Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Disability Incidence and Recovery Models

Abstract. Disability insurance valuation depends on movements between eligibility, disability, partial capacity, recovery, relapse, death, and contractual termination. These movements are measured in distinct populations and on different time scales: insured person-years before claim onset, life-months of exposure after a claim, and policy or administrative intervals defining payable benefits. This monograph establishes a mathematical and empirical account of disability incidence and recovery models, grounded in actuarial experience studies, continuous-time event-history analysis, multistate probability, and applicable professional standards. It distinguishes medical and functional conditions from insured claim events; cause-specific recovery hazards from aggregate terminations; elapsed disability duration from attained age and calendar period; and statistical survival from benefit continuance. It reconstructs the evidentiary scope of the Society of Actuaries–LIMRA group long-term disability studies, US individual disability valuation work, and the UK Continuous Mortality Investigation income-protection series. Original analytical work consists of conditional derivations of claim survival and discounted benefit values, a reproducible numerical example, a demonstration of how competing decrements alter cash-flow estimates, and an evaluation of exposure-denominator and transferability errors. Results are conditional on explicitly stated assumptions, and public aggregate study summaries are not presented as insurer-level observations. The central conclusion is that a defensible disability model must align the event definition, exposure unit, transition process, contract, population, and valuation objective before calibration can carry evidentiary force. No new authorial theory or proprietary model is proposed.

Keywords: Disability insurance; claim incidence; recovery hazards; relapse; claim termination; duration dependence; multistate models; semi-Markov processes; competing risks; life-years exposed; claim life-months; individual income protection; group long-term disability; experience studies; loss reserving; present value; data quality; credibility; model validation; occupational disability; actuarial science.

Introduction

The scientific object of disability incidence and recovery modeling is a sequence of contingent movements affecting insured earnings protection. An insured person can remain at work, experience a condition that prevents specified employment, satisfy an elimination period, enter payable disability, return to partial or full work, suffer a subsequent claim, die, or lose benefit eligibility under the contract. The probability law of these movements influences premiums, incurred claim reserves, future liability, reinsurance, capital, and the duration of financial protection. An assessment that counts each event as an interchangeable “disability rate” loses the necessary units and legal definitions.

The WHO classification of functioning (World Health Organization, 2001) locates functioning and disability in the interaction of health conditions, activities, participation, and environmental factors. An insurance contract asks a narrower question about specified insured inability to work and covered benefit entitlement. A public disability program may apply yet another test. These concepts are related through the observed condition and employment consequences, but the thresholds of recognition, dates of entry, and events recorded as cessation depend on their respective systems. Accordingly, a model fitted to paid claims estimates an insured-claims process conditional on the rules that produced those claims; it does not directly estimate all disability in the general population.

The public 2015–2022 group LTD incidence study (SOA Research Institute and LIMRA, 2025) describes roughly 294 million life-years exposed and 1.2 million claims from 19 contributing companies. The newer 2015–2024 LTD termination study (SOA Research Institute and LIMRA, 2026), released September 30, 2026, covers more than 43 million disabled life-months, approximately 937,000 recoveries and 181,000 deaths from 18 contributors. These are distinct empirical investigations with different numerators, denominators, cohorts, and observation windows. Their headline volumes are descriptive of their own studies and cannot be divided into a universal incidence-recovery formula. The publicly described detailed data packages have subscription restrictions; the present study uses the disclosed summaries and documentary methodological evidence without claiming access to withheld microdata.

The mathematical apparatus follows established event-history theory, including counting-process models (Andersen et al., 1993) and competing-risk and multistate analysis (Putter, Fiocco, and Geskus, 2007). Professional quality requirements are examined through the US Data Quality standard (Actuarial Standards Board, 2016), Credibility Procedures standard (2013), and Modeling standard (2019), while recognizing their applicable professional scope rather than assigning them universal legislative force.

The chapters proceed from entry rates and their exposure definitions to the distribution of spell duration, competing transitions and relapses, empirical source construction, expected benefit values, and validation. The resulting synthesis focuses on the precise implications of assumptions rather than an independent newly named theory.

Incidence Rates

The event that enters the numerator

An inception is an event defined by the study protocol. Candidate dates include medical onset, first inability to perform the insured occupation, onset of earnings loss, reported claim, satisfied elimination period, approved liability, and first payment. Those dates can differ substantially. For group long-term disability (LTD), an inception study may restrict observations to claims reaching a minimum elimination period and surviving required eligibility rules. In that case the incidence rate estimates entry into the study-defined claim class, conditional on exposures and coverage, and excludes disability episodes that resolve before its threshold. A short-term sickness model and a long-term paid-claim model therefore require different event clocks and often different denominators.

Let the insured active state be A. For a policyholder in A at time t, with age x, insured occupation o, contract design c, elapsed coverage duration p, and relevant covariates z, the instantaneous incidence intensity is

λAD(t∣x,o,c,p,z)=lim⁡h↓0Pr⁡{X(t+h)=D∣X(t)=A,Ht}h,\lambda_{AD}(t\mid x,o,c,p,z)=\lim_{h\downarrow0}\frac{\Pr\{X(t+h)=D\mid X(t)=A,\mathcal H_t\}}{h},

when that limit exists and the history and eligibility conditions are understood. The unit is expected onsets per active insured life-year when time is measured in years. The estimate based on counts divided by exposure is a rate, not exactly a probability: with constant intensity (\lambda) over one year, the onset probability in a single-decrement active state is (1-e^{-\lambda}). For small (\lambda), the rate and the corresponding probability are close; with multiple exits from A they require competing-decrement adjustment.

In a counting-process formulation, N_AD(t) counts eligible A-to-D transitions and Y_A(t) is the number or weighted exposure at risk immediately before t. The aggregate exposure estimator over an interval is

λ^AD=∑iNAD,i(T)−NAD,i(0)∑i∫0TYA,i(u) du.\widehat\lambda_{AD}=\frac{\sum_i N_{AD,i}(T)-N_{AD,i}(0)}{\sum_i\int_0^T Y_{A,i}(u)\,du}.

The estimator has an interpretation only after checking whether the exposure includes partial periods, cancellations, waiting periods, and persons ineligible to enter the claimed state. A binomial model of each covered life observed for a full year can be useful, but is not identical to this event-time estimator in the presence of staggered observation or changing covariates. The distinction between risks, exposure and counting processes is developed in the survival-analysis monograph (Andersen et al., 1993).

Conditional inception and the elimination period

For a condition beginning at time zero, let T be its duration until resolution, death, or another competing exit; let e denote the elimination period. A simplified relation between onset and a claim surviving e is

λqual(t)≃λonset(t)Pr⁡{T>e∣onset at t},\lambda_{\mathrm{qual}}(t)\simeq\lambda_{\mathrm{onset}}(t)\Pr\{T>e\mid\text{onset at }t\},

provided that the same eligible population, event definitions, reporting rule, benefit contract, and competing exits apply. This expression is an idealized factorization; actual claims may be reported late, adjudicated retrospectively, or allowed after nonconsecutive waiting days. Consequently, a calibration should not multiply a generic general-population disability onset rate by a generic disability-survival fraction and label the result an insured LTD incidence assumption.

The SOA/LIMRA incidence methodology specifically investigates elimination period, industry, case size, geography and diagnosis. These factors affect both risk and the observation mechanism. If an employer changes the elimination period, the modeled incidence may change even when the underlying medical occurrence of work-limiting conditions does not. Similarly, own-occupation versus any-occupation eligibility affects event recognition and continuity, so definitions must travel with rates.

Selection, composition and the risk set

Actuarial incidence experience depends on the insured population. Underwriting, employer participation, benefit caps, wage distribution, occupation, age, sex where permitted and appropriate, job security, and contract participation may affect the at-risk set. A rate averaged across groups is a weighted composite, not a natural constant. If occupations o have rates (\lambda_o) and exposure weights (w_o), the portfolio rate is (\sum_o w_o\lambda_o). Changing the weights changes the composite without changing any occupation-specific rate. This is a direct arithmetic explanation for why commercial mix shifts can resemble morbidity trends.

A fair comparison therefore fixes its standard population, or explicitly decomposes movement into within-stratum rate change and exposure-mix change. Stratification increases interpretability but reduces cell sizes; credibility methods assist in choosing which detail the evidence can support. The ASOP 25 guidance on credibility requires careful selection of relevant experience and attention to homogeneity within its U.S. professional scope. An external benchmark cannot be assumed exchangeable with a contract whose benefit definition, workforce, and elimination period differ.

Calendar period, coverage duration and diagnoses

Inception intensity may change with calendar time, employment conditions, policy tenure, or diagnostics. A coherent regression could write (\lambda_{AD}(t\mid z)=\lambda_0(t)\exp(\beta^\top z)), where t's meaning and baseline time scale are made explicit. When duration since policy issue and calendar period jointly matter, the two clocks must be distinguished. An age-period-cohort model faces classical dependency between age, period, and cohort; identification requires constraints and transparent interpretation. Improvements in claim coding do not automatically represent changing disease prevalence. The WHO ICF framework further explains why changes in environmental accommodation can change participation outcomes independently of diagnostic labels.

The empirical implication is that incidence rates are conditional summaries of observed insured events. Their transfer to future exposure must state what population, eligibility rule, calendar horizon, and change in case mix is being projected.

Claims occurrence, notification and incurred-but-not-reported events

The inception date of an insured disability episode and the moment it becomes available in an administrative dataset are different random variables. Let T_o denote qualifying onset and T_r the reporting date. A claim report at time T_r contributes to observed reports in the calendar period containing T_r; its disability incidence belongs to the period containing T_o when that onset date can be reconstructed reliably. A claims database extracted on a fixed date omits some recent onsets for which T_r exceeds the extraction cutoff. This creates right truncation in reported-onset data and can masquerade as declining incidence in the latest calendar periods. An analyst estimating claim incidence must state whether claims are arranged by occurrence, report, approval, or payment date, and how the delay distribution is handled.

A basic onset-report decomposition is (N_{\mathrm{reported}}(s)=\int_{t\le s}N_{\mathrm{onset}}(dt)\Pr(T_r\le s\mid T_o=t)), understood as an expected-value convolution under the specified conditional reporting process. A reporting delay may depend on diagnosis, occupation, employers' human-resource systems, elimination period, and time since the contract was issued. Estimating it from already approved claims alone can introduce selection bias when long-pending claims are systematically less likely to have been adjudicated. Development triangles and individual event histories offer complementary perspectives, but both require stable definitions and complete extracts. The status of a report is therefore a measurement variable and cannot automatically be taken as proof of onset.

For a hypothetical study with 10,000 life-years exposed and 40 qualified onsets identified after complete reporting development, the exposure-based incidence rate is 0.004 per life-year. If only 34 of those onsets were recorded at the original extraction date, the provisional observed rate would be 0.0034. The gap of 0.0006 reflects reporting maturity in this example, rather than a real change in underlying insured risk. This is an original arithmetic illustration; no external company experience is inferred from these fictional counts.

Model purpose and estimated quantity

An incidence model may serve pricing, active-life reserve calculation, scenario design, or claims workload forecasting. The estimand for workload is often approved claim arrivals per month, which includes reporting and adjudication mechanisms; the estimand for actuarial risk may be contract-defined qualified onsets per eligible life-year. A rate useful for one purpose can be misleading for another. To establish the target, identify the event clock, covered population, required elimination duration, claim decision rules, and whether repeats are counted by person or by episode. A retrospective study may use attained age at inception, whereas a forward projection updates age continuously. A coherent model makes these operations explicit before fitting its regression or graduation model.

Disability Duration and Recovery

Duration as a second time axis

Once claim disability begins, the relevant clock often becomes duration u since inception or since commencement of paid benefits. The recovery intensity at duration u, given disabling history and other covariates, is

ρ(u∣x0,o,c,z)=lim⁡h↓0Pr⁡{D→R in [u,u+h)∣D at u}h,\rho(u\mid x_0,o,c,z)=\lim_{h\downarrow0}\frac{\Pr\{D\to R\text{ in }[u,u+h)\mid D\text{ at }u\}}{h},

where x_0 is age at disability and R is the model's defined recovery state. This is a cause-specific hazard with inverse-time units, not a cumulative probability. It may vary sharply with duration: short claims often have different exit patterns from claims persisting several years. The 2008 GLTD experience study and the subsequent 2012 GLTD valuation table document the importance of duration and disability definitions in established industry tables.

Suppose recovery is the only exit and its hazard depends on u. Then the probability of no recovery through duration u is

SR(u)=exp⁡[−∫0uρ(s) ds].S_R(u)=\exp\left[-\int_0^u\rho(s)\,ds\right].

With a second cause-specific death hazard (\mu_D(u)) and other exit hazard (\kappa(u)), still conditional on being in D, the probability of remaining in D is instead

SD(u)=exp⁡[−∫0u{ρ(s)+μD(s)+κ(s)} ds].S_D(u)=\exp\left[-\int_0^u\{\rho(s)+\mu_D(s)+\kappa(s)\}\,ds\right].

The recovery cumulative incidence is (F_R(u)=\int_0^u S_D(s)\rho(s),ds), while the death cumulative incidence is (F_M(u)=\int_0^u S_D(s)\mu_D(s),ds). The sum of all properly exhaustive cumulative exit incidences is (1-S_D(u)). An independent Kaplan–Meier analysis that treats competing deaths as ordinary independent censoring estimates a hypothetical net recovery survival, not the actual probability of recovery before death. This distinction follows the established competing-risks tutorial (Putter et al., 2007).

Recovery as an observed contractual event

“Recovery” may denote medical improvement, functional improvement, paid work resumption, no longer satisfying the definition of disability, or an administrative code indicating claim termination. These outcomes must be mapped explicitly. A claimant may return to partial work yet receive a residual benefit; an insurer may close a claim at maximum benefit duration even though disability persists; a transition to a different occupational test may terminate eligibility independently of medical recovery. The historical group LTD valuation work explicitly treats recovery and death assumptions separately and incorporates disability-definition changes in the architecture of its table.

Dates also matter. Claim termination is frequently recorded at a monthly processing date, while actual recovery could precede it and be backdated. If data supply only end-of-month states, a continuous-time hazard cannot be treated as observed exactly without accounting for interval censoring or a defensible within-month convention. The analysis of paid claims and of underlying disability duration should not be merged by a change of label.

Choosing a duration distribution

A constant recovery hazard yields an exponential distribution, a convenient benchmark with a memoryless property. That property is rarely a satisfactory empirical assumption over every duration of disability. Piecewise-constant hazards can capture early, middle, and late phases and integrate easily into benefit calculations. A Weibull specification (\rho(u)=a b u^{b-1}) permits increasing or decreasing recovery hazard depending on b, but makes a strong global shape assumption; a flexible spline, segmented table, or semiparametric model may fit heterogeneous periods better. A duration model must specify whether hazards are measured from clinical onset, elimination-period completion, claim opening, or benefit commencement.

The expected duration spent in a disabled state before exit, if payments and observation are conceptually indefinite, is (E[T_D]=\int_0^\infty S_D(u)du), assuming integrability. A contract with a fixed maximum benefit period L has an expected payable occupancy of (\int_0^L S_D(u)du), subject to payment and offsets. The right-hand side measures occupancy of the payable state rather than expected clinical illness duration when eligibility restrictions are present.

Markov, semi-Markov and latent heterogeneity

A continuous-time Markov chain uses transition intensities determined by the current state and current calendar time. If recovery depends on u, the time elapsed in the disabled state, then a model on A and D alone generally fails the Markov memorylessness requirement. It becomes semi-Markov or Markov after augmentation of the state with elapsed duration. This difference is structural: using a Markov transition matrix while externally applying duration-dependent recovery percentages can produce inconsistent probability propagation.

Unobserved severity is another mechanism generating declining observed recovery hazards. Even if each latent type has a constant hazard, fast-recovering individuals exit early, leaving an increasingly severe residual claimant pool. Let type j have weight (w_j) and hazard (\rho_j). Survivor composition at u is proportional to (w_j e^{-\rho_j u}). The aggregate observed hazard (\sum_j w_j e^{-\rho_j u}\rho_j / \sum_j w_j e^{-\rho_j u}) tends to shift toward low-recovery types with increasing u. The result is a direct selection effect, not proof that duration itself causally reduces recovery. Individual claimant heterogeneity must be distinguished from genuine duration dependence during calibration.

Discrete observations and interval-censored recovery

Many claims systems record status only at monthly benefit adjudications. A return to work may happen between assessments and be recorded at the next payment cycle; the event time is then interval censored. Treating all exits as if they occurred precisely at month-end mechanically lengthens observed spells. In a piecewise constant-hazard model with interval length (\Delta), probability of remaining continuously disabled during an interval is (\exp[-(\rho+\mu_D+\kappa)\Delta]). The probability of recovery during that interval is (\rho(1-e^{-g_D\Delta})/g_D), with (g_D=\rho+\mu_D+\kappa), when the intensities are constant and recovery is one competing exit. The analogous formula for death substitutes (\mu_D). These interval probabilities sum with disability continuance and other exits to one; they are not just (\rho\Delta) when durations are material.

A fitted model using discrete month-end records can employ transition-probability likelihoods rather than pretending to observe precise continuous event times. Alternatively, a counting-process analysis may use midpoint approximations with a stated error bound. The choice depends on actual data resolution. Contractual termination can also be observed exactly while clinical recovery remains interval censored, generating unequal temporal precision across causes. Identical calendar timestamps in an administrative extract do not imply identical event-generating precision.

Dependent event mechanisms and causal discipline

Recovery and mortality hazards may both depend on age, diagnosis, severity, accommodation, treatment, benefit definition, and duration since onset. Conditioning on observed claim survivors induces a selected risk set: the composition of people disabled after five years differs from that of entrants even if every individual's transition law is constant in time. Consequently, an aggregate declining recovery rate with duration can emerge from selection among heterogeneous claimants as well as genuine duration dependence. A frailty-mixture model can illustrate this: (S_D(u)=\sum_j w_j e^{-g_j u}), with (w_j\ge0) and (\sum_jw_j=1), produces an aggregate hazard (\sum_jw_j g_j e^{-g_j u}/S_D(u)) that typically falls as the lower-hazard strata dominate the surviving stock. This formula proves a mixture-composition mechanism; its empirical importance requires claimant-level evidence.

A rehabilitation program may be associated with later recovery, yet claimants who enter it can differ from others in health, motivation, coverage, workplace opportunity or severity. Estimating a change in hazard conditional on participation therefore establishes an association unless an identification design supports causal interpretation. Administrative terminations, claimant motivation, and selection into a program must be handled consistently. The claims model alone cannot establish the causal efficacy of treatment, rehabilitation or benefit design from observational termination differences.

Relapse and Mortality Interactions

Multiple transitions and recurrent spells

A minimal model distinguishes A (active and eligible), D (disabled and potentially payable), R (recovered or returned to work), and M (dead). The recovery state R can be combined with A only if its subsequent incidence risk and contractual rights are equivalent to those of an never-disabled active insured. For recurrent conditions, relapse intensity from R to D may differ from first-onset incidence from A to D. Residual benefits, recurrent-disability provisions, waiver of elimination periods, and partial work require additional states when materially relevant.

For a time-homogeneous illustrative model, rows of a generator Q sum to zero. With transition intensities (a=A\to D), (d_A=A\to M), (r=D\to R), (m_D=D\to M), (\ell=R\to D), and (d_R=R\to M), one possible generator is

Q=(−(a+dA)a0dA0−(r+mD)rmD0ℓ−(ℓ+dR)dR0000).Q=\begin{pmatrix} -(a+d_A)&a&0&d_A\\ 0&-(r+m_D)&r&m_D\\ 0&\ell&-(\ell+d_R)&d_R\\ 0&0&0&0 \end{pmatrix}.

This deliberately treats A and R as distinct histories. For fixed Q and time horizon t, the transition matrix is (P(t)=\exp(Qt)), where every entry is nonnegative, rows sum to one, and M is absorbing. For time-varying Q(t), forward transition probabilities solve (\partial_t P(s,t)=P(s,t)Q(t)), with (P(s,s)=I). The historical Aalen–Johansen transition-matrix article (Aalen and Johansen, 1978) develops a product-integral approach for transition probability estimation under censoring; the later counting-process treatise supplies broader technical foundations. This representation is useful for expected occupancy and aggregate valuation when state definitions are homogeneous enough to support it.

Relapse and episode definitions

A short return to work can be treated as the end of one claim, a pause in the same claim, or the beginning of a recurrent claim when disability returns. The insurance policy may define recurrent disability within a fixed time window as a continuation of the previous episode, possibly waiving a second elimination period. Research data therefore need two parallel identities: clinical or employment episodes and insurer-defined claim episodes. Without both, a transition attributed to “new incidence” might be a reopened claim. Such misclassification simultaneously distorts counts of inception and termination.

The relevant likelihood uses the observed transitions, exposures and eligibility constraints. For transition class j\to k, with intensity (\lambda_{jk}(t;\theta)), an idealized fully observed individual likelihood has the product of observed intensities at event times times (\exp[-\int\sum_{k\ne j}\lambda_{jk}(t;\theta)dt]) over time spent in state j. Recurrent episodes within the same person are statistically dependent; robust standard errors, shared frailty, or individual random effects may be needed. Absence of a recorded relapse may mean remission, loss to follow-up, a different insurer, or benefit exhaustion. Right censoring and loss of observation must not be treated as clinical recovery.

Death as a competing decrement

Death during disability removes the person from the recovery risk set and ordinarily ends disability-income payments, though survivor riders or accrued benefits can create separate cash flows. A recovery rate estimated as recoveries divided by all initial claims for a long interval is not a cause-specific instantaneous hazard, because the denominator evolves through both death and recovery. When hazards are constant, the probability of a recovery before time L is (\rho(1-e^{-(\rho+\mu_D+\kappa)L})/(\rho+\mu_D+\kappa)); the corresponding death probability replaces (\rho) with (\mu_D). These formulas demonstrate why a single combined termination rate cannot separately identify recovery and death rates: identical sums can arise from different compositions, implying different future disability, death benefits, and medical interpretations.

The newest SOA/LIMRA termination summary reports recoveries and deaths separately. That published distinction is actuarially substantive, while its headline counts alone cannot yield duration-specific hazards without properly aligned cause-specific exposure. The SOA individual claim termination investigation likewise compares observed individual disability claim terminations over 2006–2014 to the 2013 IDI valuation table, underscoring the importance of a separate individual-market basis.

Recovery does not erase lasting economic effects

The OECD's Disability, Work and Inclusion report (2022) documents labor-market exclusion and persistent employment and income gaps among persons with disabilities. Return to work therefore belongs in the social and economic interpretation of recovery, but insured claim closure does not prove restored wages, full participation, or permanent health improvement. More detailed models might allow a recovered partial-earnings state, a repeated disability state, and an external labor-market employment state, if the policy and data justify them. Here they are analytical extensions of established multistate methods, not a separate commissioned authorial system.

Experience reconciliation and a reproducible mapping exercise

Suppose two hypothetical datasets cover the same 100,000 lives for one year. Dataset A counts 150 new insured onset episodes, including episodes shorter than a 90-day waiting period. Dataset B counts 60 benefit-entitled claims satisfying 90 days, after excluding withdrawals and noncovered events. The raw observed incidence rates are 0.00150 and 0.00060 per life-year, respectively. Their ratio of 2.5 measures a change in observation definitions and cannot be interpreted as a relative population risk without reconstructing which onsets survive the waiting period and whether the exposures are identically eligible. If 40 of the 150 initial episodes remain continuously impaired beyond 90 days, yet 60 payable claims appear in B, the discrepancy would itself be a diagnostic sign of a population, cohort, adjudication or attribution mismatch. Counts drawn from different clocks cannot be reconciled simply by comparing denominators.

An experience reconciliation should start with a contract map. Each dataset receives an observation unit, event definition, inclusion population, exclusion policy, exposure denominator, duration clock, time window and censoring rule. The analyst then traces events through eligibility, onset, elimination, approval, payment, partial return and termination, retaining multiple event dates where possible. Source tables should be joined only on verified unique person or claim identifiers, with claim reopening and duplicated submission checks. This gives an auditable basis for deciding which differences are substantive and which are data production effects.

International evidence transfer and the ecology of classification

Even after correct denominator standardization, insured claims are selected from labor markets with distinct sick-pay systems, occupational structures, health access, contractual definitions, and social-protection interactions. A UK individual income-protection inception table, a US employer-sponsored LTD termination experience study and a US federal disability-roll statistic observe different institutions. Their terms disability, recovery and termination therefore map to different events. Transferring a rate requires a parameter-by-parameter justification: which benefit trigger, eligibility period, duration, age basis, diagnosis coding, payment status, benefit offsets and competing exits carry across the target population? A demographic standardization alone cannot establish equivalence of contract or adjudication rules.

Observed experience may also change under new documentation requirements or computer-assisted decision systems. A classification change can produce apparent onset or recovery breaks even if the health conditions of claimants remain similar. The appropriate response is to preserve versioned data dictionaries and test for discontinuities around implementation dates. Subgroup comparisons are scientifically strongest when accompanied by denominator disclosure, uncertainty, reporting lags, and the governing coverage rules. This is a methodological synthesis from the source characteristics, not a claim that all surveyed jurisdictions have identical regulatory classifications.

Experience Data

Different exposure units in the SOA studies

The SOA/LIMRA group LTD incidence study covers calendar years 2015–2022, approximately 294 million life-years of fully insured group LTD exposure, and approximately 1.2 million qualifying claims from 19 contributors. Its public documentation identifies several excluded business types, including administrative-services-only groups and policies below stated waiting or benefit-duration thresholds. A crude ratio of these rounded totals is approximately (1.2/294=0.00408) claims per exposed life-year; this number is merely an approximate aggregate ratio calculated from rounded public figures. It cannot be substituted for the study's segmented incidence tables, classified as a general-population disability incidence, or assumed to incorporate a particular case's risk mix.

The SOA/LIMRA termination study covers 2015–2024, over 43 million life-months of disabled exposure, roughly 937,000 recoveries and 181,000 deaths, with 18 contributors. The month-based numerator and denominator structure has a different target from life-years before disability. A hypothetical conversion of disabled exposure to life-years by division by twelve changes units only; it does not make conditional claimant hazards equivalent to incident claim probability. Counts summed across duration bands would also mask the survivor selection and varying mix of diagnoses. Claims already active at the study's entry date are left-truncated at the observation boundary and require duration reconstruction to avoid mislabeling them new claims.

Comparison across the two studies requires caution over different contribution sets, observation windows, products, and exclusion criteria. The public summary descriptions establish large data volumes, not individual record verification. Source audit therefore records the high-level published counts and source access limitation rather than asserting derived actual-to-expected (A/E) ratios or unobserved diagnosis-cell estimates.

US table development and the valuation objective

For group disability, the 2008 GLTD Experience Table preceded the 2012 valuation table development. Experience tables characterize observations in a specified contributing population; valuation tables can incorporate prescribed margins, regulatory applicability, and simplifications for minimum reserve purposes. A statutory table's adequacy for a regulatory floor does not establish that it is an unbiased prospective best estimate for another product or calendar year. For individual disability income (IDI), the 2013 IDI Valuation Table was approved by the NAIC in August 2016 as the new statutory minimum morbidity reserve basis described by the SOA. This distinction illustrates why group and individual experience should not be pooled merely because both are called disability insurance.

The 2006–2014 individual termination study offers a later benchmark against 2013 IDIVT base termination rates. Comparison to expected must use exactly the relevant table version and adjustment definitions. If a dataset captures only paid claims, a reserve model requiring all disability episodes should distinguish the unobserved population rather than backfill it from another source without validation.

UK income protection and documented source correction

The UK Continuous Mortality Investigation (CMI) has collected income-protection experience over decades; its income protection investigation history describes the shift from earlier sickness approaches toward claim inception and termination analysis. The final IP11 inception and termination rates, originally issued in 2020, were subsequently accompanied by a correction warning: overstated inception exposure had understated inception rates, while termination experience was unaffected. The CMI methodological correction paper (2021) explains the issue and publishes accessible descriptive guidance, although the substantive paper itself has authorized-user access restrictions.

This is a direct empirical lesson in source governance. Under the elementary estimator (\widehat\lambda=N/E), if recorded exposure is inflated to (E'=E(1+\varepsilon)), the published estimate becomes (\widehat\lambda'=\widehat\lambda/(1+\varepsilon)). The relative downward bias measured against the correct rate is (\varepsilon/(1+\varepsilon)), holding the numerator fixed. For an illustrative 10% exposure overstatement, the estimate is about 9.09% too low. That arithmetic is an analytical illustration of a known error mechanism; it is not a claim that the actual CMI exposure discrepancy equaled 10%.

The subsequent CMI Working Paper 203 (2025) covers 2021–2023 individual income protection experience and includes analysis by policy duration and commencement year alongside occupation, sex, deferred period, age, smoker status and calendar year. Its publicly available description states that inception comparisons adjust for the earlier underlying data issue. The CMI rate-table tool documents calculation of rates and claim annuity values and explicitly notes that the IP11 tool itself was not retroactively updated for the issue. Neither authorized-user-only results nor detailed curves are reproduced or inferred here.

Public disability benefits and international non-equivalence

The US Social Security Administration's 2025 disability statistical report, released October 2026, records approximately 633,307 disabled-worker awards and 762,526 disabled-worker terminations in 2025. These are administrative benefit events and cannot be interpreted as private group LTD inception or medical recovery rates. The termination-by-reason table separates death, age-related conversion and other benefit exits, establishing why aggregate termination counts require event coding before actuarial use. The continuing-eligibility guidance also identifies periodic reviews and benefit rules tied to observed work activity. Changes in review policy and the timing of decisions can alter administrative termination patterns independently of underlying medical health trajectories.

Across Europe, Eurostat health measurement methodology describes survey-based long-standing activity limitations under EU-SILC and EHIS. Such survey measures are relevant to population functioning but differ from eligibility and contract-specific paid-claim inception. The WHO ICF provides common conceptual language rather than harmonizing national disability policies, waiting periods, or private policy definitions. The OECD's cross-national disability labor-market analysis adds evidence that work participation is a major outcome dimension, not a substitute for either insured claim status or diagnosed disease.

A reproducible experience-data schema

A source-to-model data specification should carry: a stable pseudonymous person ID; policy ID; individual claim episode ID; in-force dates; coverage and benefit definition version; occupation class and insured earnings basis; disability onset and report dates; waiting-period rules; benefit commencement; all dated state changes; cause-coded exits; attained age; diagnosis classification and version; claimant payment history; premium or coverage status; entry and censoring dates; calendar-period and reporting adjustments; and source and revision provenance. Person, policy, and claim IDs are distinct because one insured can hold multiple policies and have repeated claims. Data suppression and privacy controls govern access to sensitive claims records.

Data review should reconcile exposure at each state boundary, events against valid risk sets, calendar and duration units, missing exit causes, duplicate episode IDs, out-of-order dates, claims reopened after closure, and legal benefit expiry. The adopted ASOP 23 data-quality guidance provides a documented professional framework for selecting, reviewing, using, and disclosing data limits. In this study the schema is a transparent analytical checklist, and no claim is made that an actual confidential insurer data warehouse was audited.

Model specification through continuous-time rewards

For a general finite-state process with state-reward vector (b(t)), initial state distribution (\pi_0), transition matrix (P(0,t)), and deterministic discount force (\delta(t)), the prospective benefit present value under the specified model is

EPV=∫0Hexp⁡[−∫0tδ(s)ds]π0P(0,t)b(t) dt.\mathrm{EPV}=\int_0^H \exp\left[-\int_0^t\delta(s)ds\right]\pi_0P(0,t)b(t)\,dt.

This representation can accommodate active exposure, payable total disability, payable partial disability, unpaid waiting periods, recovery, recurrence, and death. The benefits in b(t) must be defined by contract and possibly depend on time in state or elapsed duration; when a reward is path-dependent, the state space must be expanded or the semi-Markov process used explicitly. For instance, two claimants both currently disabled may have different future benefit limits because one is at month six and the other at month 60. A state vector that records disability status alone loses this liability information. Markov valuation is therefore exact only for a state description sufficient to determine future conditional payments.

The relationship between claim counts and financial liability can be expressed by grouping prospective new claims by onset time and contract class. Under an illustrative constant force of new-incidence (\lambda), no active-state decrements, a constant rate discount (\delta), and fixed conditional new-claim value (V_D), expected present value of claims entering during H years is (\lambda V_D(1-e^{-\delta H})/\delta) per initially active life under a Poisson-arrival idealization. If only one qualifying incidence can occur before exit from A, active-state survival changes the factor to (\lambda V_D[1-e^{-(\lambda+\delta)H}]/(\lambda+\delta)). These expressions follow different event-generating assumptions and must never be substituted for one another merely because incidence is small. Recurrence and policy termination require an expanded model.

Reserve components and the role of model boundaries

A benefit reserve for an existing disabled claimant differs from the aggregate liability for healthy active policyholders. Existing claim reserves integrate the conditional remaining benefit process from the valuation date, given current disability duration and information known about the claimant. Active-life reserves add the probability of future onset and future claim values. Gross premium reserves can additionally include future premiums, expenses and other contractual cash flows; statutory reporting may impose additional margins, prescribed bases, or financial-reporting constraints. A claim model cannot be deemed compliant with a jurisdiction's reserve regime solely because its transition intensities are statistically calibrated.

Sickness, injury, mental-health, and musculoskeletal claims can have materially different time patterns under a given eligibility structure, and the same medical diagnosis can have different occupational earnings effects. A reserve by occupation and diagnosis must use credible segmentation and report the direction of sensitivity if cells are sparse. Duration-conditioned claim reserves should preserve survivors' selection profile; replacing them by a fixed inception duration curve at every valuation date introduces timing bias. This follows directly from conditional survival algebra: (\Pr(T>u+s\mid T>u)=S(u+s)/S(u)), which generally differs from (S(s)) when the hazard depends on duration.

Benefit Valuation

The cash-flow object

A disability model contributes to pricing and reserving through expected future cash flows, not merely through event counts. Let b(u) be the benefit intensity payable at disability duration u, (v(u)=\exp[-\int_0^u\delta(s)ds]) a discount factor, and (I_{\mathrm{pay}}(u)) an indicator of contractual benefit eligibility. Conditional on being in the disabled state at valuation time zero, an expected present value can be written

VD=E[∫0Lv(u)b(u)Ipay(u) du∣X(0)=D].V_D=\mathbb E\left[\int_0^L v(u)b(u)I_{\mathrm{pay}}(u)\,du\mid X(0)=D\right].

Here L is the model horizon or a maximum benefit-duration limit, and payment frequency is represented continuously for analytic convenience. If payment is fixed at annualized rate b, eligibility coincides with residence in D, and discount force (\delta) and cause-specific exit intensities are deterministic, then

VD=b∫0Lexp⁡[−∫0u{δ(s)+ρ(s)+μD(s)+κ(s)} ds]du.V_D=b\int_0^L\exp\left[-\int_0^u\{\delta(s)+\rho(s)+\mu_D(s)+\kappa(s)\}\,ds\right]du.

For constant (g=\delta+\rho+\mu_D+\kappa>0), (V_D=b(1-e^{-gL})/g). The formula is a conditional expected value, not a statutory reserve without further allowance for expenses, taxes, benefit offsets, inflation escalation, margin requirements, interest-rate methodology and additional policy provisions. The 2013 individual disability valuation documentation and 2012 group LTD valuation documentation confirm that insurance valuation bases are product-specific.

Verified hypothetical numerical illustration

Consider a hypothetical eligible disability claim paying $2,400 monthly, represented as a continuous annual benefit of $28,800, with a maximum remaining benefit term L=5 years. Assume independent constant recovery intensity (\rho=0.45) per year, disabled-state mortality intensity (\mu_D=0.035), additional contractual termination intensity (\kappa=0), and constant force of discount (\delta=0.035). Then g=0.520 and

VD=28,8001−e−0.520×50.520≈$51,271.V_D=28{,}800\frac{1-e^{-0.520\times5}}{0.520}\approx\$51{,}271.

The five-year discounted disability occupancy weight is (e^{-2.6}\approx0.0743), while the physical probability of continued disability excludes discounting. The probability of recovery before five years in this two-exit example is ((0.45/0.485)(1-e^{-0.485\times5})\approx0.846), and the probability of death before recovery by five years is ((0.035/0.485)(1-e^{-0.485\times5})\approx0.0658); the remaining probability is continued disability at five years, (e^{-2.425}\approx0.0885). Discounting appears only in the present-value calculation and not in the state-transition probabilities. The physical claim-continuance probability is (e^{-2.425}). The discounted occupancy weight, used for valuation, is smaller because it incorporates the time value of money. This distinction is essential for actuarial interpretation.

A further sensitivity calculation changes recovery intensity while retaining all other assumptions: a recovery intensity of 0.30 gives g=0.370 and a five-year value approximately $65,599; a recovery intensity of 0.60 gives g=0.670 and a value approximately $41,477. These values are hypothetical, calibrated to no insurer dataset. If there are 1,000 eligible active lives with a hypothetical one-year inception intensity of 0.004 per year, an elementary first-order expected inception count is four. Multiplying four by the first conditional claim value yields about $205,084 of expected claim value at the times of incidence before onset-time discount, existing deaths, policy attrition, premium effects or repeated-claim risk. A fully prospective reserve requires the correct inception-time integration and the joint state process instead of that shortcut.

Prospective valuation from active exposure

Let (p_{AD}(0,t)) denote the probability of being in the model's new-disability event configuration at t, or express first-entry probability density as (S_A(t)\lambda_{AD}(t)), where (S_A) accounts for all exits from A. When claims begin at t, the present value of a future new claim stream over time horizon H is schematically

VA=∫0He−∫0tδ(s)dsSA(t)λAD(t)VD(t) dt,V_A=\int_0^H e^{-\int_0^t\delta(s)ds}S_A(t)\lambda_{AD}(t)V_D(t)\,dt,

provided that each first incidence generates one separate future liability (V_D(t)) and the risk set is appropriate. If relapses, return-to-work states, or multiple policies contribute future liabilities, a full multistate reward model is preferable. With state reward vector r(t) and transition matrix P(0,t), a general expected discounted benefit is (\int_0^H v(t)\pi(0)P(0,t)r(t)dt) under the stated Markov construction. Claim reopening, waiting-period restart, contingent partial benefits and survival-dependent offsets become explicit state rewards or transition cash flows.

Contractual rules create important discontinuities. Own-occupation protection may change to any-occupation after a specified duration; a claim reaching retirement age may cease to pay; a residual earnings benefit may shrink as employment income resumes; an indexation rider may increase b(u). These are modeled as rule-based boundaries, state transitions, or changing reward functions as appropriate. A universal recovery table with a single unchanged benefit rate cannot capture these provisions by a casual scaling factor.

Recoveries, discounting and sensitivity

For the constant-hazard case, the derivative of (V_D) with respect to recovery intensity is negative when benefit is positive: (\partial V_D/\partial\rho=-b\int_0^L u e^{-gu}du<0). A similar derivative holds for death and other exits under the same cash-flow assumptions. A higher recovery hazard reduces expected disability payments; it may also increase insured return-to-work incentives or generate rehabilitation expenses not captured in the benefit stream. A decline in discount rates raises value under a positive benefit stream, while increased mortality can reduce pure disability-income outgo but may change other life or survivor liabilities. These are conditional model effects, not a recommendation for social or claims policy.

The relative sensitivity to duration assumptions becomes especially pronounced for long-duration claims because early recoveries remove a large part of the occupancy distribution. Model validation should report sensitivity by elapsed duration, diagnosis, occupation, benefit horizon and disability definition rather than hiding the structure in an aggregate reserve percentage.

Model Uncertainty and Validation

Structural and statistical uncertainty

An actuarial model is exposed to measurement error, stochastic claim volatility, estimation variance, omitted variables, model-form error, parameter drift, contractual ambiguity, and distribution shift. Structural uncertainty includes whether relapse is a new incidence or continuation; whether death and recovery are separate exits; whether missing claims are censored or terminated; whether benefit payments continue during partial employment; and whether the chosen state aggregation remains valid. Statistical uncertainty covers confidence intervals for exposure-based intensities, model coefficients, duration curves and future cash flows conditional on chosen structure. Uncertainty in interest rates adds an additional asset-liability component.

The ASOP 56 modeling standard identifies the need to consider intended purpose, model structure, data, assumptions, controls, validation and communications within its scope. The ASOP 23 data-quality standard addresses limitations of underlying data; and the ASOP 25 credibility procedures address the predictive value of sparse or heterogeneous experience. These institutional standards support process-level obligations; they cannot certify the predictive calibration of a specific claim model in the absence of valid evidence.

Numerical and logical validation tests

A complete technical check should establish the following invariants: transition probabilities are nonnegative and sum to one across exhaustive states; cause-specific cumulative incidences plus state-survival sum to one; the generator has rows summing to zero; probability of benefit payment lies between zero and one; present value of bounded nonnegative payments is nonnegative and at most the undiscounted maximum contractual payments; for fixed positive benefit stream, higher exit intensity weakly lowers value when all else is held equal; the claim value converges to (bL) as discount and exit intensities jointly approach zero; and an infinite horizon with positive constant g has value (b/g). For piecewise hazards, continuity of S(u) and integration over all interval boundaries must be checked. In a multistate model, the product-integral or matrix-exponential solution must agree with independent numerical integration to documented tolerance.

The internal numerical example is independently recalculated using standard exponential functions and automated assertions of closed-form arithmetic, including the distinction between physical continuance and discounted occupancy. No evidence is offered for an empirical performance measure without model fitting and holdout data. Claims of predictive superiority would require an actual out-of-sample experiment.

Validation with observed experience

An effective empirical protocol begins by preserving time-ordering, coding valid exposure, and separating inception cohorts from existing disabled claims. It then partitions experience by calendar periods, claimant cohorts and elapsed durations; fits cause-specific hazards or a consistent multistate likelihood; evaluates the calibration of observed against expected transitions; and assesses goodness of fit within major contract and diagnosis classes. Because claim history is longitudinal, random splitting at the claim-month level can leak information from the same claimant into both training and validation. Split by person or independent claim cohort as appropriate, and use forward-time validation to test temporal drift.

The distinction between net and crude recovery probabilities must be tested by comparing cause-specific cumulative incidences to the full state-occupancy model. If the model uses Aalen–Johansen estimates, its assumptions about observation and independent censoring require review; a multistate analysis tutorial explains the practical relationship among state probabilities, transition probabilities and cumulative incidence. Standardized mortality comparisons alone cannot establish appropriate recovery calibration. The current group LTD termination research is an appropriate external benchmark at the level of its actually accessible definitions and strata, but paywalled data are not silently treated as available.

Stress testing, credibility and governance

Sensitivity analysis should vary incidence, cause-specific recovery, disabled-state mortality, relapse, diagnosis and occupation mix, benefit offsets, interest rates, waiting periods, maximum duration and administrative practice. One-variable stresses reveal directional effects; combined scenarios reveal interactions. For example, worsening incidence combined with slower recovery increases both the arrival of new claims and length of benefit occupancy; a projection that stresses only the first factor understates the joint adverse scenario. The total effect remains portfolio- and contract-dependent.

Low-frequency strata need credible uncertainty ranges and appropriate partial pooling, not arbitrary smoothness presented as precision. The actuarial practitioner should document reference data, blending weights, assumptions, overrides and version history consistent with credibility practice guidance. A change in insured occupational mix or claim adjudication practice can be tested as a structural break before attributing it to disease incidence. A comparison between a group LTD pool, a UK individual income-protection table, and a public disability program needs an explicit transferability bridge; none has a general entitlement to serve as the others' direct rate basis.

Ethical and methodological interpretation

Observed disability and recovery are filtered through contractual definitions and assessment processes. Treating claim termination as complete physical recovery can misrepresent the claimant's continuing limitations and weaken research reliability. Differences among diagnosis groups, employment circumstances and benefit definitions may reflect genuine risk, administrative measurement or social barriers. The WHO ICF framework supplies a functioning-sensitive vocabulary; the OECD analysis of disability and work contextualizes return-to-work outcomes. The appropriate scientific conclusion is to label each measured outcome precisely and assess its intended use without conflating disability status, claim status and social participation.

Scope of this research and reproducibility

This dossier supplies primary-institution comparisons, verified publicly disclosed historical and 2026 experience-study identities, mathematical derivations and a hypothetical numerical evaluation. It does not estimate a new proprietary incidence or termination table, analyze restricted insurer-level records, conduct a causal intervention study, or certify future actuarial rates for a specific jurisdiction. Subscription-only CMI and SOA data were examined at the level of publisher descriptions rather than their undisclosed tables. The cited scholarly books and journal articles support established methods based on publisher metadata, abstracts and accessible institutional descriptions, without a claim that every paywalled page was inspected in full.

The reproducible arithmetic and source-to-claim records are documented in the accompanying EVIDENCE.json, and metadata integrity is recorded in PASSPORT.json. In a future empirical implementation, anonymized event-history inputs, covariate definitions, scripts, environment versions, and observed validation outputs would be necessary for a stronger empirical replication claim. This analysis makes no new theoretical system; its contribution is the explicit integration of established methods with contract and evidence boundaries, and the demonstration of their consequences for disability benefit valuation.

Conclusion

Disability incidence and recovery are distinct transitions in a contractual and historical risk process. Their observed rates depend on eligibility definitions, exposure denominators, claimant and policy selection, waiting periods, elapsed disability duration, competing exits and censoring. The governing probabilistic relationships are explicit: incidence intensities originate in eligible active exposure; recovery and death hazards compete within the disabled state; duration-specific survival integrates all outgoing hazards; recurrence requires person- and episode-aware histories; and expected disability payments are discounted rewards of time spent in payable states.

Three empirical findings ground the synthesis. First, the SOA/LIMRA incidence and termination studies are large but distinct studies of active and disabled exposures, so their aggregate counts require correctly matched units and populations. Second, the US group and individual disability valuation tables embody different experience and reserve purposes; regulatory minima and historical tables are not automatically current best-estimate forecasts. Third, the CMI IP11 exposure correction demonstrates that an apparently precise graduated rate can be materially distorted by an exposure denominator even when its mathematical calculation is internally consistent.

The original analytical derivations show the direction and magnitude of disability claim value sensitivity under declared assumptions, and distinguish physical transition probabilities from discounted occupancy weights. The working numerical example establishes how recovery, mortality and interest contribute differently to prospective liability. The broader scientific consequence is a requirement of alignment: every modeled transition, exposure, benefit and supporting source must describe the same insured process. This is an evidence-based synthesis of actuarial science and risk research intended for a full research dossier, without a claim of a new authorial theory or clinical, financial, or individual insurance advice.

References