Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Disability Insurance Claim Incidence Trend and Rating Credibility

Abstract. The estimation of disability insurance claim incidence joins three inferential problems: definition of the insured event, measurement of the population at risk, and prediction under limited experience. These problems become especially consequential in small employers, narrow occupational classes, newly insured portfolios, and rapidly changing labor markets. This monograph establishes a source-grounded framework for separating contractual incidence from clinical disability, observed claim counts from ultimate claim emergence, population composition from temporal change, and statistical credibility from regulatory or social acceptance of rating variables. It analyzes public long-term disability experience, including the SOA Research Institute–LIMRA study covering January 2015 through June 2022, and places that evidence alongside professional actuarial standards, individual disability experience, public disability statistics, and insurance accounting. It derives incidence estimators, Poisson and negative-binomial uncertainty, elimination-period-standardized comparisons, classical credibility and hierarchical partial pooling, and a reproducible Gamma–Poisson illustration. The main finding is that credibility is conditional on the quality and relevance of the exposure, claim definition, calendar maturity, and complement of credibility: a large number of policyholders does not itself validate a causal trend or an occupational factor. The work contributes an original analytical synthesis and explicit validation criteria, while proposing no new actuarial doctrine or empirical claims beyond the inspected sources.

Keywords: Disability insurance; group long-term disability; claim incidence; insured life-years; elimination period; experience rating; Bühlmann credibility; Bayesian shrinkage; Poisson regression; reporting lag; occupational exposure; employer risk; actuarial trend; risk classification; insurance pricing; model validation.

Introduction

An insurer that observes two new long-term disability claims in one hundred covered workers has learned something about its insured group; it has not necessarily learned that its prospective annual disability incidence is twenty claims per thousand. The observations may extend over several years, the insurer may have counted administrative notifications instead of qualifying disability onsets, or an individual claim may have arrived long after its disability date. Even once counts and exposure align, two events yield a highly uncertain estimate. The problem is both statistical and institutional: insurance contracts decide which forms of work incapacity become eligible claims, while enrollment arrangements determine whose lives generate exposure.

The SOA Research Institute and LIMRA 2015–2022 group LTD study provides an unusually informative case. Its public report records 293,814,050 life-years and 1,183,962 claims in fully insured group long-term disability business, with an aggregate incidence of approximately 4.030 per thousand life-years. The study covers January 2015 through June 2022, with the latter boundary chosen because claim reporting and long elimination periods made later 2022 observations immature. The study's operational definitions and omissions establish the boundary of valid inference; they do not authorize the extrapolation of its rates to every disability program or to an individual employer with unlike benefits.

The central object of this dossier is the prospective rate of contract-defined qualifying disability claims per insured exposure unit and the uncertainty associated with its estimation. The study distinguishes four quantities: underlying onset intensity, recognized or payable incidence under policy provisions, reported incidence as of a valuation date, and disability-related population prevalence. Its core method couples careful observation definitions with transparent likelihood-based models and a documented reference population. The argument proceeds from definitional foundations to exposure accounting, trend identification, pricing credibility, and verification protocols.

The interdisciplinary scope is deliberate. WHO's functioning and disability classification characterizes limitations and environmental contexts of functioning; private long-term disability policies translate particular work incapacity into benefits. The ILO disability protection analysis and Eurostat activity-limitation indicator measure different questions about eligibility, effective coverage, or participation. These statistics become relevant covariates or external plausibility checks only after a justified mapping of populations, years, and definitions. By contrast, the SOA's individual disability incidence study belongs to a different insurance product class and cannot be substituted without adjustment for an employer's fully insured group LTD experience.

Conceptual Foundations and Historical Development

From morbidity tables to contract-conditioned event rates

Disability is a complex state of functioning and ability to work. For private disability income insurance, claim incidence is most usefully defined by a specified entry event into a contract-governed claim state during a defined exposure interval. A diagnosis, a day off work, a request for claim forms, an accepted claim, the end of an elimination period, and commencement of cash benefits can represent distinct dates and populations. A single observational data set may legitimately support more than one incidence rate, provided each rate identifies its event and denominator. The WHO ICF framework identifies functioning and participation together with environmental factors; those domains are broader than insurance adjudication.

Historically, actuarial credibility emerged from the question of how much weight to attach to an employer's or insured group's own experience relative to the experience of a larger pool. Professional synthesis in ASOP No. 25 identifies subject experience, relevant experience, credibility procedures, and the conditions for applying the resulting weights. The CAS credibility chapter reviews classical and empirical approaches, including the requirement that the complement of credibility be selected for predictive relevance. Those methods remain useful in disability insurance when adapted to contract-specific claim frequencies and heterogeneous covered populations.

The contemporary actuarial task is more demanding because the observed administrative event is mediated by contract terms, employment and claim-management practice. Fully insured group long-term disability, employer self-insured administrative-services-only arrangements, voluntary plans, and individually underwritten disability policies differ in risk sets and incentives. The group LTD study's exclusions remove administrative-services-only business, individual forms, certain short-elimination-period or short-benefit-duration groups, converted policies, and selected other products. This is a study design choice, not an estimate of their disability incidence. The 2006–2014 individual income study therefore supplies methodological contrast rather than a directly exchangeable base rate.

Incidence, prevalence, frequency, and termination

Incidence refers to newly occurring events among those eligible to experience that event over a stated interval. Prevalence counts persons in a state at a specified time or over a period. Duration measures time spent in a claim state, with exits through recovery, death, benefit expiry or other contract outcomes. An insurer's number of in-force claimants is affected by incidence and duration; it is not an incidence numerator. The 2026 SOA–LIMRA group disability termination study reports disabled life-months, recoveries and deaths, reflecting a different denominator and different transition events than the 2015–2022 incidence study.

For a conceptual multistate representation, let A denote active insured employment, W disability in an elimination period, P payable disability, R recovered or returned to work, and D death. A transition A→W can begin before a W→P eligibility decision. Contract-defined incidence may attach to A→W with subsequent eligibility conditions or to entry into P, depending on the study protocol. Transition intensity, count of accepted claims, and stock of in-payment claimants should never be silently combined. A relapse may constitute a continuation of an existing claim under contractual recurrence provisions rather than a fresh incident case; data dictionaries must state the rule.

When the evaluation horizon is short, the time origin matters. Calendar-year incidence by date of disability differs from notification-year incidence and decision-year incidence. Claim records assigned to a notification calendar year can rise because adjudication accelerates without any change in underlying morbidity. Conversely, a slower claim reporting process may mimic a falling incidence trend. This observation is a direct implication of administrative event sequencing, reinforced by the study's choice to truncate immature late-2022 experience.

Unit of analysis and comparative jurisdiction

The unit at risk can be an insured employee, certificate, employment spell, employer contract, or exposure-weighted benefit amount. A person with two overlapping employer policies may generate two contractual exposures, yet remains one biological person. For insurer valuation, certificate-based benefit exposure can be relevant; for epidemiologic interpretations, it may double count individuals. The denominator must follow the estimand. The SOA exposure calculations materials provide the methodological background for converting enrollment periods into consistent experience rates, including complications of partial period boundaries.

The international scope of this inquiry requires a clear separation of private insured entitlement from public disability protection. The ILO reports that effective periodic disability-benefit coverage varies widely by geography and social-protection architecture in its 2024–26 report. Eurostat's disability employment gap is calculated from self-reported daily-activity limitations and employment in the EU-SILC population under its documented definition. The U.S. BLS reports disability status through the Current Population Survey in its 2025 labor-force release. None counts newly payable private LTD claims per covered employee-year. These sources illuminate the wider labor market and potential inequalities, but cannot supply the contractual event denominator.

Research scope and evidentiary hierarchy

The strongest direct evidence for the present subject is a reproducible experience file with policy-month exposure, claims by disability date, eligibility provisions, stable diagnosis definitions and sufficient development time. The second tier consists of audited summary tables and methodological reports; the third consists of population indicators for external comparison. Professional standards govern suitable practice but are not empirical studies of claim incidence. This dossier uses the publicly available full 2025 industry report and its detailed tabulations; additional proprietary interactive experience dashboards are not claimed as reviewed. The study authors themselves warn against treating aggregate reports as any single insurer's experience. The distinction is an essential condition of scholarly interpretation.

Exposure and Incidence Credibility

Person-time and policy-month reconstruction

For covered individual i in interval t let Y_i(t) equal one while the individual is eligible and exposed under the operative policy definition, and zero otherwise. Define E_g as the sum of eligible covered days for group g divided by 365.25, or, where trustworthy individual records do not exist, as the sum of validated group covered-life months divided by twelve. The incidence rate is r_g=N_g/E_g and the conventional display is 1,000r_g claims per life-year. The reference interval, rounding, leap-day convention and changes in coverage must be specified. The expression is an intensity approximation when repeated events or competing exits require additional assumptions.

At the study aggregate, 1,183,962 divided by 293,814,050 yields about 0.004030 claims per covered life-year. Its close match to the published 4.030 rate is a basic reconciliation, not a claim of individual-level replication. There is a major limit to the source: the industry study worked at aggregate group level; the authors expressly state that individual employee censuses were not supplied and that specific employee age, sex, salary and occupation incidence rates were not calculated. Accordingly no individual occupational hazard ratios or standard errors by those variables can be extracted from the public report.

Employment spells require entry and exit handling. A worker joining midmonth contributes only the period actually covered; an employer dropping its plan should not supply exposure after termination. Paid leave, continuation of coverage, and premium-waiver arrangements need documented definitions. If administrative data provide only an average monthly count, that approximation should be justified by reconciliation to total person-days where a census exists. Overlapping employer plans, midyear benefit redesign and retroactive claims adjustments also require explicit treatment.

Numerator integrity and claim maturity

The claim numerator should count unique qualifying first-event observations under the agreed rule. Duplicate files, re-opened claims, separate benefit components, and claimant movement across systems can inflate frequency. In the LIMRA data validation procedures, claim records were examined for duplicate policy ID, birth date, disability date and sex combinations; missing policy-month records and plan provisions were investigated or corrected according to contributor guidance. Approximately one quarter of exposures lacked a covered-benefit value and approximately four percent lacked covered payroll, limiting certain benefit-weighted analyses. These deficiencies make it especially important to avoid treating missingness as a new underwriting effect.

Define N_{t,v} as claims with disability date in period t that are known by valuation date v. Let p_{t,v} be the probability that a ultimately included claim from t has emerged and met adjudication criteria by v. A maturity-adjusted count may be estimated as N_{t,v}/p_{t,v} under suitable homogeneity and correctly estimated reporting-development patterns. This is a conceptual relationship; p must be measured from comparable mature triangles, claims systems and policy provisions. Unobserved late claims cannot be recovered by applying a universal factor. If a calendar period remains immature, transparent exclusion may be superior to an unstable development correction.

The SOA–LIMRA report excluded July–December 2022 because the June 30, 2023 claim valuation date left insufficient development for some longer elimination periods in its processing methodology. This establishes a crucial interpretation of the 2022 annual rate: it is a first-half-year estimate, not a fully comparable full-year incidence observation. Comparisons across all years must preserve this selection and assess remaining seasonal factors. Trend estimation should treat 2022 as partial rather than blindly extrapolating 3.747 per thousand to a complete calendar year.

Heterogeneous eligibility and elimination periods

Longer elimination periods often reduce recognized LTD incidence because more episodes resolve before benefits become eligible. The experience table by elimination period gives 4.388 per thousand for three-month periods and 3.566 for six-month periods; this descriptive reduction of roughly 19 percent is conditioned by the contract and its enrolled population. It does not establish a causal treatment effect of lengthening the elimination period. Other conditions, including group size, occupation and participation, can differ between plans. A rare one-month category has a much higher rate and should not be forced into a monotone empirical dose response.

The study's actual-to-expected comparison uses expected rates weighted for the composition of elimination periods within each subset. For stratum j, denote its observed count O_j and expected count X_j=Σ_k E_{jk} r_k^ref where k indexes compatible elimination-period categories and r_k^ref is the corresponding reference rate. Then A/E=O_j/X_j. This is a rate-standardization exercise, not a full age/sex or occupation adjustment. The industry report explicitly cautions that its A/E ratios do not adjust for demographic differences because age and sex were not consistently populated. An A/E of 120 percent therefore describes a conditional deviation from the selected reference, not a causal premium relativity.

Sampling variation and overdispersion

Under a homogeneous Poisson approximation N_g∼Poisson(E_g λ_g), the maximum-likelihood estimate is N_g/E_g, and its estimated standard error is √N_g/E_g. Exact Poisson confidence limits for a positive observed count N are based on χ² quantiles; these give wider, asymmetric intervals when N is small. The coefficient of variation of an observed Poisson count is approximately 1/√N, so four events imply about 50 percent relative standard error before any sampling, measurement, or model uncertainty. At zero events, the naive estimate zero is compatible with a broad range of positive true rates; an upper one-sided 95 percent Poisson bound is −ln(0.05)/E, about 2.996/E.

The Poisson variance assumption often understates disability claims variability across employers, industries and time. If employees share occupational hazards, benefit plans or a claims administrator, counts can be correlated. A negative-binomial model represents extra-Poisson variation, for example Var(N_g|E_g)=μ_g+φμ_g² with μ_g=E_g λ_g and φ≥0, while an employer random effect explicitly models variation between groups. Models must distinguish correlation from omitted covariates and exposure errors; an overdispersion parameter cannot repair a wrong event definition. ASOP No. 56 emphasizes model purpose, structure, testing and material overfitting risk.

Classical credibility as an estimation problem

Let r_g be the subject rate, r_0 a relevant experience rate and Z_g a predictive weight between zero and one. A common credibility estimate is r*_g=Z_g r_g+(1−Z_g) r_0. The complement is a substantive assumption about the relevance of the external pool. If the complement is based on unlike contracts or industries, a high-quality subject observation may be blended with a biased benchmark. ASOP No. 25 instructs actuaries to examine the suitability of relevant experience, data homogeneity, and professional judgment. It does not designate one universal disability credibility threshold.

For simple Poisson frequency, the sampling coefficient of variation is approximately 1/√N. A classical requirement for a two-sided relative margin k at confidence level α can be derived as N_full≈(z_{1−α/2}/k)² under normal approximation. At 95 percent and k=5 percent, the benchmark is approximately 1,537 expected claims. This calculation is illustrative and assumes homogeneous, independent Poisson counts with a stable underlying rate. It should not be interpreted as a legally prescribed or empirically validated disability minimum. In small groups the number of covered lives can be misleading because relevant information accrues as expected claims and exposure adjusted for the hazard level.

Gamma–Poisson partial credibility illustration

An analytically transparent comparison uses a gamma prior for λ with shape a and rate b, combined with N∼Poisson(Eλ). The posterior is Gamma(a+N,b+E), with posterior mean (a+N)/(b+E). This equals a weighted average of the raw rate N/E and the prior mean a/b, with data weight E/(b+E). Here b represents prior exposure-equivalent strength and a/b the relevant pooled annual frequency. This is an established conjugate Bayesian calculation rather than a newly proposed model. The same form illustrates credibility shrinkage when exposure is sparse.

Choose a hypothetical relevant rate of four claims per thousand life-years, represented by a=4 and b=1,000 life-years. A 100-life-year employer with one qualifying claim has a raw rate of 10 per thousand but posterior mean 5/1,100=4.545 per thousand, with weight Z=100/1,100=0.09091. The same exposure with zero claims yields 4/1,100=3.636 per thousand, rather than zero. A 500-life-year employer with three claims has raw rate six per thousand and posterior mean 7/1,500=4.667 per thousand, with Z=1/3. These figures are constructed examples, checked arithmetically, and are not observations from the SOA/LIMRA portfolio. Their substantive use depends on how a, b and group comparability are estimated and validated.

Posterior uncertainty also matters. In the first example, Gamma(5,1,100) has standard deviation √5/1,100, or about 2.033 claims per thousand. An estimate of 4.545 per thousand without this uncertainty would overstate precision. Empirical Bayes procedures estimate between-group heterogeneity from many employer groups; the Gamma illustration uses a declared hypothetical prior instead. Fully integrating uncertainty in the prior and trend generally widens posterior predictive uncertainty beyond a plug-in calculation. The classical credibility account and ASOP No. 25 support this methodological distinction between credibility selection and its numerical implementation.

Trend and Macroeconomic Drivers

What a calendar trend measures

A true prospective trend is a temporal change in the expected frequency for a specified reference population and stable contract definition. The observed change between two calendar-year aggregate rates combines at least four components: altered health or work incapacity risk, altered composition of insured workers and employers, changed contract provisions and claims practices, and changes in maturity or reporting. Even if all but one of these were held fixed, random sampling error remains. Accordingly the trend parameter requires a target standardized population and a stated period and temporal scale.

The SOA–LIMRA annual table records 4.169 per thousand in 2015, 3.949 in 2017, 4.021 in 2020 and 4.177 in 2021, followed by 3.747 in the first six months of 2022. The 2015-to-2021 point-rate ratio is about 1.002, an almost unchanged endpoint comparison, while intermediate years vary. The report's own year-by-year analysis describes a relatively flat 2016–2020 sequence and notes that the changing exposure mix can affect year-to-year rates. It would be misleading to call the 2022 half-year observation proof of a durable annual decline or to fit a single exponential trend through all observations without maturity and seasonal testing.

For a standardized trend, define r_t^{std}=Σ_c w_c^* r_{ct}, with fixed reference weights w_c^* over homogeneous classes c, while raw r_t=Σ_c w_{ct}r_{ct}. Their difference separates pure changes in within-class rates from changes in class composition, to the extent classes are fully observed. If both within-class rates and weights change, a decomposition using fixed base and end weights can display interaction effects. With only aggregate calendar marginals, this separation is unidentifiable. A group-level table cannot be inverted into individual age, occupation or demographic effects absent source-level records.

Count regression and identification

A general log-rate specification for employer g, class c and time t is log E[N_gct]=log E_gct+α+β_t+γ_c+u_g+δ'X_gct. Here log exposure is an offset; β_t is a time effect, γ_c captures observable classes, u_g an employer effect, and X includes pre-specified contract and economic variables. A smooth time trend can replace β_t by a coefficient times t, while segmented trends can represent a demonstrable structural break. A negative-binomial model may account for extra-Poisson variation, and random effects support partial pooling. Estimated annualized rate factor exp(β) must be reported with standard error and a defined unit of time.

Identification demands sufficient overlapping observations. If one industry's entire contribution appears only after a calendar date, an industry effect and a time effect may be indistinguishable. An employer fixed effect can control stable employer differences but does not repair time-varying selection of covered workers. When every plan with a longer elimination period belongs to a different industry, the model needs additional variation to distinguish the effects. The original SOA data request planned incidence by contract and employer characteristics precisely because aggregate trends alone lose this resolution. ASOP No. 23 requires assessing whether selected data are appropriate for their intended purpose.

Model fitting should use pre-declared estimands. If the purpose is next-year claims per covered life-year, test next-year out-of-sample log predictive density, mean calibration and prediction intervals. If the purpose is employer-specific premium indication, validate within suitable employer and industry strata, including unseen employers. Calendar holdouts and employer-cluster holdouts answer different questions. Cross-validation that randomly divides records from the same employer and policy across training and testing can leak stable policy risk attributes. The modeling standard demands procedures proportionate to the model's intended use, including input reconciliation and testing against prior runs.

Macroeconomic and labor-market mechanisms

Economic conditions influence workforce selection, job demands, access to care, wage replacement and insurance take-up. Changes in unemployment can affect the employed exposure denominator and the claims propensity of those who remain insured. Employer workforce reductions can concentrate physically demanding tasks among remaining workers or remove employees with greater disability vulnerability from an eligible risk set. Claimant behavior, medical evidence and vocational assessment may also respond to economic conditions. These are plausible mechanisms requiring identification; they are not quantified causal findings of this dossier.

The BLS 2025 disability labor-force release reports employment-population and unemployment measures for persons with disabilities in a national household survey. That is a stock and labor-market status measure, not a claim-event rate. Eurostat's EU-SILC disability employment gap similarly compares employment rates by self-reported activity limitation. These population statistics can motivate questions about labor-market conditions or external representativeness but should not be inserted directly as private LTD frequency multipliers. The ILO comparative social-protection evidence shows why public benefit entitlement and access vary internationally, with implications for coordination and claimant choices, yet these differences require local documentation before any incidence adjustment.

Other candidate covariates include aggregate real earnings, job turnover, industry composition, sick-leave arrangements, rehabilitation access, benefit offsets, and medical coding practice. A time series with few annual points cannot support dozens of predictors without severe overfitting. The strongest design for a causal question would exploit well-documented policy or administrative discontinuities with a credible comparison group; otherwise, the analysis should remain predictive and descriptive. Even a highly significant association between recession indicators and private claim incidence would not establish the sign or magnitude of a structural economic effect.

Pandemic and diagnostic reclassification

COVID-19 created a sudden, temporally concentrated source of illness and administrative disruption during the study period. The study's diagnosis breakdown records changes in infectious-disease and COVID-labeled claims, along with temporal variation in mental and nervous categories. Those observations arise in the study's particular coverage population, diagnosis mapping, and claims processing systems. A diagnostic shift may represent changes in clinical classification or claimant composition as well as true incidence. Cause-specific trends must therefore document ICD mapping, diagnosis precedence rules, unknown categories, and comparability across calendar revisions.

A prospective stress scenario could separately alter true disability onset, delayed reporting, average claimant age, occupational mix, rehabilitation capacity and elimination-period design. Reporting only a pandemic period dummy compresses these channels into an unstable coefficient. The historical report supports descriptive associations, while scenario parameters require independent support. In particular, drawing a COVID effect from post-2019 rates without controlling for claim maturation and affected industries invites confounding.

Trend uncertainty beyond point forecasts

There are at least three uncertainty layers: process volatility conditional on model parameters, parameter uncertainty from finite claims, and specification uncertainty over models and future structural change. A Poisson prediction interval given a known rate captures only the first layer. Trend estimates require joint uncertainty over regression coefficients and employer heterogeneity; alternative contracts or scenarios require model-selection sensitivity. When a small employer has two claims, adding a mechanically fitted pandemic trend to a raw experience rate compounds noise rather than creating information. A transparent report should state the base rate, trend factor, selected horizon, calendar calibration, exposure shift and joint prediction interval.

Pricing Uncertainty and Fairness

Frequency is one input into claim cost

A price for disability income coverage combines frequency with expected claim duration and benefit cash flows, expenses, contractual options, capital costs, and a target financing basis. In a simplified one-year setting, expected claim cost per life-year may be approximated by λ×V, where λ is eligible incidence per life-year and V the discounted expected cost of a newly incurred qualifying claim. In a multi-state setting V depends on recovery, mortality, recurrence, benefit limits, offsets, and salary replacement. The 2026 SOA–LIMRA termination experience study highlights this separate survival and termination dimension. No pricing analysis should mistake a lower reported incidence from a longer elimination period for an equivalent reduction in total claim severity or capital requirements without cash-flow modeling.

For accounting, IFRS 17 Insurance Contracts frames insurance liabilities using current fulfillment cash-flow estimates and an explicit treatment of nonfinancial risk within its applicable reporting regime. This is distinct from the narrower question of selecting an employer incidence rating factor. A best-estimate frequency parameter, a prudence or risk adjustment, and an expense/profit contribution play different roles; their separation makes it possible to test rather than hide uncertainty. National statutory pricing or rate-review requirements may require a different presentation than IFRS financial reporting.

The complement of credibility is a choice of target population

Suppose an employer's raw incidence is materially above a national mean. The proper blend depends on the employer's benefit design, participation, payroll distribution, occupation mix, claims-maturity profile, geography and observation years. The 2025 study's case-size table reports 3.440 per thousand for groups under 25 lives when aggregated, versus 4.425 for groups over 5,000, a ratio of about 1.286. Those are descriptive grouped data, and the groups differ in elimination period and other characteristics. Their observed contrast is therefore not a directly transferable small-employer surcharge or size effect. The report's elimination-period A/E figures materially refine, but do not exhaust, adjustment for case mix.

Credibility weighting cannot turn a biased reference into an unbiased estimate. If r_0 is based on industries with very different physical demands, the posterior mean may be stable and systematically wrong. The higher fidelity approach is to construct relevant class benchmarks with sufficient pooled exposure, verify cross-class overlap, and disclose an interval for the target population after model uncertainty. The credibility standard directs attention to the predictive value of the selected subject and reference experience; the risk-classification standard addresses the actuarial rationale and review of classes. Both are pertinent, but neither substitutes for jurisdiction-specific legal approval of rating features.

Occupation, employer size and correlated risk

Occupation may be a proxy for physical exposures, work environment, benefit design, earnings profile, and access to accommodations. It is also associated with social distribution of work and income. A significant employer or occupation coefficient must be interpreted in light of data quality and structural correlation: a single employer's observed high incidence may arise from temporary hazardous work, while another may reflect delayed reporting or a different elimination period. Employer-level clustering means a hundred workers from one workplace are not necessarily as informative as a hundred unrelated independently distributed lives. This is why statistical precision should be indexed to effective information and estimated heterogeneity, not simply headcount.

The SOA–LIMRA industry table reports material differences between blue-, gray-, and white-collar industry groupings, but these are broad industry categories, not measured individual occupations. The document explicitly lacks individual census-based occupation rates. A proposed occupation factor derived from industry aggregates would risk ecological inference: within-industry occupational composition can vary and so can the mechanism linking work demands to benefit eligibility. For multinational comparative work, industry code mappings, local employment arrangements and disability definitions require harmonization before transfer.

Fairness and legitimate classification

Fairness has distinct legal, statistical and social dimensions. Statistically, a rating model should predict the selected future event accurately and be calibrated across material subgroups. Legally, permitted factors and disclosure obligations depend on governing jurisdiction, contract and product class. Socially, one should examine affordability, exclusion and whether observed claim differences reflect preventable workplace conditions or access barriers. ASOP No. 12 addresses risk classification as an actuarial professional practice; it is not a universal global legal safe harbor. A factor can be strongly predictive yet prohibited by applicable law; another can be allowed but unsupported by credible local experience.

A fairness audit should compare observed versus expected incidence by relevant groups with sufficient exposure, report confidence or posterior intervals, test missingness and differential claims emergence, and examine whether proxy features transmit prohibited or institutionally inappropriate distinctions. Blindly equalizing all premiums can create cross-subsidy consequences, while uncritically transmitting every observed risk contrast into price can entrench exclusion. The proper scientific report displays these tradeoffs, the objective, constraints and measured impacts. It does not infer individual ability or social value from a disability insurance claim label.

The ILO disability benefits chapter supplies a broader access context, and the BLS disability survey reveals a population labor-market dimension. Neither provides evidence that a particular private employer class factor is lawful or fair. Those determinations require the target country, contract structure, supervisory rules, consent architecture and evidence of disparate consequences. The present comparative work therefore specifies tests rather than asserting universal rating legality.

Capital, uncertainty and rate stability

The credibility premium has a time dimension. An aggressive raw-experience surcharge following one claim can generate volatile renewals, unstable take-up and experience selection. Conversely, smoothing that ignores a durable occupational hazard can produce sustained underpricing. A transparent renewal formula should separate ordinary credibility development from documented changes to risk, insurer targets and policy terms. The rate filing or customer communication should give a stable definition of the observation window, the complement population, the trend period and the uncertainty allowance.

To illustrate frequency–severity interaction, suppose a fully hypothetical claim's expected present-value cost is 80,000 currency units. At annual incidence 0.004 per covered life-year, expected claim cost is 320 units before expenses and other adjustments. Increasing the assumed frequency by ten percent while increasing the average claim present value by fifteen percent gives 320×1.10×1.15=404.80, a combined increase of 26.5 percent. These are transparent arithmetic illustrations, not estimates for any product or employer. The example shows why a single claim-incidence trend factor cannot determine an insurance premium. Actuarial communication standards support clear statement of assumptions and limitations.

Integrated Findings and Research Agenda

Reconciliation of direct evidence and method

The most important public benchmark in this dossier is the joint SOA–LIMRA 2025 detailed incidence study. It supports a precisely bounded statement: its included fully insured group LTD observations comprise 293,814,050 life-years and 1,183,962 claims across January 2015–June 2022, corresponding to 4.030 qualifying claims per 1,000 life-years. It supports descriptive comparisons by elimination period, calendar year, case size and broad industry, subject to the authors' case-mix and classification cautions. It does not support direct individual-age or occupation-specific incidence estimates because a seriatim employee census was unavailable. It does not convert an annual aggregate into a causal time trend by itself.

The next inference is method-derived: employer pricing needs an estimate for a prospective population. Raw N/E estimates, exact small-count uncertainty, a stable reference complement and a calibrated partial pooling method each answer different pieces of that problem. Classical credibility and Gamma–Poisson shrinkage stabilize small portfolios, but the predictive benefit depends on reference-group relevance and honest uncertainty reporting. Temporal modeling must preserve event maturity and adjust for exposure composition. These are reproducible mathematical and statistical conclusions, not discoveries of a new mortality or disability law.

The report's six-month elimination-period rate of 3.566 per thousand compared with 4.388 for three-month periods, and its case-size contrasts, indicate the significance of policy mix. The insured denominator, eligibility rules, and development of claims are therefore part of the incidence estimand itself. For international applications the same discipline applies even though insurance and labor-law systems vary. The ILO coverage comparison and Eurostat indicator help explain why international disability protection statistics cannot be treated as directly interchangeable premium inputs.

Reproducible calculations and interpretation table

The following calculations have deliberately limited evidentiary status. The first row reproduces a descriptive statistic from the public source; subsequent rows apply standard probability or arithmetic identities under the assumptions stated. Numerical values are rounded only in presentation.

Quantity Expression or input Computed result Interpretation
Industry-wide published frequency 1,183,962 / 293,814,050 × 1,000 4.030 per 1,000 Observed insured claims in source study
Three- versus six-month EP ratio 3.566 / 4.388 0.813 Descriptive association, not causal effect
2021 versus 2015 rate ratio 4.177 / 4.169 1.0019 Raw aggregate endpoints; does not identify trend
Hypothetical 100 life-years, 1 claim Gamma(4,1,000) updated by N=1, E=100 4.545 per 1,000 Credibility-adjusted mean, Z=0.0909
Hypothetical 100 life-years, no claims Gamma(4,1,000) updated by N=0, E=100 3.636 per 1,000 Zero experience leaves nonzero prospective estimate
Hypothetical 500 life-years, 3 claims Gamma(4,1,000) updated by N=3, E=500 4.667 per 1,000 Credibility-adjusted mean, Z=0.3333
Zero-event 95% Poisson upper bound −ln(0.05)/100 person-years × 1,000 29.957 per 1,000 Illustrates wide uncertainty, not estimated insurer risk
Frequency–severity joint stress 320 × 1.10 × 1.15 404.80 Combined factor 1.265, before expenses and capital

These computations are finite arithmetic checks, not simulations of an observed employer dataset. The sample prior in the credibility examples is chosen for illustration and its estimated strength has not been inferred from insurer records. The large separation between a zero-event empirical estimate and its one-sided upper bound captures why a single year's experience may offer little basis for a precise renewal rate. Appropriate confidence or predictive intervals depend on the chosen likelihood and model assumptions.

Falsifiable hypotheses for future research

Future empirical work can test whether elimination-period-standardized annual trends differ from raw trends after demographic and industry mix controls; whether employer-level partial pooling improves held-out log predictive score and interval coverage over classical square-root credibility; and whether a model with reporting-delay adjustment materially changes pre/post-shock comparisons. Each is a testable question, not an asserted discovery. A coherent study would secure consented insurer microdata, pre-register the outcome and exclusions, split by employers and calendar periods, report the full claims-maturity pattern, and retain declined or zero-claim employer exposures.

Further work should quantify how much of reported group-size A/E variation remains after contract design, industry and group participation adjustment, while investigating potential selection into employer-paid versus contributory plans. The source report's group-size association is strong enough to warrant analysis but insufficient for causal attribution. International comparative work would need a contract-to-contract mapping framework and local legal validation rather than importing U.S. group LTD classifications unchanged. A separate research question concerns whether claims acceptance delays vary systematically by subgroup and therefore bias apparent incidence fairness metrics.

The evidence needed to advance these questions is concrete: individual or consistently aggregated longitudinal exposures, adjudication dates, occupational classifications, contract provisions and time-stamped claims records; confidentiality and ethical governance; an explicit model selection plan; and genuinely external testing. Until such evidence exists, the public report and mathematical illustrations support measured descriptive and methodological conclusions, with limits made explicit in the manuscript and evidence register.

Conclusion

Disability insurance claim incidence is a contract-conditioned event rate measured over a carefully reconstructed insured risk set. Its numerical value depends on the definition and timing of a qualifying claim, exposure denominators, benefit provisions and the maturity of observed records. Contemporary industry evidence demonstrates the magnitude of these dependencies: the large SOA–LIMRA fully insured group LTD sample produces a useful benchmark while disclosing missing plan and payroll attributes, a lack of individual census data, and partial-year truncation to address reporting lag.

Credibility is the problem of predicting with incomplete and heterogeneous information. The principal analytical result is that an employer's prospective incidence cannot be determined from its raw claims ratio or membership count alone. Credibility weighting requires a relevant complement, coherent exposure units, uncertainty estimates, and tests of transportability to the employer and contract at issue. A low or high observed group rate can reflect statistical variation, policy mix, reporting maturity, genuine risk, or their interaction. Structured comparison and partial pooling improve estimation only when these sources are modeled and disclosed.

Trend analysis extends the same principle through time. Within-class change must be distinguished from changing portfolio composition, administrative definitions, economic selection and temporary disruptions. Pricing then places incidence alongside recovery and claim-duration assumptions, cash flows, expenses and capital. Risk classification must remain empirically supported, transparent and compliant with applicable jurisdictional rules. The combined conclusions establish a reproducible research and editorial standard for evaluating sparse disability claim incidence without presenting any new proprietary actuarial theory or unsupported causal finding.

References

  • Society of Actuaries Research Institute and LIMRA (2025). 2015–2022 Group Long-Term Disability Incidence Study Report. Experience Studies PRO, September 2025, full public report. Official source.

  • Society of Actuaries Research Institute and LIMRA (2025). 2015–2022 Group Long-Term Disability Incidence Study. SOA Research Institute, research overview, October 2025. Official source.

  • Society of Actuaries Research Institute and LIMRA (2023–2025). 2015–2022 Group Long-Term Disability Incidence Study Data Request. SOA Research Institute and LIMRA, study design and variable list. Official source.

  • Actuarial Standards Board (2013). Actuarial Standard of Practice No. 25: Credibility Procedures. Revised December 2013; effective May 1, 2014. Official source.

  • Actuarial Standards Board (2016). Actuarial Standard of Practice No. 23: Data Quality. Adopted December 2016; effective April 30, 2017. Official source.

  • Actuarial Standards Board (2019). Actuarial Standard of Practice No. 56: Modeling. Adopted December 2019; effective October 1, 2020. Official source.

  • Actuarial Standards Board (2005). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). Adopted December 2005; effective May 1, 2006; subsequent editorial revisions. Official source.

  • Society of Actuaries Research Institute (2023–2024). Experience Study Calculations Educational Tool. SOA exposure calculations paper and accompanying educational workbook; updated March 2024. Official source.

  • Society of Actuaries Individual Disability Experience Committee (2019). Analysis of Claim Incidence Experience from 2006 to 2014. Society of Actuaries, individual disability experience study. Official source.

  • World Health Organization (2001). International Classification of Functioning, Disability and Health (ICF). WHO classification, endorsed at World Health Assembly, May 22, 2001. Official source.

  • Society of Actuaries Research Institute and LIMRA (2026). 2015–2024 Group Long-Term Disability Termination Experience Study. SOA Research Institute, September 30, 2026. Official source.

  • International Labour Organization (2024). World Social Protection Report 2024–26: Chapter 4. ILO, disability benefits and social protection coverage. Official source.

  • Eurostat (2026). Disability Employment Gap by Level of Activity Limitation and Sex. EU-SILC indicator tepsr_sp200, dataset update September 17, 2026. Official source.

  • U.S. Bureau of Labor Statistics (2026). People with a Disability: Labor Force Characteristics—2025. News Release USDL-26-0364, March 3, 2026. Official source.

  • U.S. Department of Labor, Employee Benefits Security Administration (2018). Employee Retirement Income Security Act (ERISA): Disability Benefit Claims Procedures. Official ERISA claims-procedure rule collection and 2018 applicability materials. Official source.

  • IFRS Foundation / International Accounting Standards Board (2017–2026). IFRS 17 Insurance Contracts. Issued May 2017; applicable from January 1, 2023; official standard overview. Official source.

  • Casualty Actuarial Society (2019–2021). Basic Ratemaking: Chapter 12, Credibility. CAS educational ratemaking text, Chapter 12 (online PDF). Official source.

  • Actuarial Standards Board (2010). Actuarial Standard of Practice No. 41: Actuarial Communications. Revised December 2010; effective May 1, 2011. Official source.