Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Life Insurance Cause-of-Death and Excess Mortality Analysis

Abstract. Cause-of-death information and deviations from expected all-cause mortality enter life insurance through distinct channels. Cause-specific tabulations describe the medically certified distribution of deaths, subject to selection rules, diagnostic practices, classification changes and delayed reporting. Excess mortality compares observed deaths with a modeled counterfactual and measures the net impact of concurrent changes across causes. Neither measure, alone, determines the death-claim frequency or monetary loss of a life insurer. Contract eligibility, underwriting selection, age composition, insured exposure, policy duration, death-benefit amounts, timing, reinsurance and the dependence of deaths across insured lives must also be considered. This monograph develops an integrated critical analysis of the existing statistical and actuarial evidence. It compares the World Health Organization's mortality-classification and global excess-death methods with Human Mortality Database surveillance, national statistical approaches, and Society of Actuaries insured-life experience. It derives exposure-weighted cause-specific rates, a transparent expected-claims decomposition, alternative counterfactuals, claim-severity covariance, and bounds under incomplete cause attribution. Numerical examples are explicitly hypothetical and independently recalculated; no insurer microdata or new causal estimates are claimed. The central conclusion is that a coherent life-insurance assessment requires separate measurement of death occurrence, cause attribution, expected mortality, exposure and insured monetary loss, followed by reconciliation across these layers. The monograph supplies a reproducible research workflow, identifies transportability and identification failures, and explains the implications for underwriting evidence, pricing assumptions, reserves, capital stress, reporting and scientific interpretation.

Keywords: Life insurance; cause of death; underlying cause; multiple causes; excess mortality; all-cause mortality; mortality baseline; standardized mortality ratio; actuarial exposure; insured selection; claim frequency; death benefit; mortality catastrophe; mortality surveillance; data provenance; competing risks; pandemic; uncertainty; international comparability.

Introduction: the scientific object and evidentiary problem

Life insurance creates a contractual financial claim triggered by an insured death, according to the applicable policy wording and law. Mortality research informs the probability of that trigger and its dependence on age, calendar period, underwriting history and external shocks. Cause-of-death statistics supply a different kind of evidence: they classify a certified death into a biomedical reporting system. Excess mortality supplies yet another: a comparison with deaths expected under a specified counterfactual. Scientific inference becomes unreliable when an observed rise in aggregate deaths is treated as a direct estimate of a particular insurer's death benefits, or when a shift in cause-coded counts is treated as a causal effect without verifying the coding and denominators.

The WHO cause-of-death guidance establishes the connection between medical certification and underlying-cause selection. WHO's excess-mortality methodology instead works from observed all-cause mortality and modeled expected counts. The Society of Actuaries insured-life study illustrates the additional requirement of an insured portfolio-specific experience base. These evidence types answer related questions, while their statistical populations, error mechanisms and financial endpoints differ.

The research object is the inferential passage from cause-specific death occurrence and excess mortality to life-insurance claim exposure. The inquiry uses documentary comparison, statistical derivation, explicit hypothetical stress calculations and critical appraisal of insured-life and population evidence. It covers international measurement architecture, with U.S., European and global sources used as illustrative documented cases. It makes no cross-jurisdictional assertion that a particular medical code, exclusion, payment deadline or regulatory treatment governs every contract. It introduces no new named actuarial model or doctrine.

Six questions organize the analysis. What is certified and what is counted? Which population occupies the denominator? How was the no-shock baseline constructed? Can a population change be transported to insured lives? How do death-benefit amounts and dependent shocks convert frequency to aggregate loss? Which uncertainties remain even when the arithmetic is internally correct? The answer to each question conditions the next, and an explanation that fails one step cannot be rescued by precision in the subsequent step.

Definitions, historical and documentary foundation

Death, cause, exposure and claim as separate observations

A death event is the occurrence of death in a defined person, place and time. An all-cause death count registers events without depending on one selected medical cause. A cause-specific death count allocates those events under specified certification and classification rules. An insured death is a death of a person covered by an identified contract during an applicable period; a reported death claim is an insurer-side administrative record that may occur later and may contain multiple policy records for one death. A paid death benefit is a monetary outcome after application of coverage conditions, benefit structures, offsets and other contract-specific provisions. A single event can therefore correspond to zero, one or multiple policies, and to multiple claim transactions.

For a population cell defined by age a, sex s, time t and relevant geography g, let D(a,s,g,t) denote deaths and Y(a,s,g,t) person-years exposed. The central death rate is m = D/Y. A probability of death over an interval is a distinct quantity, often denoted q, and requires life-table or survival assumptions to translate from a central rate. For insured data, policy-years, life-years, face amount exposed and sum-at-risk exposed are different denominators. The choice must follow the scientific question. A comparison of claim counts with population death rates is invalid when the numerator and denominator have mismatched temporal eligibility or risk classification.

Cause shares D_c / D measure the recorded distribution among deaths. They are not probabilities of death from cause c for a living cohort. The cause-specific mortality rate m_c = D_c/Y shares an exposure denominator across mutually exclusive underlying causes. Where multiple contributing causes are counted, one death may contribute to more than one cause category; such counts cannot simply be summed to reconstruct all-cause mortality. This elementary accounting rule becomes especially important in analyses of infection, chronic disease, injury and multimorbidity.

An insurer's economically relevant exposure depends on benefit amounts. With insured person i, death indicator I_i and contractual amount payable B_i, aggregate gross death benefits equal L = sum_i B_i I_i under a simple one-period death-triggered benefit abstraction. Actual product payments require contract-specific definitions and may involve reserves, accelerated benefits, riders and exclusions. The expression is used here for analysis, not as a substitute for claims administration or valuation rules. A portfolio with fewer deaths can experience a higher aggregate loss if the affected policies have sufficiently higher sums assured.

Historical standardization and international data architecture

Mortality classification permits large sets of clinical descriptions to be compared by categories with documented selection rules. The WHO ICD overview places the International Classification of Diseases in the longstanding tradition of comparative mortality statistics, and records the adoption of ICD-11 by the World Health Assembly in 2019 with an effective date of January 1, 2022. Effective international status does not establish immediate universal adoption in national death-registration systems. WHO's implementation FAQ expressly allows continued ICD-10 use during transition and cautions against treating simple crosswalks as definitive data conversions.

The statistical consequences are substantial. A change in diagnostic sensitivity, death-certificate wording or mortality coding can shift reported cause shares without an identical change in the physiological risks facing insured lives. The researcher must record the jurisdiction's classification revision, implementation date, national modifications, automation methods, comparability studies and changes in civil registration. Historical trends spanning classification changes should be analyzed on a documented comparable basis or separated into distinct measurement regimes.

The Human Mortality Database short-term series provides standardized weekly death counts and selected age-specific rates. The associated STMF data description explains that the data emphasize countries with comparatively complete registration, that historical coverage differs by country, and that recent weekly observations are revised. It is useful for detecting shocks and evaluating trends within its coverage, while its country selection limits direct generalization to low-registration settings. Insurer death-claim systems have different selection and timing problems: a single multi-country insurer may require different mappings and reporting-lag adjustments in each jurisdiction.

Units, populations, time and analytical scope

Any comparison needs a declared risk set: resident population, registered deaths, policyholders, insured lives, employee groups, policy contracts or claims. Duration since underwriting matters when newly selected insured cohorts have lower mortality than mature in-force blocks. Age and sex matter because portfolio composition can dominate apparent aggregate mortality changes. Calendar week, month and year matter because cause mix, seasonality, epidemics and reporting latency occur on different timescales. Country of death and country of residence may differ from policy jurisdiction. The relevant date could be death occurrence, registration, notification, admissibility decision or payment. These dates must not be interchanged without reconciliation.

A formal source record for each estimate should therefore contain a population definition, age range, exposure unit, period, coding version, treatment of missingness, baseline version, territorial coverage, source retrieval date, data revision and financial basis. The same metric name may otherwise mask incomparable numbers. Comparative statistics should prefer age-specific rates or fixed-standardized rates when population structures differ and should display unstandardized counts where capacity and cash-flow volume are the object of decision.

Cause-of-death coding

Certification, underlying cause and contributing causes

The WHO medical certification framework distinguishes the sequence leading directly to death from other significant conditions contributing to death. Underlying-cause tabulation selects one cause following international rules, while multiple-cause systems retain additional medically recorded conditions. The principal analytic distinction is between a unique statistical allocation of a death and a full medical description of multiple contributing pathologies. A single-code allocation simplifies tabulation but may conceal clinically relevant comorbidity. The same death can be described differently for surveillance, clinical inquiry, claims evidence and litigation, without any one record carrying all the information needed by every institution.

A death certificate is an expert-reported observation, produced under institutional and clinical constraints. Its causal chain is not a randomized attribution experiment. Coding rules govern selection among recorded statements; coding consistency alone cannot establish that the certified underlying condition was the sole biological determinant. Conversely, the absence of a particular diagnostic code may reflect insufficient testing, incomplete information or reporting practice rather than proof that the condition played no causal role. The WHO pandemic mortality research explicitly identifies differences in diagnostic capacity and certification among reasons reported COVID-19 deaths diverged from modeled pandemic-associated excess mortality.

For the insurer, cause may matter for a cause-conditioned benefit, an exclusion, a rider, an occupational classification, a mortality study or an unusual-death investigation. The conditions for a particular death-benefit claim remain those of the applicable contract and legal system. An observed cause-code sequence never substitutes automatically for the actual policy trigger. Research use of certificates should be purpose-limited and subject to lawful access and confidentiality requirements; aggregate statistics cannot reconstruct a beneficiary's entitlement.

Classification revision, ill-defined causes and bridge coding

Cause-specific series can break when revisions reassign cases across categories. The ICD-11 transition guidance warns that mappings between ICD-10 and ICD-11 are not a lossless automated transformation. A scientifically sound bridge study applies two coding regimes to an appropriate common set of records and estimates cause-level comparability ratios, together with uncertainty and subgroup differences. A ratio estimated for one setting cannot be assumed invariant across different cause mixes, certifier practices or years.

Ill-defined and pending causes require an explicit analytic category, rather than silent redistribution. The U.S. NCHS technical notes explain that provisional cause-of-death estimates may contain unknown or pending classifications and that delays vary by cause. A researcher comparing injury with cardiovascular mortality in the latest weeks is especially exposed to differential completion. Reallocating all pending deaths proportionally across known causes assumes missingness independent of cause after conditioning on observed strata. That assumption should be evaluated, not treated as a data-cleaning identity.

If a share u of deaths is unclassified, a simple identification interval for cause c is D_c <= D_c,true <= D_c + U, where D_c is known classified deaths and U is the unresolved count. These are intentionally broad deterministic bounds; clinical investigation, registration metadata or plausible redistribution constraints can narrow them, but only after introducing and documenting additional assumptions. A reported cause percentage should specify whether its denominator includes the unknown-cause category.

Misclassification, missingness and delayed registration

Write the vector of true mutually exclusive cause-specific counts as T and observed coded counts as O. A misclassification matrix M with entries M_{jc} = P(observed cause j | true cause c) gives E[O] = M T when classification probabilities are stable and all deaths are captured. Recovering T requires information about M and adequate rank; estimated or assumed matrices introduce uncertainty. Under-recorded deaths require a separate registration component, so a more realistic observed process also incorporates capture probabilities and timing. An inverse of a noisy estimated M can magnify error severely, particularly for rare causes. Reporting a corrected point estimate without a corresponding error interval exaggerates what the evidence can establish.

The NCHS provisional death surveillance states that recent counts are revised as later death certificates arrive and that reporting speeds differ across jurisdictions. A reproducible nowcast should select a data vintage, fit lag-distribution parameters using historical revisions, and validate against held-out finalized periods. Pandemic surges may disrupt normal reporting dynamics, so the lag model requires stress testing. Deaths should be grouped by occurrence date when the mortality event is the question, while claims cash-flow analysis should separately model notification and settlement timing.

For insurer claims, delay has at least three components: time from death to notification, notification to adjudication and adjudication to payment. Medical cause revision may occur independently of these intervals. Combining all delays into one correction factor prevents analysts from identifying whether late reporting changed apparent mortality, claims settlement or both. Record linkage can improve information by connecting insured life histories with reliable external vital records, but the linkage's false matches, unmatched records and permissions must be explicitly audited.

Cause-specific and competing-risk estimators

Under mutually exclusive first-death causes, cause-specific hazard lambda_c(t) measures the instantaneous event rate from cause c among lives surviving to t. The all-cause hazard satisfies lambda(t) = sum_c lambda_c(t), and the survival function is S(t) = exp[-integral_0^t lambda(u) du]. The cumulative incidence of cause c is F_c(t) = integral_0^t S(u) lambda_c(u) du. Thus, an increase in one cause affects the number surviving long enough to experience another. Cause-specific hazards cannot be read as independent probabilities that simply add over a finite period. This derivation follows ordinary competing-risk probability theory; no novel framework is asserted.

For an insured portfolio, estimate D_c / Y in appropriately homogeneous strata, with confidence or predictive intervals reflecting sparse counts and exposure uncertainty. A Poisson event-count approximation with mean Y lambda_c can be useful for short homogeneous periods; it becomes inadequate when deaths exhibit overdispersion, dependent hazards or material variation across policyholders. Negative-binomial, hierarchical, multistate and survival analyses may improve fit but require substantive justification and out-of-sample validation. The presence of a detailed mathematical model does not repair weak cause data.

Baseline excess mortality

Counterfactual deaths, rates and standardization

Let O_t be observed all-cause deaths in period t and E_t^0 expected deaths under a specified counterfactual without the shock under investigation. Absolute excess mortality is X_t = O_t - E_t^0; its proportional counterpart is P_t = (O_t - E_t^0)/E_t^0 for E_t^0 > 0. A negative value denotes fewer observed deaths than expected. Both are properties of the chosen counterfactual; neither is a directly recorded field on a death certificate. The WHO global excess-death explanation explicitly describes increases and decreases from different direct and indirect mechanisms.

A baseline that multiplies age-sex-specific pre-event rates by projected exposure is preferable to a raw average of prior annual death counts when population size and age structure change materially. Express E_t^0 = sum_{a,s,g} Y_{a,s,g,t} m_{a,s,g,t}^0. The counterfactual rates m^0 may include seasonality, trend and specified demographic evolution. This expression also clarifies that an unexpectedly large number of deaths may coexist with declining age-standardized rates if a population becomes older or larger. Crude excess counts can remain necessary for hospital capacity, funeral-service demand and insurer cash flows, while standardized rates address comparative intensity.

The U.K. Office for National Statistics methodology explanation documents the move from a five-year count average to model-based expectations that recognize population aging and mortality trends. This is an empirical example of how the baseline is part of the statistical result. Changing baselines may revise the excess estimate without changing any observed death. For life insurance, expected deaths should similarly be derived from the in-force composition and contract exposure rather than imported unadjusted from a national count series.

Baseline construction, shocks and sensitivity analysis

Three common classes of baseline are a moving average of historical death counts, age-structured rate extrapolation, and statistical forecasting with seasonal, secular and contextual covariates. A simple average is transparent and can be useful in stable populations, but it conflates mortality rates with population change and can incorporate prior shock years. Trend extrapolation can be sensitive to endpoint selection and to pre-shock stagnation or acceleration. Flexible forecasting introduces greater model dependence and must document selection, diagnostics and prediction intervals. The Human Mortality Database STMF documentation and EuroMOMO methods FAQ provide complementary contexts for surveillance of short-term fluctuations.

A baseline sensitivity analysis should vary training-window endpoints, demographic exposure projections, seasonality specifications, excess outliers and the treatment of prior epidemics. It should present multiple plausible expected-death trajectories and identify whether conclusions about direction and magnitude are robust. Repeated fitting after viewing the shock period can induce hindsight bias; a prospective validation exercise fits baselines using only pre-event information and scores predictions on held-out periods. The absence of a single internationally agreed method, acknowledged in the ONS explanation, makes declared methodology a prerequisite for comparison.

A persistent mortality shock may create later mortality displacement: individuals who died during the shock could otherwise have died in a later period, altering the pool at risk. Lower later deaths can reflect this depletion, broader health improvement or changing social conditions. The mechanism cannot be inferred from a low later aggregate count alone. Long-run interpretation requires cohort tracking or a well-specified multistate approach rather than mechanical subtraction of an earlier peak from later deficits.

Statistical uncertainty, displacement and attribution

Observed counts vary stochastically, and expected counts are estimated. A basic variance calculation for X = O - E^0 includes Var(O), Var(E^0) and their covariance where relevant. Conditional Poisson reasoning for O with a fixed baseline omits uncertainty in estimated trends, exposure denominators and registration. More realistic uncertainty propagation can use the posterior predictive distribution or bootstrap replicates that refit the counterfactual and reporting-lag model. For insurer capital purposes, cross-cell dependence and tail behavior matter more than a narrow interval around a total observed-count ratio.

Excess deaths are an all-cause net measurement. A pandemic can increase deaths directly from infection and indirectly through delayed treatment, while restrictions may decrease traffic injuries or other infections. WHO's 2023 updated methods explicitly accounts for such countervailing effects. Therefore X = O - E^0 does not identify unique medical causes, and the difference between X and a disease's death certificates is not itself an estimate of miscoding. It combines divergent mechanisms, observational differences and counterfactual uncertainty. Attribution to a single exposure or intervention requires additional design, confounder control and credible counterfactual reasoning.

Cause-specific excess, X_c = O_c - E_c^0, can be computed if classification and comparability are adequate and each cause has an appropriate baseline. It is particularly fragile when diagnosis or coding changes during the shock. If causes are mutually exclusive and both baselines and denominators are coherent, the sum of cause-specific excesses equals all-cause excess algebraically. The equality fails as a meaningful diagnostic when different cause series have incompatible coverage, when multiple-cause counts are added, or when baseline counterfactuals are assembled inconsistently. An unexplained reconciliation gap is a data-governance finding, not residual proof of a new disease mechanism.

Comparisons from global and national evidence

Msemburi et al. (2023) WHO estimates in Nature report approximately 14.83 million global excess deaths for January 2020 through December 2021, compared with 5.42 million reported COVID-19 deaths, with a 95% uncertainty interval of 13.23–16.58 million for modeled excess. These are estimates for a defined period and estimation system, not a perpetual or insurer-specific multiplier. The paper uses modeled deaths where complete records are unavailable and describes the limitations of missing national registration. WHO's 2025 explanatory Q&A reiterates the distinction between reported causes and the estimated all-cause counterfactual.

The Lancet systematic analysis represents an alternative global modeling exercise. Estimates across international projects should be compared with attention to country coverage, estimation date, source vintage, national consultation, baseline specification and model uncertainty; numerical differences by themselves do not establish which mortality estimates are intrinsically valid for insured populations. The STMF series publication makes data-vintage and reporting differences particularly visible at a weekly scale.

For U.S. interpretation, the 2023 mortality data brief reports an age-adjusted decline in overall mortality from 2022 to 2023, while the NCHS technical notes explain that provisional cause-specific counts are subject to revision. These two observations illustrate separate challenges: a finalized annual rate does not determine a prior shock's causal attribution, and a rapidly available cause-specific series may not yet have reached final classification. Neither source, alone, supplies an insurer-specific projection.

Insured portfolio concentration

From population deaths to policyholder claims

The observed mortality rate of a national population is the outcome of a particular mix of ages, socioeconomic positions, occupations, health states and exposure environments. The insured risk set has been formed through demand for coverage, access, underwriting, policyholder choices, group employment, lapse and duration effects. The SOA individual-life COVID-19 study compares insured mortality with general-population experience; its disclosed sample contains data from 29 U.S. companies and approximately 3.2 million death claims in the analyzed period through 2022. The study's reported industry face-amount coverage is a representation measure, not a random-sampling guarantee for every segment or a statement of global transportability.

For stratum k, let Y^I_k be insured person-years and m^I_{k,0} the counterfactual insured mortality rate. Expected insured deaths are E^I_0 = sum_k Y^I_k m^I_{k,0}. If the population mortality shock multiplier r^P_k = m^P_{k,1}/m^P_{k,0} is applied as a scenario, the implied insured scenario expectation is E^I_1 = sum_k Y^I_k m^I_{k,0} r^P_k. This is a stress assumption, not a measured fact about insured mortality. Its credibility depends on similarity of risk pathways and on whether underwriting, occupation, residence, access to care and protection against the shock change relative risk. A single national P-score imposed uniformly on every insured stratum suppresses all these conditions.

A more informative empirical comparison estimates the insured ratio r^I_k from observed death experience, calibrated exposures and a stable pre-event insured baseline, and then compares it with r^P_k. Where sample sizes are small, hierarchical pooling or credibility weighting can stabilize estimates but can also obscure a genuinely concentrated shock. Stratification should be designed by scientific relevance rather than solely by convenient administrative partitions. Mortality effect heterogeneity cannot be inferred simply because the aggregate ratio differs between two insurance books; compositional differences must first be addressed.

Selection, benefit-weighted loss and concentration

In a standard death-benefit abstraction, let p_i be the one-period probability of death and B_i the gross benefit payable conditional on an eligible death. Then E[L] = sum_i B_i p_i. If policies have variable benefits or exclusions, the expectation must include the relevant conditional payment probability and amount, not just the existence of a certificate. Expected excess benefit loss relative to a specified baseline is Delta E[L] = sum_i B_i (p_i^1 - p_i^0), provided contract payoffs are held fixed; policy lapses, new business and benefit changes require an expanded dynamic model.

The distinction between lives and amount exposed is decisive. Consider two insured strata. Stratum A contains 80,000 covered lives with average death benefit $100,000, baseline mortality probability 0.002 and stressed probability 0.003. Stratum B contains 20,000 lives with average benefit $500,000, baseline probability 0.004 and stressed probability 0.005. Expected excess deaths equal 80 + 20 = 100. Expected excess death benefits equal $8 million for A and $10 million for B, or $18 million combined. Stratum B generates only one-fifth of excess death counts but more than one-half of excess benefit loss. These figures are hypothetical. The arithmetic is transparent and illustrates why cause- and age-specific claim counts cannot replace insured amount-weighted analysis.

For life insurance, concentrated benefits matter alongside concentrations of people. Geographic employer clusters, occupations, travel patterns, common buildings or institutions, and co-resident insureds may generate joint exposure to a hazard. A single catastrophic event can affect many policies, and a widely distributed pandemic can also create dependence across otherwise geographically diversified insured lives. Whether a higher share of a cause of death produces higher losses depends on the intersection of cause distribution, coverage conditions, benefits and joint exposure. Risk accumulation should be investigated at the insured life level and at the contract and counterparty levels.

Cause-of-death groups do not carry a universal financial severity weight. Under conventional death cover, an eligible death from one cause may produce the same benefit as an eligible death from another. Cause still provides useful information for stress identification, trend inference, claim verification and the relevance of policy riders or exclusions. When cause changes the payment amount or eligibility, the analysis must cite the actual product terms and jurisdictional rules. Such conditions cannot be assumed from a general mortality study.

Correlation, tail loss, capital and reinsurance interfaces

The variance of aggregate simple death-benefit loss is Var(L) = sum_i B_i^2 p_i(1-p_i) + 2 sum_{i<j} B_i B_j Cov(I_i,I_j). The covariance term is zero under independent deaths but can dominate under common shocks. If conditional death probabilities vary with a common random event intensity Z, conditional independence still yields positive unconditional dependence when individuals' conditional probabilities increase together with Z. Consequently, a portfolio's expected excess deaths and its loss distribution tail answer different questions. A well-calibrated expected-loss ratio can coexist with a materially underestimated catastrophic tail.

The European Solvency II mortality risk rule describes an instantaneous permanent 15% increase in applicable mortality rates for the mortality risk sub-module. The distinct life-catastrophe rule specifies an instantaneous increase of 0.15 percentage points to the rates applicable over the following 12 months, subject to the rule's scope. A relative 15% stress and an absolute increase of 0.15 percentage points are mathematically different: if a baseline annual death probability is 0.002, the former adds 0.0003 while the latter adds 0.0015. These regulatory tests cannot be interchanged, and standardized formulas are not historical observations of actual insured catastrophes. The IAIS supervisory framework supplies the broader institutional context for credible prudential risk assessment across jurisdictions.

Mortality reinsurance changes the insurer's retained financial position, not the physical death event. Gross claims, net claims after reinsurance, reinsurance recoverability and capital requirements should therefore be maintained as distinct accounting and risk quantities. Attachment points, reinstatement, aggregate limits, event clauses, pandemic wording and counterparty creditworthiness determine net effects. Their precise assessment belongs to the governing treaty and applicable regulation. This monograph addresses the informational basis of accumulation and reserves, while dedicated reinsurance and solvency dossiers analyze contract and capital architecture in full.

Governance of underwriting and financial projections

The insurer should retain an inventory of mortality assumptions distinguishing experience-derived level, long-term trend, near-term shock loading and catastrophe scenarios. Each assumption should specify its source population, calibration period, insured selection regime, geographic reach, age treatment, face-amount weights and uncertainty interval. Experience studies with incomplete coverage or a changing mix require explicit adjustment. A mortality assumption can be empirically successful for aggregate pricing yet poorly calibrated for a high-benefit concentration, making subgroup and tail validation indispensable.

Observed claim changes require reconciliation against exposure changes before a reserve or pricing conclusion is reached. New business alters in-force ages and underwriting duration; lapse may change both risk composition and benefit volume; inflation and product revisions may change payment severity. A rise in reported death-claim amounts may reflect a larger total insured amount without a comparable rise in mortality probability. Conversely, a high mortality shock can be partially masked by a concurrent fall in active policies. The financial basis—gross, ceded, net, undiscounted or present-valued—must be stated.

Cause-of-death research improves life-risk governance when it identifies which assumptions deserve scrutiny. For example, unusually high respiratory mortality can motivate enhanced near-term surveillance, but management should not convert a cause category directly into a causal prediction for selected lives. Appropriate actions include validation against the insurer's own experience, exposure and claim amount reconciliation, scenario analysis, review of sources and independent actuarial challenge. The role of evidence is to improve decisions while preserving the distinctions on which those decisions depend.

Empirical tests and comparative research

Insured-life studies and population comparability

The SOA individual life experience report and its group life mortality survey are valuable precisely because they use insured experience rather than assuming that national all-cause changes transfer unchanged. Individual-life and employer-sponsored group-life insured populations are different selection mechanisms. The SOA group-life survey series reports analyses by broad age group, industry, region and cause. These are published research programs, but particular survey results require their own period, carrier coverage and methodology before numerical comparison. A landing page is sufficient to verify the documented scope, not to reproduce unseen underlying datasets.

The SOA general-population comparison analyzes excess mortality by age and sex in U.S. population evidence. The researchers explicitly state that this work does not analyze vaccination status or establish causal vaccine effects. The same evidentiary rule applies to any attempted assignment of mortality change to a vaccine, policy action, environmental exposure or medical intervention: aggregate excess deaths do not uniquely identify a cause. Population analyses remain informative for calendar shocks and common risk factors, provided their inferential limits are retained.

A credible insured-population comparison should define pre-event and shock periods identically, use consistent attained-age groups, account for policy duration and changes in underwriting, distinguish individual and group business, and separate claim incurred date from paid date. Where a source offers only life claims but not policy-years or benefits, only a subset of the desired actuarial metrics is identified. That is a limitation of data access, not a license to introduce an undocumented denominator.

Worked example: cause allocation under pending coding

Suppose a hypothetical month records 1,000 insured deaths, of which 700 are coded into noninfectious underlying-cause groups, 200 into an infectious cause group and 100 are still pending. The infectious share among all deaths is currently 20%, while the known-coded share is 200/900 = 22.22%. Under no further evidence, the final infectious share lies between 20% and 30%. Filling all 100 pending deaths with an imputed 22.22% infectious distribution produces approximately 222 infectious deaths, but this point estimate rests on the assumption that pending classification resembles completed classification. If investigation delays are systematically longer for one cause, the assumption can fail.

For a stable cohort exposure of 50,000 person-years, the infectious-cause rate based on confirmed cases is 200/50,000 = 0.004 per person-year. The classification bounds imply 0.004 to 0.006 per person-year, without accounting for deaths missed by the insurer or by vital registration. An apparently precise 0.00444 based on proportional allocation has a narrower displayed precision than the information supports. A proper experience table would preserve the unclassified category, record data vintage and update the estimate when pending cases are resolved.

Worked example: baseline selection and age composition

Suppose 2026 all-cause deaths in a hypothetical insured population total 1,200, and one plausible pre-event extrapolation predicts 1,000 deaths. The resulting excess is 200 and the P-score is 20%. A second plausible age-adjusted baseline predicts 1,100 deaths, giving 100 excess deaths and a P-score of approximately 9.09%. The observed deaths are identical. Differences are generated entirely by counterfactual construction. Reporting only the higher P-score as an empirical discovery conceals the sensitivity that matters for underwriting and capital assumptions.

Now suppose the underlying insured annual mortality probability remains 0.004, but insured exposure expands from 200,000 life-years to 250,000 life-years. Expected death counts increase from 800 to 1,000 while the mortality probability is unchanged. A raw count increase of 200 is therefore not evidence of excess mortality unless expected deaths account for exposure. Growth or aging of an insured portfolio can produce the same visible pattern that a poorly constructed baseline labels as a shock.

Worked example: population transport and benefit concentration

Consider an insurer with baseline expected deaths of 500 in a defined year. A population study suggests 15% excess mortality, but the insured cohort may differ in age, occupation, employment and preexisting health selection. Using 15% as a stress assumption yields 575 expected deaths; it does not establish that 75 actual additional insured deaths occurred. Suppose instead there are 600 observed insured deaths and an independently estimated insured no-shock baseline of 540. The insured excess is 60 deaths, or about 11.11%, relative to that baseline. The insured and population figures can both be internally correct because their populations and counterfactuals differ.

If the 60 additional insured deaths carry a mean benefit of $250,000 and the 540 baseline deaths carry a mean benefit of $100,000, the expected or observed increment in amounts requires a carefully defined comparison; the simple product 60 × $250,000 = $15 million describes the extra-death component under an assumption about those 60 deaths, but does not equal total actual-to-expected financial loss unless the distribution of claim amounts in the counterfactual and realized portfolio is reconciled. In particular, changes in severity among all 600 claims can change the aggregate monetary deviation independently of extra death count. This nuance matters for interpreting claim reports from insurers with shifting benefit distributions.

Reproducibility and arithmetic audit

The examples were recalculated using direct arithmetic, with no reliance on proprietary datasets. In the two-stratum example, additional deaths equal 80,000(0.003−0.002) + 20,000(0.005−0.004) = 100. Additional benefits equal 80,000 × 100,000 × 0.001 + 20,000 × 500,000 × 0.001 = $18,000,000. For the age-adjusted baseline example, (1,200−1,100)/1,100 ≈ 0.090909. For the insured example, (600−540)/540 = 0.111111. These are deterministic illustrative checks, not inferential tests, fitted parameters or estimates from insurer records.

An empirical replication package would need machine-readable count tables, exposure denominators, code lists, source-vintage metadata, baseline specification, reproducible estimation code and revision history. Confidential insured microdata may require restricted access; disclosure-safe stratified tables can still support independent comparison if denominator and suppression rules are published. A public chart without documented denominator, selection design or reporting completeness supports descriptive context only.

Robustness, falsification and source limitations

The leading challenge in causal inference is that counterfactual mortality is unobserved. Before claiming an association, compare several plausible baselines; evaluate age-standardized and crude results; inspect cause coding during and before the event; test reporting-delay revisions; and conduct placebo forecasts on pre-event periods. Seek external variation where appropriate, while acknowledging its own selection and confounding risks. If a finding changes sign across reasonable baseline choices or depends on one delayed cause category, it requires an interval or qualification rather than a confident unique number.

A second challenge is ecological inference. Different regional excess-mortality rates may coincide with differences in social deprivation, health services, diagnostic systems and exposure, and do not identify an insured person's cause of death. Selection into insurance may depend on the same variables. Stratification reduces some compositional mismatch while potentially introducing sparse-cell variance or selection bias. The scientific response is to document which contrasts are descriptive, predictive or causally identified, and to avoid converting adjusted associations into medical-certification claims.

A third challenge is aggregation. Age-standardized population rates may be suitable for comparing trends, but insurer cash loss requires the actual benefit-weighted risk mix. A portfolio average can appear stable while a small, high-face-amount subportfolio experiences extreme exposure. The reverse can also occur: high excess death counts may concentrate in relatively low-benefit certificates. Reconciliation by cause, age, geography, occupation, policy form, duration and amount-at-risk is analytically superior, provided that the resulting small-cell confidentiality and uncertainty are handled rigorously.

A fourth challenge is source revision. Global modeled excess estimates, provisional national counts and privately submitted insured-claim studies do not share a common update schedule. Version-stamped comparisons should retain the reported population, cut-off and publication date. It is unsound to combine a recent national rate with an older insured baseline without disclosing their differences, or to recalculate old numbers with new population denominators while presenting the result as if it were the original published estimate.

International and intertemporal interpretation

The WHO's global estimate is intended to provide broad coverage where national data completeness varies, while the HMD short-term data prioritize standardization in higher-quality registration systems. The WHO study in Nature uses statistical prediction where mortality observations are missing; the HMD data article emphasizes documented coverage limitations and revisions. These projects are complementary rather than interchangeable. The correct research design selects the source whose population, temporal resolution and uncertainty model fit the question.

Likewise, the EuroMOMO surveillance approach uses standardized signals and describes elevated-mortality thresholds in terms of z-scores over successive weeks. A high z-score is a signal relative to a model's expected variation; it is not a direct count of death benefits payable. U.S. vital-statistics technical guidance highlights differential cause reporting lags. The statistical object differs even before one introduces insured selection or policy terms.

Jurisdictional comparisons of life-insurance data need to reconcile contract form, underwriting, benefits, claim maturity, nationality and residence definitions, reinsurance and supervisory accounting. Solvency II regulations provide explicit European stress parameters; they cannot be generalized as universal prudential calibration. The IAIS Core Principles support internationally framed supervision but permit national implementation. Scientific comparability requires a common analytical dataset and methods; legal comparability additionally requires recognition of locally applicable statutory and contract rules.

Scientific synthesis and conclusions

Integrated findings

The literature and formal analysis support five interrelated findings. First, death certification defines what medical cause is reported, and the selected underlying cause depends on classification and decision rules. Second, all-cause excess mortality is the difference between observed deaths and a specified modeled counterfactual; it is not a direct causal allocation to one medical condition. Third, insured-life mortality must be estimated using insured exposures or transported from population data under transparent, testable conditions. Fourth, life-insurance monetary loss adds death-benefit weights and contractual eligibility to mortality frequency. Fifth, portfolio tail loss adds dependence across insured events, concentrations and stress-relevant financial structure. Each conclusion follows from an established research object and standard mathematical accounting rather than a newly asserted independent theory.

The WHO's mortality certification standards, its pandemic excess-mortality methods, the HMD short-term data series and the SOA insured claims study jointly substantiate why an insurer must link medical classification, population counterfactual, insured risk set and contract benefits. No one cited system contains all the variables needed to infer claims losses for every portfolio. Their connection is an analytical task that must preserve the provenance and limitations of each layer.

Decision-oriented research procedure

A complete study begins by freezing the risk set and target question: causes among insured deaths, expected excess insured deaths, aggregate death benefits, net insurer loss, or capital tail. It then records the mortality source, classification revision, observed and pending cause counts, occurrence and registration dates, claims lag and exposure basis. The analyst establishes a pre-event baseline using age-specific insured experience where available, compares alternative counterfactuals and makes the target shock period explicit. Cause-specific and all-cause totals are reconciled under a single coding system, with missing classifications and revisions retained as uncertainty. Insured estimates are stratified or appropriately standardized, and comparisons with general-population data are labeled as observations, assumptions or modeled transports according to their actual basis.

The next stage attaches insured amounts at risk, verifies policy eligibility and assesses dependence or common-cause accumulation. Results are presented separately for deaths, gross benefits, retained benefits and regulatory capital exposures. Scenario outputs should show how conclusions change under baseline and insured-selection alternatives, cause-misclassification assumptions and reporting lag. A final review identifies which decisions are robust, which rely on unverifiable counterfactuals, and which require new confidential data or legal analysis. This is an ordinary auditable research workflow synthesized from the literature, not an invented named protocol.

Research agenda and limitations

The strongest empirical opportunity is the reproducible linkage of incident mortality records to carefully documented insured-life cohorts while protecting personal and medical information. Research could estimate how changes in cause distributions differ between newly underwritten individual policies, older in-force policies and employer group cover. It could test whether a mortality shock's age, geographical and occupational gradients persist after adjustment for portfolio composition, and how changes in face amounts alter the divergence between excess deaths and excess death benefits. Methodological comparisons could quantify sensitivity to counterfactual windows, ICD transitions, cause-specific reporting lags and probabilistic linkage error. Such projects require microdata, data governance approvals and appropriate statistical designs; they have not been performed in this monograph.

The present work is a conceptual, methodological and critical synthesis with transparent illustrative calculations. It does not analyze insurer administrative microdata, reproduce proprietary SOA datasets, estimate a new national mortality series, perform a causal intervention study or compute live insurance solvency capital. National mortality source coverage, classification adoption, finalized versus provisional reporting and contract rules vary. The online research pages accessed for selected actuarial studies establish their declared coverage and methodology, not every unpublished table. A complete prospective calibration would need raw exposure, claims, death-benefit amounts, duration and cause-coded records and a fresh model validation exercise.

Conclusions

Cause-of-death analysis, excess-mortality estimation and life-insurance loss analysis are three connected but different inference problems. Their reliable integration depends on knowing what the medical record classifies, which deaths were observed, which deaths the counterfactual predicts, which lives were insured and which monetary benefits become payable. A cause code cannot substitute for an identified population counterfactual; a population excess-death rate cannot substitute for the insured experience of a selected portfolio; and a death count cannot substitute for amount-weighted insurer loss. The evidence examined here makes those distinctions operational through coding-quality controls, demographic adjustment, stress-specific assumptions, financial exposure reconciliation and explicit uncertainty.

The practical consequence is a more disciplined interpretation of mortality shocks. Analysts should report baseline sensitivity and cause uncertainty alongside central death estimates, validate selection and duration effects in insured cohorts, analyze benefit concentration and dependence, and distinguish gross from retained losses. Scientific conclusions become stronger when the source lineage and counterfactual assumptions remain inspectable. The proper endpoint of the research process is an auditable explanation of observed deaths and insured financial consequences under stated assumptions, with remaining causal questions reserved for designs capable of answering them.

References