RESEARCH DOSSIER
Life Tables and Survival Functions
Abstract. A life table translates observed or estimated mortality into an internally coherent sequence of conditional death probabilities, survival probabilities, survivors, deaths, person-years lived, and remaining life expectancy. Its mathematical identities are elementary; its interpretation requires scrutiny of population definition, exposure measurement, selection, observation, temporal reference, graduation, and the terminal age interval. This monograph reconstructs the relationship between deaths and person-years at risk; derives discrete and continuous survival functions and conventional complete and abridged life-table quantities; compares aggregate demographic construction with policyholder experience estimation and censored individual-level survival analysis; and differentiates period, cohort, projected, and contract-specific mortality. Official evidence from the Human Mortality Database, the United Nations, the World Health Organization, and U.S. vital statistics is read alongside actuarial valuation material and foundational statistical literature. A reproducible synthetic five-year calculation demonstrates survival recursion, the distinction between death rates and death probabilities, and the opposing mortality sensitivities of term assurance and life annuities. The central analytical result is that the transformation from observed events to a valid insurance valuation requires an explicit chain of definitions and assumptions. A mathematically consistent life table can remain economically unsuitable when its experience population, underwriting selection, calendar-year basis, or benefit timing differs from the insured obligation. The manuscript presents a critical integrative synthesis, mathematical derivations, and worked examples; it introduces no new actuarial theory and makes no claim to original mortality estimation from an external dataset.
Keywords: Life table; survival function; mortality; hazard; force of mortality; exposure to risk; death probability; central death rate; actuarial science; life insurance; annuities; Kaplan–Meier estimation; period life table; cohort life table; censoring; mortality experience; demographic statistics; longevity risk.
Introduction
Life tables are among the oldest enduring quantitative architectures in insurance and population science. Their central operation is to take mortality measured at ages and transform it into a hypothetical trajectory of a fixed initial number of lives. The result provides a transparent accounting of survival and death; it can also be used to calculate remaining lifetime, the expected duration of a benefit, and present values contingent on mortality. The distinction between the table as a statistical representation and its application to a particular contract governs every interpretation in this dossier.
The historical connection between observed deaths and life-contingent finance is documented in Edmond Halley's 1693 Breslau mortality and annuity study. The Royal Society archival description identifies its manuscript and the city mortality materials on which it drew. Early demographic tabulation and present-day insurance pricing share the objective of converting age-related survival into quantities with an explicit reference population, but modern methods must also account for age definitions, migration, selection, observational completeness, and temporal change.
A contemporary life table may describe a national population during a specified calendar year, a birth cohort across many decades, a class of underwritten insured lives, or a standardized regulatory mortality basis. These tables can share symbols while answering materially different questions. The Human Mortality Database overview explains how nationally registered deaths, births, census observations, and population estimates enter harmonized demographic calculations. The Society of Actuaries 2015 valuation tables instead exemplify a mortality-experience basis organized for insured-life valuation work, with separate dimensions such as smoker classification and age basis. Treating these two sources as interchangeable would substitute different selection mechanisms, denominators, and institutional objectives.
This investigation distinguishes mortality rate from death probability, survival at an exact age from survival within a bounded period, actual cohort experience from a period synthetic cohort, and observed experience from an assumed or projected valuation basis. It then follows the consequences into insurance: death benefits are generally sensitive to deaths occurring during coverage, while life annuities are sensitive to the duration of survival. The equations alone do not decide which mortality table a contract should use; that choice involves a population bridge, an appropriate projection horizon, and a valid financial and contractual basis.
The research procedure is documentary and mathematical. Primary or institutionally authoritative descriptions supply the factual methods and regulatory examples; statistical sources supply survival estimators and censoring assumptions; derivations and a synthetic numerical table test mathematical consistency. The 2025 HMD methods protocol and the U.S. 2024 national life tables provide distinct but complementary examples of data reconstruction and life-table production. Relevant institutional publications are cited in the body with short contextual links and recorded in full at the end.
Deaths and Person-Years at Risk
Defining the event, age, and population
A life-table death is the event of death during a specified interval in a specified population. That plain description contains several operational choices. Age may be measured at last birthday, nearest birthday, exact attained age, or duration since entry. A person may be observed as a resident within a national population or as an insured exposed under a policy. The calendar time of death, the entry time, and the point at which observation ends all affect the denominator. A change of any of these choices changes the estimand.
For population records, the first requirement is to align death counts and exposure populations. Let represent deaths observed between exact ages and during calendar time . Let represent the associated central exposure measured in person-years. The central mortality rate is
Deaths are counts and exposure is time lived or time at risk; therefore has units of deaths per person-year. The ratio is a rate, not a bounded probability. For a one-year age interval and a mortality rate of 0.01 per person-year, stating that the one-year death probability is exactly 0.01 silently adopts an approximation. The approximation may be small, but its assumptions and cumulative effects matter in actuarial calculations.
In demography, death registers may cover completed ages while the numerator is classified by year of event and the denominator derives from population estimates at fixed calendar dates. The HMD calculation workflow identifies six stages: births, deaths, population size, exposure-to-risk, death rates, and life tables. Its explanatory data notes distinguish one-year age and calendar formats from five-year groups and Lexis-triangle decompositions. These conventions support comparability, but each reported rate still belongs to its exact age-period cell.
Lexis time, person-time, and the exposure denominator
A Lexis diagram places age on one axis and calendar time on another; the life trajectory of an individual travels diagonally. A death in a calendar-year and completed-age square can belong to one of two neighboring birth cohorts. Distinguishing those cohorts becomes important when birth volumes change abruptly, migration is large, or age at death and age at exposure refer to different boundaries.
The HMD Methods Protocol, version 6, revised August 5, 2025, explicitly addresses both the estimation of infant and the use of monthly birth distributions to improve exposures. The latter change responds to distortions that can arise when births are assumed uniformly distributed across the year despite historical concentration in particular months. This example demonstrates a consequential fact: an apparently innocuous interpolation convention can appear in the final mortality table as a spurious cohort pattern.
In an individually observed life portfolio, person-time can be accumulated directly. For person with observed entry and exit , with time measured consistently on an attained-age or policy-duration scale, exposure in interval is the duration of overlap:
The total exposure is , after enforcing appropriate inclusion rules, death definitions, and observation windows. An observed death contributes to the numerator in the interval in which it occurs; a lapse or administrative end can contribute exposure up to exit without being counted as a death. This is a conventional construction, with limitations discussed below: policyholder lapses need not be statistically independent of future mortality.
A portfolio's exposure basis must also resolve multiple policies on the same person, currency amount versus life count, effective dates, and reinstatement. If the unit is policy-years rather than distinct life-years, the table describes policy exposure, and duplication must be considered when interpreting a life-level death probability. For an amount-weighted mortality experience study, the numerator and denominator should likewise be expressed on compatible amount-weighted scales. A weighting rule is part of the estimand, not merely a reporting preference.
Rate, probability, and the within-interval assumption
Let be the probability of dying before age , conditional on being alive at exact age . Let be the mean number of years lived in that interval by those dying there. In a closed -year interval, the standard identity
connects the survivors , interval deaths , and person-years lived . Since , algebra gives
This conversion assumes the specified faithfully represents where deaths occur inside the interval. It also presumes that the rate and probability refer to the same closed age interval and coherent life-table representation. Under the uniform distribution of deaths over a one-year interval, and . Under constant force , the exact one-year death probability is , while in the ideal stationary exponential survival interval; these two conversions yield slightly different numerical results because they encode different within-interval survival shapes.
Infant mortality is a conspicuous exception to casual uniform-death approximations. The average time lived by infants dying before their first birthday depends on neonatal and postneonatal mortality composition. The 2025 HMD infant exposure revision is especially relevant to analysts who apply an adult-age convention at age zero. Proper treatment requires an empirically appropriate infant separation factor or a documented estimation procedure.
Exposure under migration, selection, and missing observations
A life-table rate requires consistent population accounting. In population data, migration affects the number of persons exposed and may affect the age composition of survivors. In insurance, people enter after underwriting, change coverage, lapse, surrender, or transfer between policy states. A population table conditioned on birth and a portfolio table conditioned on underwriting acceptance therefore represent different selection histories. Even where their ages and calendar dates coincide, their death rates need not.
If death registration is incomplete, the numerator is biased downward relative to the correct exposure unless an explicit adjustment is made. If the population denominator is underestimated, the rate can be biased upward. If an error varies by age, it also changes the curvature of the survival function and the inferred remaining lifetime. The UN World Population Prospects methodology distinguishes empirically supported mortality estimation from model-based reconstruction when data are insufficient. An apparently complete numerical table can therefore contain unobserved, modeled age segments. Those segments must be understood before the table is treated as direct evidence.
The analytical consequence is straightforward: the event definition and exposure definition determine what the rate means. Only after those definitions are established can a mortality rate be converted into a survival probability suitable for a life table.
Life-Table Notation
The life table as a system of linked quantities
A complete discrete life table is often written for one-year age intervals. It begins with a radix , a hypothetical initial number of survivors, customarily 100,000 in many demographic tables. Let be survivors to exact age , the probability of death before the next birthday, and the probability of surviving one further year. The recurrence is
No empirical population must ever have contained exactly 100,000 individuals: the radix is a scaling convention. Dividing , , , and by a common positive factor does not change , survival ratios, or life expectancy. This fact allows the same underlying probability distribution to be represented on several scales.
For a closed one-year interval, is the person-years lived by the synthetic cohort from exact age to . With uniformly distributed deaths, . More generally, . Total remaining life-years at age are
where the upper limit includes the terminal open interval or a specified, coherent extrapolation. Remaining life expectancy is when . It is measured in years and is an expectation, not an age everyone is forecast to attain.
The CDC 2024 life-table report provides practical implementations of these functions. Its period tables are estimated from final 2024 death registrations and population estimates, together with Medicare observations for specified older-age groups and modeled adjustments where data are inadequate. Its life-table construction shows why the terminal interval, old-age smoothing, and source quality matter even after the familiar recurrence equations have been written down.
Conditional and multi-year survival
For successive one-year probabilities,
Here is remaining lifetime conditional on life at age . The probability of survival for years and then death during the following years is . This identity is important in deferred assurance: survival to the deferral date and death in a later covered interval are both required for payment.
The age-specific survival curve relative to is . It satisfies , is nonincreasing and right-continuous for a continuous-time lifetime distribution under the usual survival convention, and takes values between zero and one. For integer , . A national life-table column is consequently a survival function measured from birth under that table's assumptions.
The probability of a life aged 60 surviving five years is determined by the mortality assumptions for ages 60 through 64, in the appropriate calendar-year projection. It is not automatically determined by a single observed value . Conversely, an annual death probability at age 60 cannot be recovered from life expectancy at 60 alone without additional information about the mortality distribution. Multiple survival curves can share the same expected remaining lifetime while assigning very different probabilities to early and late death.
Continuous-time hazard and integrated mortality
For an absolutely continuous remaining-lifetime distribution, the instantaneous force of mortality is
With , , so the cumulative hazard yields
The force of mortality has dimension inverse time; the cumulative hazard is dimensionless; survival is a probability. A positive hazard is not constrained to be less than one. Confusing hazard with annual probability is especially problematic when exposure intervals become wider or hazard rates become large at very old ages.
Life expectancy under a continuous survival curve satisfies the tail-integral identity
provided the expectation is finite. This gives a direct conceptual connection between the area under a survival curve and the conventional remaining life-years in a discrete table.
Internal consistency tests
A constructed table should pass a set of arithmetic constraints. Every closed interval must have , , , and for interval width . When the table closes with all hypothetical lives eventually dying, . Remaining person-years satisfy for unit intervals, with . Survivor probabilities must multiply consistently across adjacent intervals.
Passing these equations is necessary but does not establish data validity. A table built from the wrong insured class may pass every recurrence while still being unfit for pricing. Mathematical QA and population representativeness are distinct requirements, and both are necessary in applied actuarial work.
Survival and Death Probabilities
Direct estimation from fully observed lives
If a closed cohort of individuals is observed completely from one exact age to the next, with deaths and no lost observations, the simple estimator is . Under an independent, identically distributed binomial approximation with common , its estimated standard error is . This formula does not apply unchanged when the underlying mortality probabilities differ materially across insured classes, individuals contribute correlated policies, or some lives have incomplete observation.
With full person-time but events occurring at different times, a central rate can instead be estimated as . The conversion to must then identify a within-interval model. A Poisson counting approximation commonly gives when the relevant conditional intensity and exposure assumptions are satisfied. Its uncertainty is associated with estimating a rate from events and person-time, not with the abstract certainty of a table's algebra.
Right censoring, delayed entry, and Kaplan–Meier survival
Insurer experience files rarely observe every insured life until death. A policyholder may leave the portfolio, be lost to follow-up, or remain alive when the administrative observation period ends. A record then includes the time observed and an indicator of whether death was observed. Right censoring conveys that the death time exceeds the observed duration; it should not be treated as a death or as proof of indefinite survival.
Kaplan and Meier's 1958 product-limit estimation method constructs survival from observed event times and risk sets. Let be ordered distinct death times, the number at risk immediately before time , and the deaths at that time. The estimator is
A right-censored individual contributes to the risk set while observed and then exits it at censoring. The estimator requires that censoring permit valid inference about the event-time distribution, conventionally expressed by an appropriate independence or conditional independence condition. Censoring induced by serious illness, lapse tied to health, or selective withdrawal may invalidate a naïve estimate. The Klein–Moeschberger treatment distinguishes right censoring from left truncation, in which a person enters an observable risk set only after surviving to an eligibility or enrollment time.
Insurance underwriting makes left truncation especially important. An insured first entering at age 45 cannot contribute to the risk set at age 20, even though the person's eventual policy file might contain a date of birth. Correct analysis conditions on survival and entry eligibility; otherwise survival may be overstated by crediting unobserved time before enrollment as guaranteed survival time.
The approximate Greenwood variance of the product-limit estimate is
at points where the denominators are positive and the usual large-sample conditions apply. Near the last observed failure time this approximation can be unstable. Confidence intervals are uncertainty statements about estimated survival, not bounds on the actual lifetime of a given insured person.
Cumulative hazards and regression adjustments
An alternative nonparametric summary is the cumulative-hazard estimator , with survival often approximated by . Aalen's 1978 counting-process inference framework grounds cumulative-intensity estimation when subjects enter and leave observable risk sets. The Nelson–Aalen approximation and the Kaplan–Meier product are closely related at small event fractions, but their finite-sample forms differ.
When analysts seek to control covariates rather than merely tabulate a marginal curve, Cox's 1972 regression models and life-tables describe a multiplicative covariate effect on an unspecified baseline hazard. Such a model can accommodate age, sex, attained duration, or medically relevant risk characteristics, subject to the data, validity of the proportionality assumption, and proper handling of censoring. It does not automatically create a contract-approved mortality table; the resulting survival predictions still require calibration, group definition, documentation, and suitable valuation governance.
Mortality selection and heterogeneity
An insurer's insured population is selected at underwriting and subsequently by continued policy participation. When low-risk lives remain longer in a risk pool, the composition of survivors can change with attained age or policy duration. A mixture of different hazard levels can exhibit a declining average hazard relative to the hazards of its constituent classes, because high-hazard individuals tend to leave the surviving population earlier. Such patterns are relevant to select-and-ultimate mortality tables and to the difference between entry-age and attained-age analyses.
For example, if individuals have positive unobserved multipliers on a common baseline cumulative hazard , aggregate survival is . Aggregating before calculating the hazard yields a survival-weighted mixture of individual hazards, which can evolve as the surviving mix changes. This elementary calculation explains an empirical concern without introducing a new theory or establishing a universal direction for every portfolio. The relevance of selection depends on underwriting, lapse, class definitions and follow-up.
Competing exits and attribution
A death from any cause is the usual decrement of a basic all-cause life table. A portfolio may also contain lapses, disability incidence, recovery, retirement, and contractual termination. Those non-death transitions require separate definitions. If death and lapse are treated as competing events, the observed probability of a death while a policy remains in force differs from the hypothetical probability of death in a world where no lapse occurs. Likewise, a cause-specific death rate and the probability of death from that cause are not interchangeable without accounting for competing causes.
The WHO global life-table series reports the core all-cause functions . This institutional classification is useful precisely because the basic table's decrement is death from all causes; other benefit triggers and multi-state insurance problems require extensions beyond the scope of the elementary single-decrement construction.
Complete and Abridged Tables
One-year and grouped-age construction
A complete life table typically reports one-year age intervals; an abridged table may use the infant interval, ages one through four, and subsequent five-year intervals. The central identities survive the change in interval width, but the assumption about where deaths occur inside the interval becomes more consequential.
For a five-year closed interval, , , and . If is justified by a uniform distribution of deaths within the interval, . This is a convention, not a universal empirical law. Abridged infancy and older-age groups frequently require special estimation or closure methods.
The HMD data format documentation identifies 1×1, 1×5, 5×1, 5×5 and other age-period resolutions. Higher temporal resolution makes changes visible but also increases sparse-cell variability. Grouping reduces noise and improves coverage of small populations, while hiding differences within the group. A comparison between abridged tables is only valid after the boundaries, interval widths, sexes, territories, time periods and terminal age treatment have been harmonized.
An open terminal interval is written . Its eventual death probability is one if all survivors ultimately die, but that fact alone does not determine the number of person-years lived beyond . Under a constant terminal force , ; if a published method instead models the old-age hazard, it produces a different terminal person-year total. Setting and treating the entire surviving radix as dying instantly at the start of the open interval would incorrectly discard all subsequent lifetime. Old-age closure therefore affects and at every preceding age.
Reproducible synthetic five-year life table
Consider a fully hypothetical population of 100,000 lives surviving to exact age 70. Assume five successive one-year death probabilities of 0.010, 0.015, 0.022, 0.030 and 0.040. They are demonstration inputs, not mortality observations, institutional recommendations or estimates for any jurisdiction or underwriting class. Apply , , and , using an illustrative uniform-deaths convention.
| Exact age | Annual | ||||
|---|---|---|---|---|---|
| 70 | 0.010 | 100000.000 | 1000.000 | 99000.000 | 99500.000 |
| 71 | 0.015 | 99000.000 | 1485.000 | 97515.000 | 98257.500 |
| 72 | 0.022 | 97515.000 | 2145.330 | 95369.670 | 96442.335 |
| 73 | 0.030 | 95369.670 | 2861.090 | 92508.580 | 93939.125 |
| 74 | 0.040 | 92508.580 | 3700.343 | 88808.237 | 90658.408 |
All table quantities except are expressed in hypothetical lives or person-years on the radix of 100,000; fractions arise from expected-value arithmetic and have not been rounded into fictitious observed people. The five-year survival probability is
Thus . The survivor count at exact age 75 is , equal to the terminal survivor value generated by the recurrence. The five-year survival probability and the displayed row totals are independent arithmetic checks on the table.
This five-row table cannot determine remaining life expectancy at age 70 because no mortality assumptions have been supplied after age 75. Taking computes expected person-years lived within the first five years, a restricted mean survival time, not unrestricted . This distinction prevents a common truncation error in small illustrative or incomplete empirical tables.
From a death rate to a probability: a separate check
Suppose a one-year central death rate is 0.010 per person-year. With uniform distribution of deaths, . Under constant force of mortality of 0.010 per year, . Both are approximately 0.995%, but their distinction is methodologically real. Larger rates, broader intervals, and structured infant death timing generate larger discrepancies. It follows that a table should publish the interpolation or hazard assumption that links rate to probability.
Complete versus abridged sources in practice
The 2024 U.S. period life-table report constructs complete tables using final registered deaths, population estimates and supplemental older-age data; the method differs across some population groups because the data sources permit different degrees of direct observation. Meanwhile, the WHO global life-table methodology seeks country and international comparability from multiple national and modeled data sources. The UN 2024 population methodology similarly uses empirical and model-based mortality approaches depending on data availability and historical coverage.
These institutional tables may be compared to understand methods, but their estimates should not be mixed mechanically row by row. A national rate from one institution and an infant separation factor from a different institution may have been jointly calibrated to different mortality levels and source populations. A reproducible comparison records the edition, age definitions, all-cause scope, population boundaries, calendar periods and revisions before interpreting numerical differences.
Cohort and Period Tables
The period synthetic cohort
A period table uses age-specific mortality rates observed or estimated for a particular calendar interval, then applies them simultaneously across all ages to a hypothetical life-table cohort. It answers a counterfactual question: what survival sequence would result if a cohort experienced that period's age schedule throughout life? The different ages in one period table typically correspond to different actual birth cohorts.
The 2024 CDC life-table introduction distinguishes cohort and period tables explicitly. Its published 2024 period values show life expectancy at birth of 79.0 years overall, up from 78.4 in 2023. These values describe the published period schedule and its underlying methodology; they do not imply that every newborn in 2024 is expected to live exactly 79.0 years under the mortality rates that will actually prevail across the newborn's life.
Because a period table holds mortality conditions fixed across an artificial lifetime, it is particularly useful for describing contemporaneous mortality conditions, historical change and standardized comparisons. Its future-life interpretation is conditional on the period schedule being sustained. In times of epidemics, wars, abrupt mortality improvement or data revisions, that assumption can be implausible as a forward forecast.
The actual birth cohort
A cohort table follows individuals born in a given year through successive ages and calendar years. Its death probability at age 70 may occur seventy years after its death probability at birth. The HMD cohort data notes explain that cohort death rates require sufficiently long data records and that complete cohort life tables are available where cohorts have been observed to extinction or near-extinction under documented rules. Recent birth cohorts inherently lack observed old-age outcomes and therefore require projection for lifetime valuation.
A cohort lifetime distribution should be written with the time dimension visible. Let denote the death probability during age interval and calendar year . For an individual aged in year , future survival over years under a projected mortality surface is
A period synthetic survival probability instead applies a single year's schedule for every future age. The difference is the diagonal travel of an actual life through the age-calendar mortality surface.
Mortality improvement and projection risk
Insurance liabilities extend over years or decades. A contemporaneous observed table is a starting point; using it for long-term liability valuation requires a choice about future improvement, deterioration, shock scenarios, longevity trends and uncertainty. Different contract types have opposite exposures: a decline in future mortality generally reduces expected pure death-claim amounts for many life assurance contracts while extending expected annuity benefit periods.
The Society of Actuaries pension mortality resources list distinct retirement mortality tables and mortality improvement scales. Their separation illustrates that current mortality level and future improvement are related but distinct inputs. A historical population curve, a pensioner mortality curve, and a life-insured underwriting curve should not share improvement assumptions without evaluation of population differences and the intended valuation standard.
An improvement adjustment may be represented schematically by where is a specified proportional improvement factor. This is an illustrative transformation, not a complete projection method. It requires the projected rates to remain between zero and one, the full surface to be coherent, and the assumptions to be validated against relevant experience. Substantial mortality shocks, migration, product selection or structural changes can defeat a naïve extrapolation.
International comparison and revision dependence
The World Population Prospects 2024 methodology explains that data support differs among countries, with empirical estimation for sufficiently reliable mortality records and model-based approaches where observations are weaker. The WHO global health estimates combine national registration, specialized technical programs and partner estimates into internationally comparable series. These systems are valuable for international analysis, but their tables are estimates governed by specific revisions rather than raw observations of a single uniform kind.
Consequently, changing an institution's source data, model version or population boundary can alter historical estimated life expectancy even for a year whose actual deaths occurred long ago. An analyst must identify the release used and should compare like-for-like revisions. This is especially important when an insurance study uses historical mortality curves to calibrate assumptions or to claim a change in life expectancy. Observed mortality changes and estimation-method changes can both contribute to apparent time-series differences.
The HMD protocol update in 2025, including improved newborn exposures and infant-age treatment, is a concrete case of methodological revision. It strengthens research comparability in its intended datasets but means numerical series from different protocol versions need provenance before being pooled.
Sex, age, occupation, and insured selection
Mortality is heterogeneous across demographic and insured groups. Segment-specific tables can inform a scientific comparison when the grouping is explicitly defined and reliably observed. Insurance categories additionally incorporate underwriting and policy-duration effects. The 2015 VBT insured-life material distinguishes sex and smoking classifications and alternative age conventions. This technical organization is evidence of distinct statistical bases, not proof that the same segmentation is permitted or desirable for every legal or commercial use.
Population subgroup disparities must also be interpreted through registration, exposure, classification and access to insurance. A socioeconomic mortality association observed in a national dataset cannot be assumed to have the same magnitude among underwritten policyholders. A portfolio-specific survival function must account for selection into the insurance product and persistence after enrollment.
Data Quality and Insurance Uses
Data provenance and reproducibility
An evidentially credible life table should carry enough metadata for another researcher to reconstruct what it represents. Core fields include geographical and population coverage; attained-age convention; calendar or birth cohort; start and end of observation; death/event definition; numerator and person-time sources; completeness and revision date; exposure adjustments; censoring and truncation rules; smoothing and graduation method; open-age closure; uncertainty estimates; and any mortality-improvement projection. These fields explain the table's meaning and practical use.
The HMD explanatory notes and input files document such distinctions, including age-period formats, input data, and population or territorial adjustments. The CDC 2024 data methods identify final vital-registration deaths, population denominators and older-age Medicare sources. The WHO life-table resource identifies rate, probability, survivor, deaths, person-years and expectancy series by age. These primary materials provide complementary models of source transparency.
For small datasets, an arithmetic check may detect implementation errors before any substantive modeling review. Reconstruct from , compare it with the input , sum deaths, reproduce person-years under the stated interval model and confirm that . Repeat the comparison after changing the radix; probabilities and life expectancies should remain invariant. For grouped ages, test and total person-years across a partition. For a calendar-cohort surface, verify that the projection reads the correct diagonal cells.
Selection and mortality experience in insurance
An insurer's relevant exposure is a set of lives or benefits with particular underwriting characteristics, policy terms, risk classifications, durations and observation windows. A national mortality table is not automatically an unbiased estimate of mortality for that set. Insured lives may be healthier at entry than the general population, and subsequent lapses may change the remaining mix. The simple identity remains useful, but it estimates the rate of the actual population at risk during its observation period; the analyst must justify extrapolation to the liability population.
The SOA 2015 Valuation Basic Tables and 2017 CSO tables illustrate two different institutional products. Valuation basic experience and commissioners' standard mortality tables are not automatically identical. U.S. regulations impose specific bases for specified contracts and dates: IRS Notice 2016–63 explains transition rules for the 2017 CSO mortality tables in its federal tax context, while Washington's CSO mortality regulation addresses a minimum valuation standard for qualifying policies issued in that state. Neither document is a general international mandate; each illustrates why legal applicability must be established separately from statistical fit.
Mortality selection can also depend on the measurement unit. An experience study weighted by death benefit amounts can differ from a study weighted equally by distinct insured persons. Large-benefit policyholders can materially shift the amount-weighted result. An insurer using such data must document whether its intended valuation concerns number of lives, insured amounts, or the risk distribution of cash flows.
A transparent five-year assurance calculation
Return to the hypothetical ages 70–74 with . Suppose a term assurance pays 100,000 currency units at the end of the policy year of death, if death occurs during the next five years. For a purely illustrative effective annual discount rate of 3%, let . The expected present value of the death benefit is
This is the discounted expected death benefit under the specified synthetic probabilities. It excludes expenses, commissions, taxes, capital charges, margins, non-death decrements, adverse selection, mortality improvement and contract-specific provisions. Consequently it is a net actuarial illustration rather than a premium quote or a regulatory reserve.
The timing convention matters. A payment at the instant of death would require integrating the discounted mortality density or making a within-year timing assumption. The end-of-year expression is appropriate only to the explicitly stated payment schedule. A changed benefit amount or interest structure would change the valuation even if the survival table remained identical.
A transparent five-year annuity calculation
Suppose instead that a life annuity-due pays 1,000 units at the beginning of each of the next five years while the life survives to the respective payment date. Using the same synthetic mortality and discount assumptions,
The first payment at time zero is certain for the life alive at entry; later payments require survival. The five-year term assurance benefits from death during coverage, while the temporary annuity benefits from survival. A reduction in within this illustrative horizon generally reduces the assurance's expected death benefit and increases the annuity's expected value, holding the discount rate and other assumptions fixed. This is a direct consequence of the cash-flow definitions rather than an independent theory.
When applying actual actuarial tables, the chosen survival model must respect payment frequency, exact age, deferral, joint-life status where applicable, benefit guarantees, policy termination and the governing valuation method. A period national life table and a prospective insured-life cash-flow model coincide only under restrictive assumptions. The SOA retirement mortality references help demonstrate why annuity practice maintains dedicated mortality and improvement bases.
Model risk and quality assurance
There are several distinct classes of error. Source errors affect reported deaths, population and dates. Statistical errors arise from censoring assumptions, unsupported interpolation, inadequate event volume, or incomplete observation. Mathematical errors violate table identities and probability constraints. Transfer errors arise when a valid source table is imposed on a materially different insured population. Contract errors confuse payment timing, survival contingencies, termination or benefit amounts. Governance errors omit applicable jurisdictional rules or incorrectly describe a regulatory table as a general economic forecast.
The minimum useful QA sequence is therefore: verify source and edition; reconstruct age-and-time exposure; inspect mortality estimation and selection; check survival recursion and terminal closure; evaluate uncertainty and sensitivity; match the table to the policy's insured class and calendar horizon; reproduce benefit timing calculations; and retain a verifiable record of each decision. Source traceability and mathematical QA should be completed before the manuscript is designated ready for release.
The book-length statistical literature places these issues within a larger inferential setting. Chiang's life-table construction treatise establishes the classical statistical object and its applications; the censored and truncated survival framework treats incomplete individual observations. Reading them together with contemporary demographic and actuarial source protocols clarifies why life tables are simultaneously statistical estimators, accounting identities and inputs to institutional decisions.
Integrated analytical synthesis
The evidence supports a four-stage interpretive chain. First, observation defines deaths, lives, time at risk and the source population. Second, estimation turns these observations into age-time rates, including treatment of incomplete records and modeled cells. Third, a survival representation converts rates or events into probabilities, survivors, person-years and expectations. Fourth, an insurance valuation applies an appropriate survival representation to legally and contractually specified cash flows.
Each stage introduces assumptions. A later stage cannot repair an undocumented population mismatch earlier in the chain. A sophisticated survival estimator cannot make biased death registration accurate; a perfectly calculated mortality table cannot determine an annuity's value without payment and discount assumptions; and regulatory adoption of one mortality table does not make it representative of all countries or product classes.
This analytical chain is a synthesis of established actuarial and statistical relationships. It is offered as a practical explanation of how to evaluate evidence and transfer it into valuation, not as a new theory, model, doctrine or priority claim. Its principal value lies in making every consequential bridge between measured mortality and insured benefits explicit and independently testable.
Conclusion
Life tables express mortality through a disciplined set of linked quantities. An age-specific death count acquires meaning from the exposure that produced it; a central death rate becomes a death probability only through a specified within-interval survival structure; consecutive probabilities become a survival curve through conditional multiplication; and the curve produces person-years and remaining lifetime through summation or integration. These transformations can be tested exactly, and their institutional implementations can be documented.
The most consequential scientific distinction is between numerical coherence and substantive suitability. Population period tables, observed birth cohorts, insured-life experience tables and statutory valuation bases carry different histories of observation, selection and purpose. A valid insurance application must match its insured cohort, age convention, forecast horizon, benefit trigger and payment schedule. Where a projection is needed, its assumptions should be assessed separately from the observed mortality base. Where records are censored or truncated, estimation must honor the corresponding risk sets and inference conditions.
The primary institutional materials demonstrate the scale and diversity of present practice. The Human Mortality Database emphasizes harmonized historical source reconstruction and explicit protocol revisions. The United Nations and WHO combine empirical and model-based mortality information in international frameworks. The U.S. National Center for Health Statistics constructs detailed period life tables using specified national data sources and older-age adjustments. Actuarial experience studies and commissioners' standard tables address distinct insurance populations and regulatory purposes. These sources complement one another when their definitions and editions are retained.
The reproducible hypothetical table establishes the algebraic relationship among and multi-year survival. The paired assurance and annuity calculations demonstrate the economic direction of mortality sensitivity for two opposite benefit contingencies. The examples are exact within their assumptions and have no claim to empirical mortality calibration. This separation between arithmetic illustration and statistical evidence is a scientific requirement.
The resulting editorial conclusion is that the life table is best evaluated as a traceable chain from observation to contractual cash flow. Reliable decisions require a demonstrable match at each link: event definition, population exposure, estimation assumptions, survival representation, time horizon, and benefit design. This monograph contributes an integrated and independently checkable interpretation of those established relationships, with its evidence limits and numerical assumptions explicitly stated.
References
Aalen, Odd O. (1978). “Nonparametric Inference for a Family of Counting Processes.” The Annals of Statistics 6(4): 701–726. https://doi.org/10.1214/aos/1176344247
Arias, Elizabeth, Jiaquan Xu, and Betzaida Tejada-Vera. (2026). United States Life Tables, 2024. National Vital Statistics Reports 75(5), August 25, 2026. National Center for Health Statistics, Centers for Disease Control and Prevention. https://www.cdc.gov/nchs/data/nvsr/nvsr75/nvsr75-05.pdf
Chiang, Chin Long. (1984). The Life Table and Its Applications. Malabar, FL: Robert E. Krieger Publishing Company. ISBN 9780898745702. Bibliographic and publisher information: https://books.google.com/books/about/The_Life_Table_and_Its_Applications.html?id=T_PrAAAAIAAJ
Cox, D. R. (1972). “Regression Models and Life-Tables.” Journal of the Royal Statistical Society: Series B (Methodological) 34(2): 187–202. https://doi.org/10.1111/j.2517-6161.1972.tb00899.x
Halley, Edmond. (1693). “An Estimate of the Degrees of the Mortality of Mankind; Drawn from Curious Tables of the Births and Funerals at the City of Breslaw; with an Attempt to Ascertain the Price of Annuities upon Lives.” Philosophical Transactions of the Royal Society of London 17: 596–610. https://doi.org/10.1098/rstl.1693.0007
Human Mortality Database. (n.d.; accessed November 14, 2026). “Overview.” University of California, Berkeley; Max Planck Institute for Demographic Research; and partner institutions. https://www.mortality.org/Project/Overview
Human Mortality Database. (n.d.; accessed November 14, 2026). “Explanatory Notes.” Human Mortality Database. https://www.mortality.org/Data/ExplanatoryNotes
Internal Revenue Service. (2016). Notice 2016–63: Guidance Concerning Use of 2017 CSO Tables Under Section 7702. Internal Revenue Bulletin 2016-45. https://www.irs.gov/irb/2016-45_IRB
Kaplan, E. L., and Paul Meier. (1958). “Nonparametric Estimation from Incomplete Observations.” Journal of the American Statistical Association 53(282): 457–481. https://doi.org/10.1080/01621459.1958.10501452
Klein, John P., and Melvin L. Moeschberger. (2003). Survival Analysis: Techniques for Censored and Truncated Data. 2nd ed. New York: Springer. https://doi.org/10.1007/b97377
Royal Society Archives. (1693; catalog consulted November 14, 2026). Paper, “Of the Breslau Bills of mortality” by Edmond Halley, CLP/21/45. London: The Royal Society. https://makingscience.royalsociety.org/items/clp_21_45/paper-of-the-breslau-bills-of-mortality-by-edmond-halley
Society of Actuaries. (2015; revised materials through 2018). 2015 Valuation Basic Report and Tables. Society of Actuaries Research Institute. https://www.soa.org/resources/experience-studies/2015/2015-valuation-basic-tables
Society of Actuaries. (2015–2016). 2017 Commissioners Standard Ordinary (CSO) Tables. Society of Actuaries. https://www.soa.org/resources/experience-studies/2015/2017-cso-tables/
Society of Actuaries. (n.d.; accessed November 14, 2026). Pension Mortality Resources. Society of Actuaries. https://www.soa.org/communities/retirement-practice/mortality-resources/
United Nations, Department of Economic and Social Affairs, Population Division. (2024). World Population Prospects 2024: Methodology of the United Nations Population Estimates and Projections. UN DESA/POP/2024/DC/NO.10. New York: United Nations. https://population.un.org/wpp/assets/Files/WPP2024_Methodology.pdf
Washington State Office of the Insurance Commissioner. (2017; current online rule consulted November 14, 2026). Washington Administrative Code Chapter 284-74, Section 284-74-535: 2017 CSO Mortality Table. https://app.leg.wa.gov/wac/default.aspx?cite=284-74&full=true
Wilmoth, John R., Kirill Andreev, Dmitri Jdanov, Dana A. Glei, and Tim Riffe. (2025). Methods Protocol for the Human Mortality Database. Version 6, revised August 5, 2025, with named contributors. Human Mortality Database. https://www.mortality.org/File/GetDocument/Public/Docs/MethodsProtocolV6.pdf
World Health Organization. (2026; available content accessed November 14, 2026). Global Health Estimates. Geneva: WHO; global, regional and national mortality series and documentation for 2000–2023. https://www.who.int/data/global-health-estimates
World Health Organization. (2026; accessed November 14, 2026). Global Health Estimates: Life Expectancy and Life Tables. Global Health Observatory, WHO. https://www.who.int/data/gho/data/themes/mortality-and-global-health-estimates/ghe-life-expectancy-and-healthy-life-expectancy/