Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Mortality Data Provenance and Coding Quality in Life Insurance

Abstract. Life insurance mortality analysis depends upon a sequence of observations and transformations that begins with the establishment of insured exposure, passes through death occurrence and notification, and may continue through civil registration, medical certification, cause-of-death coding, claims adjudication, and construction of actuarial experience. Each operation has its own population, legal purpose, dates, definitions, selection process, uncertainty, and revision history. This monograph evaluates how the provenance of those observations and the quality of diagnostic and administrative coding affect the measurement of insured mortality. It integrates civil-registration standards, international disease classification, demographic data production, probabilistic record linkage, incomplete-data statistics, and documented insurer experience studies. Formal estimands distinguish population mortality from insured exposure rates, all-cause death from cause-specific incidence, and observed events from reported or matched records. Analytical examples establish the direction and magnitude of selected ascertainment and coding biases, demonstrate that highly accurate aggregate coding can coexist with large cause-specific distortions, and separate statistical comparability from contractual claim eligibility. The research establishes practical audit criteria and sensitivity-analysis procedures, with explicitly stated identification limits. Its contribution is an original comparative synthesis and reproducible analytical exposition; it proposes no new proprietary theory or classification system. The work is based on published primary and methodological documents, rather than individual-level insurance microdata.

Keywords: Mortality data provenance; life insurance; death registration; cause-of-death certification; ICD-10; ICD-11; probabilistic record linkage; missing data; insured exposure; experience studies; misclassification; ascertainment; actuarial data governance; measurement error; auditability.

Introduction

An insurer's measured mortality rate is a statement about a specified set of insured lives, a specified interval of risk exposure, a specified definition of death, and a specified process for discovering and recording events. Every component matters. A death may occur during valid coverage yet enter the claim system much later; a claimant may report an event that is outside the covered interval; an official death record may be present but fail a person-level match; and a correctly identified death may have an uncertain underlying medical cause. A single percentage called “data completeness” cannot distinguish these problems. The scientific object is the chain of custody and transformation that connects the person at risk to the actuarial observation and, where required, to the insurer's contractual obligation.

The international context makes the problem especially consequential. The World Health Organization's civil-registration assessment estimates that almost 40% of deaths globally are unregistered, with major regional variation (WHO, 2024). That global figure describes public vital registration, not an estimate of missing death claims in any particular insurer. In a mature market, a well-maintained policy administration system may establish policy status accurately while still failing to connect certain external deaths. Elsewhere, claims systems may establish contract-level payment events even when the surrounding population registration system has limited completeness. The relationship among these sources must be investigated, rather than inferred from national averages.

Cause information requires a separate evidentiary chain. WHO distinguishes the medical certification of the sequence leading to death from the statistical selection of an underlying cause under classification rules; the WHO mortality certification guidance makes both operations explicit. The US National Center for Health Statistics likewise distinguishes a single selected underlying cause from the additional conditions recorded as multiple causes in its cause-of-death definitions (NCHS, 2026). Insurance cause-of-death exclusions, accidental-death riders, occupational classifications, underwriting records, and claims manuals can employ different contractual definitions. Epidemiologic code and policy entitlement therefore require a documented mapping, not an assumed equivalence.

The investigation uses documentary comparison, mathematical derivation, bias analysis, and illustrative synthetic data. It reads technical material from the United Nations, WHO, ONS, Eurostat, NCHS, Human Mortality Database (HMD), Society of Actuaries (SOA), and statistical literature. These sources are evaluated for what each actually measures. The examples are internally calculated demonstrations of possible error, not newly observed empirical findings about any insurer or country. The temporal reference point for source assessment is November 14, 2026; a “current” coding standard is distinguished from the version actually implemented by a source producer at a specified date.

Definitions, Historical and Documentary Foundation

The event, the policy, and the statistical observation

Let an eligible insured life be identified by index i. Let coverage begin at entry time aᵢ and end at the earliest of termination, exit from observation, and the study cutoff bᵢ. Let Tᵢ be the true date of death where applicable. The all-cause mortality event for the experience analysis is Dᵢ = 1 when death occurs during the relevant exposure interval and the study's eligibility conditions are met. An administrative event Rᵢ = 1 records that a death has reached the specified dataset by extraction date; a linkage event Lᵢ = 1 states that it has been correctly associated with the eligible person. A cause assignment Cᵢ has a code system, revision, selected cause, certifier, and processing date. These variables answer different questions.

An insurer can measure mortality per life, per policy, per sum insured, or per benefit unit. A joint-life contract, multiple policies on one person, and a whole-life policy with riders can yield more than one financial record for a single death. The denominator must specify the unit. Person-level death incidence normally requires deduplicated lives; claim-frequency and financial-loss analyses may count separate contract entitlements while preserving their person link. The distinction guards against accidentally interpreting two death claims as two deaths.

The common central death rate over a time cell is

m^x,t=Dx,tEx,t,Ex,t=∑i∫t0t1Yi(s) ds,\widehat m_{x,t}=\frac{D_{x,t}}{E_{x,t}},\qquad E_{x,t}=\sum_i\int_{t_0}^{t_1}Y_i(s)\,ds,

where D counts correctly defined deaths and E is life-years exposed, with the at-risk indicator Yᵢ(s) accounting for entry, cessation, and death. The resulting central rate is a hazard approximation; a one-year probability of death qₓ requires a life-table convention or within-year hazard assumption. With constant force of mortality over the year, qₓ = 1 − exp(−mₓ). The HMD Methods Protocol, Version 6 describes population-based death-rate and exposure conventions, including adjustment of dates and age structures (Wilmoth et al., 2025). It provides a methodological comparator, not a substitute for insurer entry and termination histories.

A provenance chain as a documentary requirement

The United Nations vital-statistics recommendations distinguish the institutional production of a civil registration record from the statistical system that derives aggregated vital statistics (United Nations, 2014). A comparable distinction exists in life insurance among policy administration, beneficiary notice, claims adjudication, mortality experience compilation, and reserving systems. A provenance description must therefore identify both the original record and every transformation that can alter its meaning.

A reproducible case-level lineage typically contains a source-system identifier; an original event identifier; an immutable or versioned copy of the input; applicable policy or person identifier; dates of occurrence, notification, registration, processing, and extraction; the jurisdiction and definition of the event; coder and classification revision where cause is used; linkage method and decision; edits and overrides; and the version and approval of the final experience extract. For an individual who moves between insurers or administrative platforms, entity resolution, product history, and merged identity keys become especially important. Preserving only a final mortality table is insufficient to reconstruct which deaths were added, excluded, recoded, or duplicated.

The Human Mortality Database input notes give a useful public example of provenance-aware statistical design: input files are accompanied by background documentation, data-source and reference files, and methodology for converting raw material into complete series (HMD, n.d.). HMD populations are geographically defined, whereas insured populations are selected through purchase, underwriting, distribution, affordability, and persistency. Comparability must be explicitly evaluated by age, sex, calendar period, residence, selection duration, and other observed risk characteristics. A large public database is not automatically the best denominator for a selected contract portfolio.

Historical changes in the evidence infrastructure

Civil registration has increasingly been digitized, and mortality classification now depends on medical records, training, electronic certification, specialist coding software, and statistical dissemination systems. This produces a reproducibility challenge: a dataset's numerical contents may change when a certification error is corrected, a delayed record arrives, a coding rule changes, a dataset is remapped, or the definition of a reportable population changes. Each is a distinct revision mechanism and should be captured in a revision log.

Insurance experience research supplies an instructive institutional example. The SOA's 2009–2013 individual life experience report describes the underlying data collection, cleansing, field definitions, and associated empirical output (SOA Individual Life Experience Committee, 2017). Later the SOA's 2019 individual life mortality report documents that the data-supply role moved from MIB to the National Association of Insurance Commissioners for 2018 onward within the 2012–2019 reporting window (SOA, 2024). A discontinuity in reporting source or data custody demands a field-by-field continuity assessment even where the resulting time series retains its familiar title.

Documentary scope and legal boundaries

A civil death certificate, a claims file, a physician's note, a reinsurance bordereau, and an actuarial experience extract have different formal authority. A regulatory requirement to submit mortality information must be interpreted in the applicable country and reporting period. A claim adjudication may need a prescribed standard of proof and contractual definitions which differ from the statistical definition selected for cause-of-death analysis. Personally identifiable health and mortality data may be restricted by purpose, secrecy, consent or other legal basis, access controls, retention periods, and cross-border transfer rules. The European Union's GDPR, where applicable, supplies legal requirements concerning accuracy, purpose limitation, data minimization, special-category data, and security (European Parliament and Council, 2016). These requirements can constrain linkage datasets and release of analysis files even where an actuarial analysis would benefit from more granular information.

The practical scientific implication is that the data dictionary, contractual interpretation, privacy authority, and analytical method belong in the same case-specific documentation, while retaining their distinct roles. National classification rules should be cited by revision and implementation date. An assurance based only on a contemporary website that describes the latest codebook may misstate what a prior-year institution actually applied.

Record Linkage and Missingness

What must be linked

The simplest life insurance experience study begins with an authoritative roster of covered persons and their dated eligibility windows. It then seeks deaths from claims files, external vital registers, insurer inquiries, reinsurer reports, or administrative feeds. Each source has its own selection mechanism. Claim notification depends upon the awareness and behavior of beneficiaries, contract terms and notification processes. Civil registration depends upon the jurisdiction and its population coverage. Reinsurance reports may identify particular ceded risks but omit business outside the treaty. A successful person-level match is evidence that records describe the same subject; it is not, by itself, evidence that the death occurred within an insured coverage period.

Linkage is affected by name changes, transliteration, missing date of birth, duplicate names, typographic errors, migrant status, nonstandard identifiers, and deliberate anonymization. An exact match on a purported unique identifier performs well only when identifier issuance, uniqueness, and capture quality are established. Otherwise, deterministic matching can silently discard valid deaths or attach events to the wrong insured life. A particularly serious failure arises when several policies share one person and a merge process mistakes policy identifiers for unique persons.

Probabilistic linkage and decision uncertainty

The classical Fellegi–Sunter record-linkage theory treats pairs of records as matching, nonmatching, or indeterminate using comparison patterns and controlled decision errors (Fellegi and Sunter, 1969). For a comparison vector γ, the likelihood ratio is

Λ(γ)=P(γ∣M)P(γ∣U),\Lambda(\gamma)=\frac{P(\gamma\mid M)}{P(\gamma\mid U)},

where M denotes a true match and U an unrelated pair. Under an expressly specified comparison model, log ratios can be formed from field-specific agreements and disagreements. High scores imply evidence favoring a match, but their calibration depends upon the actual distributions of names, dates, and errors in the portfolio. The decision process requires an estimate of false-match frequency, missed matches, and uncertain matches, ideally assessed against clerically reviewed or independently verified pairs. Jaro's operational census-matching study illustrates the practical problems of linking partially independent enumerations (Jaro, 1989).

Let true in-scope deaths be D, with linkage sensitivity s = P(L=1\mid D=1), and let false-positive matches be F. If correctly specified coverage exposure E is known, the observed death count satisfies approximately

Dobs=sD+F,m^obs=sD+FE.D_{obs}=sD+F,\qquad \widehat m_{obs}=\frac{sD+F}{E}.

This is an expected-count relationship, not a deterministic law for a realized sample. Correcting it by D̂ = (Dobs − F̂)/ŝ requires independently supported estimates of sensitivity and false positives. If s varies with age, residence, cause, policy duration, or insured amount, a single aggregate correction can distort subgroup trends. Weighting unmatched or late-linked deaths using estimated group-specific capture probabilities is defensible only where the probability model and record-recovery mechanism are adequately supported.

Registration, notification, and extraction missingness

At least four distinct sources of under-observation matter: an actual death is unregistered in the civil system; a registered death is absent from the insurer's accessible external feed; an available death record fails to match the insured person; or a properly linked record is excluded by the experience extraction rule. A fifth problem is incomplete or inaccurate exposure, which may either increase or decrease estimated mortality. Records with a missing cause can still be complete for all-cause death incidence. Conversely, a detailed medical cause coded on a record provides no assurance that all unreported deaths have been captured.

Temporal missingness requires a data-vintage perspective. The NCHS provisional mortality documentation explains that recent death counts are revised as certificates are completed, transmitted, and processed (NCHS, n.d.). An insurer may similarly receive delayed claims or amendments. The correct analysis states both date of occurrence and data extraction date. A fixed delay cutoff can create an artificial improvement in mortality in the most recent months if late-reported deaths are disproportionately missing. The direction changes when duplicate early claims, late reversals, or erroneous records are present.

Define reporting delay W = notification date − occurrence date. Where observed data and censoring information support it, estimate a development factor by occurrence period and subgroup; retain the uncertainty in the distribution of W. Development factors estimated on stable pre-disruption periods may fail when a pandemic, IT migration, administrative reform, or claims-handling change alters the reporting process. A fully matured cohort is a better calibration target than merely waiting a fixed number of calendar days when development remains uncertain.

Missing-data mechanisms and identification

Little and Rubin's analysis of missing data distinguishes missing completely at random, missing at random conditional on observed variables, and missing not at random under the specified data model (Little and Rubin, 2019). In insured mortality these distinctions are substantive. Missing cause may depend on region and death place, which are observed; it may also depend on the very cause that was not certified, which is unobserved. Likewise, absence of a matched death may depend on migratory status or health characteristics that are incompletely measured in the insurer's roster. An administrative extract rarely identifies its own selection mechanism.

Maximum-likelihood estimation with latent records can draw on the EM algorithm developed by Dempster, Laird, and Rubin (1977). Multiple imputation, inverse-probability weighting, and selection models have their uses, but all require explicit assumptions about missingness, linkage and model specification. Imputation must not be presented as independent evidence of the deaths it fills in. Where missingness plausibly depends on unobserved information, sensitivity analysis over substantive mechanisms is essential, including bounds that remain valid without choosing a single missing-at-random model.

Dependency of multiple ascertainment sources

Two-source capture–recapture can be a diagnostic aid when source overlap is observable. With n₁ cases on the first list, n₂ on the second, and m shared cases, a familiar independence-based estimate of total events is N̂ ≈ n₁n₂/m. It is vulnerable to dependent capture, heterogeneous inclusion probabilities, linkage errors and a moving population. Beneficiary notifications and insurer claims may be strongly dependent because the claim itself initiates the registry inquiry. Capture–recapture should therefore be labeled as a modeled estimate, accompanied by coverage definitions and at least one credible sensitivity scenario. Reliable direct reconciliation against a verified cohort may be stronger evidence than a complex correction fitted to two dependent administrative lists.

A worked all-cause ascertainment example

Consider a purely hypothetical portfolio with E = 100,000 life-years of exposure and D = 500 true covered deaths. If the data pipeline captures 90% of true deaths and creates five false-positive person matches, the expected observed count is 0.9 × 500 + 5 = 455. The observed central mortality rate is 4.55 per 1,000 life-years, versus the underlying 5.00 per 1,000. Under these assumptions, the absolute bias is −0.45 per 1,000, or −9% relative. If a separate audit estimates capture sensitivity at 0.90 and five false positives, the algebraic adjustment yields (455 − 5)/0.90 = 500. This agreement is built into the example and demonstrates a mechanism, not the accuracy of an estimator when sensitivity or error counts must themselves be inferred.

If the denominator is also wrong—say exposure is overstated as 103,000 life-years—the observed rate becomes approximately 4.4175 per 1,000, which combines numerator under-ascertainment with denominator inflation. Data quality evaluation must examine numerator and denominator jointly; correcting only the mortality count leaves the estimated risk distorted. Cohort-specific linkage audits should report uncertainty in both quantities and avoid treating error estimates as known constants.

Coding Standards and Revisions

Certification precedes classification

The immediate condition recorded by a physician often describes the final clinical mechanism, while the underlying cause seeks the disease or event that initiated the causal sequence. A death certificate may state several successive conditions, with contributing factors recorded separately. Statistical systems then apply internationally specified selection and modification rules to derive a single underlying cause. WHO's medical certification and coding resources describe this division of work. It follows that a coding audit must start by testing the clinical statement and its sequence, then assess the coder's application of rules, rather than validating the final code against itself.

Certification error and classification error have different remedies. If the physician records an incomplete causal sequence, a coding specialist cannot reliably reconstruct the original clinical narrative from the final ICD code. If the certificate is adequate but the wrong version of a coding rule is applied, a properly archived certificate can be reprocessed and compared under the correct revision. If person identity has been linked incorrectly, an otherwise accurate cause code may be attached to the wrong policyholder. Quality controls must therefore separate source capture, medical certification, person matching, and rule-based coding.

Alipour and Payandeh’s systematic review of death-certificate errors reported substantial frequencies of missing time intervals, incorrect causal sequences, and errors in underlying-cause statements across its included investigations (Alipour and Payandeh, 2021). Those pooled or descriptive error rates measure the reviewed certification settings, not a life insurer's current error prevalence. They demonstrate the importance of validation against the source certificate and the local certifier process. They also show why the absence of an invalid-code flag is not evidence that the certified medical cause is true.

ICD revision and implementation are time-indexed variables

WHO states that ICD-11 came into effect internationally on January 1, 2022, while countries may continue using ICD-10 during transition. The WHO ICD-11 implementation guidance was updated in December 2025 and explicitly recognizes variable national transition periods. Therefore, a data extract dated 2026 cannot be labeled “ICD-11” solely because ICD-11 is internationally effective. It may contain ICD-10 records, ICD-11 records, local derivative categories, or retrospective mappings. The reporting producer's implementation notice and code-level metadata decide the matter.

For every cause-specific analysis, retain at least: ICD family, revision and edition; applicable national modification; classification year; software engine and version; adopted selection and modification rules; version of any cause-group shortlist; manual coding proportion; date of reprocessing; and provenance of retrospective mapping. The WHO ICD-11 overview describes the digitally oriented international standard, but the statistical break created by implementation must be evaluated at the level of a specific dataset rather than presumed to disappear through digital interoperability.

Calendar-year labels can conceal two different dates: the date the death occurred and the date the certificate was registered and coded. If revised software applies only to records processed after a changeover, a statistical series grouped by registration year may show a break at a different point from a series grouped by occurrence year. Historical back-coding can improve comparability when original source detail is available. Crosswalks that map only an aggregated ICD chapter often cannot resolve distinctions introduced or removed at lower code levels, and may produce different effects across countries with different certification practices.

Empirical evidence from dual coding in England and Wales

The Office for National Statistics provides directly relevant original evidence. In its January 2020 software comparison, ONS dual-coded 42,413 historical death records using the earlier IRIS 4.2.3 and newer MUSE 5.5 methods (ONS, 2019). More than 98% of the cases remained in the same ICD-10 chapter, yet the number assigned to the “certain infectious and parasitic diseases” chapter fell 19.8%. Reassignment of certificates listing sepsis was a major contributor. This is a decisive methodological observation: a high percentage of apparently concordant records at a broad level can coexist with a material change in a smaller category of interest.

The subsequent January 2022 coding comparison evaluated the move from MUSE 5.5 to MUSE 5.8. ONS reported that 99.2% of dual-coded deaths retained their ICD-10 chapter and 98.6% their leading-cause designation (ONS, 2022). The first figure is not a certificate-validity estimate; it is agreement between two coding implementations at a specified aggregation level. The second is not a universal guarantee that rare causes, exclusions, or rider-trigger conditions were stable. It is a direct comparison of the classification results under two software versions on a sample.

ONS's mortality-statistics methodology maintains an explicit history of these rule and software changes. Comparability must accordingly be judged by affected cause and time span, with dual-coded bridges where available. A single adjustment factor applied to every cause would overlook heterogeneous switching patterns. A code family whose definition remained stable may need only ordinary sampling uncertainty; one involving altered causal selection may need reprocessing of historical certificates or a modeled comparability ratio.

Coding as a measurement problem

Let dₖ be the true count in underlying-cause category k under a chosen clinical reference definition and yⱼ the count assigned to reported category j. With a classification matrix M defined by Mⱼₖ = P(reported = j | reference = k), expected observed counts are

E⁡[yj]=∑k=1KMjkdk,\operatorname{E}[y_j]=\sum_{k=1}^{K}M_{jk}d_k,

where the column probabilities sum to one if all events are detected and assigned to one of the enumerated categories. This column-sum assumption is often violated when an “unknown,” “pending,” or unreported class has been removed from the extract. For such data the model needs an additional unobserved outcome or a separate ascertainment process. Coding misclassification and incomplete death registration should never be conflated in the same unqualified factor.

For a two-cause illustration, suppose 100 deaths have reference cause A and 200 reference cause B. An algorithm assigns A correctly in 80% of true-A cases and assigns B correctly in 95% of true-B cases. The expected reported A count is 0.80 × 100 + 0.05 × 200 = 90; reported B is 0.20 × 100 + 0.95 × 200 = 210. Overall assignment accuracy is (80 + 190)/300 = 90%. Yet cause A is underestimated by 10%, while B is overstated by 5%. If cause A corresponds to a narrowly defined accidental-death rider, the 90% global coding agreement is an inadequate assurance of the rider's risk estimate. A further adjudication step is necessary because the contract may not define “accidental” in the same way as a cause-code category.

A classification matrix can in principle be estimated on a clinically reviewed dual-coded sample and inverted where it is identifiable and numerically stable. The inversion can greatly amplify sampling error if columns are nearly indistinguishable. For sparse causes, regularized estimation, transparent aggregation, and uncertainty intervals may be more defensible than unstable numerical correction. Validation must document the reference definition, sample selection, reviewer independence, and disagreement adjudication. The exercise remains a sensitivity analysis when the reference clinical classification itself has uncertain accuracy.

Cause-group quality, ill-defined deaths, and plausibility

Public health systems evaluate data quality through distributions of ill-defined causes, biologically implausible age or sex patterns, and inconsistencies across places and periods. The WHO ANACoD3 analytical tool highlights errors, estimates reporting completeness, and reports cause-specific distributions (WHO, n.d.). Such diagnostics are valuable for an insurer's external benchmark, but they do not independently certify that the underlying records match the insurer's policies, and population-level “garbage code” measures have limited relevance to an all-cause death incidence series that does not use medical causes.

The Eurostat cause-of-death metadata show that international statistical exchange depends on distinctions between residence and place of occurrence, reported age and sex, code shortlists, and country-specific production details (Eurostat, n.d.). An insurer issuing cross-border coverage must also distinguish country of underwriting, country of policy administration, country of residence, and country of death. These dimensions can differ for the same insured life and should not be collapsed to a single country field without a documented analytical purpose.

Classification quality and insurance decisions

All-cause term-life mortality can often be estimated without any ICD cause, provided death occurrence, date, insured eligibility and exposure are established. Cause-of-death coding becomes central for cause-specific research, accident and disease riders, exclusions and catastrophe attribution, but an insurer's claims decision may additionally require medical evidence, legally prescribed tests, policy interpretation, and claims investigation. Coding can improve the organization of evidence and assist audit selection; it does not automatically determine coverage. This separation protects both actuarial inference and fair treatment of a claimant.

There is also a risk of circular validation when insurance claim adjudicators enter a “cause” based on the policy section under which benefits were paid and analysts later treat it as an independent medical cause. A field called “claim cause” must be defined as either a clinical event, a coded official cause, an operational loss cause, or a coverage category. Its authority and version must be documented. For multi-cause analysis, the original multiple-cause fields should be retained where legally and technically permitted, rather than reducing all evidence to the selected underlying cause prematurely.

Source Uncertainty

A structured account of uncertainty

Source uncertainty encompasses error in the coverage roster, time at risk, death identification, link assignment, demographic attributes, medical certification, classification, delay estimation, and statistical model. These components are related: a missing date of birth changes age grouping; an incorrect closure date changes exposure; an invalid person match can import both a death date and an apparently precise ICD code; a software transition can make a cause trend appear to shift. An uncertainty statement should specify which components are measured independently, which are modeled, and which remain unidentifiable from the available records.

Random sampling uncertainty is only one part of the problem. For a count D drawn under a homogeneous Poisson approximation with known exposure, the approximate standard error of the central rate D/E is √D/E. This approximation is useful for orientation but does not measure bias from omitted deaths or excess exposure. It becomes less reliable under clustered risks, heterogeneous hazards, multiple events per person record, or substantial parameter uncertainty. A narrow confidence interval around a biased estimator is a precise measurement of the wrong target.

Consider E = 100,000 life-years and Dobs = 455. The Poisson-based standard error is about √455/100,000 = 0.0002133 deaths per life-year, or 0.2133 per 1,000. This standard error is considerably smaller than the 0.45 per 1,000 ascertainment bias in the earlier synthetic example. A formal uncertainty interval for corrected mortality must include uncertainty in reporting sensitivity, false positives, denominator error, and, where applicable, age or policy status.

Partial identification when a capture rate is unavailable

Suppose false positives can reasonably be bounded by F ∈ [0, 10] and audit evidence supports sensitivity s ∈ [0.85, 0.95]. With 455 observed deaths and fixed exposure of 100,000 life-years, possible true counts under the earlier simple measurement model lie between (455 − 10)/0.95 ≈ 468.42 and 455/0.85 ≈ 535.29. Corresponding central rates are about 4.684–5.353 per 1,000. This interval is conditional on the chosen bounds, does not reflect binomial or count sampling fluctuations, and is not a conventional statistical confidence interval. It communicates the dependence of inference on empirical bounds more honestly than applying an arbitrary single “data quality adjustment.”

For heterogeneous cohorts, the same calculation should be performed within coherent strata, or a hierarchical model should retain differing sensitivity and capture uncertainty. A single global envelope is informative only if its assumptions really cover every subgroup; otherwise it may omit worst-case outcomes. Severe missingness in a small, high-sum-insured class can be financially material even while overall missingness is low.

Attribution of risk and benefit amounts

The actuarial meaning of coding error depends upon the benefit design. In a uniform-benefit synthetic example with 100,000 in-force lives, one-year death probability q = 0.005 and a $100,000 death benefit, the expected undiscounted death-benefit amount is 100,000 × 0.005 × $100,000 = $50 million. If a flawed observed death rate were 9% below the true risk and then used without adjustment as the pricing or reserving assumption, the corresponding expected undiscounted amount would be $45.5 million, a $4.5 million shortfall in this simplified comparison. The assumption that all coverages, dates, benefits, and deaths are identical is deliberately artificial; practical valuation incorporates product rules, benefit timing, discounting, expenses, lapses, margins, risk adjustment, and other information. The example isolates measurement sensitivity rather than computing any insurer's actual reserve deficiency.

The distinction also matters for expected claims cost by amount rather than lives. Two groups may have the same mortality rate by lives but different concentrations of insured sums, causing financially weighted mortality experience to differ. A death record must be reconciled both to the correct person and to the correct coverage amount at the insured event date. Out-of-date face amounts, missing riders, post-claim reversals, and partial claims settlements are distinct financial data-quality risks. The SOA valuation basic tables illustrate why precise risk-class definitions and age conventions must accompany mortality rates (SOA, 2015–2018).

Materiality differs by use case

For an all-cause experience analysis, missing deaths and incorrect exposure are usually central. For a cause-specific underwriting or epidemiologic study, cause coding becomes material in addition to event ascertainment. For an accidental-death rider, policy language and individual claim proof may dominate category-level ICD agreement. For an IFRS financial measurement, the entire future cash-flow assumption must be assessed in light of current relevant information; the IFRS 17 standard overview describes the principles of risk-adjusted fulfilment cash flows and current information (IASB/IFRS Foundation, 2017/2023). This does not mean one observational correction mechanically produces the same amount in premium, statutory reserve, IFRS liability and economic capital. Each measurement basis has a distinct purpose and method.

The absence of microdata constrains what this monograph can establish. It can demonstrate bias mechanisms, compare actual methodological revisions and identify audit conditions; it cannot quantify a particular market's insured-population under-registration, determine real portfolio coding-matrix elements, or estimate individual insurer reserve effects. Those quantities require authorized case-level data, definitions, exposure histories, matched validation records, and appropriate uncertainty calculations.

Empirical Tests and Comparative Research

Research design for an insurer portfolio

A defensible empirical study begins by freezing a roster of eligible covered lives, policy and benefit histories, and observation windows as of a declared extraction date. It then separately freezes each death ascertainment source and records the source date, event date, data type and update history. Linking is performed with documented deterministic and probabilistic strategies; a stratified sample of pairs classified as matched, unmatched and uncertain is examined by independent reviewers. The review targets both false matches and missed matches. Duplicate persons, repeated contracts, concurrent policies and invalid exposure segments are analyzed before rate calculation.

Next, the analyst constructs at least three distinct outcomes: (1) all-cause occurrence among eligible insured lives; (2) cause-coded deaths as supplied by the selected public or medical source; and (3) claims or benefits established under contract definitions. Agreement among these outcomes is informative only after their populations and dates have been harmonized. For example, a benefit may be payable on an insured person's death in one period but adjudicated and paid in the next; a medical record may be coded much later; and a valid death outside the period of coverage may be present in the medical data but excluded from the insured-exposure numerator.

The SOA 2009–2016 individual life report provides a public illustration of structured insurance mortality experience files and data specifications, while the SOA mortality study index demonstrates that successive studies have different observation years and supporting material (SOA, 2021; SOA, 2026). A historical validation must use the actual versioned data specification for each submission year. A familiar field name can preserve syntax while its eligibility or cleaning rules change.

Minimum validation domains

A practical insurer-specific QA plan assesses coverage, timeliness, identity linkage, coding, and reproducibility, then connects each domain to an actuarial impact. For coverage, reconcile person counts, policy-years and life-years to administration data by product, age, sex, underwriting class, duration, jurisdiction, face-amount band and termination reason, insofar as those fields are valid and lawful to process. For timeliness, evaluate death occurrence-to-notification, occurrence-to-registration and registration-to-extract delays. For identity, measure positive predictive value and recall of the linkage decisions on a reviewed sample, and report nonresolved pairs rather than silently dropping them.

For coding, measure the proportion with medical certification, unknown or ill-defined causes, missing ICD editions, invalid code/version pairs and causes that fail age/sex plausibility checks. Draw a stratified sample for independent dual coding, assess both broad and high-impact cause groups, and retain cross-tabulated transitions. For reproducibility, verify whether a second analyst can regenerate the exact death counts, exposures, grouped rates and tables using the archived inputs, transformation code, extracts and documented policy rules. A successful overall reconciliation should not suppress failures in a high-risk subgroup.

Time series validation requires two passes: one on a consistent historical definition where possible and one on the reported original vintages. Disagreement between them is evidence of definition, source or processing change, not necessarily a true change in health risk. The ONS 2024 deaths bulletin records coding changes affecting the interpretation of certain cause patterns (ONS, 2025). A data change log should be examined alongside apparent excess mortality peaks or troughs.

Comparator populations and selection

The Human Mortality Database and WHO population mortality sources are useful benchmarks for trend plausibility and age-calendar composition, but an individual life insurer's risks are selected. Underwriting eligibility, smoker distinctions, acquisition channel, policy affordability, lapse behavior and duration since issue can produce substantial differences from the population average. A ratio of insured to population mortality should be interpreted after standardizing relevant variables and explaining residual selection. Even a well-matched population reference may fail to capture policyholder selection during a pandemic or rapid change in the insured cohort.

The HMD project overview documents the purpose of comparable population mortality series, while the HMD methods and explanatory materials make their input conventions explicit. The meaningful comparison is between estimands after appropriate adjustment. A life-insurance study of persons aged 30–70, for instance, cannot infer coding validity from agreement with a national rate that includes different ages, sex distributions and exposure conventions. Agreement at one aggregate rate can result from offsetting numerator and denominator errors.

International comparisons face further problems of different civil registration completeness, medical certification infrastructure, ICD adoption, migration, national cause-group shortlists and insured product selection. The WHO's global CRVS fact sheet and Eurostat's published metadata indicate why cross-country comparisons should specify the source jurisdiction and observation framework. Pooling several countries without country-specific reporting and coding controls can yield false precision and spurious trends. The solution is explicit stratification, hierarchical estimation where warranted, and sensitivity analysis around country-specific ascertainment and classification changes.

Evidence hierarchy and competing explanations

The strongest validation uses original policy and death-event records plus independently checked linkage and certification samples under a documented governance process. Next come institutionally reconciled aggregates with preserved revision histories. Official public releases with robust methodological documentation can benchmark external mortality, while general research literature explains mechanisms but ordinarily cannot supply an insurer's true undercount fraction. The hierarchy is functional: a peer-reviewed publication can provide excellent statistical theory yet be unsuitable as a source of case-specific contract facts; an internal claims ledger can define a contract payment while offering no independent clinical cause validation.

A mortality trend anomaly should initially be examined against several possible explanations: true risk change, cohort composition, lapse-selection shift, exposure error, late death reporting, source transfer, duplicate records, correction of identity matches, and ICD or software transition. Multiple hypotheses can be compared with source-vintage and cohort-level diagnostics. A method that presupposes every unexplained change is a medical phenomenon risks circular reasoning. Conversely, a documented coding change does not prove that no real mortality change occurred during the same period.

Comparative sensitivity illustration

Consider two years of equal true all-cause deaths, 500 each, and identical life-year exposure. In year one, 95% of deaths are detected without false matches, yielding 475 observed. In year two, ascertainment falls to 90%, yielding 450. An unadjusted comparison suggests a decline of 25/475 ≈ 5.26% in mortality, despite equal true risk. If coding software also shifts 10% of a particular cause group into another category, the cause-specific trend may diverge further from the all-cause trend. The first distortion is missingness; the second is classification. Both can coexist with an unchanged underlying risk distribution.

Suppose independently audited capture rates have uncertainty of ±3 percentage points in each year. A single-point adjustment might yield a flat corrected trend, but the range of compatible trends would be wider. A robust report would state results under plausible reporting scenarios, analyze changes in certified cause distributions within matched deaths, and avoid using the coded-cause result as an automatic validation of the ascertainment correction. The latter would assume away the distinct mechanism under study.

Scientific Synthesis and Conclusions

Integrated interpretation of the evidence

The documentary comparison yields a coherent account of the insured-mortality evidence process. The United Nations and WHO describe the public registration and medical-certification infrastructure. HMD documents the construction of population rates from traceable input data and exposure conventions. NCHS and ONS demonstrate the significance of provisional reporting, underlying-cause selection, coding revisions and dual coding. SOA experience research shows how private insured-population evidence depends upon underwriting and participant-specific data, changing custodians and table versions. Statistical record-linkage and incomplete-data literature provides methods to estimate and bound uncertainty under explicit assumptions. These sources converge on a requirement for traceable definitions, but they address different observational populations and objectives.

Three results follow from the formal analysis. First, an accurate ICD code on a detected death does not measure how many deaths are missing from the insured experience sample. Second, high broad-group agreement between coding versions can coexist with a materially distorted small cause category, as the documented ONS software comparison illustrates. Third, a seemingly accurate all-cause aggregate rate does not establish record quality: numerator and denominator errors, or subgroup errors of opposite signs, can cancel. The required validation is therefore multidimensional and linked to the decision for which the data are used.

The comparison also identifies an important distinction in how evidence is regarded. A statistical death is an observed or reconstructed event under a population and time definition; an insurance death claim is a contractual and administrative event; a cause-of-death code is the output of a medical-certification and classification process. The same person's death may appear in all three, yet their alignment needs an explicit record-level explanation. This conclusion is an analytic interpretation of established sources and definitions, not a proposed new theory or independent legal standard.

Conditions for an auditable mortality estimate

An insurer can support a reproducible mortality estimate when it has: a documented cohort and unit of analysis; verified exposure histories; source and extract dates; a death ascertainment plan; quantified linkage errors; treatment of unknown and late deaths; a coding version ledger for any cause-specific outcome; documented selection and comparability constraints; uncertainty and sensitivity analyses matched to plausible data defects; and an auditable transformation history. The requirements need proportionate application. All-cause mortality pricing may need intensive death and exposure reconciliation with modest cause-coding work; a cause-specific rider study requires detailed clinical and contractual crosswalks.

Quality assurance should preserve both the original evidence and a reasoned account of corrections. Retroactive changes to a death file should be reported with effective dates and impacts on counts, rates and liabilities, not silently inserted into historical analyses. Comparison tables should distinguish “as originally observed” from “restated under the common method.” A source can be suitable for one purpose and unsuited to another. That judgment is part of the scientific conclusion and should be preserved in the study passport.

Research agenda supported by the analysis

Further empirical work could compare administrative mortality extracts with independent vital registration in several legally accessible insured cohorts, using documented dual-source validation to quantify capture sensitivity, false matches, delay development and population selectivity. Separate dual-coding studies of insured deaths could estimate cause-specific misclassification matrices for material product features. These studies would need to stratify by source change, contract type, country, age, sex, underwriting class, policy duration and era of coding implementation, while maintaining lawful handling of health information. None of those estimated results is available in the public source material reviewed here; their expected value lies in testing the mechanisms derived and documented in this monograph.

Conclusion

Mortality evidence in life insurance becomes scientifically interpretable when the chain from insured exposure to death occurrence, registration, linkage, classification and actuarial use is documented and tested. The quality of a final count or code cannot be inferred from its formal appearance. Completeness of death ascertainment, identity-linkage accuracy, exposure integrity, medical certification, code-version stability, and contractual meaning are related but distinct empirical questions. The original sources reviewed here establish that public mortality registration varies across jurisdictions, provisional data develop over time, and coding-software changes can materially redistribute selected causes even when aggregate agreement is high.

The central analytical consequence is practical and generalizable: every mortality estimate must be tied to a clearly stated population, observation window, source version, transformation history, and use case, with uncertainty appropriate to the evidence available. Validated mortality data can support pricing, assumption governance, risk monitoring, financial measurement and research. Unexamined source uncertainty can travel from apparently minor record-level defects into material portfolio-level conclusions. The appropriate response is reproducible analysis, independent source checking, and explicit sensitivity assessment. These are conclusions supported by the documented research and mathematical demonstrations, without a claim of newly discovered empirical mortality rates or a new doctrinal system.

References

  • Alipour, J., and Payandeh, A. (2021). Common errors in reporting cause-of-death statement on death certificates: A systematic review and meta-analysis. Journal of Forensic and Legal Medicine 82, article 102220. DOI 10.1016/j.jflm.2021.102220; PMID 34325081. Source.

  • Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum Likelihood from Incomplete Data via the EM Algorithm. Journal of the Royal Statistical Society, Series B 39(1), 1–22. Source.

  • European Parliament and Council of the European Union (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union L119, 1–88; consolidated text available EUR-Lex. Source.

  • Eurostat (n.d.). Causes of death (hlth_cdeath): Reference metadata. Eurostat, European Commission. Source.

  • Fellegi, I. P., and Sunter, A. B. (1969). A Theory for Record Linkage. Journal of the American Statistical Association 64(328), 1183–1210. Source.

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

  • Human Mortality Database (n.d.). Project Overview. Human Mortality Database, international demographic collaboration. Source.

  • International Accounting Standards Board (2017, effective 2023). IFRS 17 Insurance Contracts. IFRS Foundation, IFRS Accounting Standard and explanatory overview. Source.

  • Jaro, M. A. (1989). Advances in Record-Linkage Methodology as Applied to Matching the 1985 Census of Tampa, Florida. Journal of the American Statistical Association 84(406), 414–420. Source.

  • Little, R. J. A., and Rubin, D. B. (2019). Statistical Analysis with Missing Data, Third Edition. Wiley Series in Probability and Statistics, John Wiley & Sons. DOI 10.1002/9781119482260. Source.

  • National Center for Health Statistics (2026). Cause of death. Health, United States: Sources and Definitions, updated September 16, 2026. Source.

  • National Center for Health Statistics (n.d.). COVID-19 Death Data and Resources. National Vital Statistics System (NVSS), provisional mortality data guidance. Source.

  • Office for National Statistics (2019). Cause of death coding in mortality statistics, software changes: January 2020. ONS technical report, released November 29, 2019. Source.

  • Office for National Statistics (2022). Cause of death coding in mortality statistics, software changes: January 2022. ONS technical report, January 7, 2022. Source.

  • Office for National Statistics (2025). Deaths registered in England and Wales: 2024. Office for National Statistics release, 2024 reference year. Source.

  • Office for National Statistics (n.d.). Mortality statistics in England and Wales: Quality and Methodology Information. ONS methodological guidance. Source.

  • Society of Actuaries (2015–2018). 2015 Valuation Basic Report and Tables. SOA tables and report; report updates through September 13, 2018. Source.

  • Society of Actuaries (2026). Individual Life — Mortality Experience Studies. SOA Research Institute, online study index. Source.

  • Society of Actuaries, Individual Life Experience Committee (2017). 2009–2013 Individual Life Insurance Mortality Experience Report. Society of Actuaries, 2017, updated October 23, 2017. Source.

  • Society of Actuaries, Individual Life Experience Committee (2021). 2009–2016 Individual Life Insurance Mortality Experience Report. Society of Actuaries Research Institute. Source.

  • Society of Actuaries, Individual Life Experience Committee (2024). 2019 Individual Life Insurance Mortality Experience Report. Society of Actuaries Research Institute; study years 2012–2019. Source.

  • United Nations Statistics Division (2014). Principles and Recommendations for a Vital Statistics System, Revision 3. Statistical Papers, Series M No. 19/Rev.3. United Nations, New York. ISBN 978-92-1-161572-2. Source.

  • Wilmoth, J. R., Andreev, K., Jdanov, D., Glei, D. A., and Riffe, T. (2025). Methods Protocol for the Human Mortality Database, Version 6. Human Mortality Database, revised August 5, 2025; with collaborators. Source.

  • World Health Organization (2022). ICD-11. WHO fact sheet. Source.

  • World Health Organization (2024). Civil registration and vital statistics. World Health Organization, Fact sheet, February 27, 2024. Source.

  • World Health Organization (2025). ICD-11 Implementation. WHO, frequently asked questions; updated December 2, 2025. Source.

  • World Health Organization (n.d.). Cause of death. WHO classifications and mortality certification guidance. Source.

  • World Health Organization (n.d.). Analysing Mortality Levels and Causes of Death (ANACoD3). WHO cause-of-death data quality service. Source.