Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Long-Term Care and Disability Insurance Experience Data Harmonization

Abstract. Long-term care (LTC) and disability insurance experience studies depend on distinct legal, clinical, functional, occupational, and actuarial definitions of an observable event. Their datasets describe different risk sets: insured lives eligible to incur a claim, claimants exposed to recovery or death, beneficiaries eligible for payment, and persons receiving formal or informal care. The same word, incidence, can refer to onset of functional limitation, onset of work incapacity, submission or approval of a claim, satisfaction of an elimination period, or start of reimbursed services. Their event counts cannot be combined without reconstructing the associated denominators, time origins, contractual predicates, and observation mechanisms. This monograph investigates how insurer experience, health and care assessments, public benefit records, and statistical surveys can be reconciled while preserving the information required for actuarial inference. It analyzes the Society of Actuaries (SOA) long-term care intercompany experience collections, its 2015–2022 group long-term disability incidence study and 2015–2024 claim termination study, alongside WHO functioning classifications, administrative care assessments, health-data standards, legal eligibility provisions, and established statistical literature. The analytical contribution is a reproducible account of comparability conditions, event-state crosswalks, exposure reconstruction, selection and missing-data diagnostics, competing-risk estimation, and validation. Worked examples show how pooled crude rates reverse under standardization and how unobserved outcomes create identification bounds. Harmonization is treated as a claim-specific scientific operation: data sources become comparable for a defined estimand only after their event definitions, time-at-risk, observation rules, and populations align or the remaining differences are explicitly modeled. The work offers a substantive synthesis and practical analytical recommendations without announcing a new theory, doctrine, or proprietary classification system. It uses publicly documented sources and provides no insurer-level reanalysis of restricted microdata.

Keywords: Long-term care insurance; disability insurance; actuarial experience; data harmonization; incidence; claim termination; exposure; functional status; activities of daily living; return to work; competing risks; missing data; data provenance; selection bias; standardization; longitudinal data; risk set; care utilization; clinical and insurance data; cross-jurisdiction comparability.

Introduction

Long-duration morbidity risk occupies several observational systems at once. A person may develop a health condition, experience limitations in ordinary activities, cease working, require assistance from another person, enter an insured benefit period, receive public support, and subsequently recover some capacity while continuing to need other forms of care. Each transition can leave a record with a different date, identifier, institutional purpose, and definition. The actuarial task begins by identifying which transitions generate insured liabilities and which observations legitimately estimate their frequency, duration, or monetary consequence.

The World Health Organization's International Classification of Functioning, Disability and Health (WHO, 2001) explicitly links functioning to bodily, activity, participation, and environmental dimensions. That breadth is important for understanding the underlying phenomenon. Insurance policies and statutory programs apply additional eligibility conditions that determine whether observed needs become payable claims. Accordingly, clinical disability prevalence, receipt of a long-term care service, private disability claim incidence, and eligibility for a social insurance benefit answer separate empirical questions even when they concern overlapping individuals.

The Society of Actuaries' 2000–2011 LTC aggregated databases distinguish claim incidence, termination, and utilization rather than offering one universal morbidity rate. By contrast, the SOA/LIMRA 2015–2022 group LTD incidence study concerns claims arising in covered group disability business, while the 2015–2024 LTD termination study describes recovery and death experience while claimants are exposed to termination. These studies provide unusually clear public demonstrations of the need to distinguish a life-year incidence denominator from a claimant-month termination denominator. Their public summaries also illustrate why data purchased from one experience-study vendor should not be assumed to have the same observation specification as other products from that institution.

The investigation uses documentary analysis of primary institutional definitions, transparent mathematical treatment of rates and state transitions, constructed numerical examples, and published methodological research. Empirical observations are attributed to their actual source, period, and study population. The numerical examples are original illustrations with explicit hypothetical inputs. No insurer-specific confidential table, complete paid study dataset, external peer-review process, or original patient record is represented as examined. The comparison aims to determine what can be made commensurable, what must remain differentiated, and how uncertainty should be expressed when harmonization requires assumptions.

Conceptual Foundations and Historical Development

Distinct insured objects and events

Long-term care insurance addresses financial consequences of needing assistance or supervision, often through contractual triggers based on activities of daily living (ADLs) or severe cognitive impairment, and may finance home or institutional care. Disability income insurance primarily responds to a contractual loss of occupational earnings capacity and may involve own-occupation or any-occupation rules, waiting periods, income offsets, and return-to-work provisions. Public disability benefits use statutory tests. The NAIC long-term care insurance overview describes evolving American LTC product design, including multiple service alternatives. A health diagnosis can contribute to either outcome while remaining insufficient to identify the insured claim itself.

The five observations that should be kept separate are: functional limitation, qualifying event under a specified contract, administrative claim, benefit payment, and use of a care service. In formal terms, for individual i at time t, the observed health state H_i(t), contractual eligibility E_i(t;v), claim administration status A_i(t), payment P_i(t), and service use U_i(t) are distinct variables; v indexes the contract or legal-rule version. A person can have H=impaired and E=0, E=1 and A=pending, A=approved and P=0 because a waiting period applies, or P>0 while U is underreported in a separate administrative record. Collapsing these into one binary disability flag generates structural measurement error.

The US Social Security Act's disability insurance benefit provisions, Section 223 of Title II, define disability and entitlement through an employment-activity standard, insured status, duration, application, and waiting-period rules. The criteria differ from private policy wording and from functional assessment. The German Federal Ministry of Health describes statutory long-term care insurance under Germany's long-term care framework, including care grades based on independence and abilities. The Dutch government's Long-term Care Act guidance describes a CIZ entitlement assessment grounded in need for constant care or supervision. A comparative database therefore records a jurisdiction, a product class, and a versioned entitlement rule for each insured event.

The evolution of intercompany experience studies

The intercompany actuarial tradition uses pooled submissions to make sparse or heterogeneous claim experience informative while retaining statistical and product-specific caveats. In its 2000–2011 LTC study database description, the SOA states that information from 22 insurers was assembled into incidence, termination, and utilization databases, following claims to termination or the 2011 cutoff. The related LTC basic table development constructed three multiplicative models using this underlying experience; the SOA describes the steps of data collection and cleansing, database development, and basic experience tables. This provenance matters because a published model factor is a processed estimate with assumptions, not a raw observed count.

In October 2025 the SOA/LIMRA released the group LTD incidence study for January 2015 through June 2022. The study's incidence specification reports approximately 294 million covered life-years and 1.2 million claims from 19 contributing companies. Its public report restricts the study to fully insured group LTD business and excludes administrative-services-only contracts, individual forms, and several other categories. In September 2026 the group LTD termination study reported more than 43 million life-months of claimant exposure, approximately 937,000 recoveries, and more than 181,000 deaths for 2015–2024, contributed by 18 companies. The approximate figures and different contributor counts should remain linked to their respective denominators and should not be blended into a synthetic observed market portfolio.

A further SOA/LIMRA/NAIC new LTC industry study announcement in August 2025 outlined planned examination of mortality, persistency, claim incidence, and claim termination. The launch was a program announcement, not publication of the contemplated final experience tables. This distinction prevents the use of anticipated datasets as if they already supplied validated rates.

Population statistics and contract experience

Population surveys have a different target. Eurostat's disability-data measurement guidance distinguishes impairments, activity limitations, participation restrictions, ADLs, and the Global Activity Limitation Indicator (GALI). The CDC's functioning and disability data guidance describes the Washington Group Short Set's functional questions. These instruments measure dimensions of participation risk or functional difficulty in surveyed populations; neither observes the full set of underwriting, contractual eligibility, premium payment, and claim filing requirements for a particular insurance block.

The OECD's 2024 long-term care affordability study discusses older people's care needs, income, support and payment burdens across countries. It is informative about economic exposure to care costs, but it is not an incidence study of insured LTC claims. The ILO's 2024–2026 social protection assessment reports disability-benefit coverage, which is an administrative and social-protection outcome. A harmonized account should preserve each source's target population, numerator, denominator, coverage criterion and survey or administrative selection process.

The scientific question of comparability

The relevant question is whether two observed rates estimate the same statistical quantity for the same target population and horizon. A compact definition is useful. Let D denote a dataset, T its target population, R(t) its risk set, N_e its count of events matching event definition e, and E its accumulated exposure. The crude rate N_e/E is interpretable only conditional on the definitions of e, R and E. Harmonization for estimand θ requires either a shared definition or an estimable mapping from each source's observation process into θ. Standardized file labels, code systems, or field names are necessary for machine handling but do not establish that mapping.

Two observations can legitimately yield different event rates because the underlying processes differ. The challenge is to determine when differences result from actual morbidity or contract outcomes and when they arise from thresholds, data capture, selection, or changing eligibility. The mathematical distinction is then translated into an evidence requirement: each comparison must state its target, event, origin, exposure unit, censoring rule, selection process, and uncertainty treatment. Without those properties a numerical comparison may remain a description of the datasets, but it cannot sustain an inferential claim about relative risk.

Coding and Data Provenance

Event ontology and source record boundaries

At least six date-bearing events may arise in one episode: functional onset, cessation of work, claim notice, satisfaction of the elimination period, claim approval, and payment start. A seventh event can be service commencement. Their ordering may vary, some dates may be retrospectively amended, and the analyst may observe only a subset. A genuine harmonization record should carry distinct fields for these events with their sources and revision history. It should also distinguish denial, withdrawal, administrative closure, clinical recovery, contractual exhaustion, aging out, death, and recurrence as separate operational outcomes.

The WHO ICF classification framework organizes domains of health-related functioning with environmental context. It supplies a foundation for describing limitation without supplying the actuarial contract trigger by itself. In LTC contracts a need for assistance with bathing or transferring may be relevant; in disability income coverage a documented inability to perform occupational duties may be relevant. A crosswalk from a functional code to an eligibility code is therefore conditional on the specific policy, rule version, assessment practice and effective date. A one-to-one lookup table that ignores these conditions introduces false equivalence.

The minimum data dictionary

For a source dataset to support reproducible inference, the following fields have explicit semantic definitions rather than merely matching column names:

Domain Required record or metadata Analytical reason
Provenance source organization, source file/version, refresh date, system, legal access Reconstruct origin and permissible reuse
Population jurisdiction, product form, group/individual status, eligibility and participation dates Define target and selection
Person and contract pseudonymous person, contract, coverage, and episode keys Prevent duplicates and preserve multiple episodes
Clinical/function instrument and version, diagnosis or functional code, assessor, observation date Separate health observation from entitlement
Trigger applicable policy wording or statutory rule, threshold and elimination period Identify insured event
Chronology onset, notice, approval, payment and exit dates, precision and correction history Preserve time origin and delay
Risk exposure entry and exit from at-risk state, earned exposure fraction, left truncation Specify denominator
Claim exits recovery, death, exhaustion, cancellation, other, unknown and reopening Model competing outcomes
Benefits eligible amount, paid amount, currency, cost basis, payment frequency, offsets Interpret utilization and value
Missingness field-level status, reason, timing, record accessibility Diagnose selection and identification

The table is an analytical documentation checklist, not a newly named data standard. Clinical exchange formats can support parts of it. HL7 FHIR R5 defines the Observation resource, including an effective time for when the measurement is asserted to hold, and the QuestionnaireResponse resource, which preserves structured questionnaire items and responses. Their design helps represent dated functional assessments. Insurance-specific eligibility and contractual episode fields may require separate controlled structures rather than pretending that a clinical Observation is a paid insurance claim.

Observation periods, time stamps, and event lineage

Exposure cannot be derived from all calendar days between a person's earliest and latest medical record. The OMOP Common Data Model 5.4 documentation carefully explains that an observation period is a span in which clinically relevant events are expected to be captured, and that absence of records outside such periods cannot be interpreted as absence of events. An insurance risk period is a further, contract-dependent interval. A member may be clinically observable while no longer covered, or covered while provider records are unavailable. A harmonized source should therefore distinguish clinical observation coverage, insurance eligibility, claim-active exposure, and service-claims capture.

A time stamp must retain local time, UTC offset where appropriate, date precision, occurrence versus processing status, and retrospective amendments. ISO 8601-1:2019 supplies information-interchange representations for dates and times; adopting a uniform representation improves parsing without resolving the distinction between an event's onset date and the date it was entered into a system. For long-term trends the dataset should retain both, because claims reported after a cutoff can change apparent incidence for an earlier year and because a database refresh can restate prior experience.

A reproducible extraction records the steps from the original files to derived exposures and events. The W3C's PROV-O provenance recommendation (2013) provides vocabulary for entities, activities, and agents; it can describe which transformation generated a particular derived table and from which input. A provenance record should include software and rule versions, transformation order, rejected records, duplicate matching decisions, input and output checksums, and known limitations. Such information is required for replicating a rate, especially when a code mapping is amended between data vintages.

Crosswalks across clinical, administrative, and insurance systems

Crosswalks should preserve the direction of inference. A hospital diagnosis can support an assessment of likely assistance needs; an ADL item supports a measured limitation; a policy trigger converts specified evidence into an eligibility decision; a claim administrator applies coverage provisions; and a payment record establishes cash flow. Each inference has a different conditioning event. A crosswalk must declare whether its output is an observed fact, a rule-derived predicate, or a model-estimated probability.

The distinctions are especially clear in American nursing-facility records. CMS maintains MDS 3.0 technical specifications and the Resident Assessment Instrument manual. These sources document how particular facility assessments are structured and versioned. An MDS assessment may capture performance in a facility, but it does not by itself demonstrate eligibility under an unrelated disability-income policy; conversely, a group disability claimant may never have a facility assessment. A geographic or temporal harmonization requiring MDS codes must record the item-set version and any code changes that affect longitudinal comparability.

Consider a source term dependent. It can mean assessed ADL assistance, a person receiving care, an individual entitled to payment, or a covered spouse or child in an insurance record. These are four distinct data concepts. A safe coding method retains the original variable name, its full definition and original source value, and creates separately named derived variables for each target meaning. Ambiguous terms are coded unknown until documentary evidence resolves them; assumptions used for partial mapping are stated and tested.

Provenance and disclosure controls

Privacy and quality are connected. Removing direct identifiers can permit lawful comparative research, but it can also obstruct detection of repeat claims and movement between providers. The analytical response is a governed pseudonymization and linkage protocol with documented false-match and missed-match risk. In the European Union, GDPR Article 9 treats health information as a special category of personal data subject to specified lawful conditions and safeguards. A study's legitimate interest in complete episodes cannot replace an applicable legal basis, purpose specification, access governance, or disclosure controls.

Source quality should be stratified by what was directly inspected. Public study summaries can establish stated sample size, population exclusions, and headline results; restricted insurer-level databases and paid visualization products require separate access and checking. In this monograph public institutional pages and accessible source descriptions were directly reviewed; no numerical coefficient from an inaccessible paid database is reproduced as an independently verified estimate. Reproducibility at the source-summary level is distinguished from replicating the underlying study analysis.

Period and Cohort Comparability

Risk sets and denominators

For incidence, define Y_i(t)=1 when an insured is eligible, covered, under observation under the study's rules, and has not yet experienced the first event under study; otherwise Y_i(t)=0. Total exposure over interval [a,b) is E=Σ_i∫_a^b Y_i(t)dt, in life-years when t is measured in years. The event rate is I=N/E, with N counting the study-defined eligible incidents. Where the risk is recurrent, the first-event definition must be replaced with an episode-reentry convention and associated at-risk intervals. This explicitly differs from prevalence, which measures the proportion in a state at a date or over a period.

Once a person becomes a claim participant, a termination analysis uses an active-claim risk set Z_i(u), where u is time from the contractually defined claim-duration origin. Its denominator is accumulated claim-time, such as claimant-months; its numerator can be recovery, death, or a specified other exit. The 2026 SOA/LIMRA termination experience is publicly presented using life-months exposed and separate counts for recoveries and deaths. Treating these exposures as the same 294 million covered life-years used for 2015–2022 LTD incidence would be dimensionally and substantively incorrect. Converting months to years by dividing by twelve changes units, not the underlying risk set.

LTC utilization introduces a third denominator. A ratio of paid service days to eligible benefit days, or paid amount to available daily benefit, depends on eligibility and claim status. The SOA's three-part LTC experience architecture reflects this distinction. It is therefore inappropriate to call a quotient a claim rate without specifying whether it is per covered person-year, per claimant-year, per eligible service-day, per paid day, or per monetary unit of exposure.

Calendar time, cohort time, and duration

Experience has multiple clocks: calendar year, policy issue year, coverage duration, claimant age, attained age, claim duration, and calendar year of claims adjudication. Incidence may depend on attained age and elapsed policy duration; continuation may depend on elapsed claim duration and cause; a legal reform may apply only to policies issued after a specified date. A harmonization that groups solely by calendar year can conceal differences in duration, morbidity mix and benefits design.

Claims reported in calendar year 2025 may have onset dates in 2024; calendar-year reporting can therefore introduce development lags. Observational follow-up ending on December 31 does not imply a claim terminated on that date. Right censoring at the study cutoff preserves the exposure up to the cutoff without creating a spurious exit. Similarly, a person who entered a claims database before the observation window is left truncated for the analysis of claim-duration rates: the person contributes only conditional on surviving in the relevant state until entry. Claims selected because they remain open at the study start may be systematically longer than new claims, so survivor and incident cohorts require different analyses.

Partial exposure is computed from actual eligible time. For example, an individual covered for ninety-one days during a 365-day reference year has an illustrative exposure of 91/365 life-years, subject to study conventions on inclusive dates, leap years, terminations and reporting. Counting that person as a full year of exposure biases the rate downward when the numerator includes their events. When exposure is available only in rounded months, the allowable approximation and its effect on estimates should be documented.

Multi-state processes and competing exits

A useful statistical description treats employment and care needs as multi-state processes with specified transitions. Candidate states may include employed without a qualifying disability claim, waiting or elimination period, active disability benefit, partial disability or return to work, and exited benefit; separate LTC states may describe no qualifying dependency, eligible home care, institutional care, and cessation. Each state's definition is governed by the source records. Death is an absorbing transition. Recovery, death, exhaustion of contractual benefits, lapse of coverage, and administrative closure are separate causes, not identical indicators of health improvement.

For claim-duration time u, let λ_k(u|x) be the cause-specific instantaneous rate of exit type k among claims still active immediately before u, conditional on covariates x. The survival probability in the active-claim state satisfies S(u|x)=exp[-∫_0^u Σ_k λ_k(s|x)ds] in the continuous-time formulation. The cumulative incidence for cause k is F_k(u|x)=∫_0^u S(s−|x)λ_k(s|x)ds. This formulation shows why treating deaths as ordinary non-informative censoring while estimating the actual probability of recovery inflates the interpretation of recovery: death removes an individual from the recovery risk set through a competing event.

The standard competing-risk literature includes Fine and Gray's 1999 study, which models the subdistribution hazard for cumulative incidence. Cause-specific and subdistribution hazards answer different questions and require their own interpretation. When transitions among more than two states are of interest, estimated transition probabilities require a fully specified state space, entry rules, and assumptions about the transition process. Aggregating all exits into one termination rate destroys information needed to distinguish actuarial payment duration from clinical and occupational outcomes.

Selection, mix, and standardization

Unadjusted differences across insurers arise from differences in ages, occupation, benefit design, underwriting, cohort entry, claim adjudication, geography, sector, and diagnostic mix. In the 2025 group LTD study, the public study description notes analysis by elimination period, industry, case size, location and diagnosis. These are not merely presentation choices; they are potentially material determinants of observed claims and of study inclusion. Even a high share of industry sales in an intercompany study does not imply probability sampling of the insured population or of the population at large.

Direct standardization fixes a reference distribution w_g across mutually exclusive risk strata g and computes I_std=Σ_g w_g I_g, with nonnegative weights summing to one. Comparisons then refer to the same specified standard mix, subject to common support, valid stratum-specific estimates, and compatible outcome definitions. Indirect standardization uses expected events under a reference rate applied to the observed exposure distribution. Neither method removes confounding from unobserved covariates or corrects a mismatch in contractual event definitions.

Worked analysis of a rate reversal

Consider two entirely hypothetical portfolios in the same observation period. Portfolio A has 75% of exposure in a lower-risk age group with 1.0 events per 1,000 life-years and 25% in a higher-risk group with 8.0. Portfolio B has 25% lower-risk exposure at 0.8 per 1,000 and 75% higher-risk exposure at 7.0. Within each age stratum B has lower incidence. The crude incidence of A is 0.75×1.0+0.25×8.0=2.75 per 1,000; B's is 0.25×0.8+0.75×7.0=5.45. Crude comparison suggests much higher B incidence. Standardizing both populations to 50% in each group gives A=4.50 and B=3.90 per 1,000, reversing the ordering.

This constructed example establishes a mathematical possibility and a validation obligation. An analyst who concludes that B's disability risk is intrinsically higher from the crude figure confuses risk composition with within-stratum experience. The reversal could arise from age, elimination period, job class, or another measured stratifier. It does not establish that age adjustment is sufficient in actual insurer data. The proper response is to publish both crude and standardized rates, the reference weights, every stratum numerator and denominator, and uncertainty intervals, then assess additional case-mix and selection factors.

Calibration, Validation, and Missingness

Internal consistency and reconciliations

Validation starts with identities rather than a model. For each closed accounting period, opening active claims plus new active claims and reactivations, less all closed claims, should equal closing active claims after adjustments for corrections and transfers. Every exit is assigned exactly one primary study category while secondary administrative reasons may be preserved separately. The total earned exposure in all mutually exclusive subgroups should reconcile to the study-wide exposure, subject to explicit residuals and rounding conventions. A source-to-target reconciliation identifies dropped, duplicated, recoded and unmatched records with counts and rates.

A careful cross-source audit compares the set of eligible covered lives with the set of persons who can contribute claims, checks the temporal ordering of onset, report, approval and payment, and tests impossible overlaps of active states. It also verifies that reported payment amounts have a currency and accounting basis, and that paid benefits during litigation or appeal are recorded in the same manner across insurers. A correction changing an earlier claim's effective termination date should change exposure and event allocation consistently, not only the status label.

Incomplete observations and identification

Missingness arises when a claimant's occupation is unrecorded, a historical ADL assessment is inaccessible, an insurer does not know exact onset time, or a care provider fails to send utilization data. Such patterns are not interchangeable. Rubin's 1976 analysis of missing data distinguishes inference conditions according to the missingness process. The practical implication is that a missing category alone is a description of incompleteness, not proof that complete-case analysis is unbiased. Where missingness is associated with severity, claim status, location, insurer, or reporting lag, the complete observed subset may be selected on variables related to the outcome.

For a binary event rate with unobserved outcomes, one can establish transparent bounds without pretending to identify the unobserved fraction. If observed records carry reliable denominators, reported upper and lower limits can be constructed by assigning all unobserved outcomes respectively to event and nonevent, subject to the relevant inclusion rules. In practice administrative missingness can also affect the denominator itself, making even those bounds too narrow unless exposure completeness is audited separately. Multiple imputation and inverse probability weighting require models for the missingness and outcome processes; analysts should test sensitivity to deviations rather than present model-based completion as an observed fact.

Competing-risk and duration validation

A claim episode may be right-censored, left-truncated, interval-censored, or repeatedly interrupted. The dataset must identify which situation applies. Recovery time estimated from claim closure dates can overstate medical recovery if claims continue for administrative reasons; it can understate recovery time if closure occurs before durable return to work. Death records may be linked with a delay. Competing-risk methods should be evaluated against event counts, cumulative incidence by cause, restricted mean claim-active time, and predicted versus observed state occupancy.

For illustration, if 1,000 active-claim months contain 20 recorded recoveries and five deaths, the crude observed recovery intensity is 20/1,000=0.020 recoveries per claim-month and the all-cause exit intensity is 25/1,000=0.025 events per claim-month. These quantities are rates, not probabilities of eventual recovery; they require additional assumptions to translate into cumulative incidence. The example also shows why a rate expressed per claimant-month cannot be compared numerically to incidence per covered life-year without specifying the event process and the risk set.

Calibration under transfer and change

Within-source model fit is weaker evidence of external transportability than independent validation. A transfer study should identify the target population, reweight exposures, harmonize eligibility rules, and compare observed-to-expected event counts in materially relevant strata. For a rate model predicting expected events μ_g in stratum g, the aggregate observed-to-expected ratio is Σ O_g / Σ μ_g; this can equal one while particular diagnosis, age, benefit, and care-setting groups show serious miscalibration. Predictive residuals, calibration plots, cause-specific comparisons, and interval estimates should therefore accompany aggregate ratios.

Longitudinal transfer raises additional difficulties: changes in underwriting, inflation, work patterns, social benefits, clinical diagnostics, caregiving supply, telehealth, and adjudication may alter both the target outcome and its recording. A current claims study cannot be applied to a past cohort as though observation and eligibility were invariant. Structural break analyses should preserve the historical definition of each claim and, where possible, calculate a second bridged series using a prespecified mapping. Data-driven break points require uncertainty and multiple-testing treatment if inference is attempted.

Worked analysis of missing outcomes

Suppose an illustrative study has 1,000 eligible equal-exposure records, of which 800 have a confirmed outcome and 200 have an unknown outcome. Of the 800 observed, 80 record the event. A complete-case estimate is 80/800=10%. Under the strong assumption that the missing outcomes have the same risk as observed outcomes, the all-person estimate is also 10%. If all 200 missing records are nonevents, the overall incidence is 80/1,000=8%. If they are all events, incidence becomes 280/1,000=28%. Thus the identification interval from the information alone is [8%, 28%], for this simple equal-risk record design.

The interval is deliberately wide because it expresses what the recorded outcomes establish without unverifiable completion assumptions. Narrowing it requires external evidence about missingness or appropriate assumptions and sensitivity tests. For insured experience measured in person-time, the same logic must be adapted to the actual exposure contributed by missing records. This example supplies an explicit limit of inference; it is not an estimate of missingness in any SOA, national survey, or insurer dataset.

Integrated Findings and Research Agenda

A cross-source interpretation of the evidence

The combined evidence supports a precise substantive conclusion. The SOA's LTC experience studies organize claim incidence, termination and utilization as different statistical objects; its later group LTD studies again distinguish the covered-life incidence process from the active-claim termination process. The LTC intercompany database documentation, 2025 LTD incidence study, and 2026 LTD termination study collectively establish the institutional basis for this separation. WHO, Eurostat, national authorities and the OECD add observations of impairment, functional needs, benefit entitlement and economic costs. These sources expand explanatory context but do not transform different measures into one event count.

Harmonization should accordingly proceed from estimand to evidence, not from file format to pooled estimate. If the question is How often do covered employees first satisfy a specified insured LTD claim definition?, the appropriate records are covered lives, insurance qualification and onset under the same trigger. If the question is How long do approved claimants continue receiving benefits before recovery or death?, the risk set is active claims and the transition outcomes must be distinguished. If the question is How much personal-care assistance does the population need?, representative functional instruments and service records become central, with financing and informal care incorporated into interpretation. Crosswalking among these questions is a research exercise, not an arithmetic conversion.

The numerical examples clarify two independent failure modes. Crude cross-portfolio rates can reverse under a common risk composition even with correct event ascertainment. Conversely, missing outcomes can leave a very wide identification interval even when the population and event are nominally aligned. Together they show why valid comparability demands both correct target definitions and evidence about the data-generation process. A perfectly curated codebook cannot repair selection bias, and a sophisticated survival model cannot recover unknown contract criteria from data that never recorded them.

Minimum conditions for valid comparison

A scientific comparison should meet the following operational conditions, each documented and auditable: (1) a single explicitly named estimand; (2) the same eligible event or a tested versioned crosswalk; (3) a compatible at-risk population and exposure unit; (4) coherent inception, exit, recurrence and censoring rules; (5) known data capture and linkability; (6) an explicit common reference distribution or justified transport model; (7) cause-specific event accounting; (8) quantified missingness, measurement and model uncertainty; (9) permissible data access and protected reporting; and (10) the ability to trace any published estimate back to the contributing source versions. A failure of one condition does not invalidate all descriptive use of a dataset, but it changes which inferential conclusions the data can bear.

One practical consequence is that research data should publish not comparable as an affirmative, documented result when a bridge lacks evidence. This classification preserves information and prevents false certainty. Similarly, partially comparable under declared assumptions is scientifically meaningful when researchers report the estimand, assumptions, bounds and sensitivity. These are editorial judgments about the strength of an empirical comparison, not the introduction of a new theory or regulated reporting standard.

Research priorities and validation questions

Future empirical studies should prioritize permissioned, reproducible insurer-level linkage of eligibility intervals, functional assessments, claim administration and cash flows; independent testing of the stability of ADL and occupational trigger definitions; cause-specific duration modeling with delayed entry and reopening; and joint analysis of formal and informal care without inferring missing family work from insurer payments. Public intercompany databases should, where permitted, publish richer versioned dictionaries for denominators, censored episodes and claims-stage transitions. Comparisons across jurisdictions should explicitly retain their statutory and contractual eligibility rules instead of collapsing national programs into one disability category.

Further questions include the sensitivity of historical risk estimates to policy revisions, the magnitude of selection caused by waiting periods, the conditional relationship between functional trajectories and insured claims, and the degree of transportability across employment structures, ages and care settings. Progress can be measured by independent validation and the narrowing of justified uncertainty intervals. No universal direction or size of any of these effects is asserted absent applicable underlying data.

Conclusion

Long-term care and disability insurance experience data harmonization is a problem of scientific identification. To estimate the same quantity from two sources, researchers must preserve the same insured or functional event, define the same time at risk, and reconcile population selection, contractual eligibility, observation and exits. The institutional experience evidence considered here demonstrates that claim incidence, continuation, recovery, death, utilization and benefit payment belong to related but different processes. International functional classifications and care assessments establish valuable observations of need, while legal and policy provisions determine which observations become payable events.

The analytical results follow directly. Calendar-year matching and uniform column names do not establish comparability. Event-specific risk sets, versioned source crosswalks, cohort standardization, competing-risk analysis, missingness bounds and independent calibration make the degree of comparability assessable. A pooled statistic becomes defensible when the underlying observation and selection mechanisms support the same estimand or when model-dependent bridging assumptions are plainly exposed to validation. Where those mechanisms remain unknown, the scientifically valid output is a bounded or restricted conclusion.

The monograph provides a documented cross-disciplinary synthesis with reproducible derivations and hypothetical counterexamples. It advances no new proprietary theory and makes no empirical claim beyond the inspected public evidence. Its practical result is a standard of evidentiary discipline for future insurer-level research: preserve provenance, retain the meaning of every risk-set denominator, and connect each rate to the event it actually measures.

References