RESEARCH DOSSIER
Long-Term Care Actuarial Models
Abstract. Long-term care (LTC) actuarial valuation concerns the timing, duration, setting, intensity, and contractual financing of assistance required by people experiencing physical or cognitive dependency. Its mathematical object is a set of contingent cash flows generated by transitions among functional states, eligibility decisions, care settings, and policy states. A single frequency-severity representation conceals mechanisms decisive for long-duration liabilities: recovery, death before a claim, death during a claim, recurrence, waiting periods, service substitution, mortality selection, persistency, price inflation, and capacity constraints. This monograph derives a coherent approach to those mechanisms using multistate transition models, competing-risk survival analysis, duration-dependent extensions, state-contingent benefits, prospective premiums, and reserve formulas. It distinguishes insurance claims from observed population care prevalence and separates the contractual definition of disability from clinical need. The source base integrates published actuarial standards, Society of Actuaries (SOA) experience-study documentation, UK multistate disability research, US statutory eligibility, OECD comparative evidence, and financial reporting guidance. A transparent hypothetical three-state case produces an expected 15-year benefit present value of $53,532.77 and a continuous net level premium rate of $6,097.89 per in-force policy-year under expressly stylized assumptions. Joint sensitivity to incidence, exit rates, and nominal care-cost growth raises the modeled benefit present value to $94,160.88; neither figure is a portfolio estimate. The principal finding is an analytical one: defensible LTC valuation requires explicit, internally consistent links among underlying dependency, observable experience, benefit design, and financial assumptions, with empirical uncertainty disclosed at every mapping.
Keywords: long-term care insurance; actuarial models; multistate models; competing risks; incidence; recovery; dependency; claim duration; utilization; morbidity trends; pricing; reserves; care inflation; model validation.
Introduction
Long-term care is an exposure to persistent assistance needs. The insured event may be the onset of specified functional limitations, severe cognitive impairment, use of particular services, or satisfaction of another policy-defined condition. The ensuing financial obligation depends on how long the condition persists, whether the insured survives and remains covered, whether home or institutional services are used, whether the contract reimburses expenses or pays a fixed amount, and when prices and payments change. These interactions justify an actuarial treatment that jointly models eligibility, survival, recovery, utilization, and contractual payment.
The scientific task is to determine how the probabilities of transitions into, through, and out of dependency generate a distribution of policy cash flows. The empirical task is to estimate relevant probabilities from observed experience while recognizing that records of care recipients, paid claims, clinical disability, and all persons at risk are distinct populations. The valuation task converts those probabilities into premiums, reserves, uncertainty margins, and transparent scenario results. Regulatory and accounting tasks add particular definitions and objectives, which must be identified rather than conflated with the underlying stochastic calculation.
The distinction between health-state occurrence and claim recognition is fundamental. An individual can have material need while awaiting assessment or receiving unpaid family support. A claimant can recover enough to cease payments and still face subsequent care needs. A covered person can die before a first LTC claim. A claim can end through exhaustion of contractual benefits even though dependency remains. The SOA's aggregate LTC experience documentation maintains separate incidence and termination data; this distinction supplies empirical support for modeling the flows separately. The LTC experience-study catalog documents continuing research into claims, mortality, lapses, and persistence.
The underlying research tradition is substantial. Rickayzen and Walsh's (2002) UK multi-state disability analysis combines prevalence, transition, and trend evidence to assess future LTC needs. Brown and Finkelstein's (2007) study of the private LTC market locates valuation inside a wider structure of prices, incomplete coverage, and insurance demand; their (2009) review of US evidence further documents the market's institutional complexity. These works support the use of multiple analytical lenses without implying that a particular transition model is uniquely correct in all insurance settings.
The monograph uses American English and international comparisons. The official US definition of tax-qualified LTC is treated as an example of a contract-adjacent legal eligibility rule, not as a universal definition of disability. The OECD's international LTC access comparisons address people receiving services under divergent public systems; such observations cannot be directly converted into private-policy incidence rates. Professional standards apply within their stated practice and jurisdictional scopes. The aim is an independently reproducible synthesis, accompanied by analytical conclusions and worked calculations. No proprietary actuarial theory or empirical causal estimate is claimed.
Method and evidence discipline
The research combines primary professional standards, insurance industry documentation, statutory materials, intergovernmental statistics, established actuarial scholarship, and transparent mathematical derivation. First, the conceptual state system is specified. Second, continuous-time intensities and competing hazards are translated into transition probabilities. Third, the resulting state occupation is mapped into claim cash flows and premium exposures. Fourth, numerical examples independently check the equations. Fifth, the operational limits of experience data and institutional comparability are examined. Where a source supplies a catalog, summary, or publisher abstract, the resulting claim is confined to that accessible material. An insurer's proprietary claim-level database, a fitted historical transition surface, a new population forecast, and an external peer review are outside the evidence obtained for this manuscript.
Care States and Eligibility
Defining the insured risk
The variable "needing care" must be operationalized before a probability is estimated. Functional impairment can be represented by activities of daily living (ADLs), instrumental ADLs, cognitive status, supervised support needs, or a validated severity score. Contractual payment may require certification, a specified number of ADL limitations, prior elimination of a waiting period, qualified providers, or expense documentation. Clinical deterioration, legal qualification, and insurance-payment eligibility are consequently three related observations of an individual's trajectory.
Under US tax law, the qualified LTC insurance provisions define an individual as chronically ill, for relevant purposes, through qualifying functional limitations expected to last at least 90 days, severe cognitive impairment requiring substantial supervision, or another statutorily specified comparable disability, subject to certification rules. The listed ADLs are eating, toileting, transferring, bathing, dressing, and continence. The statutory requirement that determination consider at least five ADLs for a tax-qualified contract must be distinguished from the actual benefit wording. Certification, care-plan, and contract rules can delay or restrict payments beyond biological onset. An actuarial state model therefore needs either explicit eligibility states or a documented mapping from dependency states into claim authorization.
US Medicare offers another essential boundary. Its long-term care coverage explanation states that it generally does not pay for custodial LTC. Medicare-financed skilled nursing and custodial personal assistance differ in purpose and eligibility. A portfolio projection that uses all long-stay nursing-facility costs as a private insurer's exposure would mix payment sources and benefits. Conversely, a private benefit that reimburses some home services will have claim incidence and utilization patterns different from a benefit limited to institutions. Public programs may partly substitute for private coverage and also change observed choices of where care is received.
A robust baseline uses states such as H (alive, no claim-level dependency), L (alive, low-level assistance), C_H (eligible and in home/community care), C_I (eligible and in institutional care), and D (dead). Additional policy states record active coverage, lapse, benefit exhaustion, premium waiver, and claim in payment. The care-state and policy-state classifications serve different purposes: a policy can lapse while dependency persists, and a claim can exhaust while the insured continues to receive services.
State granularity and the risk of false precision
Every additional care state creates transition parameters. With five living states and death, a fully connected continuous-time generator would permit up to 25 living-to-living or living-to-death rates, most of which may be poorly observed. Parsimony is governed by observable information and the materiality of state distinctions for benefits. If home assistance and institutional residence trigger materially different daily limits, separating the states can improve valuation. If claims data contain only an aggregate claim flag, finely differentiated disability states can produce spurious identifiability: many transition surfaces yield the same observed aggregate claims.
A practical hierarchy begins with three states: H (alive and not in care), C (alive and in qualifying care), and D (dead). Recovery C → H is permitted. H → C captures first and recurrent onset in a deliberately simplified model. H → D is pre-claim mortality and C → D is claimant mortality. Extended models separate disease progression, recovery quality, service location, pre-eligibility delay, and exhausted benefits. A three-state model is a transparent pedagogical basis; a production model requires matching the actual product.
The ASOP 18 long-term care standard applies to specified LTC benefit-plan actuarial work, including design, pricing, funding, and liability evaluation. It calls attention to actual and reasonably foreseeable future experience. A model choice should be evaluated relative to that purpose. A concise model suitable for examining a benefit cap may be insufficient for a statutory capital projection extending decades; the same mathematical structure can remain useful while its required calibration and uncertainty controls change.
Population states, payment states, and exposures
Consider three measurement clocks. Attained age measures biological and cohort exposure; calendar time represents period conditions, including inflation or practice change; duration since entering C measures the claimant's spell. A policy-duration clock tracks time since issue, which matters for underwriting selection, elimination periods, premium payment patterns, and policy termination. The transition rate from H to C may be written λ_HC(a,t,u,z), with attained age a, calendar time t, policy duration u, and covariates z. A claim exit rate may additionally depend on time already in the current care spell d.
The denominator of an incidence estimate is exposure among persons at risk of an eligible new claim. It is not all insured policy-years if those already in claim are ineligible for a new first claim. The denominator of a claim termination rate is open-claim exposure under the study's claim-episode convention. Benefits paid per claimant-day require further exposure distinctions: days eligible, days receiving services, days with an approved expense, and days paid after waiting periods are not interchangeable.
International comparison and the institutional boundary
The WHO integrated-care guidance emphasizes physical and mental capacity and care coordination; it offers a clinical and public-health perspective, not an insurance rate table. In OECD (2025) comparisons, around 12% of people aged 65 and over received LTC services in 2023 across 31 reporting countries. This recipient share across public systems is a prevalence-style service measure, not a claim entry hazard, and its interpretation depends on service availability, care definitions, population selection, and informal support. Modeling should preserve the same distinction when comparing US commercial policyholders with beneficiaries of European social LTC programs.
A useful representation separates the biological transition matrix, an observation/eligibility mapping, and a payment function. This is an analytical organization of established components, not a newly proposed theoretical system. It clarifies which features can be studied from claims records, which require external disability or population data, and which depend on contract wording.
Incidence and Duration
First-entry incidence and death as a competing event
Let T_C denote time to first entry into claim-qualifying care and T_D time to death before that entry. In an initially healthy covered population define cause-specific intensities λ_HC(t) and λ_HD(t), conditional on remaining in H and under the selected exposure definition. The probability of remaining alive and without qualifying care through t is
The cumulative probability of first care entry by t is
The corresponding pre-care death probability replaces λ_HC with λ_HD under the integral. Together, F_C(t)+F_D(t)+S_H(t)=1 when these are the only possible exits from H. Dividing care entries by all lives originally insured, or treating deaths as non-informative censoring for the absolute probability of care, can overstate claim entry. A cause-specific hazard answers a conditional instantaneous question; cumulative incidence answers the insurer's absolute event-probability question. The multistate and competing-risk tutorial demonstrates the Aalen–Johansen treatment that respects competing destinations.
Incidence curves are sensitive to age, issue cohort, selection, underwriting class, sex or gender classification where law and data permit, benefit provisions, benefit triggers, waiting periods, calendar year, and the source's incident-claim definition. The SOA's incidence trend methods report documents observation-based and actual-to-model approaches to detecting calendar trends. An analyst must specify whether a changed observed incidence rate arises from the risk of dependency itself, changing claims behavior, improved recognition, or changes in exposure composition.
Prevalence is not incidence
Prevalence at age a is the probability or population share currently occupying a care state. Incidence is a transition flow into that state. In a simplified stationary two-living-state setting, prevalence depends on incidence and exit rates; in realistic aging populations it also reflects mortality, entry cohorts, calendar trends, recovery, and immigration. One may observe increasing prevalence despite declining onset risk if care survival lengthens. One may observe decreasing prevalence despite increasing onset if claim durations shorten or deaths accelerate.
This matters for projections from cross-sectional surveys. A disabled population sample conditions on surviving long enough to be observed, giving greater weight to long spells. Recoveries may be missed if the observation interval is too long. Nursing facility records omit care delivered by relatives at home. Paid claims data omit uninsured care and may truncate benefit-exhausted episodes. The appropriate calibration problem is therefore reconstructing flows from the particular observation process, not simply substituting a stock measure into an entry-rate equation.
Duration and competing exits within care
Conditional on entering care at time 0, let μ_CR(d) and μ_CD(d) be cause-specific exit intensities after care duration d for recovery and death. Under an isolated care episode with no other exit, its survival-in-claim function is
Its mean episode duration is , provided the integral converges. Exit-by-recovery probability is ; death exit replaces the hazard accordingly. If both intensities are constant, the mean is .
For a deliberately hypothetical annual recovery intensity of 0.12 and care mortality intensity of 0.18, total exit intensity is 0.30 and mean spell length is 3.3333 years. Recovery accounts for 0.12/0.30 = 40% of modeled exits and claimant death for 60%. Treating the recovery hazard alone as the episode-ending rate would produce 8.3333 years, a materially different quantity. These are exponential competing-risk calculations, not SOA or OECD observed termination rates. The example demonstrates why claimant mortality must enter duration and reserves even when the insured benefit is described primarily as morbidity coverage.
The SOA's 2000–2016 experience databases explicitly separate incidence and termination records and define aggregate exposure. Their methodology also imposes data-privacy and granularity constraints; published aggregates cannot resolve every mixture of cohort, calendar year, benefit design, severity, and claim duration. Study definitions govern whether close-in-time interruptions merge into one claim. A recovery estimated from one claim-episode rule cannot be imported unchanged into a model with a different recurrence definition.
Elimination periods, truncation, and remaining duration
An elimination period e should be implemented as a policy rule. If care is continuous through e, payments from a flat cash benefit b begin only thereafter. With constant care-exit total hazard k and discount force δ, the discounted value at care entry of benefits from e to e+L, where L is a maximum subsequent payment period, is
This expression assumes continuous eligibility, a payment horizon measured after elimination, no return to eligibility within the same episode, and a flat nominal benefit. A calendar-day deductible, an accumulation-based elimination period, or a dollar benefit pool produces a different function. In practice, the observed date of a payable claim and the inferred date of disability may differ materially. Left-truncated spells (observed after onset), right-censored spells, death, and exhaustion must be distinguished during estimation.
A maximum payment period is not a biological termination hazard. It is a contract-limited endpoint for payment, potentially leaving the person in care. A duration model must also separate lapse while healthy from claim termination after onset. The 2026 NAIC LTC insurance market briefing describes how historical underestimate of incidence and claim duration, together with lower-than-expected lapses, contributed to premium inadequacy in older blocks. That history shows the consequence of compounding assumption errors, although it does not establish a numerical rate applicable to another portfolio.
Transitions to and from Care
Continuous-time multistate representation
Let take values H, C, and D. Under a time-homogeneous Markov illustration, the intensity generator is
where α is H → C, β is H → D, ρ is C → H, and γ is C → D. Units are events per person-year. Rows sum to zero, non-diagonal entries are nonnegative, and D is absorbing. The transition probability matrix is . If the insured initially occupies H, the first row gives probabilities of H, C, and D after t. Inhomogeneous age-and-time-dependent generators instead require solving the forward equation , initialized at , or a time-ordered product over short intervals.
Recovery permits recurrent care spells. As a result, is the probability of occupying C at time t after possibly several entries, not the probability of first entry by t. This is one reason a state-occupation integral is appropriate for recurrent-claim benefit cost, while a first-entry cumulative-incidence function is appropriate for first-claim incidence. Confusing the two can misstate both claim counts and claim-year exposure.
The landmark actuarial paper by Rickayzen and Walsh (2002) examines multi-state disability projections using prevalence, transitions, and trends. It illustrates why a multi-state framework can be attached to aggregate forecasts, while its UK-specific parameter inputs and demographic context must not be transplanted into other countries or insurance pools without new evidence.
Observational estimation
For observed transitions from i to j in a small interval, a crude cause-specific intensity estimate is , where is the number of transitions and is person-time genuinely exposed in i. Age-specific smoothing, piecewise constant intensities, proportional hazards, or flexible duration surfaces can improve estimation when supported by sufficient events. Under incomplete observation the data likelihood must reflect interval censoring and competing destinations. The Aalen–Johansen estimator constructs transition probabilities from observed cause-specific event increments, conditional on its assumptions; the survival-model tutorial explains its relationship to Kaplan–Meier estimates and cumulative-incidence curves.
Exposure validation is as important as fitting a regression. Each contract must be assigned entry, coverage, lapse, age, and claim dates, and repeated claim episodes must have a consistent definition. If a claimant switches providers while eligibility persists, administrative records may indicate two bills but a single claim spell. Death notices, lapses and benefit exhaustions may be systematically delayed. Claim status not reported by an insured can create ascertainment effects that differ across contracts and calendar periods.
The ASOP 23 data-quality guidance calls for appropriate selection, review, and disclosure of limitations in actuarial data. Record-level quality checks should therefore address impossible dates, overlaps between mutually exclusive states, missing termination reasons, inconsistent care-setting codes, inflation-base anomalies, and cohort-specific changes in underwriting. At the aggregate level, observed-to-expected statistics must carry denominators, credibility, and model-version metadata. An actual-to-expected ratio without exposure, sampling uncertainty, or segmentation can give a false impression of model accuracy.
Markov, semi-Markov, and mixture alternatives
The Markov assumption states that future transition probabilities depend on the current state and time variables, not the elapsed spell history after those variables are conditioned upon. In LTC this can be restrictive: newly disabled persons may have a higher recovery rate than persons continuously disabled for years; the risk of institutional transfer may rise with duration; cognitive decline can change the probability of recovery. A semi-Markov specification replaces a constant ρ or γ with duration functions ρ(d) and γ(d). A further alternative expands the state space into duration bands, approximating duration dependence through a larger Markov generator.
Covariate heterogeneity generates an additional phenomenon: even if each person's hazard is constant within a risk class, the aggregate duration hazard can decrease as high-exit-rate cases leave early. Fitting one aggregate exponential to a mixed cohort will therefore conflate true duration dependence with unobserved selection. Frailty models or stratified curves provide alternative descriptions, but no dataset in this monograph establishes the need for one specific frailty distribution.
Time-inhomogeneity has two forms. Aging changes each individual's hazard because attained age increases; period changes affect all current ages during a calendar year through medical practice, care capacity, disability prevalence, mortality improvement, and claim administration. An issue-year cohort can further reflect different underwriting or benefit terms. An apparent improvement in calendar incidence may actually be a changing mixture of issue ages, underwriting eras, and lapse selection. Estimation should cross-tabulate at minimum age, observation year, issue cohort, benefit design, and available severity fields, then test whether a cohort-period-age separation is empirically identifiable.
Constraints, calibration, and diagnostic governance
The fitted generator must generate valid probabilities: rows of P sum to one, entries lie between zero and one up to numerical tolerance, and state occupation evolves coherently with observed exits. A model fitted independently to each transition can violate substantive consistency if population deaths implied by its healthy and care states conflict with independent mortality exposure. External life tables from the Human Mortality Database provide harmonized population mortality series, but the database expressly limits its coverage to populations with sufficiently complete records. Insured lives have selected mortality and cannot simply inherit general-population rates.
Calibration diagnostics should include state-specific observed-to-expected transitions; early versus late spell termination; first versus recurrent care; care setting; age, sex, and cohort stratification; and claim payment per exposed care-year. Holdout tests should use genuine later time periods or independent blocks when possible, with safeguards against leakage from reprocessed claims. The ASOP 56 modeling standard links model specification, assumptions, validation, and communications to intended use and financial significance. A sophisticated estimator does not compensate for an inadequately specified insured event or a material coverage mismatch.
Benefit Utilization
From eligible care to paid benefits
A transition into C establishes potential eligibility, not automatically an immediate insurer payment. Denote a claim's state at time t by X(t), service amount by U(t), eligible expense by C(t), remaining benefit pool by A(t), and contract rules by θ. A general nominal insurer outflow can be expressed
where g may be fixed indemnity, reimbursement capped at a daily/monthly maximum, an inflation-indexed indemnity, or a percentage of approved costs. This is an accounting representation of common product features. It does not assert that any one contract uses them all.
Under an indemnity design, a qualified eligible day may generate a stipulated payment regardless of the claimant's documented cost, subject to the agreement. Under reimbursement, a typical stylized daily rule is with precise contract accounting for a pool whose balance changes. A higher unit cost can increase reimbursement up to the cap, while a fixed indemnity is affected primarily through contractual indexing and eligibility, not actual provider prices. Utilization intensity is therefore product-specific. A fixed payout and an actual-expense pool cannot share one universal "severity trend."
The NAIC's long-term care insurance background identifies home care, assisted living, respite, hospice, and other settings within historical product choices. Those categories differ materially in service hours and unit costs. Care utilization should be modeled as days or hours of service times a rate per unit, conditional on eligibility and availability. When claimant location and severity are jointly observed, a setting-specific transition and tariff model gives a direct route to projected claims. Where the source only reports total paid claims, utilization and unit-cost effects are only partially identifiable.
Two different sources of inflation
Let denote the expected number of care units in setting j while eligible and the nominal price per unit. A useful expected cost decomposition is , with contractual adjustments applied to the resulting claim. Changes in units arise from care intensity, setting mix, technology, policy benefits, and available family care; changes in prices arise from wages, provider contracts, public funding, and general price conditions. The model's "care-cost inflation" assumption should name the component being trended and the benefit structure to which it applies.
For illustration, if an annual nominal payment in C starts at $58,400 and grows at 3% continuously, its value at t is $58{,}400e^{0.03t}$ before any cap. With a matching 3% continuously compounded discount force, the nominal growth and discount exactly offset in the expected-value integrand. This coincidence is a mathematical property of the chosen scenario, not an inflation hedge inherent in LTC insurance. If a fixed policy maximum prevents payment growth, care-cost inflation might instead increase uncovered policyholder costs or accelerate reaching a benefit pool. The distinction changes the direction and size of insurer exposure.
Waiting, service switching, and benefit exhaustion
A calendar-day elimination period, a service-day elimination period, and a monetary deductible have different stochastic cash-flow consequences. For a calendar waiting period, benefit commencement can be tied to elapsed qualifying disability. For a service-day waiting period, initiation depends on utilization and may arrive after recovery and recurrence events. A benefit pool must be represented as a path-dependent balance with state , because an identical medical state and tariff can produce different cash flows after different historical claims. Exhaustion can stop payments without terminating care, a crucial point for validating experience datasets.
A claimant may transition from home to a facility, return home, and experience recurring spells. Treating a transfer as claim termination and a new incidence event can inflate counts while leaving total claimant-days almost unchanged. The choice has direct implications for incidence pricing, duration calibration, elimination periods, and recurrence limits. Therefore each data dictionary must state whether the observational unit is a person, contract, episode, claim record, service line, or paid day.
Public systems, care supply, and substitution
A country's formal LTC recipients depend on public eligibility, family support, means-testing, provider availability, and social preferences as well as health. Brown and Finkelstein's (2011) US long-term care financing review explains the substantial interaction between private coverage and Medicaid. Internationally, the OECD's 2025 LTC spending account reports that about four of five LTC spending dollars across OECD countries arise from public sources, subject to country-level definitions and recording completeness. These institutional differences influence observable insured utilization and the scope of residual private liability.
The WHO's integrated-care recommendations reinforce that patterns of formal service need not be invariant under preventive and community programs. An actuarial analysis may test alternative scenarios in which home services delay institutional entry or improve quality, but causal effects require independent empirical designs. A transition matrix calibrated to the past cannot by itself establish that an intervention reduces the future probability of severe disability. Scientific inference must preserve the distinction between predictive association and causal effect.
Demographic and Trend Risk
Aging, prevalence and the meaning of LTC growth
Projected national care expenditure can grow when the older population increases, even if age-specific disability rates fall. It can also grow if formal services substitute for unpaid family care, unit wages rise, or more eligible individuals receive support. OECD (2025) reports formal LTC receipt among older people averaging 12% across 31 countries in 2023. The same publication reports LTC expenditure averaging 1.8% of GDP in 2023 across its comparison, with major national differences. These are valuable institutional indicators. Neither is an estimate of an insurance policy's 15-year incidence, duration, or monthly claim benefit.
Demographic projections should identify at least attained-age exposure, mortality improvement, disability incidence, recovery, claimant mortality, severity, and formal-care take-up. Healthy life expectancy measures are informative, but a rise in survival need not imply equal additional years in severe disability. The Rickayzen–Walsh (2002) disability and LTC projections explicitly consider how changes in disability patterns alter future demand. The analysis is especially relevant to distinguishing absolute beneficiary counts from care probabilities conditional on age. Forecasts are also affected by whether care transitions are measured in chronological age, time to death, or time since onset.
In addition, care is labor-intensive. The OECD long-term care workforce comparison reports a mean of five LTC workers per 100 people aged 65 and over across 31 reporting countries in 2023. Workforce definitions differ and the statistic is a headcount, not a productivity-adjusted measure of available care hours. Nonetheless, it motivates a supply-side scenario: a shortage of qualified caregivers can increase unit costs, constrain observed utilization, shift care between home and facilities, or defer claims; its effect on insurer liabilities varies by policy design and cannot be inferred from workforce headcounts alone.
Morbidity trend identification
The phrase "morbidity improvement" needs a target. It may mean fewer first care entries at a fixed age, later age of first dependency, fewer severe cases, shorter claims because of recovery, reduced care utilization conditional on eligibility, or improvement in reported functional capacity. These are different estimands. Their effects on an insured portfolio also depend on longevity: delayed care onset in a longer-lived insured population may still support extensive lifetime exposure. A source-derived incidence improvement assumption must therefore state which transition rate or benefit functional is being changed.
The SOA's calendar-year incidence research sets out distinct trend-estimation methods. The 2025 SOA–LIMRA–NAIC industry study announcement describes collection of evidence intended to include mortality, persistency, incidence, and termination across 2000–2023. The corresponding data request and scope statement excludes hybrid and combination products. These are a study launch and data request, not published new 2000–2023 estimates; using them as completed morbidity tables would misrepresent the available evidence.
A credible trend assessment asks whether the observation window includes changes in policy generation, age distribution, claim triggers, reporting lags, benefit availability, care setting, and economic incentives. It should provide uncertainty around trends rather than fit one smooth improvement factor to every hazard. The Human Mortality Database's mortality rates and life tables can supply population benchmarks for mortality scenarios, with adjustment for insured selection and local data. Mortality trend extrapolation and disability incidence trend extrapolation need not share one numerical slope.
Joint stress test of benefit cost
Return to the fifteen-year hypothetical model. Increase H → C from 0.04 to 0.05 per year, keep H → D at 0.02, reduce care recovery from 0.12 to 0.102 and care mortality from 0.18 to 0.144, and let nominal benefits inflate continuously at 3% per year while the nominal discount force remains 3%. The stressed generator is
With , its benefit EPV is 58{,}400\int_0^{15}p_C^*(t)dt=$94,160.88, versus $53,532.77 on the base basis, an increase of 75.89%. The stressed discounted premium-paying time is 8.38138757 years, giving a mathematical net continuous premium of $11,234.52 if repriced on the stressed assumptions. This is a combined deterministic sensitivity, not a probability-weighted economic forecast, empirical calibrated margin, stochastic capital estimate, or permitted regulatory rate increase. Its purpose is to expose how incidence, survival in claim, and nominal cost growth can reinforce each other.
A fully informative stress analysis varies the assumptions separately as well as jointly. Analysts should record baseline and stress exposure units, matrix construction, horizon, benefit inflation convention, discount basis, lapse and mortality changes, and whether the shock applies to current claims, new entrants, or both. If a particular policy has a dollar cap, applying cost growth to unlimited insurer benefits is inconsistent with the contract. If the policy offers benefits conditioned on eligible care days, a utilization shock may dominate a pure tariff shock. The sensitivity tells decision-makers where to obtain better data; it cannot establish the next decade's true care costs.
Model failure modes and scientific limits
Important failure modes include mortality adjusted twice after already being embedded in a state transition; recurrence counted as independent first incidence; the same claim counted as ended on provider transfer and then reopened; selection bias from observing only paid care; treating benefit exhaustion as medical recovery; inferring care duration from paid days after eliminating waiting days; and assuming medical inflation equals contracted LTC unit-cost inflation. Other risks include underestimating claims in policies with high guaranteed benefits, misstating the discount curve for long payment tails, neglecting policyholder choices under rate increases, and applying general-population health trends to underwritten insureds without adjustment.
The NAIC's 2026 market overview highlights historical US premium and reserve difficulties when claim rates, claim durations, and lapse experience departed from original assumptions. Such experience warrants disciplined monitoring, without justifying a mechanically conservative assumption everywhere. An appropriate calibration is evidence-based, purpose-specific, and periodically re-examined.
The ASOP 56 modeling guidance supports internal and external validation proportionate to use. Effective checks comprise mathematically valid transition probabilities, exact reconciliation of experience definitions, numerical integration agreement under independent methods, sensitivity surfaces, source-specific issue logs, and transparent limitations. Documentation should allow another qualified actuary to reproduce the key equations and explain why the data and benefit rules support their use.
Research agenda and limits of generalization
Future empirical work can link disability survey panels, claims records, mortality registries, and provider utilization data with appropriate privacy protections. It can then estimate whether duration-dependent recovery improves out-of-sample forecasts over a Markov baseline, assess care setting substitution conditional on severity, and characterize the distribution of claim costs rather than only the mean. Quasi-experimental policy variation could help identify service effects if strong design assumptions and institutional conditions hold. Such work requires data access, protocols, and jurisdiction-specific validation beyond a documentary monograph.
This synthesis is based on checked official webpages, accessible reports and institutional descriptions, and the publisher abstracts of selected journal works. It does not reconstruct individual SOA claim tables, audit regulated insurer filings, or certify a specific insurer's reserves. The numerical case is synthetic, and its inputs were chosen to demonstrate actuarial mechanics. Generalization to any country, product generation, provider network, or covered life cohort requires empirical calibration and a current legal and reporting basis.
Conclusion
Long-term care actuarial modeling begins with a precise account of the event that generates eligibility and the rules that generate payment. A clinically meaningful state, a qualifying contractual state, an observed paid claim, and a public-service recipient may describe different facets of the same person's experience. Keeping them distinct makes incidence, duration, utilization, and insurance valuation interpretable.
The derived equations establish several practical conclusions. First-entry care incidence requires competing-risk treatment of pre-care death; recurrent care requires state occupation rather than first-event probability alone. Expected care duration depends jointly on recovery and claimant mortality, and care-setting changes must be separated from actual termination. Product-specific elimination periods, maximums, benefit pools, and inflation terms determine how care exposure becomes insurer outflow. Premium and reserve projections additionally require premium-paying exposure, consistent discounting, and a documented measurement basis. Demographic aging, disability trajectories, workforce supply, public financing, and the selection of insured policyholders produce distinct sources of projection uncertainty.
The three-state worked example demonstrates the mathematical importance of these distinctions without pretending to estimate real incidence or care costs. Its base-case benefit value and joint stressed value are fully determined by stated generators, discounting, utilization, and horizon. The calculation's transparency is part of its scientific value: every change in financial result can be traced to a named assumption or contractual rule.
The empirical conclusion is similarly specific. High-quality LTC models require reliable longitudinal exposure, correctly classified cause-specific transitions, appropriately limited external comparisons, and repeated source-to-claim and numerical checks. Existing international standards and datasets provide essential guidance and benchmarks, while inference remains tied to their coverage and definitions. Research, pricing, reserving, and public policy should use coherent but purpose-specific versions of these models, preserving uncertainty rather than transferring a rate without its original population, period, and benefit context.
References
Actuarial Standards Board (2022). Actuarial Standard of Practice No. 18: Long-Term Care. Revised March 2022; effective September 1, 2022. Source.
Actuarial Standards Board (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Source.
Actuarial Standards Board (2016). Actuarial Standard of Practice No. 23: Data Quality. December 2016. Source.
Society of Actuaries Research Institute (2020). 2000–2016 Long-Term Care Aggregated Databases: report and databases. Source.
Society of Actuaries (2018). Researching Long-Term Care Insurance Incidence Rates Over Time. Source.
SOA Research Institute, LIMRA, and NAIC (2025). Industry Leaders Launch New Long-Term Care Insurance Industry Experience Study. August 4. Source.
Society of Actuaries Research Institute and LIMRA (2025). 2000–2023 Long-Term Care Data Request. Source.
Organisation for Economic Co-operation and Development (2025). Access to Long-Term Care. Health at a Glance 2025. Source.
Organisation for Economic Co-operation and Development (2025). Long-Term Care Spending and Unit Costs. Health at a Glance 2025. Source.
Organisation for Economic Co-operation and Development (2025). Long-Term Care Workers. Health at a Glance 2025. Source.
Human Mortality Database (accessed November 14, 2026). Project Overview and Methodological Description. Source.
Rickayzen, B. D., and D. E. P. Walsh (2002). A Multi-State Model of Disability for the United Kingdom: Implications for Future Need for Long-Term Care for the Elderly. British Actuarial Journal 8(2): 341–393. https://doi.org/10.1017/S1357321700003755. Source.
Therneau, T., C. Crowson, and E. Atkinson (2024). Competing Risks and Multistate Models. Survival Vignettes, May 2, 2024. Source.
Brown, J. R., and A. Finkelstein (2007). Why Is the Market for Long-Term Care Insurance So Small? Journal of Public Economics 91(10): 1967–1991. https://doi.org/10.1016/j.jpubeco.2007.02.010. Source.
Brown, J. R., and A. Finkelstein (2011). Insuring Long-Term Care in the United States. Journal of Economic Perspectives 25(4): 119–142. https://doi.org/10.1257/jep.25.4.119. Source.
United States Congress (current online codification, accessed November 14, 2026). Internal Revenue Code, 26 U.S.C. § 7702B: Treatment of Qualified Long-Term Care Insurance. Source.
Centers for Medicare & Medicaid Services (accessed November 14, 2026). Medicare.gov: Long-Term Care. Source.
National Association of Insurance Commissioners (2026). Long-Term Care Insurance, Insurance Topics. Updated July 23, 2026. Source.
National Association of Insurance Commissioners (accessed November 14, 2026). Model Laws: Long-Term Care Insurance Model Act No. 640 and Model Regulation No. 641. Source.
IFRS Foundation (accessed November 14, 2026). IFRS 17 Insurance Contracts: Standard Overview. Source.
International Actuarial Association (accessed November 14, 2026). International Standards of Actuarial Practice: ISAP 1 and ISAP 4. Source.
World Health Organization (2017). Integrated Care for Older People: Guidelines on Community-Level Interventions to Manage Declines in Intrinsic Capacity. Source.
Brown, J. R., and A. Finkelstein (2009). The Private Market for Long-Term Care Insurance in the U.S.: A Review of the Evidence. Journal of Risk and Insurance 76(1): 5–29. https://doi.org/10.1111/j.1539-6975.2009.01286.x. Source.
Society of Actuaries Research Institute (accessed November 14, 2026). Long-Term Care Experience Studies: Research Topic Catalog. Source.