automatically imply the same functional severity.

For valuation, a finite state space may distinguish independent functioning; IADL limitations only; one basic ADL limitation; multiple ADL limitations; severe cognitive or supervision needs; qualifying but not yet payable status; payable episode; post-episode recovery; and death. Such a classification illustrates established multi-state methodology rather than prescribing a universal actuarial taxonomy. The states must be mutually exclusive for a conventional state-occupancy model: where severe cognitive impairment overlaps ADL limitations, a recorded precedence rule or a multidimensional product state space is required. Otherwise one person is counted twice.

Insurance eligibility is often a derived indicator B_r(S,t) that depends on functional information, assessment authority, contract version, duration expectation, and perhaps a prescribed care plan. It may be preferable to retain cognition and ADL components independently and compute the eligible subset separately, because a single ordinal disability score cannot represent every legal threshold. Payability may require additional states for service receipt, elapsed elimination days, claim approval, and benefit exhaustion. An individual's transition to a qualifying functional state can therefore precede the transition to a payable insurance episode.

Multi-state intensity and transition probability

For states j and k, the cause-specific transition intensity λ_jk(a,t,z,d) expresses the instantaneous rate of a transition from j to k conditional on still occupying j. Here a is age, t calendar time, z measured characteristics, and d time already spent in state j. In a time-homogeneous Markov approximation, the transition matrix P(Δ) is exp(QΔ), where off-diagonal elements Q_jk are nonnegative transition rates and each diagonal Q_jj equals minus the sum of rates leaving j. For a time-varying or age-dependent model, transition probabilities follow the corresponding forward equations or product integral.

Putter, Fiocco, and Geskus (2007) explain the multi-state estimation framework and the practical distinction between cause-specific hazards, competing risks, and state occupation probabilities. An insurance implementation should estimate the transition intensity of entry from each source state alongside death, lapse, recovery, progression, and other defined transitions. It must then use the resulting full transition system to calculate the probability of an insured payable entry before a specified horizon.

A pure one-direction health-to-claim model is often too restrictive. People can improve, move between care settings, experience a prior short episode, or cease to satisfy an ADL threshold. Absorbing death is generally appropriate, whereas classification of recovered and payable states depends on recurrence and exhaustion conditions. The model should allow a new claim after recovery only when that is consistent with the study's episode-separation rule; a reopening of the same claim may require a transition that does not count as a new incidence event.

The Markov property imposes an observable assumption: conditional on the current modeled state and covariates, the next transition probability does not additionally depend on unrecorded past history. That assumption can fail if someone with a one-year progressive impairment and someone with a one-week post-surgical limitation receive the same current severity label. A semi-Markov specification, inclusion of duration in state, or history-enriched states can address that substantive distinction. The chosen approach should be justified through comparative prediction, clinical meaning, and data support rather than by a preference for maximum model complexity.

First-entry cumulative incidence and competing death

Suppose an active person occupies preclaim state s with a constant rate λ_P of first payable entry and a constant competing death rate λ_D. If these are the only exits, the probability of payable entry by horizon τ is

F_P(τ) = [λ_P / (λ_P + λ_D)] × [1 − exp(−(λ_P + λ_D)τ)].

This follows by integrating the claim intensity against probability of remaining in s: F_P(τ) = ∫_0^τ exp[−(λ_P + λ_D)u] λ_P du. If lapse also removes insured entry opportunity, an additional intensity λ_L appears in the relevant insured-portfolio cumulative incidence, with the important interpretive caveat that lapse is not necessarily the end of functional risk. For a public-health analysis, the treatment of lapse is different.

A useful numerical stress test is entirely hypothetical. With annual λ_P = 0.04 and λ_D = 0.08, the one-year payable-entry probability is 0.03769. Ignoring death and using 1 − exp(−0.04) would produce 0.03921. Both are lower than the corresponding instantaneous rate of 0.04, but they answer different survival models. These rates were chosen to demonstrate an identity; they are not fitted LTC industry rates. Over longer horizons and with state-dependent death, disagreement may be much larger.

Fine and Gray's (1999) subdistribution hazard method directly parameterizes the cumulative incidence of a specified competing event. Its coefficient is not interchangeable with a cause-specific transition coefficient and generally is less immediately suited to simulating all dynamic functional state transitions. A cause-specific or full multi-state model is natural for actuarial cash-flow trajectories; a subdistribution model can be useful for directly predicting an observed first-entry cumulative incidence under a fixed competing-risk definition. The target estimand and intended projection, rather than a generic claim that one method is superior, determine the choice.

Recovery, recurrence, and duration dependence

A transition model for LTC must include the possibility of recovery or improved functioning. The incident-disability outcomes described by Ankuda et al. (2020) show that recovery, death, and changing care arrangements can occur within short observation horizons. A model that treats first ADL dependence as irreversible overstates future time in dependency states for recoverable conditions. Yet a model that allows every episode to resolve independently may understate persistent cognitive or degenerative impairment. Stratification by duration, underlying cause when reliably measured, and cognitive status makes these assumptions testable.

Recurrent-event analysis is distinct from first-event incidence. Define T_1 as first qualifying episode and T_2 as a later episode that meets the agreed separation rule. The hazard of T_2 depends on survival after T_1, recovery or service interruption, available benefit limits, and whether the original policy continues. Claim-level analyses may count claim reopenings that a person-level analysis deliberately excludes. The decision must be recorded as an event dictionary, not left to an analyst's intuition.

Policy benefit periods and elimination conditions create additional clocks. If a policy requires an elimination period measured in days of qualified service, time beginning at functional deterioration is not necessarily equal to time beginning at first qualified service. If a product uses calendar days, the clock differs again. Administrative payment may be retrospective. For valuation, the distinction among incurred date, approved date, payable date, and cash-transfer date affects both measurement and discounting.

Observation schedules and interval censoring

Functional decline is often first detected at a periodic assessment, while the actual transition occurred earlier. Annual survey waves can locate an onset in an interval but not on an exact day. A person observed with one ADL limitation in year one and three limitations in year two might have crossed the policy threshold at any intermediate time. Assigning all transitions to the date of the second interview creates artificial delays and seasonality. Assigning them to the interval midpoint imposes a distributional assumption.

The NHATS methodological documentation provides documentation for a U.S. aging survey with repeated measures and sample-design considerations (National Health and Aging Trends Study [NHATS], n.d.). Survey records can inform functional-state transitions when measurement definitions align, but claims dates from an insurer may have different temporal precision. A valid linkage should preserve each data stream's precision, handle interval censoring where needed, and account for sampling weights when estimating population quantities.

Misclassification matters because a trigger based on a discrete ADL count is discontinuous. An assessor's difference between one and two ADLs may move a U.S. tax-qualified individual across a functional boundary. A change in the definition of assistance, the assessor's training, the presence of a caregiver, or a cognitive cueing protocol may change the observed state without equivalent change in health. With the confusion matrix M_oy = Pr(observed class o | underlying class y), observed transition counts reflect both real transitions and error. Estimation of a latent transition process then requires information about M, often through repeat assessments or adjudicated records; it cannot be identified from a single cross-tabulation alone.

The original analytical implication is direct: the insured-event hazard is generated by a functional transition process observed through an assessment and contract filter. Models of either process in isolation can fit their own data yet disagree about payable incidence. The empirical question is the magnitude and direction of that filter under each source and product, which requires linked observations.

Experience Calibration and Uncertainty

Distinguishing experience sources

Insurer experience data measure claims and exposure governed by historical contracts, underwriting, policy duration, and reporting systems. Aging surveys measure reported or observed functioning in a population sample. Administrative social-program data measure certified eligibility and use among covered residents. These sources overlap in subject matter but have different sampling mechanisms and legal definitions. The aim of calibration is to identify which parameters each source can support and to propagate unresolved differences rather than averaging unlike estimates.

The SOA LTC experience catalog includes incidence, persistency, mortality, and claims studies. The 2000–2016 aggregated database provides original insurer-source claim experience and documentation. The 2018 incidence-trend research contrasts an observation-based approach with an actual-to-model approach to investigating calendar-year trends (SOA, 2018, 2020). Neither source establishes an insured ADL-transition dataset by itself. Annual experience and historical policy generations should be aligned before assigning trend to deterioration in population health.

The 2025 SOA–LIMRA data request sought stand-alone LTCI claim incidence, termination, and utilization experience for contract years 2000–2023, including individual and group policies, while excluding hybrid, acceleration, and combination products (SOA, 2025). Its existence indicates ongoing industry study development, not verified results of a completed 2000–2023 experience table. As of this manuscript's research cutoff, the SOA upcoming studies listing identifies the 2000–2023 LTC study as planned for 2027 (SOA, 2026). The 2020 insurer dataset must therefore be clearly dated when used; it cannot be relabeled as a current 2026 experience study.

Group long-term disability (LTD) studies address occupational or work-disability benefits and different claim eligibility. The SOA–LIMRA 2015–2024 LTD termination study was released on September 30, 2026, using more than 43 million life-months, approximately 937,000 recoveries, and more than 181,000 deaths contributed by 18 companies (SOA & LIMRA, 2026). These observations provide recent evidence of sophisticated claim termination experience, but they must not be inserted as LTC incidence rates or used to identify nursing-home entry. The shared actuarial methods are transferable; the insured-event definition is not.

Worked hypothetical: why case mix changes the crude rate

Consider four mutually exclusive origin classes among insured, not-on-claim people, observed over a fictional one-year period. These figures are illustrative arithmetic and do not originate from any insurer, survey, or official study.

Preclaim functional classification Eligible active person-years First payable episodes Incidence per 100 person-years
Independent basic and instrumental activities 10,000 30 0.30
IADL limitation only 4,000 40 1.00
One basic ADL limitation 2,000 60 3.00
Cognitive/support risk below the specified policy trigger 1,000 70 7.00
Entire fictional portfolio 17,000 200 1.1765

The illustrative crude rate is 200 / 17,000 = 0.0117647 events per active person-year, or 1.1765 per 100. Every state-specific rate is a hypothetical conditional rate. The fourth category is expressly below the relevant trigger at the start of exposure, so that a subsequent first payable entry remains possible. The table does not claim that any particular real contract recognizes the displayed categories.

Suppose a second fictional portfolio has the identical four state-specific rates but twice the share of higher-risk origin states. Its aggregate claim incidence will be higher solely because of the different state distribution. A conclusion that one portfolio's people deteriorate faster would be unwarranted. Conversely, direct standardization to a common risk-set distribution w_s yields a comparable state-standardized rate m_std = Σ_s w_s m_s, provided each state has a common definition and an adequately measured positive exposure. If one portfolio omits cognitive classification, even this standardization is not identified without additional assumptions.

The same principle extends to attained age, issue cohort, and policy duration. Direct comparison is possible only within common support. A statistically adjusted estimate extrapolating to an age–severity cell absent from one source requires explicit model dependence and uncertainty. In sparse cells, an apparently precise fitted coefficient can conceal weak empirical identification.

Actual-to-expected analysis and credibility

Let O_g be observed claim entries in stratum g, E_g its active exposure, and q_g a basis probability or appropriately exposure-converted expected annual event rate. The expected count X_g = E_g q_g is a convenient benchmark where exposures and probabilities share a convention. The ratio A/E = (Σ_g O_g)/(Σ_g X_g) summarizes deviation from a selected model. The ratio depends on what was expected, how the risk set was constructed, and whether the expectation includes death and lapse competition. A low A/E can arise from conservative expected rates, improved underwriting, missing claims, coverage changes, or survivor selection; it is not proof of improved functional health.

For state-specific calibration, the expected claim counts can be calculated from estimated transitions to payable status conditional on starting state, including the calendar interval, enrollment duration, and competing exits. If the chosen multi-state model estimates the full hazard matrix, expected entries are integrals of state occupancy multiplied by transition intensity, rather than simply the number of insured persons multiplied by a universal factor.

Sampling uncertainty must be reported together with structural uncertainty. Under a Poisson approximation, a claim count O has standard error √O, making the relative count standard error approximately 1/√O. Sparse severe-state strata may therefore be highly variable even in a large portfolio. Overdispersion, clustering by provider or insurer, and repeated claims require alternatives to an independent Poisson assumption. Credibility blending can moderate cell noise, but the reference experience must genuinely describe a comparable event, population, and reporting convention.

Historical insurer experience can also be affected by selective underwriting and lapse. The insurance selection study of Finkelstein and McGarry (2006) is a reminder that both expected care risk and preference for coverage can be privately known. Consequently, a simple calibration factor applied uniformly to a new policy generation with different selection may appear statistically smooth and still be economically misspecified.

The distinction among trend, aging, and administration

Calendar-year trend can arise from changing population functioning, care supply, pricing or benefit utilization, legal thresholds, assessment behavior, or claims administration. An improved claims-reporting system may move entries earlier without changing underlying morbidity. COVID-19 and other disruptions can change care-service use, death, and facility admissions concurrently. A cohort's age distribution may shift even if its state-specific intensities remain stationary.

A defensible trend study records at least the following: exposure and claim-date conventions; enrollment and policy vintage; attained age and duration; ADL and cognitive definitions; treatment of care setting; benefit triggers and elimination periods; lapse and mortality; changes in claim coding or service reimbursement; and the degree of left truncation or interval censoring. Calendar effects are then interpreted conditional on the measured changes. The SOA's incidence trend methods provide a direct industry reference for asking whether an observed calendar shift persists after expected experience is modeled (SOA, 2018).

Neither the absence of a statistically significant trend nor a stable A/E ratio demonstrates constant biological dependency incidence. Statistical power, reporting delays, model choice, and offsetting mechanisms determine what can be inferred. The correct research statement identifies the exact model, observation period, and remaining competing explanations.

Valuation sensitivity and uncertainty propagation

Pricing and liability projections combine incidence, severity or utilization, episode duration, recovery, death, inflation, benefit maxima, interest, and lapse. A source-linked review of ASOP No. 18 on long-term care confirms the scope of actuarial work on pricing, funding, and evaluating liabilities (Actuarial Standards Board [ASB], 2022). The ASOP No. 56 modeling standard addresses model design, assumptions, validation, and governance more generally (ASB, 2019). These U.S. actuarial practice standards guide covered professional work; they do not supply universal empirical claim probabilities.

Let π_s(a,t) be the proportion of eligible active exposure in origin state s, and let F_s,P(τ) be the state-specific probability of a payable episode within horizon τ with competing events recognized. A population mixture gives F_portfolio(τ) = Σ_s π_s F_s,P(τ) for a fixed initial cohort under consistent membership. Expected discounted benefits then integrate first entry at time u against conditional duration, covered daily or monthly amounts, benefits already used, and discount factors. This is a conventional valuation decomposition, not a new theoretical system.

Sensitivity testing should vary both the transition parameters and the exposure-state composition. One scenario can hold state-specific hazards fixed and alter π_s to quantify selection or demographic change. A second can change transition hazards with π_s fixed to isolate morbidity assumptions. A third can alter observed eligibility through a contract-definition or assessment change while preserving underlying latent functioning. A fourth can perturb the exit structure: mortality and recovery affect occupancy and benefit duration even if first-entry hazards are unchanged. The value of the decomposition lies in making assumptions legible to review.

Empirical limitations as findings about identifiability

The presently reviewed open source record supports functioning concepts, statutory eligibility, historical insurer claim incidence, selected longitudinal transitions, and standard statistical methods. It does not supply a common international individual-level panel jointly observing repeated item-specific ADL performance, cognition, policy wording, certification, insurer claims, elimination period, benefit payment, and full mortality follow-up. A single fully identified cross-country transition table cannot be derived by combining public summaries alone.

This is a substantive result about the available evidence. More complicated estimation does not resolve the missing linkage automatically. An analyst may estimate a scenario under stated bridging assumptions, then validate it if linked records become available. Reporting that scenario as observed international incidence would replace an evidence gap with a hidden assumption. Scientific and actuarial QA require that the boundary remain visible.

Integrated Findings and Research Agenda

Entry is a multistage, state-conditioned observation

The reviewed evidence supports an integrated analytical conclusion: payable LTC incidence is a conditional transition observed through multiple classification and administrative stages. A population's functional risk, its insurance-exposure composition, the policy trigger, and the claim recording process jointly determine observed claim entries. Each stage contributes a distinct source of variation.

The functional process includes changes in ADL and IADL capacity, cognition, and environment. The insurance process selects covered individuals and determines what can count as a claim. The assessment process decides when impairment is formally recognized, and the administration process decides when benefits become payable. Mortality and lapse limit exposure before entry; recovery and death govern episode occupancy after entry. The framework is a synthesis of established probability, epidemiology, and insurance practice, rather than a proposed new system.

The central consequence is interpretive. A crude claim rate can be correct for its observed portfolio while giving little information about the state-specific rate needed for a different underwriting cohort or contract. A clinical state-transition estimate can accurately describe a population while providing no valid incidence of approved insurance benefits without a coverage and claims link.

Identifiable quantities and quantities that require linkage

Different sources identify different margins of the process. The SOA's insurer databases identify historically coded incurred claims against insured exposure, with stratification by available policy variables. Health and aging panels identify trajectories of reported or assessed functioning within the sampled population and observation schedule. Official eligibility texts identify legal thresholds and certification conditions. Government administrative series may identify certifications, benefit uptake, or expenditures in publicly administered populations.

The intersection is usually incomplete. Identification of the entire mapping from health states through certification to payable insured episodes requires that the same individual's functional measurements, coverage history, event chronology, and claims outcomes be jointly observed, or that a bridge design support cross-source mapping. Absent such linkage, a claim that a given functional transition causes a specified dollar liability would exceed the data.

This limitation has an immediate effect on how hypothetical examples and projections are presented. The illustrative portfolio in this monograph is an arithmetic demonstration, not a predictive pricing table. Its state-specific rates cannot be licensed as estimates for any real insurer. Likewise, a numerical scenario translating German care grades into a U.S. two-ADL threshold would require an explicit bridge dataset or externally validated crosswalk before its outputs can be presented as empirical.

A complete minimum data structure for future research

A linked incidence study should observe a stable individual identifier; jurisdiction and insurance contract; issue date, attained age, gender or other relevant demographic information, benefit terms and underwriting class; exact periods of coverage and on-claim status; dated item-level ADL, IADL, and cognitive assessments; assistance level and care environment; relevant certification findings; claims and appeals; actual service use; payment dates and amounts; and death or other policy exits. The dataset should record who performed each assessment and under which instrument version.

The ideal design contains repeated preclaim functional observations and a sufficiently long follow-up to distinguish persistent dependence from temporary functional decline. It would keep a prospective cohort under observation after coverage lapse for population-based dependency research, under an appropriate ethical and legal basis, even though the actuarial insured exposure ceases at lapse. It would distinguish prior claims, first claims, and later independently defined episodes. Claims data alone often cannot fulfill these requirements.

A practical staged research strategy can begin with a clear risk-set definition and existing industry experience, then add a representative functional survey and independently document the bridge assumptions. Administrative claims may help validate incidence time when real contract-level data become accessible. Validation against holdout policy years should follow, with transparent reporting of whether transitions are directly observed or model-imputed.

What must be reported in every incidence table

An incidence table is interpretable only when its label includes the event and denominator. A table headed "LTC incidence by age" omits essential information. A reproducible heading or accompanying metadata should specify whether the numerator is first incurred insured claims, first approved benefits, new ADL threshold crossings, or new public care indications; whether the denominator is all in-force or active person-years; whether already-on-claim intervals were removed; the age convention; observation dates; contract generation and benefit trigger; and the treatment of death and lapses.

Each rate should be accompanied by the relevant exposure and event count, a measure of statistical uncertainty, and an explanation of suppressed or missing cells. Where a modeled or smoothed estimate is published, distinguish it from a raw experience rate. Where a contractual threshold changed, preserve separate pre- and post-change definitions rather than producing a single undocumented trend.

For portfolio projections, the full model documentation should additionally identify the fitted state space, how exposures enter and leave each state, the mortality and lapse assumptions, the transition model's calendar and duration dependence, and the method used to transform events into discounted cash flows. The objective is a reproducible scientific account in which a later reviewer can determine exactly what was estimated.

The principal research questions remaining open

First, how much of the difference between insurer incidence and population functional-dependency onset reflects underwriting, lapse, and financial selection? The reviewed publications establish that selection exists and that private information can be multidimensional, but the magnitude of the bridge is specific to products, cohorts, and data.

Second, what is the predictive value of IADL-only and early cognitive states for first payable claims, conditional on age and competing mortality? Longitudinal functional surveys and insurer files separately contain relevant information, but public summaries reviewed here do not supply a common full transition table.

Third, how robust are claim incidence estimates to the chosen episode-gap rule, assessment frequency, and the treatment of claims that begin before the data window? The SOA's six-month grouping and exposure conventions demonstrate that these choices are material. Systematic comparisons would require the underlying person-level records or defensible interval-specific sensitivity analyses.

Fourth, how do legal trigger changes influence measured claims independently of health? The U.S., Singaporean, German, Dutch, and Japanese mechanisms provide institutional variation, but any causal comparison must account for concurrent differences in eligibility, coverage, population, and administrative practice.

Fifth, how much do measurement error and missing cognitive information affect state-based projection and pricing? This calls for assessment-validation subsamples and repeated measures. An administrative zero in a severe-state cell could reflect true scarcity, privacy suppression, eligibility, or unobserved data. These explanations must be distinguished before conclusions are drawn.

Sixth, what uncertainty should accompany the use of historical experience in a future risk regime? A source-locked historical report captures its observation period and contributor mix. Projection beyond that domain requires separately argued demographic, clinical, care-supply, policyholder-behavior, and contract assumptions. A point estimate without sensitivity ranges understates decision risk.

Methodological contribution and limits

The original contribution of this monograph is the explicit source-to-estimand reconciliation and comparative interpretation of functional transition, coverage, certification, and payment incidence, supported by a worked competing-risk derivation and a clearly labeled hypothetical state-mix demonstration. It establishes why three ostensibly similar "long-term care entry" rates can be different valid statistics, why their numerical comparison requires common risk sets and observation rules, and how an evidence audit locates unmeasured linkages.

These are analytical findings drawn from established science and source-specific documentary research. No new theory, doctrine, law of incidence, cross-country actuarial table, original empirical parameter, or demonstrated causal effect is introduced. External peer review was not commissioned for this manuscript. Proprietary insurer claim microdata were not obtained or independently replicated, and cross-country incidence validation cannot be claimed from the open records reviewed.

Conclusion

Long-term care entry incidence is scientifically meaningful only in relation to a specified origin functional state, a specified entry event, and a demonstrably eligible exposure population. The observed event may be clinical deterioration, formal certification, claim incurrence, approval, or payment. These dates, states, and risk sets are institutionally related but analytically distinct.

The reviewed literature provides a coherent basis for studying the problem. Katz and Lawton–Brody supply foundational functional assessment instruments; the WHO provides a multidimensional functioning framework; longitudinal disability research demonstrates the importance of recovery and death; SOA experience identifies actual insurer exposure and claim conventions; statutory and program records specify distinct entitlement thresholds; and established multi-state statistics accommodate changing conditions, recurrence, and competing events. Their contributions can be combined conceptually without treating unlike data as if they were measurements of one universal rate.

The actuarially defensible route is to define the target event, reconstruct eligible risk time, maintain item-level functional distinctions, recognize cognitive eligibility and competing death, validate recurrent episodes and administrative dates, and stratify or standardize against known changes in case mix. International analyses must describe each jurisdiction's observation and certification rule before comparing numerical results.

Two consequences follow. First, the predictive use of a crude historical portfolio incidence rate outside its source population requires explicit selection and eligibility assumptions. Second, reliable functional-state-specific payable incidence requires linked functional, contractual, and claims records or clearly qualified bridging models. In their absence, the correct research result is a bounded estimate or a documented identification gap, never an unsupported precision.

These conclusions give insurance science a disciplined basis for future linked experience studies, transparent valuation sensitivity, consumer-protection analysis, and comparability of public and private LTC financing. They also preserve the distinction between what existing source evidence establishes and what additional empirical research must determine.

References

  • Actuarial Standards Board. (2019). Modeling (Actuarial Standard of Practice No. 56). Actuarial Standards Board. ASOP No. 56.

  • Actuarial Standards Board. (2022). Long-term care (Actuarial Standard of Practice No. 18). Actuarial Standards Board. ASOP No. 18.

  • Ankuda, C. K., Levine, D. A., Langa, K. M., Ornstein, K. A., & Kelley, A. S. (2020). Caregiving, recovery, and death after incident ADL/IADL disability among older adults in the United States. Journal of Applied Gerontology, 39(4), 393–397. Article and abstract.

  • Central Provident Fund Board. (n.d.). CareShield Life. Government of Singapore. Scheme and eligibility.

  • Centers for Medicare & Medicaid Services. (n.d.). Home and community-based services. U.S. Department of Health and Human Services. HCBS program information.

  • Centers for Medicare & Medicaid Services. (n.d.). Long-term services and supports. U.S. Department of Health and Human Services. LTSS program information.

  • Federal Ministry of Health (Germany). (n.d.). Long-term care insurance. Federal Republic of Germany. Federal ministry guidance.

  • 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.

  • Finkelstein, A., & McGarry, K. (2006). Multiple dimensions of private information: Evidence from the long-term care insurance market. American Economic Review, 96(4), 938–958. https://doi.org/10.1257/aer.96.4.938.

  • Government of the Netherlands. (n.d.). Apply for a long-term care indication (Wlz indication). Government.nl. Wlz indication process.

  • Katz, S., Ford, A. B., Moskowitz, R. W., Jackson, B. A., & Jaffe, M. W. (1963). Studies of illness in the aged: The index of ADL: A standardized measure of biological and psychosocial function. JAMA, 185(12), 914–919. PubMed study record.

  • Lawton, M. P., & Brody, E. M. (1969). Assessment of older people: Self-maintaining and instrumental activities of daily living. The Gerontologist, 9(3, Part 1), 179–186. https://doi.org/10.1093/geront/9.3_Part_1.179.

  • Medicare. (n.d.). Long-term care. Centers for Medicare & Medicaid Services. Medicare coverage explanation.

  • Ministry of Health, Labour and Welfare (Japan). (n.d.). Long-term care insurance system of Japan. Government of Japan. Official program overview (PDF).

  • National Association of Insurance Commissioners. (2017). Long-term care insurance model regulation (Model 641). NAIC. Model regulation (PDF).

  • National Health and Aging Trends Study. (n.d.). Methods and documentation. Johns Hopkins University. Survey documentation.

  • Organisation for Economic Co-operation and Development. (2024). Is care affordable for older people? OECD Health Policy Studies. OECD Publishing. Full report.

  • Organisation for Economic Co-operation and Development. (2024). Growing long-term care needs across OECD countries. In Is care affordable for older people? OECD Publishing. Care-needs chapter.

  • Putter, H., Fiocco, M., & Geskus, R. B. (2007). Tutorial in biostatistics: Competing risks and multi-state models. Statistics in Medicine, 26(11), 2389–2430. https://doi.org/10.1002/sim.2712.

  • Society of Actuaries. (2018). Researching long-term care insurance incidence rates over time. SOA. Incidence trend research.

  • Society of Actuaries. (2020). Long-term care intercompany experience study—Aggregate database 2000–2016 report. SOA. Original technical report (PDF).

  • Society of Actuaries. (2020). 2000–2016 long-term care aggregated databases. SOA. Study and data downloads.

  • Society of Actuaries. (2025). 2000–2023 long-term care data request. SOA. Research data request.

  • Society of Actuaries. (2026). Upcoming studies. SOA. Study release planning.

  • Society of Actuaries. (n.d.). Long-term care experience studies. SOA. Experience research catalog.

  • Society of Actuaries & LIMRA. (2026). 2015–2024 group long-term disability termination experience study. SOA. LTD experience study.

  • United States. (2026). Internal Revenue Code § 7702B: Treatment of qualified long-term care insurance. Office of the Law Revision Counsel, U.S. House of Representatives. Statutory text.

  • World Health Organization. (2001). International classification of functioning, disability and health (ICF). WHO. ICF classification.

  • World Health Organization. (2020). Healthy ageing and functional ability. WHO. Functional ability guidance.