Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Annuitant Mortality Experience Data Provenance

Abstract. Annuitant mortality is measured through relationships among contractual eligibility, observed survival, recorded death, administrative transactions, and time at risk. Each relationship has a documentary provenance and may introduce measurement error. This monograph examines the scientific foundations and international evidence for producing traceable mortality experience from individual payout annuities, deferred income contracts, pensioner records, and survivor benefits. It reconstructs how original contracts and administrative data become measuring-life records, risk intervals, death counts, income-weighted exposures, and actual-to-expected mortality statistics. The analysis evaluates identity resolution, death registration, conflicting dates, late reporting, commencement and termination codes, joint-life records, left truncation, informative censoring, missingness, insurer participation, socioeconomic composition, and differing institutional definitions. Primary evidence includes the Society of Actuaries Research Institute/LIMRA 2020–2024 payout annuity collection specifications and reported results, British Continuous Mortality Investigation documentation, U.S. National Death Index guidance, international actuarial professional standards, and foundational statistical literature. Three reproducible numerical constructions demonstrate how exposure mismeasurement, death-match error, population composition, and expected-mortality weighting affect reported rates. The original synthesis establishes an auditable chain from source facts to actuarial inference while preserving the distinction between documented evidence and assumptions that can be tested only with confidential underlying data. No new proprietary theory, doctrine, or forecasting model is proposed.

Keywords: Annuitant mortality; death ascertainment; mortality experience; data provenance; life-contingent benefits; pension commencement; record linkage; exposure at risk; person-years; contract-years; joint-life annuities; missing data; left truncation; informative censoring; mortality selection; actuarial experience studies; actual-to-expected mortality; reproducibility; mortality coding; pension statistics.

Introduction

The mortality of an annuitant is an event in a human life, but the mortality experience of an annuity portfolio is a statistic produced by a sequence of institutional decisions. A contract becomes eligible under a study definition; an identified measuring life enters observation; an administrator records in-force and benefit states; a death is discovered, matched, documented, and assigned a date; and the resulting records are transformed into deaths, years exposed, and expected mortality. The scientific reliability of the statistic depends on the continuity of that sequence.

The SOA/LIMRA 2020–2024 study reports 143,190 deaths, 3.1 million contract-years exposed and approximately $33 billion in annual-income-years exposed from 23 parent company groups and 26 individual insurers. The publication states that contributors covered more than 80% of industry sales during the period. The distinct units of exposure are consequential: a contract-year, a person-year and a dollar of annual income exposed for a year measure different objects. The reported coverage of sales is not automatically equivalent to coverage of annuitant person-time or every benefit cohort. This investigation interprets the official counts within their declared study design and does not assert access to contributor-level observations.

The earlier 2014–2019 SOA/LIMRA experience study reports approximately 4.3 million contract-years and 236,000 deaths. A comparison between these published aggregates could reflect changing insurer participation, portfolio maturity, ages, contract types, deaths, observation periods, and benefits. A scientific comparison requires control of the relevant composition and measurement rules. Neither aggregate count alone identifies the direction or magnitude of age-specific mortality change.

The research object is annuitant mortality experience data provenance: the traceable origin, transformations, validation history and evidentiary authority of observations used to estimate mortality in specified annuity or pension populations. This domain connects actuarial statistics, administrative records, survival analysis, insurance contract taxonomy, data governance and access law. It is narrower than general longevity forecasting, reserve valuation or financial guarantee design. Those subjects are discussed only where they depend upon the construction of observed mortality data.

Methodologically, the monograph combines primary-source document interpretation; formal mortality and person-time definitions; examination of actual study inclusion and coding rules; comparison of insurer and pension investigations; identification of ascertainment and missingness mechanisms; reproducible arithmetic; and a transparent evidence ledger. Source-specific facts are distinguished from illustrative derivations, inferences and restricted datasets. The practice framework is informed by IAA ISAP 1, whose data-quality provisions explain the need to review source reliability, completeness, consistency and materiality and to disclose relevant limitations.

Scientific Foundations, History, and Research Identity

Definitions and observation units

An annuitant, in the operational sense of an experience study, is a person satisfying that study's defined relationship to an annuity benefit. The contractual measuring life, recipient of an income payment, legal contract holder, original purchaser and surviving beneficiary can be different persons. Their identifiers and dates cannot be treated as interchangeable. A payout annuity may remain a legal contract after a measuring life's death through a period-certain guarantee. A joint-life product may cover two lives while creating a single contract obligation, and a deferred income contract may identify a measuring life before the first scheduled benefit payment.

An experience dataset is therefore defined jointly by its contracts, lives, calendar interval, eligibility conditions, death event definition and weight system. The measuring-life year is suitable for some mortality estimands; a contract-year may be a convenient administrative reporting unit; annual-income-years can summarize the financial weight of exposed benefits. An observed-to-expected ratio must specify the unit and mortality basis used on both sides of the fraction. Annuity issue, in-force observation, income commencement, primary death, survivor transition and contract termination are separate events with independently meaningful dates.

Provenance is the documented history that connects source entities, processing activities and responsible agents to each analytic record. The W3C PROV-O Recommendation represents these relations through entities, activities and agents. Applied to mortality research, relevant entities include a versioned contract extract, an administratively verified death notice, a vital-registration match, a cleaned life-period, and the resulting aggregate death cell. Activities include linkage, status recoding, interval derivation, adjudication, aggregation and correction. The provenance framework is descriptive: its use here does not mean that the W3C has issued an actuarial mortality standard.

Historical movement from tables to linked observations

Classical mortality tables describe numbers surviving and dying between ages under defined observation schemes. As life offices, pension funds and annuity providers accumulated contracts, the denominator came to depend on the insured population's rules of entry, withdrawal and entitlement. Computerized administration has made it possible to combine large quantities of policy, pension, payment and death information, but has also multiplied potential discontinuities among source systems. Intercompany experience studies create additional transformations between contributors' local definitions and common specifications.

The 2012 Individual Annuity Reserving materials record a mortality table and associated reserving framework. Such valuation bases are distinct from observed experience. The public 2014–2019 study highlights report amount-weighted actual-to-expected mortality of 108.7% against the 2012 IAM Basic Table projected with G2. This result is a comparison against a specific expected-mortality definition; it is not an unconditional probability that a particular annuitant dies.

The British CMI pension-scheme investigation covers self-administered pension schemes and distinguishes pensioners from dependants. Its data collection framework differs from the insurer-based annuity investigation. The comparison highlights an enduring question: whether reported mortality differences arise from different underlying people, eligibility rules, observation quality or real death rates.

Primary records and contractual classification

The scientific source hierarchy begins with the actual benefit terms, dated pension or annuity contract and amendments, original source event, administrative audit history and stated study data specification. A verified death register or certificate provides evidence for mortality endpoints; linked financial records provide payment and status history. A study summary establishes published results and participation descriptions, but generally does not establish the true quality of each underlying source row.

The 2025 SOA/LIMRA data request supplies an unusually useful publicly inspectable record. Its appendices identify contract effective date, in-force date, income commencement date, termination mode and date, age or date of birth, contract characteristics, benefit class, annuitant status, study year and segment identifiers. The specification requests contributor-assigned anonymous identifiers rather than actual Social Security or policy numbers. It also defines exclusions and permissible forms of submission. The presence of a requested field provides a definition of the collection process, not proof of completeness of all submissions.

The source document specifies the purpose of comparing 2020–2024 mortality with the earlier 2014–2019 study, the 2012 IAM bases, and different benefit types. Its inclusion exclusions—such as group pension payouts and structured settlements—prevent an investigator from treating its published results as a general pensioner mortality census. Individual and group annuity populations can share the economic function of retirement income while having different contractual, selection and data histories.

Record Linkage and Mortality Coding

Linking the correct measuring life

Mortality analysis requires a stable mapping from administrative contract keys to natural persons. A contract number can identify a product record yet fail to identify a unique person because individuals may own multiple contracts. Joint-life arrangements reverse the problem: one contract may contain more than one measuring life. Consolidations, migrations, spelling variations, date-of-birth corrections and survivorship changes can create further key instability.

An auditable linkage process retains input identifiers in appropriately protected form, normalizes fields under recorded rules, assigns candidate matches, evaluates conflicts, documents human or algorithmic adjudication, and preserves the matching decision. A definitive death should require more than an unreviewed fuzzy match when the consequence of misclassification is material. Rejected matches and ambiguous candidates are part of the scientific record because their characteristics determine possible false-negative and false-positive errors.

The National Death Index in the United States provides approved health and medical researchers with selected information from death certificates, including death dates and, through its designated service, causes of death. Crucially, CDC states that NDI access is for eligible statistical research uses and excludes ordinary nonstatistical commercial administration. An insurer cannot assume this research-access mechanism is available for its routine benefit operations. Researchers must identify the precise death source, authorization, match process and limits of coverage.

The NDI search procedures distinguish search approaches and matching requirements. Even the existence of a highly comprehensive national register does not make each candidate match certain. The GAO death-record assessment documented inaccurate and incomplete death records in the Social Security Administration's different Death Master File process. A Death Master File match and an NDI match are not the same evidentiary event. Source attribution needs the institution, system, version, reference date and resolution of conflicting outcomes.

Death occurrence versus administrative processing

A central provenance requirement is to retain the date on which a person died separately from dates when death was reported, recorded, verified and applied to the annuity account. Consider an annuitant dying on December 28, 2023, whose death was first reported on January 20, 2024, and accepted by the administrator on February 2. Mortality for the 2023 experience year generally concerns the original death date if the person was eligible on that date, while payment corrections belong to a different administration timeline. Merely counting the 2024 processing transaction as a 2024 death shifts the mortality event across calendar years.

Likewise, pension payments after death may reflect a guaranteed continuation period, processing lag, bank settlement or survivor entitlement. Such payments do not prove that the original measuring life was alive. Termination for surrender, commutation, exchange, transfer or benefit expiry does not establish biological death. A file that records only one undifferentiated termination date cannot resolve these questions without additional administrative or contractual evidence.

The event history should therefore record a pseudonymous life key, contract key, segment key, event type, event time, source system, ingestion time, source version, confidence or verification status, processing operator and supersession link. If a death date is corrected, both the old and corrected assertions remain traceable. Destructive overwriting destroys information about the time at which study administrators could have known the true event.

Joint-life classification

Joint-life annuities require explicit treatment of first-to-die and last-survivor states. The SOA/LIMRA joint-life examples show coding of measuring-life segments and annual income for secondary annuitants. They also allow for contributors that cannot follow secondary lives before the death of the primary annuitant. This documented limitation matters statistically: observing a secondary life only after a transition selects a population that must have survived until the transition. Its pretransition mortality cannot be inferred by recording zero deaths during unobserved intervals.

Survivor benefit commencement must be separated from initial contract issue and from the original annuitant's first income payment. The original life may cease mortality exposure while the survivor begins or continues under a distinct risk state. The ratio of benefits payable after transition is an income-weighting variable rather than a mortality identification variable. Each transformation must preserve the contractual trigger that explains the data state.

Cause-of-death evidence

All-cause experience requires death occurrence and an eligible event date. Cause-specific work introduces death-certificate revision, jurisdictional classification and ICD coding issues. CDC's NDI service descriptions distinguish the service supplying cause-of-death information. An annuity administrative death reason should not be interpreted as a verified medical cause without the relevant underlying evidence. Original-certificate and corrected-certificate versions and ICD classifications are material to a cause-specific study but typically unnecessary for the basic all-cause annuitant rate.

In-Force Exposure Windows

Observation as an integral of eligibility

Let each measuring life i have a documented entry time a_i, exit time b_i, death time T_i, and a prescribed study period [s,e). Define Y_i(t)=1 when the life is eligible, observable and alive immediately before t, and Y_i(t)=0 otherwise. Person-time exposure measured in years is

E = Σ_i ∫_s^e Y_i(t) dt.

The eligible death count D is the sum of verified deaths at times where the corresponding life was still at risk immediately before death. A central estimate of the mortality rate is m=D/E. Both numerator and denominator require the same population and time convention. Under a constant instantaneous force μ throughout a one-year interval, q=1−exp(−μ), so a central rate and a one-year death probability are not numerically identical except by approximation.

The analytic treatment of incomplete follow-up in Kaplan and Meier (1958) and the risk-process treatment of Andersen and Gill (1982) make entry, event and censoring definitions explicit. These methods do not themselves guarantee that an administrative dataset meets their assumptions. In particular, selective withdrawal, incomplete death identification and delayed entry require scientific scrutiny.

Commencement dates and delayed entry

An annuity can have a contract effective date, an income commencement date, and a study observation date. A deferred-income annuitant may be contractually relevant long before income starts. Another study may define only years when payout has commenced. The distinction must be written into the risk indicator rather than inferred from a missing payment.

An in-force date is also distinct from data availability. If a company migrated records in 2022, records that appear to be in force from 2018 may lack reliable death or exit histories before migration. Treating the reconstructed records as fully observed over those years generates unverified person-time. Delayed-entry methods condition on survival to the start of reliable observation; the study question then becomes mortality conditional on that selected entry, not mortality from original contract inception.

Exits and eligibility

Death ends biological risk for the measuring life. Administrative censoring at the study end, surrender, transfer, commutation, and unobserved loss of contact must be evaluated separately. A missing annual file does not establish that a person has died or left contract risk. Similarly, a guaranteed period may preserve cash flow after death but cannot justify continued biological exposure for the deceased original life.

A survivor whose pension starts only after another person's death is selected on both the contractual relationship and survival to that commencement. This has consequences for comparison with an original pensioner population. The CMI SAPS collection description documents pensioner and dependant categories, which must be kept separate in exposure reconstruction.

Reproducible person-time example

Consider 1,000 eligible lives alive on the first day of a synthetic 365-day study year. Suppose 20 die exactly halfway through the year, 100 others exit eligibility one quarter through the year, and 880 remain eligible for the whole year. No one enters after the first day. With fractions of 0.5 and 0.25 years, the exposure is E=880×1+20×0.5+100×0.25=915 person-years. The central mortality rate is 20/915=0.0218579 per person-year.

Dividing 20 deaths by the beginning-of-year headcount of 1,000 instead gives 0.0200000, which is 8.5% below the correctly specified person-time rate. This difference follows entirely from the denominator convention and the explicitly given exits; it is neither an estimate of contemporary annuitant mortality nor evidence that every study must include all voluntary exits. If the 100 records were transferred while follow-up of their lives continued, the appropriate censoring point could differ. The purpose of the case is to demonstrate that eligibility and observation decisions must be documented before any rate is computed.

Counting units and financial weights

A joint-life contract can create two qualifying measuring-life periods. Its contract-time and lives-time totals may therefore differ. A dollar-weighted mortality comparison can use time-varying annual income w_i(t), producing an income exposure Σ_i ∫Y_i(t)w_i(t)dt. The units are currency-years. Whether a secondary life with no current income payment receives zero financial weight depends on the declared measurement scheme; zero weight is not equivalent to biological ineligibility.

If the expected death count under a specified mortality basis is E_D=Σ_i q_i for equally observed discrete intervals and the observed eligible deaths are D=Σ_i d_i, the count-based actual-to-expected ratio is D/E_D. With amounts w_i, a simple amount-based ratio is Σ_i w_i d_i / Σ_i w_i q_i; partial-year terms require consistent time treatment. The SOA 2014–2019 highlights demonstrate how amount-weighted and other stratified metrics are reported. Such ratios have no meaningful interpretation without their expected basis and exposure units.

Selection, Missingness, and International Comparability

Who is observed?

The purchase of an annuity, participation in an occupational scheme, conversion of a defined benefit pension, and election of a survivor benefit are different selection processes. Individuals differ in health, resources, employment history, expected longevity and information. Insurer contributors are also selected: organizations with eligible contracts and available extract systems choose or are able to participate, and the study may exclude contracts with certain characteristics.

The 2020–2024 SOA/LIMRA results show substantial submitted industry volume and more than 80% of sales coverage. That coverage percentage is an institutional statement about sales, not a guarantee of proportional sampling across attained ages, historic vintages, benefit amounts, or contract-years. A comparison to 2014–2019 therefore requires common rules, age distributions, treatment of contract exits, contributor cohorts and comparable financial weights. A simple difference between 143,190 and 236,000 deaths contains no such standardization.

Missingness and detection errors

Data can be missing because a death was never linked, a secondary life was not tracked, a person changed insurer, a contributor omitted an old contract class, or administrative systems changed. Missing completely at random, missing at random conditional on observed covariates, and missing not at random entail different inferential consequences. Rubin (1976) formalized essential distinctions for inference with missing data. The classification of a real administrative mechanism requires evidence, not merely a preferred statistical label.

Death ascertainment has two separate misclassification probabilities: sensitivity to genuine deaths and false-positive matching among survivors. Treating everyone without a positive match as alive implicitly imposes perfect ascertainment. The GAO's Death Master File testimony shows why such perfection is not an institutional fact. More observations cannot average away systematic underascertainment.

For a synthetic population of N equally exposed one-year records with D true deaths, define sensitivity s=P(recorded positive | truly dead) and false-positive probability f=P(recorded positive | alive). The expected observed number of positives is D*=sD+f(N−D). Solving gives D=(D*−fN)/(s−f), provided s>f. Set N=10,000, s=0.90, f=0.0005 and suppose 95 observed positive death matches. Then D≈(95−5)/(0.8995)=100.0556. The unadjusted observed rate is 0.95%; the adjusted estimate is approximately 1.0006%. The numbers are an original hypothetical sensitivity exercise, not an observation from SOA, NDI or CMI. A genuine application must account for uncertainty in s and f, unequal person-time and varying match quality across groups.

Inverse-probability weighting or multiple imputation can address specific missingness mechanisms only with justified conditional assumptions, available covariates and adequate support. The discussion of informative observation and selection in Hernán and Robins (2020) is relevant to formulating sensitivity analyses; it does not identify missing deaths in any particular annuity dataset. If a missingness mechanism cannot be identified, results should report the plausible sensitivity range.

U.S. insurer and U.K. pension populations

The U.S. individual payout study and British CMI analyses do not have interchangeable observation frames. British CMI data submission materials document separate collection and coding procedures. Its SAPS investigation uses data for pensioners and dependants in participating occupational schemes. The insurer annuity work is separately indexed in the CMI annuities register.

The published CMI Working Paper 209 description concerns 2017–2024 pensioner experience, with subgroup analysis. The CMI Working Paper 194 description concerns the earlier multidimensional investigation. A description or publicly accessible abstract cannot substantiate numerical findings contained solely in restricted full reports. The CMI news archive also documents an October 2025 correction to a 2025 annuitant paper. That episode demonstrates the need to identify the version of a published statistical table before relying on its values.

International comparison must map the observed population, including the start and end of eligible exposure; original and survivor lives; pensions already in payment; voluntary commutations; transfers; annuity purchase selection; benefit amount distribution; age and birth-cohort mix; and death matching. Once these variables are harmonized, researchers can examine whether within-cell mortality differs. Harmonization itself does not remove unrecorded health or socioeconomic selection.

Example of composition confounding

Suppose two synthetic attained-age categories have fixed mortality rates 1% and 4%. Portfolio A has 80% of exposure in the younger category and 20% in the older; its crude central rate is 0.8×0.01+0.2×0.04=0.016. Portfolio B has 50% in each; its crude rate is 0.5×0.01+0.5×0.04=0.025. The apparent difference of 0.9 percentage points arises entirely from different age composition in the example, with identical age-specific mortality. Standardizing both to equal 50/50 weights yields 0.025 for each. The arithmetic illustrates the need for common weights and consistent individual age measurement before interpreting period comparisons.

Legal access and preservation

Death and pension records are subject to legal restrictions, access permissions and confidentiality protections that differ across jurisdictions. The European Union's General Data Protection Regulation establishes rules concerning purpose limitation, accuracy, minimization and rights in processing personal data within its scope. The regulation's treatment of deceased persons and possible member-state provisions must be assessed specifically. CDC's NDI rules impose their own research-purpose restrictions. An anonymized intercompany experience extract should not be presumed immune to reidentification risk solely because direct names are omitted.

The appropriate provenance system keeps necessary identifying data inside authorized secure environments while preserving stable pseudonyms, authorized matching decisions, restricted source hashes, derived aggregates and a precise description of legal and ethical access arrangements. Reproducible public methods and confidential microdata are compatible when custody and audit permissions are explicit.

Evidence Validation and Model Sensitivity

From source extract to published mortality cell

The evidence pipeline begins with a dated inventory of contracts, pension arrangements and administrative systems. It records source custodians, covered periods and extraction versions, resolves which benefit types and lives are eligible, assigns stable pseudonymous keys, adjudicates deaths, determines entry and exit intervals, derives attained age and calendar year, computes lives-based and amount-based exposure, applies the chosen mortality basis, and aggregates results. Every correction is retained with its cause and approval. The PROV-O data lineage vocabulary can represent the traceability relations in this chain.

A reproducible mortality cell should have a manifest identifying its contributing source population, eligibility and exit rules, inclusion/exclusion counts, death ascertainment source, unmatched candidates, exposure convention, economic weight, table version, analyst code revision and output hash. The source extract needs an independently preserved hash and the history of derived transformations. Without the source snapshot, an auditor may be able to repeat the formula but not recover the same population.

Validation layers

At the record level, test impossible date order, duplicate life segments, unsupported sex or age codes, invalid termination modes, a death preceding reliable source observation, ambiguous match keys, postdeath risk intervals, and survivor transitions with no evidence of the initiating event. At the dataset level, reconcile the number and monetary value of in-force contracts to source administration totals, compare year-to-year age distributions, and validate that death counts and person-time recompute from the same versioned rows. At the publication level, compare the printed cell totals with the code output and check source-to-claim hyperlinks.

Section 2.5 of ISAP 1: General Actuarial Practice makes explicit the professional need to consider data sufficiency, accuracy, completeness and consistency, and to deal with deficiencies and disclose relevant modifications. Its applicability depends on adoption by the relevant professional organization or governing framework. The corresponding responsibilities are substantively relevant to experience-data research regardless of whether a particular territory has formally adopted its text.

Two broad governance standards provide context rather than specific annuity algorithms: ISO 8000-150:2022 concerns roles and responsibilities in data quality management, while ISO 8000-51:2023 addresses machine-processable data-governance policy statements. The publicly accessible official descriptions establish their subjects; the present study does not claim to have independently audited implementation of either standard at any contributor organization.

Mortality denominator and baseline testing

An actual-to-expected ratio often takes A/E=Σ_j D_j/Σ_j E_j^0, where the expected deaths E_j^0 depend on eligible exposure and a selected mortality model. A different base table, projection scale, weighting convention, survivor-state rule or observation start date changes expected deaths without changing the observed death certificates. Reporting a ratio of 110% therefore means the observed count or weighted value is 10% above its specific expected benchmark; it does not mean a ten-percentage-point increase in an individual death probability.

Sensitivity analysis should vary the following documented choices separately: use of death occurrence rather than notification date; strict versus broad match adjudication; corrections to commencement and exit boundaries; inclusion of secondary lives before and after the first death; handling of unknown and transferred records; mortality basis and improvement period; and standardization of age, calendar time and benefit composition. Results should be presented with their specific denominator and original source version.

The corrected CMI news notices emphasize version control. A revised table may change historical published estimates without any change to the true historical death events. A source-linked research record therefore needs both the observation-period date and the publication/revision date.

Limits of public evidence

The primary SOA/LIMRA study statement supplies public high-level counts and institutional participation; the full 2026 analytical package is commercial or subscriber-restricted. British CMI papers may similarly have restricted access. The present study examined accessible source specifications and institutionally published descriptions and did not audit confidential insurer microdata, proprietary match software, unpublished contract files or case-level death certificates. No conclusion here attributes a measured rate of linkage error, missingness or censoring to a participating insurer.

The four numeric cases use fully disclosed hypothetical inputs. Their calculations illustrate necessary distinctions in mortality estimation; they do not estimate the level of error in any real study. The work provides original analytical synthesis using established statistical methods and primary records. It asserts no new scientific theory and claims no external peer review.

Integrated Research Findings and Open Scientific Questions

Established findings

First, mortality is an event attached to an identified life while a contract is a bundle of contingent rights and payments. Person-time, contract-time and income-weighted time are separate statistical measures. All reported ratios require a declaration of which measure supplies the denominator and which lives are eligible.

Second, death ascertainment is a data transformation that requires documentary source identification, matching decisions, processing dates, and explicit treatment of conflicts. Official death registers improve the evidentiary basis of research but retain access limits, correction possibilities and matching uncertainty. A death record's scientific authority depends on both its issuing institution and the demonstrated connection to the particular measuring life.

Third, temporal eligibility is constitutive of mortality experience. The person-time case demonstrates that using beginning-of-year headcount in place of documented exposure can materially lower a calculated central rate even when every death is correctly identified. A reproducible result requires event-time risk intervals rather than undifferentiated flags.

Fourth, missingness interacts with selection. The secondary-life reporting exception in the SOA/LIMRA specification, transfer or withdrawal records, and variation among contributing institutions show how an experience dataset can select survivors even when each retained row is internally valid. The probabilistic matching example illustrates the dependence of mortality estimates on independently justified error parameters.

Fifth, the published U.S. and British studies can inform one another, but they measure partly different institutions, cohorts and benefit states. Their differences can be assessed through matched definitions, common age/benefit weights and documented reporting periods. The aggregate public counts alone do not identify cross-national mortality differences or the effect of a change in life expectancy.

Implications for research and actuarial governance

A mortality data-quality assessment should precede and accompany any projection of longevity, pricing of a life annuity or measurement of reserve sensitivity. A downward ascertainment bias can make a measured portfolio appear to have lower death rates and consequently imply greater longevity for valuation; an upward bias can have the reverse implication. The resulting liability effect depends upon benefit state, attained age, payment schedule, discounting and risk margin rules. It cannot be estimated from the observed error rate alone.

Experience-study publishers can strengthen comparability through versioned data dictionaries, transparent inclusion rules, common denominators, correction histories and source-linked methodological summaries. Institutional contributors can preserve confidential microdata while supporting an independently auditable chain from their original records to aggregated submissions. Researchers examining periods affected by pandemic mortality or administrative disruption should distinguish year of death, year of discovery and year of publication, especially when revised source files become available.

Research questions requiring access to original observations

Future empirical work can estimate match sensitivity and specificity using adjudicated death certificates stratified by age, insurer, calendar year, product and length of administrative follow-up. Another investigation can measure reporting lag distributions and the revision of historical death counts after delayed registrations. Joint-life research can quantify the number and characteristics of secondary lives invisible before the primary's death and evaluate the resulting survivor selection. Cross-country investigations can map insurer payout annuitants to occupational pensioners and general population references while controlling for health and economic heterogeneity.

Such inquiries require legally authorized person-level source material, documented measurement assumptions, independently reproducible statistical procedures and, where applicable, research ethics oversight. This dossier establishes the evidence definitions and analytic questions necessary for those studies without claiming that restricted observations have been obtained.

Conclusion

Annuitant mortality experience is the measured intersection of life survival, contractual eligibility, observed exposure and validated death events. The recorded deaths and years at risk arise from sources with dates, identifiers, custody histories, coding rules, jurisdictional restrictions and opportunities for correction. Their provenance determines what a mortality estimate means.

Contemporary primary evidence establishes that industry investigators distinguish contract identifiers, measuring-life segments, benefit commencement, current in-force status, termination modes and financial weights. Statistical analysis shows that correct death counts can still yield inaccurate rates when exposure time is misstated, that imperfect identification of death can change inferred probabilities, and that variation in population composition can produce apparent mortality differences without variation in age-specific risk.

The integrated scientific conclusion is that a credible observed-to-expected mortality measure requires reproducible continuity from original entitlement and death records to the eligible life-periods, reference mortality basis and released aggregate. This traceability makes international annuitant and pension mortality comparisons interpretable and supplies a sound empirical foundation for subsequent longevity valuation. The conclusion follows from primary-source examination and established scientific methods, without asserting a new theory or undisclosed empirical research.

References