Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Gompertz-Makeham Mortality Models

Abstract. Gompertz-Makeham mortality models represent the age pattern of mortality through a rising exponential component and, in the Makeham specification, an additive age-independent component. Their mathematical economy has made them central to actuarial graduation, survival analysis, life-contingent valuation, and the interpretation of adult mortality. This monograph reconstructs the historical foundations of Gompertz's 1825 survival law and Makeham's contributions of 1860 and 1867; defines the force of mortality and exact conditional survival quantities; establishes the relation between age-centered and conventional parameterizations; derives likelihood functions for grouped mortality counts, individual exact survival observations, censoring, and delayed entry; and examines estimation, identifiability, model diagnostics, graduation, extrapolation, cohort selection, heterogeneity, and stochastic mortality improvement. It then derives insurance and annuity values directly from the hazard and tests internally computed, transparently hypothetical numerical examples. The international evidence includes original nineteenth-century actuarial publications, the Human Mortality Database methods, high-age mortality studies, contemporary statistical software documentation, and the distinct 2026 U.S. life-insurance and retirement mortality-improvement research programs. The central interpretive result is that the model's actuarial authority is conditional on three separately demonstrable properties: coherent mathematical survival probabilities, a credible empirical population and estimation design, and fitness for the particular financial obligation and time horizon. This is an analytical synthesis of established science, not a claim to a new actuarial theory. The monograph identifies how a satisfactory local fit can coexist with weak age extrapolation, why an estimated Makeham constant is not directly an observed cause-of-death category, and how parameter uncertainty and future improvement affect financial values asymmetrically. Its conclusions specify evidentiary and validation conditions for responsible use in insurance and pension analysis.

Keywords: Gompertz law; Makeham law; force of mortality; mortality hazard; survival function; life contingencies; life tables; mortality graduation; parametric survival analysis; maximum likelihood; Poisson exposure; cohort mortality; longevity risk; frailty; old-age deceleration; actuarial valuation; annuities; life insurance; mortality improvements; model risk.

Introduction

An insurer promising a payment upon death and a pension provider promising income while a member survives require a quantified description of when death may occur. Gompertz-Makeham laws address one important part of this problem: the age-dependent instantaneous frequency of death. Given admissible parameters, their survival functions and financial implications follow by integration. Their appeal rests on that tractable link, the ability to summarize a broad portion of adult mortality with few parameters, and the possibility of comparing experience across populations and periods. Their actual predictive content depends on observation, exposure, age range, selection, historical period, and the intended extrapolation.

The historical sources matter because they establish the relationship between mathematical form and an actuarial problem. Benjamin Gompertz's 1825 mortality investigation developed an age-related law in a study explicitly concerned with valuing life contingencies. William Matthew Makeham's 1860 annuity tables paper sought a manageable law capable of generating calculations for several lives, and his 1867 study of mortality explicitly treated the force-of-mortality concept. These writings address computational coherence and empirical approximation together. Neither author could have established that one three-parameter hazard would universally explain population mortality under every age, cohort, and future social environment.

This dossier examines the scientific content that survives from those foundations. It distinguishes a hazard from an annual probability, a description of contemporary mortality from a future improvement assumption, a population average from an individual causal mechanism, and a fitting interval from an extrapolation interval. These distinctions govern the quality of life tables and the valuation of protection against death or prolonged survival. The methodological position is to derive the central mathematical identities from first principles, cross-check historical and current claims against identifiable publications, use hypothetical computations transparently, and state where empirical verification would require a specific portfolio or downloaded demographic dataset.

Scientific Object, Notation, and Research Method

Let age be measured continuously in years, let (x) denote attained age, and let (T_x) be future lifetime conditional on survival to age (x). For (t\geq 0), define ({}_tp_x=P(T_x>t)) and ({}_tq_x=1-{}tp_x). Define the age-specific force of mortality (\mu(x)) by the limiting conditional death probability per unit time. A standard local formula is (\mu(x)=\lim{h\downarrow0}P(x\leq T<x+h\mid T\geq x)/h), interpreting the age notation relative to an original lifetime. The hazard has units of inverse years; a death probability is dimensionless. The cumulative hazard over a future duration (t) is (H_x(t)=\int_0^t\mu(x+s),ds).

The fundamental identity is ({}_tp_x=\exp[-H_x(t)]). It gives a consistent path from a modeled intensity to survival probabilities. Because the exponential is positive, a finite cumulative hazard yields strictly positive survival; because an admissible hazard is nonnegative, survival is nonincreasing. The conditional death density, when it exists, is (f_x(t)={}_tp_x\mu(x+t)). A conditional density integrates to one if survival tends to zero as duration approaches infinity. A hazard model therefore has testable logical requirements before it is confronted with observations.

Research evidence is classified by operation. A nineteenth-century publication directly supports its historical title, purpose, mathematical exposition, and date. A demographic methods protocol supports the construction and limitations of modern exposures and death rates. A peer-reviewed empirical comparison supports its own sample, age coverage, fit criteria, and results; it cannot establish universal superiority. Software documentation establishes available model specifications and computational interfaces, not scientific validation of an insurer's own data. An original derivation establishes an algebraic implication under explicit assumptions. This source-specific reading avoids treating a formula's existence as proof of empirical adequacy.

No population dataset, pension plan file, or insurer experience study was downloaded and estimated for this dossier. The numerical examples below are explicitly constructed. Model coefficients in those examples have no empirical provenance and make no claim about an actual demographic group. This deliberate boundary allows independent mathematical reproducibility while preserving the difference between demonstration and population research.

Gompertz in 1825: Historical Proposition and Scope

Gompertz's complete Royal Society article was published in volume 115 of the Philosophical Transactions in 1825, pages 513-583. Its title joins the law of human mortality with a means of evaluating life contingencies. The digitized opening identifies the paper as read June 16, 1825, and written in correspondence to Francis Baily. That contemporary identification is distinct from the dates attached to modern digitization or indexing. The article is historically important both for its age-pattern regularity and for the attempt to turn that regularity into a computational instrument.

The modern Gompertz form is (\mu(x)=A\exp(Bx)), with (A>0) and (B>0). Its exact mathematical age slope is (d\log\mu(x)/dx=B). Thus the force itself multiplies by (\exp(B\Delta)) during an age increment (\Delta). The age-dependent force doubles over (\log 2/B) years, provided the same parameters apply through that age span. Such a regularity is a descriptive restriction on a population hazard, not a literal experiment showing that each person's internal biological clock doubles its risk at a constant pace.

The article was part of a broader historical program of actuarial calculation. The claim that Gompertz provided the complete modern science of mortality would collapse several later developments into one origin: improved death registration, graduated age-specific rates, the treatment of population composition, censored survival data, old-age model comparison, and stochastic improvements were developed in distinct research settings. The correct priority statement is narrower and stronger: Gompertz supplied an identifiable early mathematical law for the increasing adult age pattern, one that became a foundational benchmark in actuarial research.

The survival law is sometimes introduced by starting from the differential equation (dS(x)/dx=-A\exp(Bx)S(x)), then integrating. If the adult-age law is extended formally from some reference age (x_0), the resulting survival ratio is (S(x)/S(x_0)=\exp[-(A/B)(e^{Bx}-e^{Bx_0})]). This expression can be checked directly by differentiation. It supplies neither empirical evidence for infancy nor a defensible extrapolation to unobserved extreme ages without separate validation.

Makeham in 1860 and 1867: Mortality and Contingent Values

Makeham's 1860 law and annuity construction appeared in volume 8, issue 6, pages 301-310 of the contemporary actuarial journal. The opening frames a problem in valuation: exact annuity calculations for several lives were cumbersome when many separate combinations of mortality and interest had to be tabulated. Makeham sought a concise approximation that would support a more systematic computation. He also expressly recognized that a theoretical mortality law could represent observed experience only approximately. This historical caution is part of the model's origins.

The later 1867 actuarial paper, volume 13, issue 6, pages 325-358, opened by explaining the expression 'force of mortality.' Its treatment connects an instantaneous death intensity with life-contingent calculation. The exact historical citation matters: the 1860 and 1867 papers have distinct titles, journal volumes, pagination, and analytical roles. They should not be conflated merely because both are routinely cited under the author's name.

The common modern law reads (\mu(x)=C+A\exp(Bx)). The added constant (C\geq0) allows a background intensity that does not depend on attained age within the defined domain. By integrating it, the future survival function gains a factor (\exp(-Ct)) relative to the Gompertz component. The parameter does not itself prove that a separate exogenous cause of death exists or that competing causes are independent. A single all-cause hazard cannot identify cause-specific etiological mechanisms without cause-of-death data and a separate causal design. The interpretation of (C) as an 'accidental' component may be historically evocative; in modern actuarial use it is most safely treated as a component of the fitted function.

Makeham's practical innovation is best evaluated as a relation between approximation, tractability, and contingent-value consistency. A smooth analytic law can provide a complete table of probabilities, which in turn supports a consistent family of annuity prices. This does not imply that the smooth law perfectly reflects a specific insured portfolio. Indeed, valuation consistency under a wrong mortality basis can be precise and financially inappropriate at the same time.

The Force of Mortality and the Hazard-Survival Identity

The hazard (\mu(x)) is the local intensity of death conditional on having reached age (x). For a short interval (h), the conditional death probability is approximately (h\mu(x)); for a full year it must be computed by integrating the hazard. The approximation (q_x\approx\mu(x)) can be useful at sufficiently small rates and nearly constant hazards, but the precise yearly probability is (q_x=1-\exp[-\int_x^{x+1}\mu(u),du]). At advanced ages, when the hazard changes substantially within a year, the approximation creates an appreciable distortion.

The survival relationship follows from (S'(t)=-S(t)\mu(x+t)). Integrate (S'(t)/S(t)=-\mu(x+t)) from zero to (t) and impose (S(0)=1). This derivation establishes the conditional nature of the function. For ages (x) and (x+t), a period life table conditionally applies the mortality rates from one period to a hypothetical life as it ages; a genuine birth cohort traverses multiple calendar periods and may experience improving or worsening mortality. The Human Mortality Database overview distinguishes historical all-cause rates, life tables, and their underlying births, deaths, and population counts. Its data design is evidence for the necessity of specifying period, cohort, age, and population.

A central mortality rate (m_x=D_x/E_x) uses deaths divided by person-years exposed, with details depending on the chosen exposure construction. It is an observation-based rate, not inherently identical to (q_x), and it estimates a smoothed hazard only under an observation model. Errors in denominators affect any mortality curve fitted to the result. The HMD Methods Protocol, version 6, dated August 5, 2025, documents both exposure construction and the distinction between cohort and period configurations. Its revisions concerning birth-month distributions illustrate why even the exposure denominator is a scientific object requiring careful reconstruction.

The Gompertz Law: Parameters, Units, and Shape

It is convenient to center adult age at (x_0):

[\mu(x)=a\exp{b(x-x_0)},\qquad a>0,\quad b>0.]

Here (a) is the modeled force of mortality at (x_0), measured per year, while (b) measures the continuous log-slope per year. The parameterization is algebraically equivalent to (A\exp(Bx)) with (B=b) and (A=ae^{-bx_0}). The exponential slope is sometimes reported as a percentage increase over one year; the exact multiplier is (e^b), so the increase is (e^b-1), not simply (b) except as a local approximation. As (b) approaches zero, the Gompertz hazard tends to a constant and the model approaches exponential future lifetime.

For an individual alive at (x), integrate to obtain

[H_x(t)=\frac{a e^{b(x-x_0)}}{b}\big(e^{bt}-1\big),\qquad {}_tp_x=e^{-H_x(t)}.]

The exponent contains a dimensionless ratio of an annual hazard to an annual slope. With (b>0), (H_x(t)) grows without bound as (t) grows, so eventual death has probability one under the formal infinite-domain model. The formula remains mathematically defined at arbitrary ages, but demographic acceptability at high ages is an empirical question. In computation, use the stable function expm1(bt) when (bt) is small, because subtracting one from a nearly unit exponential loses numerical precision.

The model imposes monotonic growth at all ages within its application domain. That can suit a portion of adult mortality experience; it is not generally suitable as a single global hazard covering infancy, childhood, and all older ages. The MortalityLaws model catalogue explicitly distinguishes Gompertz, Makeham, Siler, Beard, Gamma-Gompertz, and other forms with different age patterns and fitted quantities. The existence of these alternatives reflects substantive differences in their empirical purposes.

The Makeham Extension and Parameter Restrictions

Use the common age-centered form

[\mu(x)=c+a\exp{b(x-x_0)},\qquad c\geq0,;a>0,;b>0.]

The cumulative hazard and survival conditional on having attained age (x) are

[H_x(t)=ct+\frac{a e^{b(x-x_0)}}{b}(e^{bt}-1),\qquad {}_tp_x=\exp[-H_x(t)].]

The constant contributes linearly to cumulative hazard; the age-dependent component contributes exponentially. At a specified future duration (t), a one-unit increase in (c) measured in inverse years multiplies survival by (e^{-t}) after accounting for the units. More generally (\partial\log {}_tp_x/\partial c=-t). The derivative with respect to (a) is (-e^{b(x-x_0)}(e^{bt}-1)/b). Both are negative for positive durations, proving that higher levels of either hazard component reduce survival when the others are held fixed.

The restrictions (c\geq0) and (a>0) ensure nonnegative intensity in the specified increasing version. Some statistical implementations use an unconstrained fitted constant, including negative values, if the sum remains nonnegative throughout a finite fitting interval. Such parameterizations demand an explicit finite age domain and verification of (\mu(x)\geq0) throughout every projected payment period. A negative background constant cannot simply be propagated to infancy or distant ages without risking an invalid survival law. The distinction between an admissible local curve and a globally admissible hazard is central to extrapolation governance.

When (c=0), Makeham reduces to Gompertz. When (b=0), the law becomes a constant force (a+c), and the two components cannot be identified individually from a mortality curve. An exactly constant (c) is a structural assumption, not evidence that external risks cannot change with medical treatment, public health, transport, violence, or environment. Time dependence, covariate dependence, and cause specificity require extensions outside the three-parameter static law.

Equivalent Parameterizations and Stability

The conventional parameters (A,B,C) and age-centered parameters (a,b,c) satisfy (A=ae^{-bx_0}), (B=b), and (C=c). Changing reference age from (x_0) to (x_1) changes (a) to (a_1=ae^{b(x_1-x_0)}) while leaving (b) and (c) unchanged. A meaningful report must specify the age center, because a number described as 'the Gompertz intercept' otherwise depends on the origin of the age scale. The practical reason for centering is numerical: at an adult age such as 60 or 70, the age-centered intercept has the scale of an observed hazard rather than a tiny extrapolated intensity at birth.

A further reparameterization employs (\sigma=1/b) and (M=x_0-(1/b)\log(a/b)), yielding the Gompertz component (\sigma^{-1}\exp[(x-M)/\sigma]). The parameter (M) is a location-like quantity for the hazard, and (\sigma) has units of years; their interpretation depends on the exact chosen distribution. Documentation for Gompertz and Makeham variants reports both direct exponential and location-scale forms. Algebraic equivalence permits a legitimate translation of fitted curves but does not eliminate different constraints or optimizer behavior.

When fitting counts, a log transformation of (a) and (b) enforces positive parameters. A log transformation of (c) likewise enforces strictly positive (c) but excludes the nested Gompertz boundary, making direct likelihood-ratio procedures near that boundary delicate. Optimizers can work with bounded parameters instead. All fit routines should identify whether an offset refers to chronological age, attained age at entry, person-years exposure, or calendar time. Confusing these roles can yield a numerically successful yet scientifically incorrect model.

Conditional Survival, Death Density, and Quantiles

Define (k_x=a\exp[b(x-x_0)]). The Makeham future survival function becomes ({}_tp_x=\exp[-ct-(k_x/b)(e^{bt}-1)]). Its density is

[f_x(t)=\left(c+k_xe^{bt}\right)\exp\left[-ct-\frac{k_x}{b}(e^{bt}-1)\right].]

The density and survival provide a complete probability distribution when (b>0) and (k_x>0), because eventual cumulative hazard is infinite. Its median future lifetime (t_{0.5}) solves (ct+(k_x/b)(e^{bt}-1)=\log 2). In general the Makeham median is obtained by a one-dimensional numerical root; for the Gompertz special case (c=0), it has the explicit form (t_{0.5}=b^{-1}\log[1+(b/k_x)\log 2]). More generally the future lifetime (p)-quantile solves (H_x(t_p)=-\log(1-p)), for (0<p<1).

The conditional remaining life expectancy is (e_x=\int_0^\infty{}_tp_x,dt), assuming convergence. A finite term restricted expectation is (e_x(T)=\int_0^T{}_tp_x,dt). These quantities do not require an asserted maximum human age. For a fixed mortality law, finite quadrature produces them directly; for a population forecast, changes in calendar-year mortality must be inserted along a cohort path. Distributional summaries should state the reference age, population basis, period, model, and duration horizon.

Mathematical quantiles reveal a useful asymmetry. A small relative change in (b) can have a modest effect on survival in a short interval near (x_0) and a large effect on the highest ages reached decades later. The derivative of (\log\mu(x)) with respect to (b) in the pure Gompertz component is (x-x_0). The farther an extrapolation moves from the reference age, the more strongly it magnifies slope error. The same mechanism propagates into late pension cash flows even when the fitted likelihood is concentrated at much younger observed ages.

Annual Death Probabilities and Mortality Tables

An annual conditional death probability under the age-centered Makeham model is

[q_x=1-\exp\left[-c-\frac{a e^{b(x-x_0)}}{b}(e^b-1)\right].]

The annual survival probability is (p_x=1-q_x). A synthetic life table beginning with any radix (l_{x_0}) follows recursively by (l_{x+1}=l_xp_x) and (d_x=l_xq_x). Alternatively (l_x=l_{x_0}\exp[-H_{x_0}(x-x_0)]). These two calculations must agree to numerical precision. The choice of radix affects the counts but not the probabilities. For death payments, the financial timing within each year also matters: end-of-year benefits, midyear approximations, and continuous payments have different discounting operators even under an identical mortality table.

A graduated table can be generated from fitted coefficients, but graduation is an inferential decision. Age-specific observed death counts remain random; a smooth curve reduces sampling noise by imposing an age pattern. It can also conceal genuine cohort differences, selection effects, and structural discontinuities. A published life table is therefore better understood as a set of estimates produced through defined data and assumptions than as a transparent transcript of raw death records. The HMD methods specify a disciplined chain from vital events and exposure estimates to mortality rates and life-table quantities.

The conversion between (m_x) and (q_x) is especially consequential at the highest ages. A central death rate based on exposure during age (x) does not become a probability simply by renaming it. When an analytic force is available, integrating it over the interval gives the appropriate probability. If only grouped age data are available, the actuarial assumption about the distribution of deaths within the interval must be stated, and the effect on the valuation tested where age-specific probabilities are large.

Mortality Compression, Dispersion, and Interpretive Boundaries

A Gompertz curve of log-hazard against age is straight; a Makeham log-hazard curve is generally convex or concave depending on the way it is plotted and the parameter magnitudes. For the age-centered Makeham law, the instantaneous slope of log-hazard is (b k_xe^{bt}/(c+k_xe^{bt})), which increases with age toward (b) when (b,c,k_x>0). Thus a statistically fitted background constant changes the age pattern of relative increase even though its own component is unchanged. The 'mortality doubling time' attributed to the total Makeham hazard varies by age and is not exactly (\log2/b) except when the constant is absent or negligible.

Mortality compression is often used for a process in which deaths cluster into a narrower interval of ages, but it has multiple measurable definitions. The change in a modal age at death, the variance of lifetime, the interquartile range, and the variability of residual lifetime are not interchangeable. A larger Gompertz slope (b), with other parameters and the age center fixed, changes the upper and lower portions of the lifetime distribution differently. A lower level (a) generally shifts survival outward. Their joint changes can generate complicated patterns in age-at-death dispersion, especially when the observed domain excludes childhood and when period improvements are large. Any claim about compression must specify a metric and the conditioning age.

A fitted Gompertz slope is sometimes described as a rate of biological aging. This language has a narrower acceptable statistical meaning: it measures a proportional age gradient in the specified observed hazard. A population rate is influenced by selection, composition, diseases, treatment, calendar conditions, and measurement. Those factors also determine the fitted slope. The original frailty framework of Vaupel, Manton, and Stallard (1979) illustrates that individual and population hazards may have different age trajectories even when their component risks are precisely specified. A population curve cannot uniquely reveal the trajectory of an unobserved individual.

Observational Designs and Exposures

Four data structures lead to distinct inferential problems. First, a period age schedule records deaths and exposure at ages observed during a specified calendar interval; people of different birth years contribute to different ages. Second, a cohort schedule follows persons born in a particular interval, with subsequent death rates observed in changing calendar periods. Third, a closed or longitudinal portfolio follows actual insured or pensioned lives, recording entry, exit, and individual characteristics. Fourth, a cross-sectional life table may already contain graduated estimates, in which case fitting a second model to the published rates would partly fit the first modeler's decisions rather than independent observations.

A responsible analysis begins by reconstructing the estimand. Is the target the all-cause age hazard of residents, the experience of underwritten policyholders, the mortality of current pension recipients, or a protected benefit population? Is age defined at last birthday or exact attained age? Are calendar-year exposures split correctly at birthdays and migration dates? Has death registration been completed? Are insured lives selected by underwriting, and does selection fade with policy duration? These questions affect the sampling design and the meaning of coefficients before any numerical optimizer is chosen.

The Human Mortality Database methods protocol supplies an institutional example of exposure accounting and cohort-period separation. Its 2025 version documents a revision using births by month to calculate certain population exposures more accurately. Such changes matter because an apparent age or cohort pattern can emerge from denominator construction. A statistical goodness-of-fit test cannot correct biased source registration or an ill-defined risk population. A complex mortality model fit to flawed exposures may produce smaller residuals while preserving the underlying demographic error.

For a large insurer, individual transaction records can support a richer model than aggregate counts. Policy issue age, duration, benefits, underwriting status, gender or sex variables where legally and analytically appropriate, geographic region, product type, claim status, and lapse can affect the target. Data minimization, lawful use, and privacy governance constrain what can be retained. Claims may be reported late, some policyholders may become unobserved after termination, and the distribution of insured amounts may differ greatly from the distribution of lives. The mortality rate needed to price one dollar of benefit may differ from a head-count mortality rate if benefit size is associated with mortality.

Maximum Likelihood for Grouped Deaths

Suppose age-year cell (j) contains deaths (D_j), central exposure (E_j) measured in person-years, and a representative attained age (x_j). For small age intervals or a carefully integrated cell, a practical model is (D_j\sim\operatorname{Poisson}(E_j\mu(x_j;\theta))), with (\theta=(a,b,c)). The log-likelihood, omitting factorial constants independent of parameters, is

[\ell(\theta)=\sum_j\left[D_j\log{E_j\mu(x_j;\theta)}-E_j\mu(x_j;\theta)\right].]

This approach treats the exposure as specified and presumes conditional independence of cell counts for likelihood purposes. In mortality portfolios, exposures may reflect the same lives over multiple intervals and deaths can be correlated through common shocks. The Poisson working model may still be an efficient estimation device, but robust variance estimation or more elaborate dependence structures may be required for uncertainty statements. Exposures of zero contribute no information and must be handled before taking a logarithm. Observed zero death counts are valid and supply negative log-likelihood contributions through the expected-death term.

For a broad age interval, the more faithful expected count is (\int E(u)\mu(u;\theta),du), where (E(u)) is exposure density across exact ages and times. Approximating this by a midpoint formula is an additional modeling assumption. The accuracy depends on interval width and hazard curvature. In a Gompertz model, rapidly increasing old-age hazard can make a single midpoint inadequate even for a cell with ample recorded experience. Computation should therefore integrate hazard against the appropriate exposure profile when exact birthdays and durations are available.

A standard deviance diagnostic compares observed and fitted death counts, with care about sampling and sparse strata. Pearson residuals use ((D_j-\widehat D_j)/\sqrt{\widehat D_j}); their behavior depends on the working variance assumption. Examine residuals by age, calendar year, issue duration, risk category, and benefit size. Age-aggregate fit can conceal offsetting errors in subgroups. Data-driven age window selection, chosen after inspecting outcomes, also changes inferential uncertainty and should be documented. Model comparison using AIC or held-out predictive deviance may help, but neither tests the social or financial consequences of using a mortality curve beyond the observed range.

Individual Survival Data, Censoring, and Left Truncation

Suppose life (i) is observed from a known entry age (u_i) until age (v_i), with (\delta_i=1) if death occurs at (v_i) and (\delta_i=0) if observation ends under independent right censoring. Conditional on entry alive, the individual likelihood contribution is

[L_i=\mu(v_i;\theta)^{\delta_i}\exp\left[-\int_{u_i}^{v_i}\mu(s;\theta)ds\right].]

The log-likelihood is the sum of (\delta_i\log\mu(v_i;\theta)) minus each accumulated exposure hazard. The survival contribution automatically conditions on being alive at the left-truncation age (u_i). It is incorrect to use a birth-origin density for every pensioner first observed at retirement without conditioning on survival to entry. Delayed entry must be reflected in the risk set, especially when portfolios contain long-lived members selected by earlier survival.

Right censoring can arise when the study ends while a life is alive; its treatment assumes a censoring process conditionally independent of the remaining lifetime given the included information. Lapses, disability transitions, and withdrawals may violate this assumption if they depend on health. Interval censoring arises when death is known only between observations; it requires a probability difference of survival functions. Left censoring is different from delayed entry, and using the terms interchangeably can result in an incorrect likelihood. The flexsurv statistical documentation describes estimation under right censoring, left censoring, and left truncation, including user-defined parametric hazard families.

A vital distinction concerns the time scale. If time-to-event is modeled as policy duration rather than attained age, a Gompertz hazard (a e^{bt}) describes duration dependence. To recover an age model, the attained age at issue must enter correctly as the individual's starting age; otherwise a person issued a policy at 40 and another at 70 would be assigned the same initial hazard under an unqualified duration specification. Age, calendar period, and policy duration can each carry important information. A single one-dimensional hazard cannot represent them simultaneously without additional modeling design.

Identifiability and the Makeham Constant

A formal model may have three named parameters while a finite dataset supplies little information to distinguish them. In Makeham, (\partial\mu/\partial c=1), (\partial\mu/\partial a=e^{b(x-x_0)}), and (\partial\mu/\partial b=a(x-x_0)e^{b(x-x_0)}). Over a narrow band of ages, the exponential is approximately (1+b(x-x_0)). In that approximation, the intercept contains roughly (a+c), while the slope contains roughly (ab). Multiple combinations of (a,b,c) can produce nearly identical hazards. This is practical weak identification rather than necessarily an exact mathematical nonidentifiability over a continuous and sufficiently broad age domain.

At the exact boundary (b=0), (\mu=a+c), so (a) and (c) are not separately identified. If fitting software reports a large confidence region, unstable intercepts, or a near-singular information matrix, a simple apparently interpretable constant should not be endowed with strong epidemiological meaning. Profile-likelihood confidence intervals, bootstrap sampling under an appropriate data scheme, and multiple starting points can reveal whether the optimum is stable. A positive constrained (c) may remain at the boundary; conventional asymptotic tests that assume a parameter interior to a smooth open space then require care.

Weak identification is amplified when the data come from a high-age-only portfolio. At those ages the exponential term can dominate, and a small background constant has little influence on the total observed hazard. Conversely, in a limited midlife range the two terms may offset sufficiently that the apparent background level depends on the chosen age window. A claim that a fitted (c) measures 'accidents' or 'external causes' thus requires independent cause-specific evidence; an all-cause fitted curve supplies no such identification.

The sensible response is a documented model comparison: fit Gompertz and Makeham on the same observations and age interval; examine predictive performance, parameter stability, uncertainty, tail behavior, and valuation consequences. If the extra constant does not add reliable information, the two-parameter model may be preferable for its parsimony. If it captures repeatable lack of fit, preserve it but avoid equating numerical improvement with causal proof.

Graduation, Diagnostics, and Model Selection

Graduation converts noisy observed mortality experience into a coherent schedule. Gompertz-Makeham provides a parametric graduation with a small number of degrees of freedom. A comprehensive diagnostic examines the original numbers of deaths and years of exposure; fitted age-specific expected deaths; standardized residuals; cumulative observed-versus-expected counts; local age trends; and the sensitivity of fitted coefficients to beginning and ending ages. The logic is comparable to analyzing a model of insurance claims: an adequate aggregate total does not ensure fair or accurate distribution across age groups.

A model can be compared by likelihood-based information criteria such as AIC, by cross-validation on disjoint cohorts or years, and by out-of-sample prediction. For forward-looking insurance use, chronological validation is particularly revealing: fit to earlier experience and evaluate later experience rather than mixing all years randomly. AIC is a penalized in-sample likelihood score; it does not directly measure the cost of a 40-year annuity hedge error. A model selected for statistical fit on annual death counts may not minimize errors in the tail of the present-value distribution.

A useful published applied example is the pensioner survival-model comparison reported in Applying Survival Models to Pensioner Mortality Data. Its reported sample covers pensioner mortality experience from 2000 through 2006 at ages 60-95, and the displayed model table evaluates Gompertz, Makeham, Perks, Beard, and combinations. The study found that more flexible alternatives could achieve markedly better in-sample fit than simple Gompertz in that particular dataset. This is a population-specific model comparison, not a general ranking that automatically transfers to other countries, contract books, or years.

Regularity is not the only quality criterion. A fitted curve might have smooth mortality but implausibly steep values just past the last observed age. A model with many parameters might improve historical likelihood but become unstable after a shock or when transferred to a smaller insurer. Diagnostics should therefore cover the observation range and the economically material projection range separately. If an insurer has few events at extreme ages, uncertainty around those projections needs to be displayed rather than removed by a smooth plot.

Age Extrapolation and Old-Age Controversies

At older ages, Gompertz imposes continued exponential growth in hazard. This can be tested only against data with validated ages and sufficiently reliable death and exposure counts. Two empirical findings show why a single universal narrative is inadequate. Feehan's 2018 old-age model comparison fitted nine mortality models to 360 high-quality cohort datasets after age 80. Models allowing deceleration generally performed better than uninterrupted exponential growth, and a log-quadratic model gave particularly consistent predictive results. Yet Gavrilov and Gavrilova's 2011 cohort mortality study, using specific extinct-cohort methods and narrowly recorded ages, reported approximate Gompertz continuation through ages around 102-105 in its selected data. These findings involve different populations, measurement methods, and model comparisons; their apparent disagreement is itself an object of source-sensitive analysis.

At extreme ages, denominators can become very small. Incorrect age reports, delayed registrations, exposure approximations, migration, and cohort mixing may manufacture either an apparent deceleration or an apparent continuation of the exponential rise. These errors are not symmetric: if survival is rare, a handful of misstated observations can have large influence on estimated rates. Feehan's findings do not establish an invariant human mortality plateau. Likewise, a fitted Gompertz curve through age 105 in one data source does not establish universal exponential growth at 115 or beyond. The empirical question must be specified by source quality, sample, country, sex, cohort, age interval, and validation design.

A key actuarial consequence is that an annuity can be sensitive to upper-age survival even when small probabilities are involved, especially for long-lived contingent beneficiaries or high-duration guarantees. A pure Gompertz model sometimes makes remote survival mathematically negligible more quickly than a decelerating alternative. The dollar effect depends on benefit type, interest, improvements, and portfolio size. A stress test should compare plausible tail models on the same credible observation basis and show the variation in survival and present value at the intended horizon.

The outcome is a scope-specific rule: extrapolation across unobserved decades is a separate assumption that must be evaluated separately from fit to observed ages. This is an inference from ordinary statistical model risk and the reviewed demographic evidence, not a claim of a new mortality principle.

Heterogeneity and Frailty

Even if individuals possess Gompertz-shaped hazards, their mixture can produce a population hazard that grows more slowly at advanced ages. Suppose each individual has a positive frailty multiplier (Z), and conditional on (Z=z) the hazard is (z\mu_0(x)). The population survival function becomes (S(x)=E[\exp(-Z H_0(x))]), where (H_0) is baseline cumulative hazard. The population hazard is

[\mu_{\mathrm{pop}}(x)=\mu_0(x)\frac{E[Z\exp(-ZH_0(x))]}{E[\exp(-ZH_0(x))]}.]

As age increases, the surviving population can become increasingly concentrated among lower-frailty individuals. This survivor composition changes the observed population hazard even if the baseline individual hazard stays exponentially increasing. Vaupel, Manton, and Stallard's 1979 analysis of heterogeneity formalized the importance of differential frailty for population and individual life tables. The effect is statistical selection; it should not be interpreted as evidence that the same person necessarily becomes biologically more resistant as years pass.

For a Gamma frailty with mean 1 and variance (\vartheta>0), the Laplace transform gives (S(x)=[1+\vartheta H_0(x)]^{-1/\vartheta}), so (\mu_{\mathrm{pop}}(x)=\mu_0(x)/[1+\vartheta H_0(x)]). If (\mu_0) is pure Gompertz, the population hazard approaches a finite asymptote (b/\vartheta) as age increases. This derivation explains how an old-age hazard plateau can emerge mathematically from unobserved heterogeneity without placing a plateau directly into each individual trajectory. The general frailty result does not establish that every observed deceleration has this cause.

Population aggregation also affects actuarial subgroup pricing. If underwriting or occupational selection changes the fraction of lower-risk members, a fitted aggregate curve can change even with stable within-group hazards. The resulting differences should be decomposed where data permit. A mortality parameter cannot be exported from a general population to a selected pension or insurance portfolio as though selection and benefit weighting were irrelevant. Stress testing can vary selection assumptions explicitly.

Modern Mortality Models and Forecasting

Gompertz-Makeham is one member of a larger family of mortality descriptions. Logistic and Gamma-Gompertz or Beard specifications can permit old-age deceleration. Siler-type models add a decreasing early-life component to a constant and an increasing component, making them better suited to some whole-life age patterns. Heligman-Pollard relationships use multiple terms to approximate the unusual infancy, young-adult, and senescent shapes in annual death probabilities. The MortalityLaws model inventory provides a documented catalogue of these alternatives and explicitly distinguishes whether each is formulated for a hazard or an annual death probability. The distinction is essential when comparing formulas.

Forecasting adds another dimension. The classical laws use age alone, with parameters assumed fixed for the exercise. A stochastic mortality model allows age-specific risks to evolve over calendar time. Lee and Carter's 1992 mortality forecasting study used a time-varying mortality index and age-specific sensitivities to project U.S. mortality, addressing a different inferential problem from a fixed Gompertz fit. Cairns, Blake, and Dowd's 2006 two-factor mortality study addressed post-60 mortality dynamics and parameter risk for longevity bond valuation. Neither specification negates the usefulness of a Gompertz base curve; both demonstrate what a static adult-age law leaves unspecified when valuation depends on decades of future change.

A time-varying Makeham generalization might write (\mu(x,t)=c(t)+a(t)e^{b(t)(x-x_0)}), but this is only a generic notation unless the dynamic laws for the parameters and their dependencies are specified and estimated. Different correlations between level and slope changes generate different tail outcomes. The resulting actuarial liability depends on the future path (\mu(x+s,t_0+s)), not simply on (\mu(x+s,t_0)). Freezing a period curve into the future is an explicit no-improvement scenario, appropriate for some illustrations and unsuitable as an unexamined universal projection.

The professional context changes over time. The Society of Actuaries Research Institute's 2026 Mortality Improvement Model update reports historical information through 2023 while retaining MP-2021 rather than issuing a new retirement projection scale. The separate RPEC 2026 mortality report explains the committee's continued use of MP-2021 as a reasonable U.S. retirement-plan best-estimate basis, particularly at older ages. These are dated methodological decisions for specific U.S. retirement work, not assertions that mortality has stopped changing.

The 2026 individual-life insurance context differs. The SOA 2026 individual-life improvement scale uses insured population experience and actuarial assumptions relevant to AG38/VM20. A pension scale and a life-insurance valuation scale may legitimately use different bases because they insure different cash flows and populations. In both fields the governing question is how a static hazard representation interacts with future mortality change and with the legal or professional valuation framework applicable to the contract. International practice must be treated through jurisdiction-specific evidence, not by projecting a U.S. practice document into a universal mandate.

Life Insurance and Annuity Valuation

A mortality law becomes an actuarial valuation basis when it is combined with a benefit definition, payment timing, and interest or discount assumptions. Let (\delta\geq0) be a constant force of interest and let (v^t=e^{-\delta t}). For a benefit of one monetary unit paid at the exact time of death if death occurs within (T) years, the net single premium is

[\overline A_{x:\overline T|}=\int_0^T e^{-\delta t}{}_tp_x\mu(x+t),dt.]

For a continuous temporary annuity paying one unit per year until earlier death or time (T), the actuarial present value is

[\overline a_{x:\overline T|}=\int_0^T e^{-\delta t}{}_tp_x,dt.]

The same Makeham force enters both integrals, but with different weighting. Death insurance pays more in scenarios involving earlier deaths; a lifetime annuity pays more in scenarios involving prolonged survival. The financial direction of mortality assumption error depends on the benefit. In particular, an understatement of future mortality can increase an annuity value while decreasing the expected present value of a death-only benefit, all else equal.

Integration by parts, using (S'(t)=-S(t)\mu(x+t)), yields the exact term identity

[\overline A_{x:\overline T|}=1-e^{-\delta T}{}Tp_x-\delta,\overline a{x:\overline T|}.]

For an infinite term with zero discounted survival at infinity, this becomes (\overline A_x=1-\delta\overline a_x). These equations allow an independent cross-check of numerical quadrature: calculating insurance directly from the death density and indirectly from survival and annuity values must agree. A failure indicates a programming or convention error. The identities do not account for expenses, capital costs, taxes, lapses, or profit margins; those are separate valuation layers and may be essential for real contract pricing.

For a discrete end-of-year death benefit, (A_{x:\overline n|}=\sum_{k=0}^{n-1}v^{k+1}{}kp_xq{x+k}), using annual effective interest (i) and (v=1/(1+i)). A temporary annuity-due has (\ddot a_{x:\overline n|}=\sum_{k=0}^{n-1}v^k{}_kp_x). With benefits financed by level annual premiums paid while alive and subject to the contract, the equivalence principle can define a net premium as a ratio of benefit value to premium-annuity value. This ratio is contract-specific and is not a recommended premium for any individual. An insurer's gross rate additionally requires explicit assumptions concerning underwriting, expenses, lapses, capital, guarantees, and risk margins.

The sensitivity to discounting is sometimes obscured by discussion of mortality alone. A lower discount rate can substantially increase the present value of distant annuity payments. A positive correlation between longevity improvement and economic conditions could change total liability risk beyond separate deterministic stresses. For a reproducible illustration, holding the mortality model fixed while varying (\delta) isolates interest sensitivity; varying mortality improvements while holding (\delta) fixed isolates a different effect. The baseline Gompertz-Makeham model provides the survival component, while asset-liability management determines how it interacts with the financial balance sheet.

Joint-Life, Selection, and Portfolio Applications

Makeham's nineteenth-century purpose included calculations involving multiple lives. For independent remaining lifetimes (T_x) and (T_y), joint-life survival is ({}tp{xy}={}_tp_x{}_tp_y), and the corresponding first-death hazard is (\mu_x(x+t)+\mu_y(y+t)). A last-survivor annuity depends instead on the probability at least one life remains alive, ({}_tp_x+{}_tp_y-{}_tp_x{}_tp_y) under independence. The resulting temporary or whole-life annuity value is obtained by integrating this survivor probability against the discount factor. In each case, independence is an assumption to be evaluated, especially for spouses who share environment, health care, socioeconomic conditions, and catastrophic risks.

For a portfolio of lives (i=1,\dots,N), an expected benefit is the sum of conditional expected values under the contract and available characteristics. Large-portfolio averages may stabilize idiosyncratic variation, but common mortality improvement, epidemics, environmental shocks, and changes in health care can create correlated deviations. Increasing the number of similar policyholders does not remove shared longevity risk. This distinction underlies the importance of aggregate stochastic mortality analysis in the Cairns-Blake-Dowd study.

Portfolio selection requires particular care. The mortality of healthy underwritten lives differs from that of the general population; retirement membership may depend on employment, disability, and benefit decisions. The Richards pensioner analysis documents the importance of socioeconomic and geodemographic predictors and the practical distinction between lives-based and amounts-based mortality estimation. A single Makeham fit applied across a mixed population may average relevant differences. That average can be inappropriate when the financial value is concentrated in a high-benefit subgroup.

The first step in a multi-life application is therefore to identify the benefit trigger: first death, last death, continued joint survival, death of a designated insured, or payment conditioned on an additional status. The second is to define dependence and conditioning. The third is to integrate discounting and survival on the same time scale. Formulaic convenience should follow these contractual definitions. A compact parametric hazard is useful precisely when its tractability can be combined with a transparent contract specification.

Worked Synthetic Calculations and Independent Checks

Consider a hypothetical person who has reached exact age 60, with a fixed age-centered Makeham force

[\mu(60+t)=0.0015+0.004\exp(0.085t),\qquad t\geq0.]

Every parameter is illustrative: (c=0.0015) per year, (a=0.004) per year, and (b=0.085) per year. The model uses a fixed future age curve, no mortality improvement, no selection adjustment, and no change in contract status. It is not fitted to any real national or insured population. Its conditional cumulative hazard is (H_{60}(t)=0.0015t+(0.004/0.085)(e^{0.085t}-1)). Its survival function is (S_{60}(t)=e^{-H_{60}(t)}). All numerical results below follow from these stated values.

Years after age 60 Attained age Force of mortality (per year) Conditional survival to age One-year death probability at attained age
0 60 0.005500 1.000000 0.005659
10 70 0.010859 0.924925 0.011205
20 80 0.023396 0.786204 0.024059
30 90 0.052728 0.548467 0.053485
40 100 0.121356 0.240986 0.118912
50 110 0.281922 0.035900 0.254861

The hazard in the third column is an instantaneous rate, while the last column is the integrated probability over the following full year. Their difference widens at higher ages, confirming why their interchange is analytically wrong. Direct calculation of the hazard multiplier at ages 60 and 80 also shows why (e^{0.085\times20}) describes the exponential component rather than the full Makeham intensity when (c) remains positive. The displayed remote-age figures are algebraically generated extrapolations, not evidence of real mortality at age 110.

For a constant annual effective interest rate of 3%, take (\delta=\log(1.03)) per year and value hypothetical continuous benefits. Independent numerical quadrature of the survival and death-density integrals gives:

Term from age 60 Continuous annuity present value per unit/year Death benefit present value per unit of coverage
10 years 8.380949 0.064038
20 years 13.959190 0.152081
30 years 17.228908 0.264773

The conditional remaining lifetime mean is approximately 30.323691 years and the median is approximately 31.638452 years. The continuous whole-life annuity at 3% annual effective interest is approximately 19.024696 monetary units for one unit paid per year; the continuous whole-life death benefit is approximately 0.437653 for a unit payment. Using the independently derived identity (1-\delta\overline a_{60}) reproduces the latter result. This agreement is a numerical validation of the formulas, not evidence of predictive accuracy for the hypothetical rates.

The formulas can be reproduced in a few lines of numerical software. The following pseudocode makes the conventions explicit:

c = 0.0015; a = 0.004; b = 0.085
interest = log(1.03)
H(t) = c*t + (a/b)*expm1(b*t)
S(t) = exp(-H(t))
h(t) = c + a*exp(b*t)
q_at(t) = 1 - exp(-(H(t+1)-H(t)))
annuity(T) = integrate(exp(-interest*t)*S(t), t=0..T)
insurance(T) = integrate(exp(-interest*t)*S(t)*h(t), t=0..T)
assert_close(insurance(T), 1-exp(-interest*T)*S(T)-interest*annuity(T))

Numerical integration must specify convergence tolerance and a sufficiently remote finite truncation if the software does not accept an infinite upper bound. For these illustrative parameters the rapidly increasing cumulative hazard makes a large truncation such as 200 future years numerically adequate, although a practical production system should verify the remaining tail bound rather than assume a universally safe age cutoff. The formula expm1 reduces cancellation in a near-zero interval. In very old-age calculations a log-survival representation can avoid arithmetic underflow.

Uncertainty, Sensitivity, and Validation

At least four uncertainties should be distinguished. Sampling uncertainty concerns randomness in death counts and censoring. Parameter uncertainty concerns the fitted coefficients conditional on the chosen model and data. Structural uncertainty concerns the selection of Gompertz, Makeham, a decelerating hazard, frailty, or a time-varying alternative. Future-state uncertainty concerns changes in mortality, interest, policyholder composition, and contractual behavior beyond the historical observation period. Reporting only a covariance matrix for ((a,b,c)) cannot encompass all four.

The quantity (\log\mu(x)) gives a simple sensitivity demonstration. In pure Gompertz, (\partial\log\mu(x)/\partial b=x-x_0); a slope perturbation (\Delta b) changes the age-dependent hazard multiplicatively by (\exp[(x-x_0)\Delta b]). A seemingly small shift (\Delta b=0.005) per year leads to a factor (e^{0.25}\approx1.284) fifty years beyond the reference age. This number is an exact algebraic stress calculation under an intentionally fixed baseline, not an estimated confidence interval. In Makeham, the sensitivity of the total force is damped by the proportion due to the exponential component at each age.

Uncertainty in survival is cumulative. If (\theta) denotes model parameters, then (\partial \log S_x(t;\theta)/\partial\theta=-\partial H_x(t;\theta)/\partial\theta). The derivative accumulates effects across all years until (t). For a long-dated annuity, parameter errors can consequently have a large influence on late payments even if the single-year model residuals are small. Numerical sensitivity tests should compute both rate errors and cash-flow-weighted valuation errors under deliberately varied parameter and structural scenarios. The observed portfolio should be tested on the same valuation basis used for the intended decision.

Confidence intervals may be estimated through likelihood profiles, nonparametric bootstrap by insured life, parametric simulation under the fitted hazard, or Bayesian posterior analysis with disclosed priors. The resampling unit should respect the data dependency structure; independently resampling each age-year count can understate variability when repeated observations belong to the same insured lives. A robust uncertainty report also includes selected out-of-sample periods, economic scenarios, and model alternatives rather than selecting the most favorable confidence bound among multiple methods.

The model should be considered validated only to the extent of documented checks. Mathematically, require nonnegative hazard, coherent conditional survival, and correctly dimensioned parameters. Statistically, examine origin and completeness of deaths and exposures, a credible population, fitted residuals, covariate/selection patterns, and predictive performance. Economically, check the specific insured or annuity cash flows, benefits, timing, currency, discounting, longevity improvement, and capital purpose. These three levels can be examined separately and brought together in a decision record without presenting them as a newly invented statistical framework.

International Practice and Governance

Mortality patterns arise in countries with different vital-registration histories, cause-of-death systems, population movements, insurance structures, and longevity trends. A general population age curve from one jurisdiction is not an internationally portable table of policyholder death probabilities. The Human Mortality Database is designed around countries and areas with comparatively complete registration and census data, as its population coverage statement explains. That design improves comparability within its stated scope; it does not claim to contain a representative insurance experience dataset for every country or global population.

International actuarial work uses many different mortality bases. Some rely on observed national tables, others on insured-population studies, employer pension data, or regulatory prescribed tables, often with separate adjustments for selection and mortality improvement. A Gompertz or Makeham curve can enter as a smoothing or comparison tool without replacing those distinct sources. Extrapolation beyond credible ages may have a different regulatory treatment from internal pricing or academic demographic research. Jurisdiction-specific effective dates and professional requirements must be checked in the actual valuation assignment.

The two U.S. institutional streams published in 2026 illustrate this contextuality. The SOA's RPEC 2026 findings refer to retirement mortality improvement and the continued MP-2021 basis. The individual-life 2026 recommendation refers to individual life insurance under specific actuarial valuation purposes and uses insured experience. Distinguishing them preserves the evidence behind each institutional decision and avoids the false implication that an improvement assumption transfers automatically across product lines. No generalized three-parameter Makeham fit would by itself satisfy all U.S. or international financial-reporting requirements.

The editorial record of an actuarial study should therefore preserve the source data version, age range, modeling period, exclusion rules, mortality improvement basis, coefficient parameterization, computational settings, model comparison, substantive findings, and limitations. Under the current HMD technical protocol, exposure changes can revise the historical age-period surface; any study using HMD-derived data should state the data extraction and methods version when reproducibility requires it. Similar version control applies to national insurer experience tables and projection models.

This dossier has a scientific publication role. It supplies public explanation, mathematical derivations, source criticism, and methodological criteria. It does not select personal insurance coverage, certify an insurer's reserving basis, or issue a professional actuarial opinion. Its recommendations concern the design and documentation of research rather than the financial circumstances of an identifiable policyholder.

Research Synthesis and Practical Decision Rules

The evidence supports an integrated reading of Gompertz-Makeham models. Historically, these laws were developed in dialogue with the valuation of life contingencies. Mathematically, they are coherent low-dimensional conditional hazard families. Statistically, they provide a parsimonious approximation whose coefficients and external validity depend on the observed age range, population, and estimation design. Financially, they produce exact modeled survival and present-value operators, while the accuracy of the resulting prices depends on whether the hazard basis is suitable for a particular population and horizon.

A central conclusion is that the Makeham constant is a fitted additive component of the all-cause force. Although descriptive labels such as 'accidental' or 'background' appear in interpretations of the law, a causal reading requires independent evidence. Deaths classified by cause, covariates, competing-risk definitions, and assumptions about cause dependence would be needed to identify etiological components. The all-cause curve alone cannot undertake that inference. This distinction prevents the common error of converting a convenient actuarial decomposition into a causal account of mortality.

A second conclusion concerns observation range. A locally adequate Makeham curve at ages 60-85 may produce impressive fitting statistics and poor liability projections at ages 100-115. The inferential bridge is especially weak if slope parameters are correlated, late-age observations are rare, or a population may display deceleration or composition effects. Feehan (2018), Gavrilov and Gavrilova (2011), and Vaupel, Manton, and Stallard (1979) describe different empirical and theoretical routes to analyzing this difficulty. The correct scientific response is to investigate age data and model alternatives, not to assume that the simplest shape or the most elaborate formula has automatic priority.

A third conclusion concerns the distinction between a life table and a longevity forecast. A period schedule is a representation of hazards under the specified period conditions. An annuity payable 30 years from now encounters future calendar-year hazards; these require explicit mortality improvement assumptions. The Lee-Carter research program, subsequent stochastic longevity work, and 2026 professional mortality reports are evidence of the substantive need to model temporal change. A static Gompertz-Makeham curve remains useful as a benchmark or valuation component when that role is declared.

The practical decision sequence is consequently clear. Define the benefit and risk population; reconstruct death counts and exposures; choose the attained-age and calendar-time variables; fit and compare admissible hazards; inspect model and data limitations; stress the tail and future improvements; compute financial values using the proper payment conventions; and preserve the model choice and evidence in a reproducible record. A fully transparent model with an acknowledged validity interval is more scientifically informative than a single apparently precise coefficient table presented without its assumptions.

This synthesis is the author's independent analytical interpretation of the checked historical, mathematical, demographic, and actuarial evidence. It does not claim discovery of a new hazard function, original mortality dataset, proprietary theory, or universally optimal insurance practice. The interpretation follows from identifying how historical purpose, probabilistic coherence, statistical identification, and future financial obligations meet in the same actuarial calculation.

Conclusion

Gompertz-Makeham mortality models remain a durable foundation of actuarial reasoning because they connect age to a coherent survival distribution and permit the direct calculation of life-contingent quantities. Gompertz's 1825 law captures exponential growth of adult mortality intensity. Makeham's 1860 and 1867 contributions provide an additive extension and a historical bridge to the systematic force-of-mortality analysis used in annuity construction. The exact modern form yields explicit cumulative hazards, survival probabilities, annual death probabilities, and contingent payment values.

The scientific significance of these laws is established most clearly when their scope is specified. A fitted hazard is conditional on a population, observation interval, age range, and statistical model. Its Makeham constant is a model parameter; its slope is an observed age gradient; its long-range predictions are extrapolations. Cohort selection, frailty, incomplete registration, uncertain exposures, and changing future mortality can alter the behavior of a population's observed hazard. Contemporary research supplies alternative age models and time-varying approaches because these questions arise from different evidentiary conditions and actuarial obligations.

The independently reproduced hypothetical calculations confirm internal consistency between survival, death density, annuity valuation, and insurance valuation. Their results are reproducible from the displayed parameters, with no reliance on invented observed death counts. The broader scientific conclusion is that formulaic consistency, source validity, and fit for financial purpose must be demonstrated together before a mortality law is used to support material valuations. That conclusion is an original analytical synthesis of established research. It neither claims a novel model nor treats a historical benchmark as a universal forecast.

Gompertz-Makeham remains useful when the analyst can explain what the curve describes, how it was fitted, which horizon it supports, which alternatives were tested, and how uncertainty enters the benefits being valued. The permanence of the law in actuarial science lies in this disciplined connection between mortality measurement and life-contingent value.

References