RESEARCH DOSSIER
Disability Insurance Recovery Relapse and Claim Termination Models
Abstract. Disability income insurance is a time-dependent arrangement in which functional capacity, earnings, contractual eligibility, benefit payment, mortality, and recurrent disability evolve along related but distinct paths. This monograph investigates the measurement and modeling of recovery, relapse, and claim termination across insured disability populations. It reconstructs the observational object of a claim episode; distinguishes clinical recovery, sustainable return to work, administrative closure, benefit cessation, death, and benefit exhaustion; and evaluates their consequences for transition intensities, recurrence estimates, incurred-claim reserves, and comparative regulation. Its empirical foundation comprises Society of Actuaries (SOA) and LIMRA disability-experience research, published occupational absence cohorts, the World Health Organization's functioning framework, the United States' disability benefit and claims-procedure rules, Canadian reinstatement provisions, United Kingdom disability and workplace guidance, and the IFRS Foundation's insurance measurement documentation. Competing-risk and multi-state mathematics are developed with explicit risk sets, cause-specific hazards, semi-Markov duration structures, observation windows, and a worked hypothetical valuation. The 2015–2024 SOA/LIMRA group long-term disability study reports over 43 million exposed life-months, approximately 937,000 recoveries, and more than 181,000 deaths from eighteen contributing companies; its publicly accessible summary does not furnish all rate cells or the underlying individual records. The synthesis establishes that reported termination experience cannot be interpreted as a single biological recovery process; that relapse cannot be estimated consistently from claims closed permanently at first termination; and that reserve sensitivity depends on both competing exits and contractual routes back into benefit payment. The proposed calculations are applications of established methods, not new actuarial doctrines or original estimates from confidential insurer files. Limitations concern nonuniform products, selected cohorts, incomplete external follow-up, changing benefits rules, restricted detailed experience access, and absence of external peer review.
Keywords: Disability income insurance; long-term disability; recovery hazard; disability relapse; recurrent disability; claim termination; competing risks; survival analysis; multi-state models; semi-Markov processes; work incapacity; partial return to work; diagnosis; duration dependence; exposure measurement; actual-to-expected ratios; benefit reinstatement; claims adjudication; contract wording; claim reserves; sensitivity analysis; selection bias; actuarial experience; disability law; IFRS 17.
Introduction: Scientific Object, Questions, and Method
A disability insurance claim is a history rather than a point event. The insured person may lose the capacity to perform essential work duties, satisfy a waiting period, receive a benefit, regain some capacity, resume work with accommodations, experience a recurrence, and again meet a contractual disability definition. Death, expiration of the maximum benefit period, an age-based limit, changing eligibility evidence, exhaustion of coverage, or an administrative decision may end payments at any stage. A dataset reporting only the date on which a claim closed compresses that history into an event code. The research task is to recover the distinctions necessary for valid explanation and prediction.
The object of this dossier is the dynamics of disability income claim continuation and termination, concentrating on recovery, relapse, and recurrent entitlement. The unit of actuarial analysis varies with the estimand: an insured life at risk of first disability; an active payable claim exposed to termination; an episode of clinically defined incapacity; a return-to-work spell; or a person with multiple episodes across contracts and calendar years. Those units must be specified before estimating a rate. The WHO International Classification of Functioning provides an internationally used account of functioning in bodily, activity, participation, and environmental domains. Insurance eligibility operationalizes a narrower contract- or program-defined question, generally tied to occupational incapacity and documented criteria.
The principal questions are: how does an eligible active claim leave its risk set, and how does the allocation among recovery, death, and other exits change with duration? What distinguishes recurrent disability from a new independent incidence event, and what observations permit identification of relapse? How do recovery and recurrence affect the actuarial present value of future payments? Which assumptions remain defensible when comparing group and individual disability contracts or public systems across different jurisdictions? What claim facts, coding rules, and sensitivity tests are necessary before assigning a quantitative model to a real population?
The inquiry applies comparative reading of official source documents, a structured appraisal of insurer experience studies, the mathematical analysis of competing risks and repeated transitions, a worked hypothetical reserve calculation, and a source-specific account of uncertainty. The SOA disability experience catalog distinguishes group, individual, and credit disability experience, including different valuation-table and study generations. This is an integrated documentary and actuarial methodological inquiry. It does not claim that an unobserved proprietary claims dataset has been examined, that a formal systematic review has been performed, or that a population-level causal effect of rehabilitation has been identified.
The manuscript uses established mathematical notation. A reported rate is treated as an empirical rate only where a source supplies observations or values. A hypothetical parameter is explicitly labeled and is used solely to show the operation of a method. The distinction matters because detailed SOA/LIMRA rate cells may be contained in subscription packages whereas the publicly available summary reports aggregate exposures and event counts. Financial, legal, and medical claims are attributed to the relevant institutional or research source at the point they enter the argument.
Conceptual Foundations and Historical Development
The Observational Unit: Claims, Persons, and Episodes
An insured life is a natural unit for incidence: persons who have not yet entered a specified disabling state contribute exposure while covered and at risk. An active claim is a natural unit for termination: persons already recognized under the relevant claim definition contribute time while the claim is eligible to terminate. An episode is a contiguous interval of defined incapacity; its beginning and ending need independent rules for gaps, reduced work, changes in diagnosis, waiting periods, and insurer decisions. A claimant may generate two episodes but one reopened claim, two claim numbers for one uninterrupted episode, or an episode that spans several administrative files.
These distinctions generate three different denominators. Incidence rates count first or qualifying new claims per insured-life exposure. Termination rates count exits per disabled-claim exposure. Recurrence rates count subsequent episodes per eligible follow-up time after a return or recovery state. Combining them creates a misleading probability. If a claimant exits the claim register at an observed date, follow-up can cease even while medical vulnerability persists. A recurrent episode then enters the records only if the claimant remains insured, files again, and satisfies the new or recurrent claim criteria.
An event dictionary should separate onset date, first absent workday, claim notification, elimination-period completion, benefit eligibility determination, first payment, functional improvement, employment resumption, final payment, contractual termination, and file closure. Each date describes a different process. Late notification or retrospective approval can yield delayed entry into the actuarial risk set. Backdated administrative changes can alter event ordering. Where a termination reason is revised on appeal, the original determination and the adjudicated correction should remain distinguishable.
Function, Eligibility, Payment, and Employment
A person can remain disabled under one legal or clinical definition while partially employed. Recovery sufficient for a particular job can coexist with impairments in daily activity. Conversely, successful treatment may improve symptoms before an employee obtains sustainable work or satisfies income thresholds. The WHO's ICF classification establishes a vocabulary for bodily function, activities, participation, and context; it does not itself impose private insurance claim eligibility rules.
For a reproducible dataset let F(t) denote measured functioning, E(t) contract eligibility, W(t) work participation, P(t) benefit payment, and A(t) administrative file status. A model of claim cessation should specify whether its event is E=0, P=0, or A=closed. Those events can occur at different times. A claimant may have P=0 during a payment suspension and E=1 subject to reinstatement; employment may resume with reduced hours while residual-disability benefits continue. Differences between own-occupation and any-occupation clauses, earnings offsets, rehabilitation incentives, recurrence clauses, and maximum benefit periods are contractual dimensions, not intrinsic biological parameters.
United Kingdom workplace return-to-work guidance explicitly recognizes phased return, adaptations, and fluctuating conditions. These practices reinforce a measurement point: binary 'working' labels discard information about intensity, accommodation, persistence, and health constraints. If an insurer records return to work as recovery, that definition needs to be documented in the statistical analysis and interpreted against the contract, not presented as a medical conclusion.
Historical Development of Termination Experience
Actuarial disability valuation developed through tables that distinguish periods since disability onset and major termination causes. The Society of Actuaries' 2008 Group LTD experience work and subsequent 2012 valuation-table framework provided established benchmarks for group disability experience. The 2013 Individual Disability Income Valuation Table was approved by the NAIC in August 2016 as a new statutory minimum reserve morbidity basis for individual disability income policies, subject to applicable implementation. Its population and intended use differ from group LTD termination studies.
The 2009–2017 Group LTD study supplied a more recent historical comparison, with published preliminary materials, data structures, appendices, and exposed methodologies. The 2026 group LTD termination update covers 2015–2024 experience and allows a prospective comparison with earlier studies. A change in actual-to-expected ratios is an empirical comparison to a particular basis with a particular classification and mix of claims. The observed difference must be decomposed before being interpreted as changing health, insurer management, labor-market conditions, or economic behavior.
Individual disability income evidence has developed on a related but distinct pathway. The SOA's 2006–2014 claim termination analysis compares experience with 2013 IDIVT termination rates. Beal and Corcoran's 2025 diagnosis analysis further separates recovery and mortality patterns by diagnosis, sex, elimination period, benefit period, occupation class, and claim year. The design explicitly reaches beyond the earlier restricted cohort and illustrates why a single pooled recovery schedule can perform poorly for differently constituted portfolios. Results from 2006–2014 must not be automatically treated as current 2026 rates.
Claim Recovery Hazard
Risk Sets, Cause-Specific Exits, and Estimands
Let T be time from a declared disability origin to first termination from a specified active-claim state. Let J indicate the mutually exclusive first exit type: R for recovery, D for death, O for other administrative or contractual cessation. For a continuously observed claim, the cause-specific intensity is
[\lambda_j(t\mid X)=\lim_{\Delta t\to0^+}\Pr(t\le T<t+\Delta t,J=j\mid T\ge t,X)/\Delta t.]
Here X may include age at onset, occupation, diagnosis, elimination period, benefit form, contract provisions, calendar cohort, and reporting history. The total first-exit intensity is (\lambda_\Sigma(t)=\lambda_R(t)+\lambda_D(t)+\lambda_O(t)). Conditional survival in the active state is (S(t)=\exp[-\int_0^t\lambda_\Sigma(u)du]). The cumulative incidence of recovery by t is (F_R(t)=\int_0^tS(u-)\lambda_R(u)du). Thus the probability of recovery is not obtained by applying the recovery hazard in isolation when deaths and other exits compete.
This distinction is central to the Putter, Fiocco, and Geskus (2007) tutorial, which develops competing-risk cumulative incidence and state-transition probabilities. In practice, a death is not ordinary random censoring of recovery: death eliminates the opportunity for subsequent recovery in the same life. Treating deaths as independent noninformative withdrawals and reporting one minus a Kaplan–Meier curve for recovery would generally answer a hypothetical net-risk question, rather than the observed crude recovery probability. Actuarial reserving of the actual portfolio requires the competing exits and the relevant payment rules.
The independent-censoring condition also deserves attention. If files disappear from follow-up because an insurer changes systems, coverage terminates, a claimant changes employer, or records become inaccessible, censoring can be related to disability severity or expected recovery. An estimator calculated as if such censoring were random can be biased. Corrective methods require information about why observation ended, not a generic statistical incantation.
Duration, Case Mix, and Exposure Construction
Duration can be measured from disabling onset, first absence, first claim notification, completion of the elimination period, first paid month, or observed database entry. An early-duration recovery curve estimated from first payment necessarily conditions on surviving the elimination period and adjudication process. Its risk set excludes persons who recovered rapidly before a claim became payable. A late-duration portfolio increasingly contains claims that have survived earlier recovery opportunities, so its diagnosis and severity distribution can change even if each individual's underlying conditional hazard remains constant.
For monthly grouped data, a cause-specific rate can be presented as (q_{j,m}=d_{j,m}/e_m) when d is a count and e a carefully defined exposed-claim denominator. When competing causes occur within the month, a rate estimation method must account for fractional exposure and the order of exits. A common first approximation based on deaths and withdrawals weighted by time at risk may be justified for small monthly probabilities; an exact likelihood or multistate life-table approach is preferable when event rates are substantial. A rate described as recoveries divided by annual starting claims is not equivalent to one divided by exposed life-months, nor to a probability of recovery conditional on surviving to the month.
Case-mix stratification should account for age, duration, cause, occupational class, sex where analytically justified, elimination period, policy design, and observation cohort. Diagnosis codes may be reassigned or consolidated; employment categories can reflect benefit eligibility rather than causal exposures. Calendar trends also combine effects of economic cycles, medical practice, benefit management, and selection of contributing carriers. Splitting data indefinitely yields unstable low-count cells; pooling indiscriminately hides heterogeneity. The model must choose a workable compromise and report the effective exposure and uncertainty of each relevant contrast.
Interpreting the SOA and LIMRA Evidence
The SOA/LIMRA 2015–2024 LTD termination study was published September 30, 2026. Its public description reports eighteen contributing companies, covering over 82% of industry sales in the study period, more than 43 million life-months of exposure, approximately 937,000 recoveries, and over 181,000 deaths. These are valuable facts about observational scale and recorded outcomes. They are not a calibrated set of monthly duration-specific recovery probabilities. The subscription-based standard package includes detailed analysis and dashboards, and the available public statement does not allow independent replication of each cell, diagnosis split, or experience adjustment.
The arithmetic ratio 937,000 / 43,000,000 is about 0.0218 recoveries per exposed life-month; 181,000 / 43,000,000 is about 0.00421 deaths per exposed life-month. Because all source figures are explicitly approximate or lower bounds and the denominator is expressed as 'over,' these ratios are only illustrative aggregate quotients, not exact published incidence measures or calibrated individual hazards. They combine many durations and contracts and cannot be annualized into a representative claim's recovery probability. They are reported here to show the dimensional scale of the data, not to estimate a portfolio basis.
The 2019 predecessor report used 2009–2017 experience. Direct comparison with the 2026 cohort must reconcile overlapping calendar years, changes in contributing insurers, data requests, exit-code classifications, duration origins, and any updates to expected rates. The study comparison is scientifically informative only when the analyst distinguishes a change in risk from a change in measurement and membership.
The SOA individual disability analysis specifically measures claim termination against an individual disability valuation framework. The 2025 Phase 2 report evaluates multiple dimensions previously omitted or restricted by shorter eligibility definitions. These investigations support a general conclusion about heterogeneity and transport risk; they do not provide permission to transplant their numerical parameters into group LTD business without product and population reconciliation.
Estimation, Uncertainty, and Model Diagnostics
With individual event histories, a nonparametric cumulative incidence curve can be estimated using cause-specific increments and the probability of still being in the active state; with sufficient data, the Aalen–Johansen construction produces transition-probability estimates. A cause-specific Cox model may describe covariate associations with recovery among those still at risk. A subdistribution-hazard approach instead targets a cumulative-incidence regression, which is useful for a different predictive question. The choice of estimand should precede the choice of estimator.
For claim durations, piecewise exponential rates or flexible splines may be preferable to a single proportional-hazards assumption. There may be a discontinuity at elimination-period completion, eligibility reassessment, an occupational-definition switch, or a maximum payment duration. Such discontinuities can reflect contract terms or adjudication events rather than physiological changes. Fit a model with event-date diagnostics and compare predictions on an out-of-time cohort to ascertain whether its structure persists.
Report uncertainty in cumulative recovery probability, death probability, and continuation probability separately, together with exposure volumes. Confidence or predictive intervals should account for dependencies among repeated claims by one individual, insurer clusters, temporal shifts, and multiple covariate comparisons when relevant. Actual-to-expected ratios must disclose the expected table, its version, the exposure standardization, and any excluded classes. Formal calibration compares predicted and observed events across duration, diagnosis, age, and calendar segments, with an explicitly defined tolerance and sufficient sample size. A visually close total curve can conceal large opposing errors in the components.
Relapse and Duration Dependence
Recurrent Episodes and Their Definitions
Recovery terminates a specific interval of insured incapacity under an agreed definition. Relapse is a subsequent loss of function or ability to work after an interval of recovery, often associated with the same condition; recurrent disability is the reappearance of a qualifying covered state. The terms should have separate operational definitions. A person whose benefit was closed after a temporary return to work and reopened within a contractually allowed window may have one recognized claim with a recurrent spell. A person who became disabled for a different diagnosis after a long healthy interval may have a new incidence event. A person whose benefit was suspended pending an income verification review has not necessarily recovered at all.
A recurrence study requires an observation system that continues after the first recovery or cessation. A termination-only dataset conditioned to stop observing at closure estimates time to the first exit. It cannot independently estimate the hazard of return from recovered employment to disability. Even if a new claim enters a different policy system, lack of person-level linkage can convert a recurrence into an apparently independent first claim. Claims administrators may reopen an old identifier or issue a new identifier as a business process; record structure cannot by itself adjudicate episode identity.
The time origin after recovery should also be defined. A return-to-work episode might start on the first paid workday, on completion of a trial work period, or on confirmation of permanent eligibility termination. Recurrence risk often depends on elapsed time since actual return, past diagnoses, accumulated number of prior episodes, and the quality of workplace adjustment. Analytically, a recurrence-free interval of three days and a five-year interval cannot be counted as equivalent person-time without a model. An episode-based count over a fixed insured cohort must account for changing coverage and calendar observation.
Semi-Markov and Multi-State Specifications
An elementary disability model has states A (active and insured), D (eligible disabled on claim), R (functionally recovered with qualifying employment or relevant activity), P (partially disabled or supported work, where the policy recognizes it), and X (dead). Additional administrative states B (payment suspended or under review) and C (coverage exhausted or contractually ended) may be needed. Each pair of legally and clinically possible transitions has its own intensity. For instance, D→R can reflect recovery, D→X death, D→P partial improvement, R→D recurrence, P→D deterioration, and P→R further improvement. C must be considered carefully: being outside one contract's coverage is an administrative status, not automatically a health state.
For a continuous-time Markov process with time-varying intensity matrix Q(t), off-diagonal elements (q_{ij}(t)) represent conditional instantaneous rates of moving from state i to j. Diagonal entries satisfy (q_{ii}(t)=-\sum_{j\ne i}q_{ij}(t)). The occupancy probability row vector evolves as (p'(t)=p(t)Q(t)). A fixed Q is a strong simplification: after disability a claimant's chance of recovery commonly depends on elapsed duration in D, while after return to work the chance of recurrence can depend on time in R. A semi-Markov model or a clock-reset transition hazard (q_{ij}(t,u)), where u is duration in the present state, captures that dependence. A history-dependent recurrent-event framework may also condition on episode number and cumulative prior disability time.
Consider a simple two-way D↔R model over short periods with death X absorbing. Let (q_{DR}=a), (q_{RD}=b), (q_{DX}=m_D), and (q_{RX}=m_R). The probability of being disabled at future time t cannot be calculated from the first recovery hazard alone, because recovered individuals may reenter D. If contract eligibility permits payment after reinstatement, the expected payment occupancy depends on both a and b and on the applicable rule for new waiting periods. If later reentry is excluded by policy exhaustion, a state representing ended coverage must absorb that transition from the covered financial process. Treating all recovery as permanently absorbing will understate expected benefits where covered recurrence is material; treating all recurrence as automatically payable may overstate them when eligibility is conditional.
The transition model also makes the identification problem explicit. Observed D→R transitions estimate recovery among continuing claims; observed R→D transitions need follow-up among recovered persons. When person identifiers fail to link records or work activity is unknown, a rich multistate mathematical object remains underidentified. The practical solution is to improve episode linkage and define sensitivity bounds, rather than add unobservable parameters until a desired curve is obtained. The general theory of transitions and competing outcomes follows Putter et al. (2007), applied here to insurance rather than hospital event histories.
Recurrence Evidence and Limits of Transportability
Published occupational evidence supports the prediction that past absence is informative about later absence. Laaksonen, He, and Pitkäniemi's Helsinki employee study followed 6,934 municipal employees and found that earlier sickness-absence episodes predicted subsequent episodes, with stronger associations for longer previous absences. A separate Dutch employee follow-up observed 53,990 workers and documented elevated future sickness absence among those with frequent or long-term prior absence. These results make recurrence a serious scientific phenomenon in work incapacity.
The evidence also illustrates a transportability boundary. Sickness absence is not necessarily a claim under a private long-term disability income contract. Public employees may receive paid sick leave for durations shorter than commercial elimination periods; medically documented absence may include employees who never seek an insurance benefit; employment retention and benefit generosity vary. The studies support questions and covariates for model design, but their ratios should not be applied unadjusted as insured LTD relapse parameters. A claim-specific study must derive recurrence probabilities from its own risk population or construct an explicit, empirically tested bridge between the populations.
The following causal pathways must be kept open as alternatives rather than collapsed into a single relapse explanation: residual disease severity; recurrent episodes inherent in a diagnosis; delayed treatment effects; occupational hazards; limitations in work accommodation; labor-market demand; loss of insurance coverage; and administrative differences in recognizing disability. Repeated absence can also be a marker of underlying vulnerability rather than a consequence of the earlier spell. Observational associations do not by themselves identify the effect of a policy incentive or rehabilitation program.
Employment Transitions and Policy Reinstatement
Disability benefit rules influence which transitions are visible. The U.S. Social Security Administration's trial-work and eligibility rules permit a test of work and, in specified circumstances, a 36-month extended period of eligibility during which payments can resume if earnings fall below the applicable threshold and a disabling impairment persists. The agency's current work guidance explains a nine-month trial work period and income limits; those values and thresholds vary by calendar year. Payment suspension in this system does not equate straightforwardly with full medical recovery.
Canada's CPP disability automatic reinstatement provides a particularly clear institutional illustration of recurrence as a legally relevant event. A beneficiary whose payments stopped after return to work, volunteering, or study may be entitled to reinstatement for the same or related recurring disability within two years, subject to stated requirements and the application window. The official explanation also describes a distinct fast-track reapplication route within five years. These are public-program provisions and are not private LTD contract clauses; they establish an example of how legal design determines whether recurrence represents reopened entitlement or a fresh application.
The United Kingdom Equality Act 2010 addresses long-term effects and recurrence in the determination of disability as a protected characteristic. Schedule 1 treats an effect as continuing in specified circumstances if substantial adverse effects have ceased but are likely to recur. This statutory treatment serves equality law and cannot be substituted for a private disability insurance definition. It nevertheless demonstrates the legal incoherence of assuming that every observable symptom-free period proves permanent absence of disability in all policy contexts.
An institutional comparison should code: original disability standard, work trial, benefit offset, minimum earnings, reinstatement window, requirement that the same condition recur, eligibility reassessment, appeal rights, and maximum benefit period. Only after mapping these fields can a model separate medical recovery, employment outcome, and financial claim termination. National regulations and private wordings create different observation processes even where underlying disease histories are similar.
Reserve and Continuation Effects
Prospective Cash Flows and Transition Probabilities
For an active payable claim valued at time 0, a simplified expected present value of future monthly benefits is
[V_0=\sum_{m=1}^{M}v_m,B_m,\Pr{\text{eligible and payable at }m\mid\mathcal{H}_0}+E_0[\text{eligible expenses}],]
where (v_m) is the appropriate discount factor, (B_m) is the benefit net of contractually applicable offsets or taxes where relevant to the insurer's obligation, M is the contractual horizon, and (\mathcal{H}_0) includes the observed claim history. This is a transparent expected-cash-flow identity; actual regulatory and financial reporting reserves may incorporate expenses, discount methodology, risk adjustment, contract grouping, reinsurance, and prudential margins. The IFRS 17 overview requires current estimates of fulfilment cash flows consistent with available information, together with an adjustment for non-financial risk in its specified group measurement. The standard's key terms distinguish the liability for incurred claims and remaining coverage.
Where recurrence is not covered after final recovery, the payment probability can be computed from the first-exit survival curve: (\Pr(D\text{ payable at }t)=S_D(t)) adjusted for pauses, benefits, and legal exits. Where reinstatement or linked recurrent episodes are covered, payment probability arises from the full transition process and the contract-dependent payment indicator. A return-to-work state can reduce payments without closing future eligibility. The actuarial value can therefore depend more on the legal definition of reinstatement than on a simple change in the estimated first recovery rate.
Benefits may also be contingent on income or function level. Let (b_i(t)) be the payable monthly amount in state i. Then (V=\int_0^T v(t)\sum_i p_i(t)b_i(t),dt) is a continuous-time version of the occupancy valuation, with an additional term for claims expenses and appropriate boundary conditions. A change in claim recovery probability affects p; a change in residual-disability benefit wording affects b; and a change in reinstatement rights changes which transitions are payable. Every valuation audit should distinguish these mechanisms.
Worked Hypothetical Calculations
Consider an illustrative, nonempirical active disability income claim with a level monthly benefit of $2,000, twelve remaining monthly benefit opportunities, a monthly effective discount rate of 0.4%, a conditional monthly recovery probability of 4%, a conditional monthly death probability of 0.5%, and a conditional probability of another permanent covered-claim exit of 0.5%. Assume independent mutually exclusive exits are modeled as a single total 5% monthly exit probability applied at month-end, and that the monthly benefit is paid at the end of each month only if the claim remained active through that month. The 0.95 active-claim persistence is chosen for illustration, not estimated from an experience study.
The resulting present value is (V=2{,}000\sum_{m=1}^{12}(0.95/1.004)^m). Setting (r=0.95/1.004), the finite geometric sum gives (V=2{,}000r(1-r^{12})/(1-r)). For a comparison assumption in which monthly recovery decreases from 4% to 3% and the other exit probabilities remain unchanged, total monthly exit becomes 4%, and (V'=2{,}000\sum_{m=1}^{12}(0.96/1.004)^m). The difference V'−V measures the valuation consequence of the chosen recovery assumption under this very restricted contract and benefit convention. The numerical values are V = $17,061.83, V′ = $18,150.53, and V′ − V = $1,088.70; the calculations use the stated monthly payment timing and rounding only at the result. It is a model sensitivity, not an estimate of any insurer's loss reserve.
For a recurrence extension, suppose instead that the recovered state can return to a payable disability state at 2% per month while the original recovery intensity remains 4%. An illustrative one-month transition calculation from a recovered state would add a 0.02 probability of disability reentry, provided the contract permits it and the observation time covers that transition. It would be incorrect to add 2% directly to the first-exit payment survival probability: subsequent reentry must be conditioned on the proportion in the recovered state, deaths, elapsed duration, waiting periods, and possible repeat recoveries. The appropriate calculation iterates a transition matrix or integrates state-occupancy differential equations. This distinction prevents double counting.
An alternative illustration makes the termination-reason distinction particularly sharp. Suppose 100 comparable active claims are exposed for one month, and four terminate for recovery, one dies, and one closes because the maximum contractual benefit period expires. The observed all-cause termination proportion is 6%; recovery contributes 4% in this elementary complete-month count. Under a competing-risk framework the share of first exits attributed to recovery is 4/6, approximately 66.7%, while the chance that an active claim recovers in that month under the stated simple assumptions is 4%. These figures answer different questions. A portfolio report displaying 66.7% as the monthly recovery probability would have confused composition of terminations with the risk of termination.
Reserve Sensitivity, Tail Risk, and Capital
Greater first recovery intensity ordinarily reduces the duration of payments in an absorbing-recovery model with fixed benefits, mortality, and contract conditions. When recovery itself increases future exposure to covered relapse, the net effect depends on reentry rates and waiting periods. A model that predicts earlier but less durable return to work can generate lower initial reserves and higher later reinstatement payments. The sensitivity should be evaluated over the full covered path and legal horizon rather than by treating the current case closure as the final economic event.
Long-duration claims can dominate incurred-claim reserve uncertainty because continued eligibility generates repeated payments. Occupation, diagnosis, age, contract benefit horizon, and inflation-linked benefits can affect the tail. Reduced mortality increases expected payment duration for some claims even when recovery is unchanged. Declining discount rates raise the present value of distant payments. A stronger operational definition of recurrence may move costs between 'new claims' and 'reopened claims' without changing the underlying economic exposure. These are reasons to reconcile both state probabilities and cash-flow allocation during experience reviews.
For governance, a reserve sensitivity grid should vary first recovery rates, mortality, recurrence, administrative exits, benefit offsets, and discounting both singly and jointly. Correlations matter: a severe diagnosis may simultaneously lower recovery, raise death probability, and reduce potential return-to-work exposure. Imposing independent marginal stresses can produce implausible histories, while ignoring joint deterioration can understate concentration. Testing two- and three-year incurred-claim runoff by duration and cause helps identify whether the model systematically accelerates closures.
The 2013 IDI Valuation Table materials describe a recognized actuarial valuation basis, while the IFRS 17 framework addresses reporting principles for insurance contracts. One should not claim that an illustrative expected present value or a SOA table, by itself, satisfies all jurisdiction-specific statutory solvency requirements or the complete IFRS 17 standard. Valuation applications must identify their governing accounting and regulatory regime, effective dates, policy groupings, and applicable margins.
Data, Legal, and Model Validation
Data Dictionary, Linkage, and Missingness
A defensible insurer dataset must specify person identifier, contract identifier, episode identifier, diagnosis and coding version, onset and notification dates, elimination period, paid period, employment activity, partial benefits, claim closure reason, reopening relationship, coverage expiration, mortality, and reporting cut-off. The person key should link multiple claim episodes; privacy controls should restrict linkage without destroying statistical reproducibility. An anonymized data specification can preserve transition structure when a direct identifier is replaced with a stable and governed pseudonymous key.
The core observation table should have one row per interval during which the claimant's measured state and covariates are stable, together with start and stop times, entry conditions, transition type, and exposure fraction. It should preserve both original and corrected adjudication statuses. A second episode table should identify diagnosis relationship, recurrence waiting rules, reinstatement approval, and intervals out of observation. Simply concatenating policy years is insufficient if the underlying identity or coverage history changes.
Missing cause-of-termination codes deserve special analysis. Some missingness is administrative; some may be informative because difficult-to-classify claims, disputes, or pending medical reviews are more likely to be unresolved. An 'unknown' event should be represented separately from a proven recovery. Sensitivity bounds can first treat unknown outcomes under alternative plausible assignments, followed by covariate-informed multiple imputation or weighting where assumptions are supportable. The missing-data mechanism must remain a stated assumption rather than a claim established by software output.
Linkage may fail when a person changes insurer, group sponsor, or jurisdiction. The absence of recorded recurrence then overstates durable recovery if interpreted as a health outcome. Calendar truncation also affects late claim episodes: a person recovering at the end of 2024 has little opportunity for observed return within a 2015–2024 study window, unlike someone who recovered in 2015. Risk sets and cumulative incidence estimators need administrative censoring and explicit follow-up horizons. Comparisons of relapse across calendar cohorts should use equivalent follow-up windows or appropriate survival methods.
Contractual and Cross-Jurisdictional Interpretation
The legal definition of a covered disability arises from the governing policy or program. The U.S. ERISA disability claims regulation establishes detailed requirements for covered plans' claims procedures, including disclosure, reasons for adverse determinations, review, and impartiality. A benefit termination based on revised medical or vocational findings cannot be converted into 'recovered' solely because the file closed. The administrative event and the underlying functional inference require separate documentation.
The SSA extended eligibility provisions can suspend cash payments during periods of substantial work while preserving a route for reentitlement; the Canada Pension Plan reinstatement rules provide a specified route following recurrence. Both make the original benefit-termination date inadequate as a generic marker of permanent health recovery. These policies also operate in public disability benefit systems whose eligibility tests, financial structures, and legal sources differ from privately insured group LTD contracts.
In the United Kingdom, the Equality Act recurrence provision concerns protected disability status, while official fit-note guidance addresses work advice and gradual resumption. Neither source proves that a particular private insurance benefit remains payable. It would be equally mistaken to conclude that a person lacks a disability protected under equality law simply because a disability insurer's policy ceased to pay. A serious comparison keeps employment protection, public benefit entitlement, clinical function, and private contract benefits on separate axes.
The comparative research problem is partly institutional selection. Different legal standards determine who becomes a claimant, who remains in a study's risk set, which events are coded as termination, and which later episodes enter the same insurer record. A country with generous early rehabilitation benefits may appear to have more recorded recoveries because more people enter the register; another may appear to have fewer relapses because reapplications have different identifiers. Standardization should start with eligibility and observation rules before model coefficients are compared.
Validation, Fairness, and Governance
An appropriate validation framework tests data integrity, statistical fit, forecasting, and decision relevance independently. Data integrity controls include chronology, impossible transitions, conflicting diagnosis codes, duplicate claim identifiers, negative exposure, contradictory death dates, and benefit payment after an absorbing state without a documented legal reason. Model checks include observed-versus-predicted recovery and death cumulative incidences, calibration by duration, diagnosis and cohort, sensitivity to censoring, and out-of-sample validation across carriers. Decision checks ask whether the selected estimand matches the reserve, pricing, or claim-service question for which the prediction is used.
Clinical and occupational variables may improve statistical calibration, but sensitive characteristics should be handled according to governing law, fairness, clinical relevance, and empirical validity. A model can make accurate aggregate predictions and still provide poor decisions for claimants whose disability is episodic, whose workplace is accommodating, or whose records are unusually incomplete. The formal statistical output is a prediction under stated observation and contract conditions, not a substitute for legally appropriate individual adjudication or medical assessment.
The strongest falsification exercises deliberately challenge the proposed model. Estimate results using duration from onset and from first payable month; classify unknown closures under recovery and nonrecovery alternatives; omit a carrier at a time; validate on later calendar cohorts; vary reinstatement windows; and compare an absorbing-termination model with a recurrent multistate model. The differences reveal whether the conclusions depend mainly on definition, selection, actuarial parameter, or contract. A transparent model card should record data vintage, population, contract form, transition definitions, calibration date, uncertainty bounds, fairness evaluation, and circumstances requiring revalidation.
Regulatory and publisher accountability require an audit trail linking assumptions to identifiable primary documents and study generations. The SOA group termination study offers a major current industry benchmark, but its paid materials are subject to access conditions. The scientific record should identify that access boundary and avoid fabricating unpublished cell rates. In a separate layer, the IFRS 17 explanatory overview provides public accounting principles but does not substitute for the full standard text and implementation review required for a regulatory accounting decision.
Integrated Findings and Research Agenda
What the Evidence Establishes
The evidence supports five integrated findings. First, 'termination' is a composite administrative and contractual outcome. In an active claim risk set, recovery, death, benefit exhaustion, and other cessations need different event codes and estimands. Second, duration dependence reflects the joint behavior of underlying recovery processes, contract features, and selection among survivors. Third, recurrent absence and disability episodes require continued person-level follow-up after first closure. Fourth, the financial effect of recovery depends on whether a recurrent episode can restore benefits; an absorbing-recovery approximation is accurate only within a specified contractual domain. Fifth, legal and administrative measurement rules are part of the data-generating process and therefore materially affect the interpretation of experience studies.
These are analytical findings derived from established survival mathematics and verified institutional sources, not assertions of a newly invented theoretical system. The 2026 SOA/LIMRA update establishes a contemporary empirical base for group LTD terminations, while the 2025 individual DI analysis illustrates heterogeneity across diagnosis and contract characteristics. The Laaksonen et al. recurrence evidence demonstrates the predictive relevance of prior sickness absence in a defined occupational cohort. The studies collectively justify careful measurement; they do not jointly identify a universal cross-market relapse rate.
An insurer can improve a first-termination fit and still make poorer reserve predictions if the fitted event codes merge actual recovery with nonrecovering terminations or fail to follow subsequent payable episodes. Likewise, a pooled model can appear highly accurate when a large lower-duration population overwhelms a small, high-cost, long-duration subgroup. The operational implication is to validate component states, transitions, and present-value payments alongside any aggregate performance measure.
Research Design for a Reproducible Experience Study
A future claim-level empirical project should preregister its target population and outcomes: group versus individual disability; first eligible claims versus all episodes; a fixed cohort or sequential accrual; first recovery, all-cause claim cessation, death, and later covered recurrence. Define a minimum person-level follow-up period and a rule for policy termination or change of insurer. Publish a codebook and deidentified analysis specification before inspecting event patterns, with legal and privacy review appropriate to each jurisdiction.
The recommended sequence is: reconstruct chronological claim histories; check competing exits; estimate descriptive recovery and death cumulative incidences; estimate recurrence conditional on a documented recovery state; compare clock-forward and clock-reset intensity models; stratify by claims duration, diagnosis, occupational class and benefit design; calculate incurred-claim payment occupancy under actual contracts; and perform temporal, insurer, and coding sensitivity validations. Each stage should have a predetermined diagnostic criterion. When a comparison cannot be identified from recorded data, specify the reason and data needed rather than filling the gap with a point estimate.
A valuable extension would link work participation, benefit receipt, and medical functioning with consent and governance. Such data could test whether operational closure corresponds to sustained employment and whether a benefit change predicts renewed claim dependence. Causal effects of vocational rehabilitation, administrative review, or contract incentives would require a defensible identification strategy, such as a policy change with credible comparison groups or a suitable prospective design. The present review provides neither such a dataset nor evidence for an untested causal assertion.
Future comparative work should focus on legal recurrence windows, income-offset structures, and cross-carrier portability because these features determine whether the same underlying health recurrence generates a reopened claim, a new claim, or no recognized claim. A study built around an identical set of transition labels, with jurisdiction-specific legal mappings, would offer a much better foundation for comparing experience than placing unrelated first-termination percentages side by side. The present dossier's contribution is this reconciled research specification and the documented implications for reserve and claim measurement.
Conclusion
Disability insurance recovery, relapse, and claim termination are dynamic phenomena whose actuarial measurement depends upon precisely defined states, observation intervals, competing exits, contract wording, and post-termination follow-up. Recovery from disability, return to work, termination of payment, and closure of a claim are distinguishable events. Death competes with recovery for active claims; benefit exhaustion and adjudicatory decisions may end payments without establishing biological improvement. Relapse and recurrent eligibility can return a person to the payable state after a recovery or work interval, but the future financial consequences depend on the benefit scheme and the legal route back into coverage.
The available evidence supports established competing-risk and multistate methods, explicit selection controls, and calibrated reserve sensitivity. Current SOA/LIMRA group experience offers substantial observed exposure and termination volume, while individual disability studies and occupational recurrence research demonstrate the importance of diagnosis, duration, and history. The institutional examples from the United States, Canada, and United Kingdom show how administrative and legal definitions condition recorded outcomes. A responsible international model treats these systems as distinct data-generating arrangements and records the translation between them.
The decisive editorial result is a reproducible analytic standard for this subject: declare the state, risk set, exit cause, observation and contractual clock; preserve person-linked recurrent episodes; estimate competing transitions with uncertainty; calculate financial cash flows under actual benefit rules; and validate predictions on independent cohorts and alternative coding assumptions. The findings are an original integration of established methods and verified documents. Their empirical extension awaits claim-level observations with adequate follow-up and applicable contractual access, and no universal numerical recovery or relapse parameter is asserted.
References
Beal, Robert W., and Thomas R. Corcoran. (2025). Report on IDI Mortality and Recovery Rates by Diagnosis Coding—Phase 2. Society of Actuaries Research Institute, May. https://www.soa.org/resources/research-reports/2025/diagnoses-impact-mortrates-phase2/
Employment and Social Development Canada. (2026, accessed November 14). Automatic Reinstatement of Canada Pension Plan Disability Benefits. Government of Canada. https://www.canada.ca/en/services/benefits/publicpensions/cpp-disability-benefit/reinstatement.html
IFRS Foundation. (2017; effective 2023, accessed 2026). IFRS 17 Insurance Contracts: Overview, Scope and Principles. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
IFRS Foundation. (2023, accessed 2026). IFRS 17 Insurance Contracts—Key Terms. https://www.ifrs.org/supporting-implementation/supporting-materials-by-ifrs-standards/ifrs-17/key-terms/
Koopmans, Petra C., Corné A. M. Roelen, and Johan W. Groothoff. (2008). “Risk of Future Sickness Absence in Frequent and Long-Term Absentees.” Occupational Medicine 58(4): 268–274. https://pubmed.ncbi.nlm.nih.gov/18390880/
Laaksonen, Mikko, Liang He, and Janne Pitkäniemi. (2013). “The Durations of Past Sickness Absences Predict Future Absence Episodes.” Journal of Occupational and Environmental Medicine 55(1): 87–92. https://pubmed.ncbi.nlm.nih.gov/23235465/
Putter, H., M. Fiocco, and R. B. Geskus. (2007). “Tutorial in Biostatistics: Competing Risks and Multi-State Models.” Statistics in Medicine 26(11): 2389–2430. https://doi.org/10.1002/sim.2712
Society of Actuaries. (2013, updated 2016). 2013 Individual Disability Income Valuation Table and Supporting Materials. https://www.soa.org/resources/experience-studies/2013/idi-valuation-table/
Society of Actuaries, Group Disability Experience Committee. (2019, updated 2020). 2019 Group Long-Term Disability Experience Study Preliminary Report: 2009–2017 Experience. https://www.soa.org/resources/experience-studies/2019/group-ltd-experience-study/
Society of Actuaries, Individual Disability Experience Committee. (2021). Analysis of Claim Termination Experience from 2006 to 2014. https://www.soa.org/resources/experience-studies/2021/analysis-claim-termination/
Society of Actuaries and LIMRA. (2026, September 30). 2015–2024 Group Long-Term Disability Termination Experience Study. https://www.soa.org/resources/experience-studies/2026/2015-2024-group-long-term-disability-term/
Society of Actuaries. (2026, accessed November 14). Disability Experience Studies: Group, Individual, and Credit Disability Research Catalog. https://www.soa.org/research/topics/disability-exp-study-list/
United Kingdom Health and Safety Executive. (2025, accessed 2026). Principles to Support Disabled Workers and Workers with Long-Term Health Conditions in Work: Sickness Absence and Return to Work. https://www.hse.gov.uk/disability/best-practice/sickness-absence.htm
United Kingdom Parliament. (2010, as amended). Equality Act 2010, Schedule 1, Paragraph 2: Long-Term Effects. https://www.legislation.gov.uk/ukpga/2010/15/schedule/1/paragraph/2
United Kingdom Department for Work and Pensions. (2023, accessed 2026). The Fit Note: Guidance for Patients and Employees. https://www.gov.uk/government/publications/the-fit-note-a-guide-for-patients-and-employees/the-fit-note-guidance-for-patients-and-employees
United States Department of Labor. (2018, as amended). 29 CFR § 2560.503-1: Claims Procedure for Employee Benefit Plans. https://www.law.cornell.edu/cfr/text/29/2560.503-1
United States Social Security Administration. (2026). Try Returning to Work Without Losing Disability: Trial Work Period and Extended Eligibility. https://www.ssa.gov/disability/work
United States Social Security Administration. (2026, accessed November 14). The Red Book: Extended Period of Eligibility (EPE). https://www.ssa.gov/redbook/eng/epe.htm
World Health Organization. (2001; current WHO classification portal). International Classification of Functioning, Disability and Health (ICF). https://www.who.int/Standards/Classifications/International-Classification-of-Functioning-Disability-and-Health