Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Age-Period-Cohort Models for Insured Mortality

Abstract. Age-period-cohort (APC) models organize mortality experience by attained age, calendar period, and year of birth. Their distinctive mathematical problem is the exact linear dependence of these three clocks; the allocation of unrestricted linear trends among age, period, and cohort is not identifiable from observed mortality alone. Life insurance adds an important institutional layer: the insured risk set is affected by underwriting, issue age, policy duration, issue-year vintage, benefit amount, policy lapses, survival selection, and changes in companies or products. This monograph establishes the mathematical limits of APC inference, studies the observational conditions of insured-life data, and evaluates the projection and valuation consequences through original analytical synthesis and transparent constructed numerical examples. It integrates the demographic and epidemiological identifiability literature, the Lee–Carter and cohort-extended stochastic mortality literature, Human Mortality Database and United Nations documentation, Society of Actuaries insured-life experience studies, and the Continuous Mortality Investigation's forecasting disclosures. The analysis distinguishes invariant fitted death-rate contrasts from constraint-dependent parameter decompositions; population trends from selected insured experience; and historical estimates from forecast assumptions. A synthetic twenty-year scenario demonstrates how alternative mortality-improvement assumptions change expected gross life claims, while a separate survival scenario illustrates longevity exposure. The central finding is that actuarially usable APC forecasting requires coherent treatment of identifiability, the insured exposure-generating process, and explicit future dynamics. No proprietary insurer data are estimated, and no new named theory, doctrine, or model is proposed.

Keywords: Age-period-cohort analysis; insured mortality; birth cohort; calendar period; attained age; actuarial science; mortality improvement; Lee–Carter; underwriting selection; policy duration; exposure-to-risk; model identifiability; survival analysis; life insurance claims; stochastic mortality; forecast validation.

Introduction

Mortality statistics are conditioned by the people exposed, the deaths counted, the relevant clock, and the institution constructing the observation. Actuarial death rates divide deaths by exposure-to-risk, so a defensible rate presupposes a precise account of how policyholders enter and leave the risk set. Age-period-cohort analysis examines three interrelated coordinates of that experience. Attained age places an insured life on an individual life course; calendar time situates deaths within a historical environment; birth cohort indexes generations with potentially different cumulative exposures. These variables are linked by an arithmetic identity, creating an enduring identification problem. Holford's (1983) analysis of estimable APC effects supplies the formal point of departure.

The insurance application is more demanding than the basic population mortality table. A policy may be selected by medical underwriting, issued under a particular historical regime, renewed or terminated, converted into another product, or paid under a specified benefit schedule. The portfolio observed at one date contains survivors of distinct entry and exit processes. An apparent birth-cohort effect can reflect real generational mortality changes together with shifts in selection, duration, sum insured, product mix, and market participation. The Society of Actuaries' 2012–2019 individual-life study identifies changes in underwriting, policy size, age, duration, and the contributing companies as considerations in interpreting observed experience.

The research object is the estimation and projection of insured death rates by age, calendar period, and birth cohort, including necessary contract-level distinctions. It is bounded from general insurance demand theory, the complete theory of life contingencies, and comprehensive accounting or solvency treatment. The present inquiry examines those fields only where a mortality model's definition, calibration, or projection directly changes its interpretation or actuarial consequence. It considers international research but does not impose one jurisdiction's prescribed mortality tables on the rest of the world. The 2017 CSO actuarial tables are official evidence of U.S. table structure and risk classifications, not global estimates.

The method has four components. First, original peer-reviewed methodological and actuarial publications are checked against their accessible publisher abstracts and bibliographic records. Second, the publicly available methodology and institutional records of population and insurance studies establish exactly which people and dates the studies cover. Third, algebraic relationships and constructed financial sensitivity examples are independently derived. Fourth, the implications are synthesized through explicit attention to alternative explanations, data provenance, model identifiability, forecast design, and relative financial materiality. The monograph's original contribution consists of interpretation and integration of verified materials, rather than a claim to discover a new statistical identification solution. For some research articles only the accessible abstract and bibliographic metadata were checked; restricted full text and confidential insurer datasets are outside the verification scope.

Its principal questions are what a fitted age-period-cohort rate surface can identify; when birth-cohort movements represent selected insured experience rather than population generational mortality; how identification constraints interact with forecasting; and what sensitivity in mortality improvements implies for life claims and longevity-contingent liabilities. The answers require keeping the observed death-rate surface, its allocation among temporal effects, and its uncertain continuation into the future analytically distinct.

Historical, Statistical, and Documentary Foundations

Age, period, and cohort as three measurement coordinates

An age-period-cohort table arranges deaths and exposure by attained age aa and observation period tt, with birth cohort c=t−ac=t-a under a consistent time convention. Exact dates and completed ages complicate this identity at the edges of yearly cells. Someone recorded as age 55 in the calendar year 2025 may have been born in either of two adjacent years depending on birthday, observation date, and age-label definition. A valid design must describe these Lexis boundaries and their treatment rather than infer a perfectly measured year-of-birth variable from coarsely rounded age alone.

An age pattern summarizes how mortality differs along the life course. A period pattern follows a calendar-time influence across ages; a cohort pattern tracks a birth-generation signal across age and period cells. All three may describe the same recorded deaths, while their individual coefficient trajectories are not automatically causal mechanisms. Clayton and Schifflers (1987) examined the conceptual difficulty in age-period and age-cohort models and then the combined age-period-cohort specification. Their work remains central because visually intuitive interpretations of temporally patterned mortality can exceed what the statistical design uniquely supports.

A calendar shock can reflect disease, treatment, war, public health conditions, reporting changes, or other influences, and need not affect all ages proportionally. A birth-cohort pattern can be associated with cumulative exposures, generational social structure, smoking, occupational risk, or medical history. These are investigative hypotheses. A fitted APC diagonal is evidence of a pattern conditional on the statistical model, while the attribution of that pattern to a particular causal exposure requires additional measured information and a stronger identification design.

Death rates, person-time, and contract probabilities

For attained-age cell aa, calendar year tt, and insured classification zz, let Da,t,zD_{a,t,z} count deaths and Ea,t,zE_{a,t,z} denote insured person-years exposed. The central observed death rate is ma,t,z=Da,t,z/Ea,t,zm_{a,t,z}=D_{a,t,z}/E_{a,t,z}. In a standard aggregate likelihood, Da,t,z∼Poisson⁡(Ea,t,zμa,t,z)D_{a,t,z}\sim\operatorname{Poisson}(E_{a,t,z}\mu_{a,t,z}), and log⁡E\log E is an offset in a log-link model. The offset records the population actually at risk; if death counts and exposure cover different individuals or time intervals, accurate numerical optimization cannot rescue the resulting rate.

For a constant force of mortality μ\mu over one year, the corresponding death probability is q=1−exp⁡(−μ)q=1-\exp(-\mu). More generally, survival through an interval is exp⁡[−∫λ(u) du]\exp[-\int\lambda(u)\,du], so the annual death probability depends on the integrated hazard. With complete one-year contract observations, a binomial representation of deaths among lives at risk may be appropriate. With staggered entry, lapse, maturity, and delayed exit, person-time exposure or individual counting-process methods can be more faithful. A central death rate and a contractual qxq_x are therefore related but separate quantities.

The Human Mortality Database overview describes national vital-statistics death counts, annual population estimates, exposure-to-risk construction, and the conversion of estimated rates into life tables. Its methodology forms a reference for demographic coherence. It also explains why national population rates are not observations of privately insured, underwritten policyholders. A national death-registration denominator includes persons without life insurance, with different socioeconomic, geographic, medical, and financial characteristics.

A policyholder may hold several contracts, creating distinct denominators for persons insured, policies in force, face amount, and amount of insurance paid upon death. Policy-count weighted results and amount-weighted results are both potentially meaningful; they answer different questions. The problem intensifies when an insured dies holding multiple policies, when coverage matures before death, or when a claim is reported after the exposure period. Each death event, contract exposure, and financial payment must be assigned consistently to the selected metric.

Four clocks in a life insurance record

The insurer records at least six related temporal quantities: birth date, attained age, calendar time, age at issue, issue calendar date, and duration since issue. Let ii be issue calendar year and d=t−id=t-i the policy duration; then issue age a0=a−da_0=a-d when measured on consistent continuous-time scales. The identities c=t−ac=t-a, i=t−di=t-d, and a0=a−da_0=a-d constrain what independently interpretable temporal functions can be estimated. Issue cohort is the set of contracts initiated around a common date; birth cohort is the set of persons born around a common date. They are fundamentally different classifications even when one is strongly correlated with the other.

An underwriter's evidence requirements, preferred-class criteria, or smoking classifications can change over issue years. Subsequently, selection changes because healthy and unhealthy lives survive, surrender, convert, or lapse at different rates. A birth-cohort APC pattern observed at a common attained age may therefore incorporate changes in entry standards, historical product design, and continued exposure. The Society of Actuaries' old-age mortality investigation explicitly considered issue-age, issue-year cohort, smoker status, benefit band, and select versus ultimate experience. It illustrates why insurance studies should retain issue information alongside demographic birth information wherever possible.

Demographic and insurance source hierarchies

Different sources provide evidence for different estimands. Population mortality databases provide long time series and broad age coverage, usually without private life-insurance selection. The HMD limits its principal national series to populations with near-complete death registration and census information according to its scope and basic principles. The United Nations' World Population Prospects 2024 methodology uses empirical or model-based mortality estimation across countries according to the available data quality. These differences in measurement foundations must accompany any cross-national analysis.

The Society of Actuaries' 2015 Valuation Basic Tables provide U.S. insured mortality tables distinguished by sex, smoker status, age convention, and other characteristics. Their relationship to the 2017 Commissioners Standard Ordinary tables is partly a valuation and regulatory matter. Neither table is an unrestricted description of mortality everywhere. Actual-to-expected experience ratios using such bases are conditional on the chosen table, product mix, and contributing company data.

The SOA's 2019 individual-life experience report evaluates observed years 2012–2019 against a specified expected table. Its description also identifies the transition from the Medical Information Bureau to the National Association of Insurance Commissioners as statistical agent. A data-collection change may create discontinuity or altered composition independently of insured mortality itself. Tracking that history is part of the scientific source record.

Development of actuarial projection models

Lee and Carter's (1992) original mortality forecasting study models age-specific log death rates through age-dependent terms and a common period index, then forecasts that index as a time series. Its parsimonious structure established a major benchmark for demographic and actuarial projection. Renshaw and Haberman (2006) added age-specific cohort terms in a cohort-based Lee–Carter extension. These model families organize the age and calendar pattern of historical rates, while their prospective behavior depends on additional choices regarding the time processes.

Cairns, Blake, and Dowd (2006) studied a two-factor stochastic mortality model for older-age UK mortality, including parameter uncertainty and longevity risk. Plat (2009) later developed a cohort-inclusive stochastic specification with emphasis on features such as age range and dependence structure. Haberman and Renshaw (2011) carried out a comparative parametric mortality study using specified England and Wales and U.S. populations and age ranges. These contributions justify comparing alternative model structures; their reported populations do not establish one globally superior method for all insured portfolios.

The historical record supports a coherent research trajectory: move from descriptive cohort tabulations through explicit identifiability analysis to stochastic forecasting and target-specific model comparison. It also cautions against treating increased parameter count as evidence of improved out-of-sample reliability. A model can describe observed cohort curvature exceptionally well and still project a poorly supported future cohort path. This problem becomes consequential when the insurer's exposed lives occupy only a narrow portion of the age-period-cohort surface.

Identifiability Constraints and Estimable Mortality Structure

Formal decomposition and exact linear dependence

Consider the descriptive log-rate surface

ηa,t=log⁡μa,t=αa+βt+γt−a,\eta_{a,t}=\log\mu_{a,t}=\alpha_a+\beta_t+\gamma_{t-a},

where αa\alpha_a is an age term, βt\beta_t a calendar-period term, and γc\gamma_c a birth-cohort term. Let hh be any real number and define the parameter transformation

αa′=αa+ha,βt′=βt−ht,γc′=γc+hc.\alpha'_a=\alpha_a+ha,\qquad \beta'_t=\beta_t-ht,\qquad \gamma'_c=\gamma_c+hc.

Because c=t−ac=t-a, the transformed sum differs by h(a−t+c)=0h(a-t+c)=0. Every fitted ηa,t\eta_{a,t} and every fitted μa,t\mu_{a,t} remains identical. Moreover, additive constants can be distributed among the three functions as long as their sum is zero. This is an exact rank deficiency: there is no unique unrestricted attribution of the aggregate linear drift to age, period, and cohort. More observations of the same kind can increase precision for estimable quantities but cannot remove a structural identity.

Holford (1983) demonstrates the invariant estimable functions arising from the APC relation. Clayton and Schifflers (1987) further establish the interpretive difficulty of temporal trends. If a graph claims that one generation's mortality improved at a particular linear cohort slope independently of the other two terms, that claim needs identifying assumptions beyond deaths, ages, and years. A different equally valid assignment of linear drift can alter the displayed cohort slope while preserving every in-sample fitted death rate.

One can verify this invariance computationally without any insurer data. Select an arbitrary fitted set of α,β,γ\alpha,\beta,\gamma, apply the transformation for h=0.01h=0.01, h=−0.03h=-0.03, and another arbitrary value, and compare the implied η\eta cell by cell. The maximum absolute difference should be zero up to floating-point tolerance. This is a test of the statistical representation rather than a fitted empirical finding. It is independent of the number of deaths in the cells, the numerical optimization routine, and the portfolio's national market.

What remains estimable

A full age-by-period mortality table can support fitted death rates and contrasts that cancel the unidentifiable linear component. For evenly spaced time indices, within-component second differences such as Δ2βt=βt+1−2βt+βt−1\Delta^2\beta_t=\beta_{t+1}-2\beta_t+\beta_{t-1} and analogous cohort curvature contrasts are invariant to the linear shift above, subject to sufficient observation support and an appropriate statistical model. These contrasts detect deviations from straight-line patterns; they do not reveal causes merely because they are statistically invariant.

An exposure-weighted likelihood can identify the mortality surface within populated cells when the design, offset, and event ascertainment are adequate. Sparsity introduces practical identification problems beyond the APC algebra. A portfolio observed for only eight years over limited attained ages has little or no empirical support for projected rates of very old ages or for cohorts not yet observed at insured ages. When these rates are generated through penalties, extrapolation, constraints, or Bayesian priors, those devices supply substantive information. Their origin must be made visible.

A useful hierarchy of inferential objects is therefore:

  • Observed counts, insured person-time, and raw rates under documented case definitions.
  • Model-fitted rates and invariant contrasts over the observed age-period support.
  • Constraint-dependent allocations of additive and linear terms among the three temporal components.
  • Out-of-sample projections that additionally require future dynamics, insured-population composition, and other assumptions.

The hierarchy states the epistemic burden of the claims. The first two can be tested against held-out observational data. The third is meaningful only with its declared convention. The fourth is a conditional prediction, and cannot be made identical with a historical rate merely by citing a standard regression package.

Restrictions, smoothing, and robustness

Common identification strategies fix selected reference cells, center components, constrain one slope, or regularize relative smoothness. A sum-to-zero convention fixes some arbitrary levels but may fail to identify all linear trends; software output should be checked through a numerical rank computation or an equivalent symbolic argument. The condition imposed should be documented as part of the model, with an account of how changing it affects interpretable quantities. Robust scientific interpretation relies on contrasts and fitted rates that survive legitimate reparameterizations.

Regularization is valuable for stabilizing sparse insured rates. A penalty on second differences can suppress erratic cohort curvature, while a prior can communicate beliefs about future improvements. But the solution changes when the penalty, its weight, or a prior distribution changes. Stability to small changes in such choices is a validation object. The existence of a stable optimizer or narrow posterior interval does not prove that the underlying data separately measured all three temporal trends; an apparently precise posterior can partly reflect a strong assumption.

Forecast invariance requires special care. APC observations do not uniquely allocate unrestricted slopes, but two equivalent parameterizations can produce the same forecast if the full forecast operator is equivariant to their transformation. Divergence arises when separately extrapolated period and cohort processes, penalties, arbitrary zero-slope conventions, or additional projection assumptions are not equivariant. The correct inquiry is therefore whether equivalent representations lead to the same projected mortality surface under the stated forecasting algorithm, not whether the APC label alone makes every future rate unknowable.

Insurance time axes increase the structural problem

Life insurance introduces a distinct policy duration d=t−id=t-i, issue calendar year ii, and issue age a0=a−da_0=a-d. In continuous time these are identities. Consequently an unrestricted model simultaneously allocating separate smooth slopes to attained age, issue age, duration, calendar period, birth cohort, and issue year has additional linear dependencies. Entering more historical time coordinates into the regression cannot by itself remove the original APC aliasing; rather, it may add further rank deficiencies.

A practical working rate model can condition a supported APC surface on duration and observed risk strata:

Da,t,d,z∼Poisson⁡(Ea,t,d,zμa,t,d,z),log⁡μa,t,d,z=ηa,t+δd+θ⊤z+ψt,z.D_{a,t,d,z}\sim\operatorname{Poisson}(E_{a,t,d,z}\mu_{a,t,d,z}),\qquad \log\mu_{a,t,d,z}=\eta_{a,t}+\delta_d+\theta^\top z+\psi_{t,z}.

Here zz denotes specified underwriting or contract covariates, and ψt,z\psi_{t,z} is an interaction included only if data support it. The formula is an illustrative standard statistical specification, not an original named model. It does not license separate causal interpretation of every time coefficient. The design matrix should be checked for dependence after the insurance variables are included, with conventions documented for any omitted or combined effects.

Policy duration can reveal select-versus-ultimate mortality patterns associated with underwriting at issue and later survivor selection, but its apparent effect can itself depend on who lapses. At extreme ages, coverage maturing or terminating without appropriate exposure adjustment can create an artificial improvement aligned with a birth or issue cohort. The SOA old-age insurance study specifically examined older-age data quality and multiple issue-age and year classifications. This evidence supports a design requirement to audit the denominator before assigning demographic meaning to the mortality surface.

Falsification and interpretation boundaries

A properly skeptical analysis tests whether observed period spikes persist within underwriting and issue-duration strata; whether cohort curvature changes when policy issue-year groups are accounted for; whether age-band regrouping changes the result; and whether the youngest or oldest sparse cells disproportionately determine fitted slopes. A claim about a historical period shock should be compared with independent mortality or causes-of-death sources, paying attention to whether the insurer's claim dates and the public dataset's death dates coincide.

Potential falsification includes a raw-rate change disappearing after the same portfolio mix is held constant, a birth-cohort signal vanishing when issue-vintage composition is modeled, or a parameter conclusion reversing under an equivalent APC identifying convention. None proves the opposite causal hypothesis by itself, but each narrows what the original descriptive association could reasonably mean. The strongest statements are reserved for estimable, replicated quantities with compatible definitions; more speculative mechanisms are explicitly identified as hypotheses.

Mortality Cohort Dynamics in Insured Portfolios

Generations, historical environments, and insured entry

A birth cohort comprises persons with shared dates of birth. An insurance issue cohort comprises contracts entering coverage during a shared underwriting or product vintage. These cohorts can overlap but represent different processes. A generational mortality pattern may reflect cumulative environmental, occupational, behavioral, medical, or social exposures; an issue-cohort pattern may reflect the insurance market's acceptance rules, distribution channel, price, age at sale, and contract guarantees. A fitted diagonal in an age-period table can mix both.

Consider two hypothetical portfolios that observe mortality at attained age 55. The first contains policies issued ten years earlier following detailed underwriting; the second contains contracts newly issued under a simplified process. Even if each person belongs to the same birth cohort, the current observed death rates can differ because the selection mechanisms, policy durations, and survivor composition differ. Conversely, when comparing the age-55 experience of distinct birth cohorts over decades, the underwriting technology and product mix may have changed. The age-period-cohort label alone does not assign the resulting contrast to an exposure experienced since birth.

The SOA older-age insured research explicitly cross-classifies attained age, issue age, issue-year cohorts, benefit bands, smoker status, and select-versus-ultimate experience. The relevance is methodological: birth cohort must not silently substitute for issue year or duration. Where source data lack these dimensions, the limitation belongs in the causal interpretation and forecast sensitivity rather than being repaired through an elaborate smooth cohort curve.

Underwriting selection and changing survivor composition

Underwriting may select applicants on information available at purchase. Subsequent mortality in the in-force block reflects not only this initial risk distribution but also survival, lapse, surrender, nonrenewal, conversion, and maturity. A fully insured person-year denominator must account for the exact interval during which an individual contract is exposed to the covered event. If a policy matures at age 100, exposure in later age cells cannot be inferred from an outdated administrative in-force flag.

Let zz index disjoint insured risk strata, wz,a,tw_{z,a,t} their shares of insured person-time at attained age aa and calendar time tt, and μz,a,t\mu_{z,a,t} their stratum-specific death rates. The aggregate rate obeys the mixture identity

μa,tI=∑zwz,a,tμz,a,t,∑zwz,a,t=1.\mu^{I}_{a,t}=\sum_z w_{z,a,t}\mu_{z,a,t},\qquad \sum_z w_{z,a,t}=1.

Changes in the aggregate rate may be due to changing mortality within strata, changing exposure weights, or both. This identity offers a direct explanation for apparent period or cohort changes generated by business composition. A standardized counterfactual holding weights at a reference distribution can separate some composition movement from within-stratum rate movement, conditional on the quality of the available strata and without claiming causal identification.

Death rates can also be weighted by sums insured. Suppose one birth cohort holds larger death benefits on average while another holds smaller benefits. A face-weighted death claim measure then differs from a lives-weighted mortality measure even with equal death probabilities. An underwriter or risk manager should be clear whether the outcome is deaths per person-year, claims per policy-year, gross claims per unit sum insured, or a present value. The distinction has both statistical and financial implications and is especially important when cohorts differ by wealth, product, and age at purchase.

An actual-to-expected (A/E) ratio is a summary of deaths measured against a reference table:

A/E=∑a,t,zDa,t,z∑a,t,zEa,t,zma,t,zref.A/E=\frac{\sum_{a,t,z}D_{a,t,z}} {\sum_{a,t,z} E_{a,t,z}m^{\mathrm{ref}}_{a,t,z}}.

The denominator is a modeled expected count on the same exposure and rate basis. If the reference table, duration convention, expected-rate improvement, or portfolio mix changes, the ratio changes even when some underlying insured risk quantities do not. The SOA's documented VBT bases show why smoking classification and age conventions cannot be overlooked. A/E histories should retain the table version and the exact improvement factors, not just the headline ratio.

Pandemic period effects and cohort attribution

A calendar-time mortality shock can affect ages and insured groups heterogeneously. The Society of Actuaries' individual-life COVID mortality research evaluates insured experience through December 2022 and includes comparisons with general-population mortality. The official study description specifies a participating-company basis and the period of claims analyzed. These data offer direct evidence for the need to account for an exceptional period, but the contributions are not a complete census of all insured people or a universal causal experiment.

If a pandemic mortality spike is absorbed into an unconstrained period trajectory, a simple extrapolation may perpetuate temporary excess rates. If the spike is removed entirely as transitory, the projection may neglect persistent mortality effects, later changes in medical care, or possible selection changes. Cohort effects estimated across these years may themselves be altered by how the period shock is handled. An actuarial sensitivity study should therefore compare explicitly declared treatments of the exceptional period and examine their implications for cohort curvature and out-of-sample prediction.

The CMI's public CMI_2025 explanation distinguishes smoothed historically observed improvements from long-run user assumptions and documents materially different age patterns during and after the pandemic. These are England and Wales population observations, and their use as an insured benchmark must account for market selection. Their methodological relevance is the explicit division between observed recent fluctuations and choices about longer-term continuation.

Population-to-insured mortality transport

Public mortality time series offer a longer and often more complete view of age, period, and birth cohort than an individual insurance portfolio. Yet direct transport of a population trend into insured pricing or reserve assumptions requires an argument about comparability. Differences arise from underwriting, socioeconomic selection, wealth, occupational composition, age at acquisition, contract duration, geography, and product termination. The SOA's 2026 population-to-insured mortality research call expressly recognizes the practical difficulty of transferring improvements from general populations to selected insured groups. Since it is a request for research, it must be described as an institutional statement of a research problem rather than evidence of a completed numerical solution.

A descriptive comparison may define a compatible rate ratio ra,t,z=μa,t,zI/μa,tPr_{a,t,z}=\mu^I_{a,t,z}/\mu^P_{a,t}. This quantity can be useful where the national and insured data share age conventions and comparable death-period definitions. But a forecast of rr requires assumptions about the insurer's future acceptance and persistence processes, market mix, and any convergence with population risk. A stable historical ratio is evidence of historical association under its data definition; it is not proof that the ratio will remain stable across future underwriting technology or a new birth cohort.

International transfer demands a further qualification. The UN 2024 population methodology documents that the quality and type of national mortality evidence differ. A U.S. industry insured table, an England and Wales population improvement model, and a country with model-based national mortality estimates refer to distinct statistical systems. Their mortality differences cannot be explained by APC coefficients alone. A defensible comparison identifies the date range, country, age conventions, exposure type, contract class, data completeness, and uncertainty in each series.

Causal mechanisms and evidence limits

Generational patterns may motivate hypotheses about historical smoking exposure, occupational risk, nutrition, treatment availability, or other determinants. These explanations require external measurements, sensible timing, control for alternative mechanisms, and an examination of how people entered insured coverage. A cohort pattern observed in aggregate insurer records remains an ecological association when individual histories are unavailable. Cross-sectional improvements can similarly reflect a changing mixture of risk classes rather than equivalent improvement for each life.

Cohort statistical significance and financial materiality are distinct. A well-identified curvature feature in low-exposure, small-benefit cohorts may have little impact on gross claims, while a modest but persistent difference in a large high-benefit block could be material. The appropriate research report therefore links cohort contrasts to exposure support and to the insurer's financial target without converting an exploratory mortality association into a universal causal assertion.

Projection Sensitivity and Insurance-Liability Consequences

Fitting historical mortality and forecasting its continuation

A historical APC model estimates or summarizes death rates over an observed age–calendar surface. An actuarial projection extends the surface into years, ages, and cohorts for which observations are unavailable. The second operation requires assumptions about the dynamics of the period factor, the continuation of cohort patterns, the maturity of the insured portfolio, near-term shocks, and changes in policyholder mix. Good in-sample fit cannot supply the missing future observations. Comparative research by Haberman and Renshaw (2011) addresses model performance for life-expectancy and annuity-value targets in defined populations; it does not convert a retrospective fit into unconditional forecast truth.

A Lee–Carter representation may be written log⁡ma,t=αa+βaκt+εa,t\log m_{a,t}=\alpha_a+\beta_a\kappa_t+\varepsilon_{a,t}, with an estimated age profile, age sensitivities, and a period index. Forecasting κt\kappa_t requires a time-series assumption and uncertainty model. A Renshaw–Haberman cohort extension adds a cohort component, but then requires a justified extrapolation for this additional process as well. Additional parameters can capture features of historic data but also create new ways for unsupported future assumptions to dominate a long-horizon forecast. The Renshaw–Haberman cohort extension is relevant to model form, while Cairns–Blake–Dowd (2006) is relevant to stochastic factors and parameter uncertainty.

The projection should disclose the underlying observations, the vintage of data, the reporting and exposure conventions, the reference population, the model's age range, the identification restrictions, and the chosen future process. It should distinguish estimation uncertainty from stochastic process uncertainty, structural model risk, and uncertainty in future insured selection. Failure to separate these can make a narrow statistical confidence band appear to represent all meaningful future mortality uncertainty, when it covers only a particular fitted model conditional on its assumptions.

Transparent constructed scenario: insured death benefits

Consider a deliberately synthetic, fixed group of 100,000 insured lives each with a $500,000 death benefit. Suppose their initial one-year death probability is q0=0.002q_0=0.002. Hold the exposure count, underwriting profile, sum insured, and all other features fixed solely to isolate the influence of future annual improvement. Scenario A assumes a 1.5% proportional reduction in annual death probability; Scenario B assumes a 0.5% reduction. Twenty years later,

q20A=0.002(0.985)20≈0.0014782729,q20B=0.002(0.995)20≈0.0018092210.q_{20}^{A}=0.002(0.985)^{20}\approx0.0014782729, \qquad q_{20}^{B}=0.002(0.995)^{20}\approx0.0018092210.

The year-20 difference is about 33.0948 expected deaths per fixed 100,000 exposed insured lives. Multiplied by $500,000, the difference in that year's expected gross claim amounts is approximately $16.5474 million. This is arithmetic based on stipulated assumptions. It is not a fitted APC forecast, a measurement from an insurer, a present value, a reserve figure, or a capital estimate. It deliberately omits changing survival counts, discounting, lapses, mortality selection, and claim-payment timing.

Synthetic year-20 quantity Scenario A: 1.5% improvement Scenario B: 0.5% improvement
Initial annual death probability 0.002000 0.002000
Annual proportional mortality reduction 1.5% 0.5%
Year-20 annual death probability 0.001478273 0.001809221
Expected deaths at fixed exposure 100,000 147.83 180.92
Expected gross claims at $500,000 per death $73.91 million $90.46 million

The synthetic example demonstrates why an apparently small long-term change in a proportional mortality-improvement assumption may matter financially. Its purpose is to isolate model-assumption sensitivity. A real valuation would require the survival-adjusted number of insured lives reaching year 20 and the contract-specific benefit and premium paths. A sensitivity statement should therefore identify the held-constant dimensions and should not be mistaken for a historical empirical finding.

Separate survival illustration for longevity risk

For a second constructed example, assume a constant one-year death probability of 0.010 for twenty successive years. Survival is 0.9920≈0.81790.99^{20}\approx0.8179. If the annual probability is instead 0.009, survival is 0.99120≈0.83460.991^{20}\approx0.8346, approximately 1.6683 percentage points higher. This is not the same scenario as the declining death probabilities above, and it does not describe a specific real cohort. It isolates a long-horizon survival consequence. The higher survival probability generally increases expected contingent payments under a conventional life annuity; for life death-benefit protection the relevant cash-flow effect differs according to benefit design, coverage term, and survival of the block.

A stylized present value for death benefits is

Vdeath=∑s=1TvsBs s−1px qx+s−1,s,V^{\mathrm{death}}=\sum_{s=1}^{T} v_s B_s\, {}_{s-1}p_x\,q_{x+s-1,s},

where vsv_s denotes the discount factor, BsB_s the benefit payable in year ss, s−1px{}_{s-1}p_x the probability of survival to its start, and qx+s−1,sq_{x+s-1,s} the probability of death in that year conditional on survival. Any mortality improvement scenario must re-evaluate both the conditional death probabilities and the preceding survival probabilities. Applying a changed death rate to the benefit in one isolated year while retaining the original survival path is a different constructed approximation, not a complete liability valuation.

A conventional annuity value aggregates discounted benefits payable while the person remains alive. Mortality improvements can increase that value and thereby create longevity exposure for the insurer. For some life insurance products, the effect of changing mortality also interacts with premium receipt, reserves, guarantees, lapse options, and coverage duration. Financial significance therefore depends jointly on rate changes, contract terms, and portfolio composition, not on which APC coefficient has the greatest numerical magnitude.

Scenario construction, risk, and actuarial governance

A useful stress exercise varies specific dimensions: short-term period shock, gradual post-shock recovery, long-run period improvement, cohort persistence, underwriting selection, future issue mix, duration-sensitive lapse, and extreme-age treatment. It should also include coherent combined scenarios because a shock may affect mortality and policyholder behavior together. The estimated covariance of stochastic mortality processes is not itself a sufficient account of future behavioral and institutional uncertainty.

The Cairns–Blake–Dowd mortality-risk review situates mortality modeling among capital-market and hedging applications, whereas the Plat stochastic specification emphasizes model features such as cohort effects and age coverage. These sources support examining different forms of model risk. An exposure-weighted life-death-claim forecast and an annuity survival forecast should be evaluated against their separate economic targets rather than through a universal mortality-fit score.

The Continuous Mortality Investigation's CMI_2025 public FAQ makes explicit that its long-term improvement rate is a user assumption, not simply a continuation mechanically dictated by the recent observed data. The CMI_2025 working-paper record explains that the model's dataset was calibrated through 2025 and notes the importance of death registration versus death occurrence timing. These documentary distinctions are necessary when an actuary compares forecast vintages after recent mortality fluctuations.

The CMI_2026 consultation was issued in October 2026 as a proposal for the next model, not as a completed released CMI_2026. The CMI's official news record additionally documents corrections to selected tables in Working Paper 211. An accurate source ledger identifies the version and erratum relevant to the numerical claim. A working-paper table superseded by a correction is not independent confirmation of its original figure.

Prescribed tables versus estimated future experience

Regulatory mortality tables can impose minimum standards for defined contracts and jurisdictions. Their existence does not determine which APC model best estimates the deaths of an insurer's own risk set. For example, Washington Administrative Code § 284-74-535 specifies use of the 2017 CSO mortality table for affected policies, including a mandatory minimum standard for policies issued beginning January 1, 2020. This is an official rule of Washington State, not a global actuarial standard and not an empirical validation of an unconstrained cohort extrapolation.

A responsible valuation analysis separately documents the prescribed legal basis where relevant, the actuarial best-estimate assumptions, and any additional margins or stress conditions. Disagreements between a prescribed basis and an estimated portfolio trend can be informative about the functions of those bases; they do not establish that one of the two data sources is automatically mistaken.

Comparative Evidence, Validation, and International Transfer

The sources describe different scientific populations

A fundamental review must preserve the population and target of each major research source. The APC identifiability papers concern vital-rate or epidemiological models and establish algebraic and statistical limits applicable to many mortality datasets. Lee–Carter and its stochastic extensions describe model construction and forecasting, often calibrated to national population mortality. The Human Mortality Database and United Nations documents concern demographic data collection and estimation. SOA experience studies describe selected industry policyholders, and CMI materials explain a mortality improvement framework primarily calibrated to England and Wales population experience. These objects cannot be treated as if each paper had fitted the same sample of insured lives.

Evidence family Directly supported object Main inference limit
Holford; Clayton–Schifflers APC linear dependence, estimable functions, trend interpretation No identification of unique causal generational mechanisms
Lee–Carter; Renshaw–Haberman Period and cohort model structure and forecasting approaches Not universal insured forecasting performance
Cairns–Blake–Dowd; Plat; Haberman–Renshaw Alternative stochastic model attributes and comparisons Population, age-range, and forecast-target specificity
Human Mortality Database and UN National mortality data, exposure construction, international methods Not observations of underwritten insurance risk
SOA mortality tables and studies Selected U.S. insured-life classes, issue and duration structure Contributor composition, product scope, table conventions
CMI studies and methods UK mortality improvement assumptions and model vintages Not a directly portable global insured cohort surface
Jurisdictional insurance rule Mortality basis required for specified contracts Legally limited geography and issue dates

This documentary comparison gives each source the authority it actually possesses. The Haberman–Renshaw comparative study evaluates particular parametric models using defined populations and age ranges, but it cannot establish their rankings on a confidential U.S. life insurer's new business. The SOA's insured COVID analysis offers documented evidence of period sensitivity among participating insurers, but its observation window and sample do not establish a general historical cohort mechanism in all insurance markets.

Reproducible insured cohort data

An empirical APC application requires a data dictionary before estimation. It should document the insured entity (person, policy, certificate, or coverage unit), applicable product and eligibility restrictions, observation start and end, complete age convention, actual birth date or derived cohort, date of issue and issue age, underwriting and smoker status, face amount, premium or benefit features, and all recorded exit types. Death date, claim notice date, claim acceptance date, and payment date serve different statistical and financial questions. Their selection can change which calendar cell contains a death.

The data audit should detect impossible dates, deaths preceding policy entry, missing or conflicting birth dates, duplicate policies per person, overlaps after lapses or reinstatements, and periods after contractual maturity. Insurance death claims should be reconciled to administrative records without assuming that a delayed payment creates new mortality exposure. Life-year exposure and policy-year exposure should be separately named. A claim count based on policies can exceed a count based on persons when one individual holds more than one contract. These definitions belong in the scientific record because the same numerator/denominator error could otherwise create a cohort pattern that appears highly statistically significant.

The 2012–2019 SOA report documents a transition in the statistical agent supplying its underlying data. The lesson is general: a source change may alter apparent trends through coverage or coding. The old-age mortality report also illustrates how unusual extreme-age experience requires exposure and maturity checks. These official sources support data governance measures without requiring an unsupported assertion that a particular unknown portfolio contains a specific error.

Comparison of models and prediction targets

A defensible candidate comparison starts with parsimonious alternatives: age-only reference rates, a period-augmented model, a Lee–Carter-type specification, a constrained APC model, and a cohort-augmented alternative when birth-cohort evidence is sufficiently populated. It must state the constraints, smoothing, likelihood, ages, available cohorts, and calendar span for each model. A gain in in-sample log likelihood from adding flexible cohort terms can reflect genuine structure or overfitting, especially in thin diagonals near the boundaries of the observation triangle.

Forecast validation should use rolling origins that simulate the information available at each historical forecast date. The evaluator fixes a training cutoff, estimates the model, projects prespecified future periods, and compares the projected deaths or rates with subsequently observed outcomes. A second and third origin check stability. Using data revisions available only years later can accidentally grant the historical model information it did not have. Explicit claims-reporting delay and source-vintage controls avoid such look-ahead bias.

The score should match the objective. For count data, Poisson deviance or proper probabilistic log scores evaluate predicted deaths conditional on exposure and distributional assumptions. For central rates, exposure-weighted errors protect against giving tiny age-cohort cells the same influence as heavily exposed cells. For death benefits, sum-insured-weighted predicted claims may be closer to the financial objective. For annuities, survival probabilities and contingent liability values are necessary targets. A model selected using population deaths may not be the best for insured claims, and one selected for death counts may not be best for tail annuity cash flows.

Predictive interval coverage should be compared with nominal coverage, taking into account the number of holdout events, overlapping forecasts, and uncertainty model. Structural sensitivity is not exhausted by a confidence interval. An important separate test changes period-shock handling, long-run cohort persistence, age closure, and selection assumptions one by one, and then assesses whether the ranking of candidate forecasts is robust. Cairns, Blake, and Dowd's review supports the broader context of mortality risk modeling; Plat's stochastic proposal demonstrates that model features and correlation structures are purposeful design choices rather than neutral extrapolation defaults.

Cohort support, sparse cells, and high ages

Age-period-cohort tables typically have uneven observed cohort lengths. Older birth cohorts are seen at later ages but may be absent from the early observation years; younger cohorts have observations at young ages but none at advanced ages. For an insured portfolio, coverage begins only after issue, adding another selection and truncation layer. Some apparent cohort differences can therefore be driven by insufficient overlap rather than a genuine common temporal gradient. A study should publish a heatmap or equivalent table of person-time and deaths by age and period, with cohort diagonals clearly indicated.

When data are sparse, pooling across neighboring years or smoothing may stabilize rates but may also conceal a genuine abrupt change. Credibility adjustment is not synonymous with APC identification: a credible fitted rate may still rely on a constraint to allocate linear drift. Conversely, a sharply estimated cohort curvature can be irrelevant to an insurer's projected liability if the associated cohort contributes little exposed benefit. The required reporting includes the local person-time, observed deaths, uncertainty, and financial weight of each supported pattern.

Extreme-age modeling deserves separate scrutiny because insured policies may mature, records may truncate observations, and national population denominators become difficult to estimate. One can compare older-age experience with an appropriately matched population baseline, but a sudden divergence must trigger an exposure and age-validation audit before it is described as a biological effect. The SOA older-age investigation specifically addresses these kinds of concerns.

Cross-country inference and legal distinctions

The HMD's documented population coverage is based on strong registration and population data. The UN's international methods statement distinguishes countries requiring more model-based mortality estimation from those with suitable empirical vital-registration evidence. Neither dataset records all the features needed to reconstruct private life-insurance underwriting selection. An international comparison should first standardize definition, age and period coverage, and exposure before comparing model parameters.

The insurance market introduces further national and historical heterogeneity in product eligibility, underwriting evidence, benefit structure, group cover, maturity rules, regulatory minimum tables, and claims reporting. A U.S. life insurance table cannot be applied to another jurisdiction simply by matching the age label. Likewise, an England and Wales population improvement series is not automatically a valid selection-adjusted trend for a different country's insured population. The analyst should state the assumed bridge, test it on local insured experience where possible, and identify what remains a scenario.

National legal instruments warrant their own time and territorial qualification. The Washington 2017 CSO provision is evidence about legally specified policy issue dates and a minimum table in that jurisdiction. Its legal status should not be confused with evidence of a global historical APC pattern. Insurance regulation and empirical mortality estimation interact in valuation practice but are separate types of evidence.

Counterevidence and negative results

An APC hypothesis gains credibility when it survives attempts to explain the same data through age-specific period changes, issue cohort, underwriting mix, and administrative coding shifts. A statistically visible cohort trend that disappears after standardization is informative evidence that part of the original association was compositional. A period-effect interpretation that fails against independent temporal mortality records may be a model misspecification. A cohort augmentation that improves in-sample fit yet degrades rolling-origin forecast scores provides reason not to prefer it for that projection task.

Negative comparisons need to be archived with the positive results. An investigator should retain model failures, the full set of plausible candidate models, sensitivity assumptions, and the exact data versions. Selection of only the visually appealing cohort plot or the best-performing holdout window would conceal model uncertainty. The central scientific objective is credible conditional prediction and clearly bounded interpretation, not maximization of apparent cohort importance.

Integrated Scientific Findings and Research Practice

Structural finding: fitted rates and component slopes have different evidential standing

The first integrated finding is a mathematical distinction. An unrestricted age-period-cohort regression does not uniquely distinguish its three linear trends because c=t−ac=t-a. An arbitrary transfer of linear drift between age, period, and cohort leaves fitted log death rates unchanged. Consequently, an invariant fitted surface and supported nonlinear contrasts have a different evidential standing from the chosen slopes of separately plotted components. Research should report identification conventions, check design rank, and emphasize conclusions that survive equivalent parameterizations. This follows directly from the original Holford identifiability analysis and the combined APC interpretation examined by Clayton and Schifflers.

The algebra does not imply that every age or cohort pattern is uninterpretable. It specifies which features the observations cannot uniquely allocate among the three time coordinates and which contrasts may still be estimated. A period shock represented by a nonlinear departure can be statistically detectable; its characterization as an infection, medical intervention, administrative change, or other causal event demands separate evidence. A cohort curvature feature can likewise survive the linear transformation yet remain compatible with multiple causal histories.

A second distinction follows for forecasting. The mortality surface estimated from historic deaths may remain invariant under a reparameterization, while its continuation into future years may depend on model dynamics and constraint-dependent extrapolation. If the forecasting operator is designed to be equivariant, equivalent APC coordinates can yield identical forecasts. Divergence under alternative parameterizations reveals assumptions embedded in that operator and should trigger a reproducibility review. The appropriate audit therefore includes both a mathematical invariance test and an explicit projection sensitivity test.

Observational finding: an insured portfolio is a selected evolving population

The second integrated finding concerns the process that produces mortality observations. Life insurance risk sets are selected at issue and evolve through survival, changes in coverage, lapse, maturity, and portfolio turnover. A time series of insured death rates therefore combines within-stratum risk and the changing composition of exposed policyholders. The exposure-weighted mixture identity demonstrated above establishes how a population-wide apparent period change may arise even when stratum-specific rates are stable, or how a change within strata may be hidden by countervailing exposure shifts.

Birth cohort, issue-year cohort, and duration must be distinguished as explicit variables or acknowledged as missing dimensions. The SOA's old-age experience report and individual-life mortality report support this methodological judgment by documenting the relevance of issue age, historical data changes, risk classes, and contributor composition. Insured mortality should not be described merely as general-population mortality with a uniform discount factor when the observable selection process differs by age, issue year, and contract.

This observation also clarifies the limits of population-to-insured transfer. A national mortality series can provide an informative prior, benchmark, or scenario, particularly where the insured cohort has few deaths. It does not measure the future underwriting mix or benefit-weighted claims of the insurer. The SOA translation research call documents that institutions continue to regard this bridge as an active research problem. Any assumed convergence, ratio stability, or improvement differential requires direct justification and testing.

Projection finding: assumptions must be attached to their financial targets

The third integrated finding is that observed mortality trends, future mortality scenarios, and valuation amounts represent distinct inferential levels. A year-20 rate can be calculated from an assumed improvement parameter, but its meaning for claims depends on benefit amounts, survival to that year, persistence, and the contractual event definition. The constructed example established a gross difference of approximately $16.55 million under artificially fixed year-20 exposure; its financial interpretation ends at that synthetic calculation. A complete present-value calculation would require year-by-year risk-set survival, claim timing, premiums, discounting, and product options.

The same improved survival that can lower near-term death claims may increase expected annuity payments. Model comparison should therefore include the relevant economic object: lives-based death counts, sum-insured-weighted claims, survival rates, annuity value, or a properly defined reserve. The mortality-risk management review provides a broader actuarial context for treating uncertainty and financial risk jointly, while the CMI_2025 FAQ illustrates explicit governance of the historical versus assumed future improvement horizon.

A reproducible research record

An empirical application should archive the dataset extract date and revision, underlying insurer cohort and contract definitions, geography, exposure construction, death and claim timing rules, age conventions, and the treatment of censoring and duplicate coverages. It should preserve the exact cohort/period support, the number of deaths and person-years in each cell, the observed versus modeled cell coverage, and information about small or missing age groups. Where disclosure of individual insurer experience is restricted, aggregate tables, mathematical specifications, and reproducible scripts can be supplied at a privacy-preserving level.

The statistical record should include every design matrix or generating formula; the rank and constraint diagnostic; the likelihood, offsets, smoothing penalties and tuning procedure; parameter uncertainty; and the invariant quantities that carry the substantive conclusions. If a model relies on prior distributions or target mortality curves, their provenance and sensitivity must be recorded. The Lee–Carter source study and Renshaw–Haberman cohort extension identify distinct model structures, but reproducing a projection requires more than citing the model names: it requires the adopted parameterization and projection rule.

The evaluation record should identify the training interval, each held-out forecast origin, the future horizon, the exact score, the financial weighting, and the source versions known at each origin. It should explicitly report whether a model's supposed improvement persists after issue cohort, select duration, and insurance mix are treated consistently. Data versions and corrections are consequential: the CMI official news and corrections show why an institutional erratum can affect the interpretation of specific reported figures.

A structured substantive audit can proceed through successive questions. Is the death count defined for the same risk set as exposure? Are exact and integer-age time conventions reconciled? Does the chosen model have enough cohort support? Does its linear decomposition depend on a constraint? Does its claimed pattern persist under selection standardization? Does the projection operator preserve equivalence under valid parameter shifts? Was its forecast evaluated on realistic holdouts? Is its uncertainty interval conditional on clearly stated assumptions? Does its financial consequence use the correct benefit and survival paths? Is every material claim linked to a source that actually studies that population or mathematical proposition? Each question targets a specific avoidable inferential mistake.

Synthesis, counterevidence, and confidence

The most reliable conclusions of this dossier are the identifiability identity, the definitions of exposure and rate, and the arithmetic of the constructed scenarios. They are established through transparent derivation. The supported methodological conclusions regarding insured selection and the need for underwriting and duration controls are consistent with official experience-study documentation. Claims about which future cohorts will experience the greatest improvement, which country will achieve the lowest insured mortality, or which APC estimator will forecast best are deliberately left to future data-driven tests. No such result is implied by the present record.

Several forms of counterevidence would alter applied interpretations. An apparent generational difference could disappear after matching issue duration and risk class. A period shock could reflect changed claims reporting rather than changed mortality. A cohort-augmented model could fit the calibration sample better while scoring worse out of sample. An observed national improvement could fail to transfer to new insured policies after an underwriting transition. Reporting these possibilities is part of explaining the actual structure of the evidence.

The monograph's authorial analytical synthesis lies in joining APC mathematical invariance, the changing insured risk set, source-population comparability, forecast-process assumptions, and benefit-weighted financial consequences within one auditable research narrative. These are integrated scholarly interpretations of established demographic, statistical, and actuarial materials. They are not represented as a new original theory, model family, or independently verified empirical discovery.

Conclusion

Age-period-cohort analysis supplies a rigorous description of mortality patterns over the life course, historical calendar time, and birth generations. Its central mathematical limitation is exact: unrestricted age, period, and cohort linear trends are related through an identity, so their individual attribution depends on conventions and assumptions. A substantive conclusion should therefore be anchored in observed exposures, identifiable mortality quantities, and explicitly stated inferential restrictions.

In life insurance, the mortality surface also reflects which people obtained cover, how coverage was underwritten, the year and age of issue, survival since issue, continuation or termination of contracts, and the composition of insured benefit amounts. These are constitutive features of the insured dataset. They can change apparent period and birth-cohort patterns and govern the interpretation of any population-to-insured mortality adjustment.

Future mortality projections require additional assumptions about shocks, improvement dynamics, cohort continuation, risk selection, and valuation horizons. Comparison of multiple specifications under realistic historical holdouts and financially relevant targets is therefore central to responsible use. The two constructed sensitivities in this dossier demonstrate the dependence of claims and survival on stipulated assumptions, without being mistaken for empirical projections.

The research conclusion is that valid insured-mortality APC work consists of a demonstrably identified description of the observed mortality surface, a documented account of insurance risk-set formation, and a forecast whose extrapolations and financial consequences are explicitly tested. That combination permits transparent reasoning and practical actuarial use while preserving the difference between established statistical identities, documented historical evidence, and uncertain future paths.

References