Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Cohort-Level Life Insurance Exposure and Credibility Adjustment

Abstract. The mortality experience of a life insurer is observed through contracts that enter, change, lapse, convert, and terminate while insured persons age and underlying mortality conditions evolve. A death count has statistical meaning only in relation to the amount and composition of exposure from which it arises. This monograph investigates cohort-level mortality exposure and credibility adjustment as a single actuarial inference problem. It defines exact time at risk, distinguishes central from initial exposure, reconstructs attained-age and policy-duration cohorts, analyzes policy-count and amount-weighted measurement, and derives the consequences of sparse observations for direct mortality estimates, actual-to-expected comparisons, and credibility-weighted assumptions. The analysis integrates Bühlmann–Straub ideas, Poisson–gamma updating, survival likelihoods, underwriting selection, lapse censoring, mortality improvement, source provenance, and professional standards. Two reproducible synthetic case studies demonstrate how an apparently adverse crude cohort difference can change after risk-standardization and how credibility weights depend on expected events, not merely record counts. Comparative evidence draws on Society of Actuaries insured-life studies, the Human Mortality Database, original statistical scholarship, and current US actuarial standards, while distinguishing those sources from country-specific valuation and accounting rules. The central finding is that trustworthy cohort adjustment requires jointly coherent definitions of the exposure denominator, target mortality estimand, reference population, variance structure, and decision purpose. The contribution is an original critical synthesis and checked calculation, without the creation or claim of a new proprietary theory or model.

Keywords: Life insurance; insured mortality; cohort exposure; central exposure; initial exposure; person-years; policy-years; credibility; Bühlmann–Straub; Poisson–gamma; empirical Bayes; actual-to-expected mortality; mortality improvement; select and ultimate mortality; underwriting selection; lapse; censoring; experience studies; actuarial standards; model validation.

Introduction

A cohort of insured lives is an analytically constructed set of risk histories. Its membership depends on the governing contract, the insured person, the observation interval, and the segmentation variables chosen for the decision. Age at observation, year of birth, policy issue year, years since underwriting, underwriting class, smoker classification, product, benefit amount, distribution channel, and jurisdiction each partition experience differently. The number of deaths recorded in a segment, standing alone, identifies neither the rate of death nor its precision. The rate is determined by the deaths together with the time during which comparable lives were exposed to death and observable under the policy definition.

Credibility adjustment assigns predictive weight to a cohort's experience relative to relevant external or pooled information. The US Actuarial Standards Board's adopted Credibility Procedures standard explicitly addresses subject experience, relevant experience, the predictive value of data, and actuarial judgment. The standard describes a professional procedure; its language does not confer a universal statistical distribution on mortality events. The statistical model must be supplied and justified by the analyst.

The practical problem becomes difficult when cohorts are small, observed for unequal periods, rapidly changing in composition, or exposed to changing mortality conditions. A mortality rate estimated from 10 deaths has different sampling uncertainty from one estimated from 1,000 deaths. The difference can persist when both cohorts have thousands of policy records. Conversely, a large nominal exposure provides limited credibility if lives are systematically misclassified, death reports are delayed, contracts are duplicated, or the assumed comparability of reference and subject populations is false.

Several documented research traditions converge here. The 2015 Valuation Basic Tables distinguish sex, smoking status, preferred risk and duration, while the Society of Actuaries' 2019 individual-life mortality report analyzes experience from 2012–2019 relative to industry tables. The demographic HMD Methods Protocol constructs exposures and mortality rates from national vital statistics, not the underwriting-selected universe of insured contracts. Bühlmann and Straub (1970) address credibility with heterogeneous volumes, while the censored-data framework of Kaplan and Meier (1958) and the regression approach of Cox (1972) clarify what can be learned from incomplete survival histories.

This dossier asks three questions. First, which exposure measure permits valid estimation of insured-cohort mortality when observation periods and eligibility vary? Second, how do underwriting selection, policy duration, contract termination and demographic change distort comparisons between heterogeneous cohorts? Third, under what assumptions do credibility adjustments improve predictive precision, and how are those assumptions evaluated, communicated, and tested? Its analytic object is cohort-level estimation. General life-contingency pricing, solvency capital, reinsurance design and comprehensive financial reporting retain their separate scientific scope.

The method combines source-critical review of original research and institutional standards; formal derivation under declared stochastic assumptions; paired examples with independently checkable arithmetic; and a qualitative cross-jurisdictional distinction between insured mortality, national population mortality, and accounting measurement. All numerical examples explicitly identified as synthetic are analytical illustrations. No proprietary insured-life file has been obtained or represented as independently audited.

Definitions, Historical Evidence, and Source Criticism

The cohort is a rule of membership

A cohort must be defined before its mortality rate is estimated. A birth cohort shares a birth interval; an issue cohort shares the calendar period of first coverage; a duration cohort shares time elapsed since issue or underwriting; a valuation cohort is the set of policies in force at a valuation date; and a calendar-period mortality cell collects portions of observation that fall in a calendar interval. A life can contribute exposure to multiple attained-age or calendar cells during one observation period, but each instant should contribute exactly once to each mutually exclusive partition of the defined risk universe.

The identity of the insured person, policy and coverage is essential. One individual with two life contracts represents one death-capable person but two insured benefit obligations. Counting two policy deaths or two face-amount claims can be correct for a contract-based analysis, while counting the same life twice is incorrect for a person-level hazard unless the model explicitly uses clustered multiple contracts. The investigator therefore records whether the outcome is a death of a person, an insured death claim, a count of policies terminating by death, or a benefit amount. The target and denominator must have the same unit.

When the relevant question is an estimate of mortality for underwritten individual policies, the data-generating universe is not the population of all residents. The 2009–2013 individual-life experience documentation identifies intercompany experience under standard individually underwritten issues and provides descriptions of data collection, cleaning and fields. A population data system such as the Human Mortality Database explanatory notes instead organizes deaths, exposures and life tables by sex, age and time for geographic populations. Neither denominator can be substituted for the other by relabeling a rate.

The historical literature has distinct objects

Credibility emerged in insurance as a method for balancing an individual experience record against a larger reference. Bühlmann's Experience Rating and Credibility exemplifies the foundational theoretical discussion. The later Bühlmann–Straub paper generalizes a credibility setting to varying volumes of experience. These are actuarial methods for estimating heterogeneous risk characteristics, rather than evidence that every individual-life mortality segment follows the same model or that policy-years are automatically independent.

Survival analysis addresses a related but distinct information problem. Kaplan and Meier (1958) studied estimation under incomplete observation; Cox (1972) showed how covariates may enter a proportional-hazards structure while leaving baseline hazard unspecified. These approaches support analysis of changing risk sets and recorded censoring, subject to their assumptions. The policyholder's lapse is not necessarily a noninformative withdrawal: the lapse decision may correlate with underlying health, financial conditions, and conversion options, altering which lives remain exposed. Treating censoring as automatically independent can therefore bias post-lapse mortality interpretation.

Primary sources, status and scope

An authoritative professional standard determines what it recommends for actuarial practice within its jurisdiction and governing professional context. The US ASOP No. 25 is listed as adopted, last revised December 2013 and effective May 1, 2014. ASOP No. 23 on Data Quality is an adopted December 2016 revision, and ASOP No. 56 on Modeling is adopted, with an October 1, 2020 effective date. Their obligations are professional-practice requirements for those within scope, with applicable law taking precedence as described in the standards. They are not international statutes or interchangeable with an insurer's product-specific valuation regulations.

The Society of Actuaries' VBT development report describes the 2015 VBT construction and the use of intercompany data from 2002–2009. It identifies $30.7 trillion of exposure measured by amount of insurance, 266 million exposure units measured by number of policies, and nearly 2.6 million death claims across contributions from 51 companies. These are reported historical study quantities, not the sample sizes of this dossier. In particular, monetary face exposure, policy exposure and person-time must retain their respective units; the report's aggregates should not be paraphrased as 266 million unique insured persons.

A published experience study aggregates definitions, exposure rules, exclusions and contributors chosen for its purpose. The 2015 VBT methodology explicitly excludes certain coverage types from its principal development scope and distinguishes risk categories; its published results cannot be presumed representative of guaranteed issue or pre-need insurance. The 2019 SOA individual-life report spans 2012–2019, with a change in its statistical-agent data supply during the period. A source-critical review therefore tracks contributor selection and changes in collection institutions before treating longitudinal differences as biological mortality change.

Minimum research metadata

For reproducible cohort-level work, retain a definition of the insured unit; contract and coverage identifiers; dates of birth, issue, termination, death and report; benefit amount history; underwriting class and class changes; observation start and end; calendar jurisdiction; exposure methodology; expected-mortality table identity and version; treatment of missing values; join rules across administrative systems; and signed data extracts. Death verification is necessary where benefits are paid following notification and investigation. Data lineage determines whether the statistical numerator actually corresponds to covered death claims arising within the same denominator interval.

A mortality study is a conditional description of its insured population. A study can be impeccably calculated and still answer the wrong question if an analyst imports an unrepresentative reference table. Conversely, sparse but well-documented experience remains useful when combined with genuinely comparable external information and its uncertainty is reported.

Life-Years Exposed and Mortality Measurement

The continuous-time risk set

Let individual life i be eligible for observation at time t when a specified insurance coverage is in force, the individual is alive, the policy lies inside the study boundary, and the event is ascertainable. Define the at-risk indicator Y_i(t) to be one in that state and zero otherwise. Over study interval [a,b], central exposure measured in person-years is

E=∑i∫abYi(t) dt.E=\sum_i\int_a^b Y_i(t)\,dt.

Exposure is measured on a consistent year basis; exact days divided by a declared civil-year convention or actuarial day count may be used. If life i enters halfway through the interval and remains covered and alive to its end, it contributes approximately 0.5 person-year. If death occurs at time d_i, exposure ceases at the event. If contract coverage lapses earlier, policy exposure ceases at lapse. This formulation makes delayed entry, right censoring, changes of age, and coverage changes explicit.

For an interval with constant force of mortality μ and independent conditionally Poisson death processes, the total observed deaths D may be modeled approximately as Poisson with mean μE. The maximum-likelihood estimator is

μ^=D/E,Var⁡(μ^∣E)≈μ/E.\widehat\mu=D/E,\qquad \operatorname{Var}(\widehat\mu\mid E)\approx \mu/E.

A person dies at most once, so a strict full-life death process is a counting process with a risk-set intensity, not an unrestricted collection of repeated Poisson events per person. The Poisson representation is a standard small-interval or piecewise-exponential likelihood approximation; an exact event-history likelihood conditions on individual observed times and risk sets. The approximation must be re-evaluated when hazards are large, event dependence is material, or the exposure itself is estimated with error.

Central, initial and policy exposure

Central exposure measures elapsed time in the observable risk set and naturally estimates a force of mortality. Initial exposure ordinarily relates a beginning-of-period life count to the probability of dying over an age interval under a specified convention. A life-table probability q_x is not numerically identical to the central death rate m_x or force μ_x. Under a constant force over a full year,

qx=1−exp⁡(−μx),μx=−log⁡(1−qx).q_x=1-\exp(-\mu_x),\qquad \mu_x=-\log(1-q_x).

An approximation q_x ≈ m_x is appropriate when the mortality probability is small and the denominator conventions match. The exact link also depends on the shape of death timing and on whether central exposure is actual person-time or reconstructed from life-table survivor counts. An apparent mortality change caused by mixing q, m, and μ is a measurement artifact.

A policy-year exposure counts eligible contracts rather than unique lives and may use fractional duration. A face-amount exposure weights intervals by sums insured, often in currency-years; death claims then require corresponding claim-amount numerator. Define F_i(t) as insured face amount. The amount-weighted exposure is E_F=Σ_i ∫Y_i(t)F_i(t)dt; an amount-weighted death rate would use sums of claim amounts attributable to deaths divided by E_F with compatible timing and benefit definition. If a large policy is disproportionately represented among deaths, amount-weighted and life-count mortality can diverge even in a correctly constructed cohort.

Lexis construction: age, period and birth cohort

Age and calendar time change continuously while birth date remains fixed. A life observed across a birthday or a year-end crosses cells on a Lexis diagram. Split exposure at every relevant birthday and period boundary before aggregation. Otherwise a 60-year-old observed across a sixty-first birthday can be assigned entirely to age 60, causing artificial age effects. The HMD methods protocol, Version 6, revised August 2025, provides a sophisticated example of age-period-cohort exposure accounting for population data, including the use of birth-month information to improve some estimated cohort exposures.

Population cohort techniques depend on national population counts, vital statistics and interpolation; insurer exposure generally derives from individual policy histories. Importing HMD exposures into an insured-life numerator is therefore a denominator mismatch. HMD's methodology is particularly useful as a conceptual benchmark for distinguishing birth cohort, age, period and exposure shape, not as a substitute for administrative contract reconstruction.

Attained age, issue age and policy duration require three separate fields: attained_age(t), issue_age, and duration(t). Underwriting selection affects early policy durations; the observed distribution by attained age changes as contracts mature. For example, a decline in crude mortality within an issue cohort may follow high-risk lives exiting through death, conversion or lapse even if the underlying age-specific hazard rises with aging. Cohort reporting must specify whether it is observing a fixed entry group, all currently in-force policies, or a dynamically replenished portfolio.

Expected deaths and actual-to-expected ratios

A standard reference with rate μ_i^0(t) yields expected deaths as

X=∑i∫abYi(t)μi0(t) dt.X=\sum_i\int_a^b Y_i(t)\mu_i^0(t)\,dt.

Define the actual-to-expected ratio A/E = D/X for X>0. Under a proportional-reference model with subject mortality θ μ_i^0(t), the expectation of D is θX and D/X estimates θ. This normalization allows comparison across age, sex, duration and risk-class compositions, provided the reference applies to those classes and is held at an explicitly stated vintage. A/E is unitless; it is not a mortality probability, and its precision depends on actual and expected events rather than on the sheer number of active contract records.

Reference assumptions become part of the result. A/E of 1.10 against one mortality table may differ substantially from A/E of 1.10 against another table because table level, underwriting definitions and improvement projections differ. Differences in reference calibration can also conceal composition change. The analyst must report the chosen table, selection basis, adjustment factors, study dates and target of inference. ASOP No. 25's relevant-experience provisions make similarity and appropriate adjustments central to professional credibility practice.

Competing exits and observation windows

A policy can terminate because of death, lapse, surrender, expiry, maturity, conversion or administrative correction. These exits have different scientific meanings. For a covered death rate, lapse usually ends subsequent policy exposure. For an underlying person-mortality estimand, a health-related lapse can make surviving observed lives unrepresentative. The distinction concerns an estimand: death while insured versus eventual death for the original applicants, regardless of their continuing coverage.

Delayed death reporting can cause right-edge undercounting, especially when the data extract closes shortly after the observation period. Backdated cancellations, reinstatements, changed benefit amounts and duplicate claims may move the numerator or denominator. The Data Quality standard supports review and disclosure of data suitability, material limitations and adjustments. Claims completeness, exposure completeness and reporting-date completeness should be measured separately rather than collapsed into a single data-quality label.

Heterogeneous Cohorts, Selection, and Survival

A life portfolio is a mixture

Suppose a segment comprises individuals with conditional death intensities μ_i(t). Its crude expected death frequency over observed exposure is a weighted average of these individual intensities. Changing the distribution of ages, underwriting classes, contract amounts, regions or durations can change the crude rate with no within-class change. This is a composition effect, familiar in the general statistical problem of standardization. Cohort comparisons must distinguish changes in conditional hazards from changes in the weights assigned to those hazards.

For mutually exclusive cells j, let E_j be exposure and μ_j its force of mortality. The crude hazard is approximately Σ_j E_j μ_j / Σ_j E_j. Choosing a common reference distribution w_j yields the standardized index Σ_j w_j μ_j / Σ_j w_j. The first responds to the cohort's realized composition; the second answers what would happen under the deliberately fixed common weights. Both are legitimate when reported accurately, and they answer different questions.

Underwriting selection and policy duration

Life underwriting changes the entering distribution of risks. Preferred nonsmoker, smoker, residual standard and simplified-underwriting groups can differ at issue; selection may change with policy duration as risk factors evolve and differential survival changes the mix. The SOA's 2015 VBT materials publish select-and-ultimate tables and distinguish preferred risk classes and sex and smoking criteria. The study's documented data and adjustment choices show why an insured-life reference must be selected for the product and classification actually under examination.

A cohort whose issue-year distribution shifts toward new, intensively underwritten business may show lower crude mortality than an otherwise comparable aging portfolio. That observation does not establish an individual-level mortality improvement. To distinguish underwriting selection from time trend, include issue cohort, attained age, sex/risk class, and duration where observed; inspect the volume of exposure supporting their interaction; and avoid overfitting narrow classes with almost no deaths.

Changes to underwriting policy may produce time-varying sample selection. A health-screening rule introduced in a new issue year changes who enters the contract portfolio. A claim-based trend estimated across the implementation date may therefore reflect both genuine mortality conditions and changed selection. Internal policy documentation, product underwriting guides and administrative data are required to distinguish the effects for a particular insurer. Intercompany tables alone cannot identify an insurer's causal underwriting effect.

Frailty and hidden mixture risk

Let an individual hazard have multiplicative unobserved frailty Z_i, so μ_i(t)=Z_i μ_0(t) after accounting for observed factors. Individuals with larger Z_i are more likely to die early, leaving a survivor population with smaller average frailty. The cohort hazard can flatten even when each individual's conditional hazard follows the reference shape. Frailty is a useful established statistical concept here; this dossier does not propose a new frailty distribution. A convenient parametric frailty model is an assumption to be tested, not an observed measurement.

The implications differ by study target. Pricing newly underwritten lives concerns entry selection and subsequent experience; valuing in-force cohorts concerns the changing survivor mix; comparing policyholders with the national population concerns an entirely different selection mechanism. The selected survivor group and the original entry cohort must not be treated as exchangeable. In sparse cohorts, evidence may be insufficient to estimate both detailed covariates and a frailty distribution; model complexity should be proportionate to observed events.

Censoring, competing risks and survival regression

Right censoring occurs when death time is unknown because observation ends or the subject exits without a recorded death. Kaplan and Meier's product-limit estimator handles censored survival histories under its underlying independence assumptions. Life-insurance lapses can be informative, especially when changing health or affordability affects termination, so standard survival estimates may fail to describe mortality of all original applicants without additional information.

Cox's proportional-hazards formulation models covariate effects multiplicatively while estimating a baseline hazard nonparametrically. Application to an insured cohort requires checking proportionality, time-varying underwriting class, left truncation, ties, delayed claims and the meaning of a recorded exit. A portfolio with only a handful of deaths cannot support a complex age-by-duration-by-class regression merely because its administrative file contains many policy records.

There are two alternative practical endpoints. One models death during active insurance, treating termination as a competing event for covered death claims. Another seeks latent natural mortality after issue; observing only insured time typically leaves this endpoint partially unidentifiable without post-lapse linkage or defensible assumptions. Any reported comparison should identify which endpoint it estimates. In particular, a cause-specific hazard among those still insured is not a cumulative probability of eventual death regardless of lapse.

Mortality trend and period shocks

Calendar time matters because mortality changes under disease, environment, treatment and social conditions. The SOA's individual-life mortality experience series publishes successive studies and valuation materials that require their own periods and data definitions. The HMD's versioned exposure and mortality files provide a separate population series with its own demographic base. Period trends must be estimated within a coherent insured reference or justified transfer adjustment, not by splicing unadjusted national and insured ratios.

An exposure cell indexed only by attained age can conflate changes in calendar period with differences in issue cohorts. Conversely, modeling every age-period-cohort interaction without external structure creates identification difficulties because age, period and birth cohort are linearly related. A practical analysis defines a target estimand and constrains or regularizes terms, tests holdout periods, and reports how much an inferred trend depends on prior assumptions. Sparse experience rarely warrants precise claims about separate biological, underwriting and economic drivers.

Credibility, Sparse Data, and Parameter Uncertainty

Credibility is conditional predictive value

An unadjusted subject estimate D/X is unstable when the expected number of deaths X is small. Credibility combines the subject estimate with relevant outside information to reduce expected prediction error under specified assumptions. In its adopted definition, the ASOP 25 terminology treats credibility as the predictive value assigned to data in an application; it recognizes both testing subject experience on its own and blending subject and relevant experience. There is no context-free credibility percentage attached to a fixed number of lives. Different claim frequencies, observation units, degrees of heterogeneity and future uses yield different appropriate weights.

The conventional linear formula is θ̂_C=Z θ̂_S+(1−Z)θ̂_R, where θ̂_S is a subject estimate, θ̂_R a reference estimate, and 0≤Z≤1 is a chosen credibility factor. Its apparent simplicity conceals two scientific questions: whether the reference is appropriate for the subject cohort and whether the chosen Z follows from an empirical or acceptable decision-theoretic criterion. A reference built partly from the same data is not independent and should not be counted as if it were an entirely external observation.

A credibility weight for a mortality ratio is most naturally tied to expected deaths rather than to raw insured counts. Suppose the reference distribution predicts X=10 deaths among 10,000 low-hazard policy-years and X=1,000 deaths among 10,000 high-hazard policy-years. Under a Poisson proportional-mortality model, the second study contains far more direct information about a common multiplicative mortality parameter. A policy-year-only formula assigning identical weights would ignore that difference. This is a consequence of the likelihood, not a matter of stylistic preference.

Limited fluctuation credibility

A classical limited-fluctuation criterion seeks a specified relative accuracy at a selected confidence level. If death count D is Poisson with mean λ, relative standard deviation is 1/√λ. A normal approximation gives a full-credibility event threshold

λ0≈(z1−α/2/ε)2,\lambda_0\approx(z_{1-\alpha/2}/\varepsilon)^2,

for desired relative error ε and approximate two-sided confidence 1−α. At 95% nominal confidence and 5% relative precision this is (1.96/0.05)^2=1,536.64, or about 1,537 expected deaths. For an expected count of 12, a simple square-root partial-credibility rule would give Z≈√(12/1,536.64)=0.0884; for 200 expected deaths the same rule gives approximately 0.3607. These are assumptions and approximate illustrative weights, not universally mandated values for life mortality. Exact Poisson intervals, model risk and correlated lives can change required exposure substantially.

The normal-approximation threshold also reveals why statements such as 'a thousand insured policies are fully credible' have no general meaning. Full credibility depends on the variance of the quantity being estimated and the targeted tolerance. It may be impossible for a narrow cohort with rare mortality to reach a specified relative precision over a practical experience window. That does not justify ignoring its experience; it supports explicit partial pooling and transparent uncertainty.

Bühlmann and Bühlmann–Straub weighting

Under the classical Bühlmann setting, a homogeneous sequence of conditional observations may be weighted by Z=n/(n+K) with K=EPV/VHM, where EPV is the expected process variance and VHM is the variance of hypothetical means across risks. The construction trades within-risk noise against between-risk heterogeneity. The Bühlmann–Straub work extends the idea to differing observation volumes: in an appropriate weighted model, Z=w/(w+K), where w is total statistical exposure weight under the model's variance assumptions. Here K inherits measurement units from the chosen weight; it should not be carried across products or mortality scales without re-estimation or re-expression.

These formulas rely on a defined portfolio of heterogeneous underlying risks and appropriate process-variance assumptions. A collection of deaths with unmeasured common shocks, repeated dependent policies, unknown underwriting change or severe temporal drift violates the simplest variance structure. The analyst should ask whether the relevant experience predicts a stable mortality characteristic of the subject cohort during the future period of use. A model that describes variation across calendar-year observations may not predict a new period after a discontinuous shock.

Bühlmann–Straub's volume-based logic is valuable for cohorts with fractional exposure: observations of unequal information content must not be averaged mechanically. The credibility measure may be based on person-years, amount-years, expected deaths, inverse variance or other explicit weights, depending on the outcome. It does not follow that monetary exposure automatically furnishes the appropriate likelihood information for death incidence, because large policies may be selected differently and may exhibit claim-amount heterogeneity. Such a change of target requires an explicit claim-severity or benefit model.

Poisson–gamma partial pooling

A transparent alternative for a common multiplicative mortality parameter uses a reference death expectation X. Suppose D | θ,X is Poisson with mean θX, and θ has a Gamma prior with shape a and rate b, so its prior mean is a/b. Conjugate updating gives

θ∣D,X∼Gamma⁡(a+D,b+X),E[θ∣D,X]=a+Db+X.\theta\mid D,X\sim\operatorname{Gamma}(a+D,b+X),\qquad \mathbb{E}[\theta\mid D,X]=\frac{a+D}{b+X}.

Set a=b=k so the prior mean mortality ratio is 1. Then the posterior mean is k/(k+X)×1 + X/(k+X)×(D/X) for X>0, with credibility weight Z=X/(X+k). Here k has the dimensions of expected deaths under the reference: it is a prior information parameter, not a magic number of policy-years. The derivation is exact for the stated likelihood and prior, and k must be selected or estimated with reasons. Under this specification the posterior variance is (a+D)/(b+X)^2, while posterior predictive deaths also include irreducible Poisson variation.

For a cohort with a priori X=12 expected deaths and D=9 observed, choose solely for illustration k=100. The raw ratio is 0.75, Z=12/112≈0.1071, and the posterior mean is (9+100)/(12+100)=109/112≈0.9732. For a second cohort with X=200, D=240 and the same prior, the raw ratio is 1.20, Z=200/300≈0.6667, and the posterior mean is 340/300≈1.1333. The first estimate is drawn more strongly toward the reference because it has less expected-event information. The difference is determined by declared parameters and is reproducible.

Prior strength is itself uncertain. Choosing k=10, k=100, or k=1,000 can materially change an estimate from a cohort with 12 expected deaths. For the first cohort the resulting posterior means are respectively 19/22≈0.8636, 109/112≈0.9732, and 1009/1012≈0.9970. These are three mathematical outputs from the same observed deaths; the appropriate one depends on defensible historical between-cohort heterogeneity and the forecasting question. Fitting k to a collection of related cohorts is empirical Bayes; modeling k with a hyperprior is fully hierarchical Bayes. Each adds estimation uncertainty and depends on how the reference cohort family is defined.

Posterior distribution, intervals and rare outcomes

A point estimate hides tail uncertainty. With a=b=100, first-cohort posterior shape is 109 and rate 112, whereas second-cohort shape is 340 and rate 300. The first has posterior variance 109/112²≈0.00869 and standard deviation about 0.0932; the second has posterior variance 340/300²≈0.00378 and standard deviation about 0.0615. These are uncertainties about θ under the model and prior, not estimates of additional sampling variance from reporting error or unmodeled systemic shocks. They should not be called unconditional confidence intervals.

For a segment with zero deaths, D/X=0 is a maximum-likelihood boundary estimate and never proof of zero risk. A gamma posterior remains positive when a>0. The relative uncertainty of a sparse death rate is asymmetric, especially at zero or one death; symmetric normal bands can extend below zero and should be replaced by exact-count intervals or appropriate posterior summaries. A credibility procedure that suppresses observed anomalies by excessive shrinkage can also be harmful if those anomalies reflect an authentic new hazard. Outlier investigation and leave-one-cohort-out validation complement shrinkage.

The Poisson likelihood treats exposure as known and events as conditionally independent given rate parameters. Administrative exposure reconstructed from incomplete records introduces measurement error not captured by that likelihood. Shared catastrophe, epidemic or geographical exposure can inflate death-count variance relative to Poisson. If Var(D|X) exceeds E[D|X], a Poisson likelihood with a fixed reference can understate uncertainty. Overdispersion and latent correlation should be examined through residuals, stratification, predictive checks and alternative models. In a hierarchical model, some overdispersion emerges from variation in θ; it must not be counted twice through incompatible extra variance terms.

Credibility for rates, ratios and improvements

A cohort mortality rate D/E has units of deaths per person-year; an A/E ratio D/X is dimensionless; a mortality-improvement rate describes relative change between periods; and a death-benefit cost may be measured per dollar of insured face or per policy. Credibility estimates for these quantities have distinct observation models. A study that provides stable evidence for relative A/E level may supply little independent information on a time slope. Fitting both level and improvement for a narrow cohort can create poorly identified parameters. A partial-pooling model should reflect the hierarchy: pooled reference levels, subgroup deviations, and temporal terms supported by observed events.

When the ultimate decision concerns future liability cash flows, a parameter estimate requires projection to the future population and its contract mix. A credible historical mortality estimate can be a poor forecast if the future product, class mix or market has changed. The IFRS Foundation's IFRS 17 description emphasizes current estimates of fulfillment cash flows and separate grouping and risk adjustment in insurance-contract measurement. This is a financial-reporting context that can consume mortality assumptions. The statistical credibility factor by itself is neither an IFRS 17 measurement policy nor a claim that a jurisdiction follows IFRS reporting.

Selection of external experience and leakage

Reference relevance is a substantive issue. The ASOP 25 selection-of-relevant-experience guidance concerns similarity of demographics, coverage and risk characteristics, and warns against double counting subject information contained in the reference. If the industry mortality table itself contains the insurer's historical experience, treating its estimate as statistically independent of that insurer's subject data exaggerates total information. A leave-one-insurer-out reference, when available, or covariance-aware modeling is preferable for an independence claim.

A hierarchical prior estimated from peer portfolios inherits their underwriting and reporting conventions. Transfer to the subject must be audited for geographic, temporal, sex, smoking, issue-age, benefit and duration compatibility. The SOA 2015 VBT report documents multiple underwriting risk categories and process adjustments, underscoring the dangers of treating a single published table as an exchangeable prior for every life-insurance book.

Comparative Evidence and Reproducible Case Studies

Insurance experience and demographic exposure

An insured mortality study and a national mortality series can both report D/E; their unit and selection mechanisms differ. In the Society of Actuaries' 2009–2015 mortality experience report, intercompany experience concerns standard individually underwritten policies, and supporting tables and files make some analyses reproducible. HMD instead calculates population exposure from demographic records using explicit age/period constructions. Its explanatory notes distinguish exposure, death rate, and cohort-life-table products. A comparison requires risk standardization, source-vintage alignment, consistent treatment of coverage exits and explicit disclosure of underwriting selection.

A superficially low insured mortality ratio relative to a national population can reflect the health and socioeconomic selection of people who obtained and retained insurance. A superficially elevated claim cost may reflect higher insured face amounts concentrated among high-risk groups. Because the underlying event and populations differ, neither finding determines the causal effect of insurance on health. This synthesis follows the documented source definitions and the mathematical distinction between person-level and contract-level estimands; it does not claim a newly measured international effect.

The 2019 SOA individual-life report is especially pertinent to inter-period comparison because it describes a change in the statistical agent supplying data. Before attributing a difference in reported A/E trends to mortality improvement, investigators should check whether field definitions, coverage of contributors, observed risk categories and extract processes changed. Cohort credibility assessment cannot correct a structural data discontinuity merely by increasing the weight on a larger but incommensurable dataset.

A related yet distinct population is payout annuitants. The SOA/LIMRA 2020–2024 payout annuity study reports 3.1 million contract-years exposed, roughly $33 billion in annual income-years of exposure and 143,190 deaths from its contributors. Its denominators and contractual economic incentives differ from those of individually underwritten life policies. It demonstrates why the choice among contract-year, income-weighted and life-count measures must follow the target; it does not supply a transferable default life-insurance credibility parameter.

Reproducible case study: exposure and ratio information

Consider two synthetic life cohorts observed for one year. The reference mortality intensity is set, solely for arithmetic simplicity, to 0.010 deaths per person-year across each cohort; this intentionally removes age and class structure to isolate the credibility calculation. Cohort A accumulates 1,200 person-years and nine observed deaths. Cohort B accumulates 20,000 person-years and 240 deaths. Both cohorts are hypothetical and carry no empirical claims about the mortality of any country or insurer.

Quantity Cohort A Cohort B
Central person-years E 1,200 20,000
Reference hazard μ₀ 0.010 0.010
Expected deaths X=Eμ₀ 12 200
Actual deaths D 9 240
Crude hazard D/E 0.00750 0.01200
Raw actual-to-expected D/X 0.7500 1.2000
Gamma information k 100 100
Predictive weight X/(X+k) 0.1071 0.6667
Posterior mean A/E (D+k)/(X+k) 0.9732 1.1333

The naive arithmetic average of the two raw cohort A/E ratios is 0.975. The combined experience ratio is instead (9+240)/(12+200)=249/212≈1.17453. The discrepancy is expected: an equal average of cohort estimates assigns the small cohort the same aggregate weight as the large cohort, whereas the combined estimate weights by expected deaths. Neither combined number is an automatic estimate of the separate cohorts' future mortality ratios, because they may have distinct underlying risk parameters. Pooling requires an exchangeability assumption.

A counterfactual data error illustrates denominator sensitivity. If cohort A's 1,200 person-years include 200 years after policies ceased being covered, corrected exposure would be 1,000; expected deaths would fall from 12 to 10 and the raw A/E would rise from 0.75 to 0.90, despite unchanged death claims. Under k=100 the adjusted posterior mean would become 109/110≈0.9909. Correcting exposure therefore changes both the subject estimate and its credibility information. This is why exposure QA precedes a final decision about smoothing.

A second scenario concerns an unreported delay: if two additional genuine deaths related to cohort A's observation year were verified after the first extract, its original-exposure raw ratio would become 11/12≈0.9167 and the posterior mean 111/112≈0.9911. The effect of late reporting differs from exposure overstatement. Conflating the two as a generic data-quality variance component would hide the actual correction mechanism.

Reproducible case study: apparent reversal after composition adjustment

Consider two synthetic cohorts comprising a lower-hazard and higher-hazard class, each observed over a common interval. For cohort C, 900 person-years belong to the lower class at an assumed hazard of 0.005 and 100 to the higher class at 0.020. For cohort D, 100 person-years belong to the lower class and 900 to the higher class. Reference-expected deaths are 900×0.005+100×0.020=6.5 for C and 100×0.005+900×0.020=18.5 for D. Suppose observed deaths are 8 and 18, respectively. Their crude empirical death rates are 0.008 and 0.018, yet their actual-to-expected ratios are 8/6.5≈1.2308 and 18/18.5≈0.9730. The cohort with the lower crude mortality rate performs worse relative to its own reference risk composition.

The example does not establish that cohort C is intrinsically less healthy than D. Its expected count is small, so the ratio is statistically noisy. Using the same illustrative gamma prior k=100 gives posterior mean ratios 108/106.5≈1.0141 for C and 118/118.5≈0.9958 for D. The observed crude contrast is dominated by exposure composition; the credible residual contrast is small under this prior. The data are deliberately simple to isolate the inferential distinction.

A different, equally legitimate question would standardize both cohorts to the same lower/higher-class weights, if class-specific death counts were available. The aggregated observations D_C=8 and D_D=18 cannot identify the two class-specific mortality parameters within each cohort. It would therefore be an error to infer an exact standardized difference from these aggregate counts. This is an example of identifiability preceding estimation: the desired comparison may demand more granular observations than the available data permit.

Sensitivity analysis and validation

A complete cohort analysis should vary plausible exposure corrections, the age/duration mix, reference table, credibility strength, reporting delay, mortality trend, exclusion rules, underwriting vintage and lapse treatment. Where resources permit, evaluate predictive performance on later calendar years or held-out issue cohorts, separately from in-sample goodness of fit. The adopted Modeling standard, ASOP 56, addresses appropriate models, data, assumptions, validation, documentation and communication. Its scope and purpose are professional practice; passing its procedural requirements cannot replace an actual numerical holdout test.

Useful diagnostics include a reconciliation of covered lives at start and end, summed fractional exposure versus aggregate in-force counts, distribution of missing fields, actual/expected residuals by attained age and duration, posterior predictive checks, drift in underwriting class proportions, and stability of the subject-to-reference ratio across successive periods. Calibration plots should be based on adequate grouped counts and should show uncertainty. A segment with three deaths should not be presented with the same apparent numerical precision as one with several thousand.

International comparison and regulatory framing

Actuarial standards, accounting rules and historical statistical datasets serve different institutional functions. The US ASOPs apply under specified professional conditions; IFRS 17 is an insurance-accounting standard effective for reporting periods beginning January 1, 2023, where adopted and applicable, and describes measurement of insurance contract groups and future cash flows; international demographic sources report population mortality, not the legal definition of a life policy. The IFRS 17 standard overview makes the accounting distinction clear. None of these documents supplies a universal legal requirement to use Z=X/(X+k) for an insured-life cohort.

A cross-country analysis must track differences in underwriting permission and practice, insured-product composition, death registration, reporting lag, policy documentation and local governance of actuarial work. A single country-specific reference cannot automatically support a universal conclusion about exposure credibility. Where direct insured-mortality datasets from a jurisdiction are unavailable, the proper result is a conditional methodological conclusion with an evidence-access boundary, rather than an invented country comparison.

Integrated Scientific Findings and Actuarial Governance

Findings from the combined evidence

The first finding is definitional: a cohort mortality rate is a ratio between an event measure and a matched measure of eligibility and exposure. Its identity must be fixed before computation. Exact exposure intervals, single-person versus multiple-contract treatment, and event ascertainment determine what the estimate means. Increasing the sample size cannot correct an inconsistent numerator and denominator.

The second finding is statistical: for a proportional mortality model, expected deaths supply the natural exposure-information scale of an A/E estimate. The Poisson variance and gamma posterior show why the number of insured policy records is a misleading standalone measure of predictive information. Credibility is conditional on an estimand, model and decision purpose. Sparse cohort information benefits from defensible partial pooling, while arbitrary or overly strong pooling can obscure real emerging risk.

The third finding is structural: mortality heterogeneity and changing membership create composition effects. Attained age, underwriting class, issue duration, policy lapse and calendar period influence which lives are observed. Standardization or a covariate model can describe differences under a specified reference, but the resulting estimates require adequate class detail, compatible data, and assumptions about censoring. Without granular deaths and exposure, some within-class comparisons cannot be identified from aggregate statistics.

The fourth finding concerns historical evidence. The Society of Actuaries' life-insurance experience study series and the HMD's population data documentation are authoritative within their respective populations and measurement conventions. Their coexistence demonstrates the importance of source scope. The VBT table is valuable evidence about constructed insured risk categories, while the national population series shows the challenges of Lexis exposure estimation. Their differences cannot be erased through a credibility weight.

The fifth finding is operational. The adopted US standards Data Quality, Credibility Procedures, and Modeling address, respectively, suitability and disclosure of data, selection and use of credible information, and modeling validation and communication. Their combined relevance supports a documented chain of analytical decisions. Scientific evidence for that chain remains the actual data, calculations and justified assumptions; a compliance label alone does not validate the estimate.

These findings constitute a synthesis of established actuarial and statistical methods. They identify logical dependencies among the components of a mortality analysis and derive their consequences in explicit examples. They are neither announced as a new actuarial theory nor attributed empirical significance beyond the reviewed evidence.

An auditable research procedure

An independent reviewer should be able to trace each mortality conclusion through the following sequence, carried out as one connected workflow: define the insurance coverage and event; reconstruct time at risk; reconcile policy/person identity; document classification and excluded contracts; map attained age, calendar time and duration; select and version a reference; calculate actual and expected deaths; evaluate uncertainty; choose or estimate a credibility method; test model assumptions and perturbations; document the target forecast; and archive source hashes and decisions.

In a ledger of substantive claims, a primary source supports a precise assertion, a derived mathematical identity is labeled by its assumptions, and a synthetic example is separated from insurer observations. A document title and DOI identify research provenance; an effective date identifies a standard's jurisdictional time frame; a cohort table version identifies the benchmark used in the calculation. The source ledger accompanies the manuscript and provides claim-level traceability independently of the reader-facing bibliography.

The Actuarial Communications standard provides an additional US professional context for disclosing assumptions, methods, limitations and intended use. In this scholarly publication, that logic is reflected through transparent definitions, limitations, explicit reference links and illustrative computations. This is a research monograph, not actuarial advice, a policy quote, or a statement of compliance for an actual insurer.

Limitations and future empirical work

This research is based on publicly accessible institutional reports, standard text, scholarly metadata or accessible article extracts, and disclosed mathematical reasoning. It has not obtained confidential company mortality files, verified individual SOA pivot-table rows, conducted a full cross-jurisdictional survey of national actuarial law, performed a survival-regression fit on microdata, or commissioned external peer review. Source titles and dates are reproduced from the reviewed authoritative records; older historical papers are discussed at the level supported by their accessible bibliographic and substantive material. These bounds concern empirical generalization, not the validity of the displayed algebra under its declared assumptions.

Further empirical investigation should evaluate held-out calibrated mortality forecasts for separately defined issue, attained-age and policy-duration cohorts; quantify informative lapse through linked follow-up where legally and practically available; compare exposure and death-reporting corrections before and after reconciliation; assess reference leakage for participating insurers; and measure uncertainty under common-shock or overdispersed death processes. The evidence needed for those questions is specific: granular cohort data, reproducible code, validated records and governance of sensitive information. Their completion would be separate empirical research, not an unstated result of the present work.

Conclusion

Cohort-level life insurance exposure and credibility adjustment is a problem of joint measurement and inference. Deaths become interpretable when matched to precisely reconstructed insured time at risk. Underwriting selection, duration, changing coverage, lapse, reporting delay and population composition determine what the observed cohort represents. Expected deaths, calculated from a documented and relevant reference, provide the information scale for a proportional-mortality A/E analysis. Credibility may then pool sparse experience with other evidence under an explicit variance or prior structure, and its predictive value must be tested for the intended future population.

The formal analysis establishes the difference between a cohort's crude death rate, its standardized actual-to-expected ratio, and its partially pooled estimate. Reproducible examples show both how a large change in raw exposure information alters credibility weights and how a crude mortality comparison can reverse in interpretation after reference-risk adjustment. These findings follow from declared likelihoods, source definitions and arithmetic; they establish the conceptual and operational consequences of existing scientific methods.

An auditable approach combines correct exposure accounting, source-relevant mortality benchmarks, cautious evaluation of heterogeneous cohorts, uncertainty analysis and professional disclosure. It preserves the distinct authority of actuarial standards, experience studies, demographic databases and accounting frameworks. The resulting research architecture supports accurate cohort-level interpretation across life-insurance applications while keeping each result tied to its data, historical period, contract population and assumptions.

References

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

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

  • Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (revised December 2016). Official adopted standard.

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

  • Academy/SOA Valuation Basic Table Team. (2018). 2015 Valuation Basic Table Report (revised September 2018). American Academy of Actuaries and Society of Actuaries. Institutional research report.

  • Bühlmann, H. (1969). Experience rating and credibility. ASTIN Bulletin, 5(2), 157–165. Original journal article.

  • Bühlmann, H., & Straub, E. (1970). Glaubwürdigkeit für Schadensätze. Mitteilungen der Vereinigung Schweizerischer Versicherungsmathematiker, 70. Digitized original and DOI.

  • Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202. Journal article and DOI.

  • Human Mortality Database. (n.d.). Explanatory Notes: Human Mortality Database. University of California, Berkeley, and Max Planck Institute for Demographic Research. Official data documentation.

  • IFRS Foundation. (2017; effective 2023). IFRS 17 Insurance Contracts (effective for annual reporting periods beginning on or after January 1, 2023). Official standard overview.

  • Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. Original journal article and DOI.

  • Society of Actuaries. (2015; updated 2018). 2015 Valuation Basic Report and Tables. Official 2015 VBT resource.

  • Society of Actuaries, Individual Life Experience Committee. (2017). 2009–2013 Individual Life Insurance Mortality Experience Report. Official intercompany study.

  • Society of Actuaries, Individual Life Experience Committee. (2019). 2009–2015 Individual Life Insurance Mortality Experience Report. Official intercompany study.

  • Society of Actuaries, Individual Life Experience Committee. (2024). 2019 Individual Life Insurance Mortality Experience Report: Experience Years 2012–2019. Official intercompany study.

  • Society of Actuaries. (2026). Individual Life—Mortality Experience Studies (research listing and archived materials). Official research index.

  • Society of Actuaries and LIMRA. (2026). 2020–2024 Individual Payout Annuity Mortality Experience Study. Official research report page.

  • Wilmoth, J. R., Andreev, K., Jdanov, D., Glei, D. A., & Riffe, T. (2025). Methods Protocol for the Human Mortality Database, Version 6 (revised August 5, 2025; with contributions acknowledged in the protocol). Human Mortality Database. Official methodological protocol.