RESEARCH DOSSIER
Mortality Heterogeneity and Frailty in Life Portfolios
Abstract. Life-insurance portfolios contain heterogeneous mortality risks even after controlling for recorded underwriting classes, age, sex, duration, contract structure, and calendar period. Unobserved heterogeneity, conventionally represented by a nonnegative frailty variable, changes the survival distribution of a cohort by systematically altering the composition of those who remain alive. The resulting survivor selection can generate attenuation of observed mortality hazards, nonproportional aggregate hazard ratios, misleading interpretations of mortality improvement, and materially different life-insurance and annuity cash-flow projections. This monograph establishes the statistical and actuarial meaning of frailty, derives the gamma-mixture representation and its implications for cohort mortality and tail survival, examines identifiability and observational equivalence, and connects mathematical selection with underwriting, lapses, annuity purchase, insurer experience data, reserving, and capital risk. Primary demographic documentation, original scholarly models, insurance experience studies, and current professional standards provide an evidence base. Original analytical contributions comprise a constructive identifiability demonstration, a numerical illustration checked directly from model equations, a separation of idiosyncratic frailty from shared mortality dependence, and an operational research-to-valuation crosswalk. They constitute an integrated synthesis of established methods rather than a new theory. The central finding is that marginal insured mortality cannot be interpreted or extrapolated reliably without identifying entry, survival, contract persistence, and data-observation mechanisms. Portfolio-specific underwriting experience and appropriate sensitivity analyses remain indispensable; aggregate national mortality and paywalled industry summaries cannot substitute for individual insured-life records.
Keywords: Mortality heterogeneity; individual frailty; life insurance; survival analysis; actuarial science; gamma frailty; survivor selection; latent risk; insured-life mortality; annuities; informative lapse; mortality improvement; identifiability; underwriting; portfolio risk; longevity risk; correlated mortality; experience studies.
Introduction
Mortality risk in a life portfolio is measured among persons who have entered an insurance arrangement, have survived to a specified observation point, and remain observable under a particular data definition. Each of those conditions can alter the risk distribution. Two individuals with the same recorded attained age and underwriting label may differ in unrecorded health history, exposure, socioeconomic circumstances, susceptibility, or risk-related behavior. When such differences persist, the higher-risk individuals tend to die earlier, so the composition of surviving members changes. The observed force of mortality therefore represents a moving mixture of risks, rather than the hazard of an unchanged representative individual.
The mathematical significance of this insight was made explicit by Vaupel, Manton, and Stallard (1979) in their analysis of individual frailty. Their research linked unobserved heterogeneity to demographic life tables, mortality differentials, and interpretations of aging. Hougaard (1986) developed survival models for heterogeneous populations based on tractable frailty distributions. Subsequent work expanded the model space to correlated frailty, alternative stochastic processes, and the identification problems that arise when different structural mechanisms produce the same observable survival curve. The books by Aalen, Borgan, and Gjessing (2008), Survival and Event History Analysis, and Wienke (2011), Frailty Models in Survival Analysis, document the breadth of established approaches.
Insurance adds separate selection stages. Applicants decide whether, when, and how much coverage to buy. Underwriters and products determine who is admitted and how risks are classified. Policyholders can change coverage, convert term contracts, surrender policies, or cease paying premiums. Annuity purchasers are selected by preferences and expected longevity, while life-claim portfolios are affected by behavior, contestability, product features, and lapse. These decisions condition the observed distribution of mortality risk. The empirical literature on the UK annuity market illustrates the contractual dimension: Finkelstein and Poterba (2002) found selection patterns in voluntary and compulsory annuities, while their later policyholder-level analysis found mortality relationships with certain annuity features rather than with annuity size alone. The implication for life portfolios is methodological: product features and selection channels must be identified before pooling lives into a single inferred mortality process.
This Research Dossier investigates unobserved mortality differences and their interactions with measured underwriting classifications, cohort dynamics, and insurer liabilities. Its primary Science is Life Insurance, with disciplinary connections to Actuarial Science, Insurance Statistics and Predictive Modeling, and Risk Theory. Life-contingent payout implications also connect the study to Annuities and Retirement Insurance, while death as a competing event links its identification questions to Long-Term Care and Disability Insurance. Broader accounts of life-table construction, mortality projection, reinsurance treaties, and general insurance regulation belong to those independent works. Here they are analyzed insofar as they change identification, interpretation, or valuation of heterogeneous insured mortality.
The evidence consists of original methodological research, official demographic methods, actuarial experience reports and publicly described data, accounting standards, and actuarial practice materials. No insurer-level policy records, purchased Society of Actuaries data packages, or restricted mortality microdata were obtained for this work. Mathematical examples below are explicitly synthetic and reproducible; they express model implications, not empirical estimates for any insurer or country. This separation determines the scope of the conclusions.
Definitions, Historical Foundations, and the Research Object
Individual mortality, marginal mortality, and frailty
For an individual exposed to the event of death, a hazard or force of mortality is the instantaneous death rate conditional on being alive at time (t). An individual survival function (S(t)) is the probability that the lifetime exceeds (t). The cumulative hazard (H(t)) is the time integral of the hazard, with (S(t)=\exp[-H(t)]) when an absolutely continuous event-time distribution is assumed. In an insurance application, time can denote attained age, years since issue, or duration since an observable entry event; those clocks have different meanings and are never interchangeable without an explicit conversion.
Mortality heterogeneity refers to dispersion of death hazards or survival prospects across persons at a common analytically defined time and conditional on the selected covariates. Some dispersion is recorded in underwriting or demographic variables; the residual may be approximated by latent variables. Frailty is a stochastic representation of this residual: in the basic proportional-hazard formulation, the individual hazard equals a baseline hazard multiplied by a nonnegative random factor. A frailty variable is not a directly observed physical or clinical measure, a diagnosis, or a synonym for functional geriatric frailty. Its interpretation depends on the model, covariates, entry mechanism, and assumed distribution.
A conditional hazard describes a particular frailty level. A marginal or portfolio hazard averages over the risk distribution among those still surviving and observable at the relevant time. The arithmetic mean of individual hazards at entry differs from the corresponding mean among survivors after positive-risk time has elapsed. A portfolio thus carries an evolving risk composition even if every individual's underlying hazard multiplier remains constant.
Observed heterogeneity is likewise distinct from shared dependence. Independent fixed frailties can make individual mortality distributions unequal while leaving death outcomes independent conditional on covariates and model parameters. A common latent factor affecting families, employers, neighborhoods, occupational groups, or an insurer's entire portfolio may induce dependence. Mortality catastrophes and systematic time shocks add further dependence mechanisms. The statistical and capital consequences of these cases differ materially.
Documentary development
The demographic problem of age-dependent mortality long predates formal frailty models, but frailty terminology sharpened the distinction between individual and aggregate dynamics. Vaupel, Manton, and Stallard (1979) demonstrated that life tables for heterogeneous populations require attention to individual risk endowments and population selection. Hougaard (1986) derived families of frailty distributions with useful Laplace transforms. Yashin, Vaupel, and Iachine (1994) discussed observationally equivalent aging models: alternative assumptions about the latent risk process can match the same old-age mortality data, making model interpretation an independent question from curve fit.
Survival-statistical foundations include Kaplan and Meier's (1958) product-limit estimator, developed to estimate survival with incomplete follow-up, and Cox's (1972) regression models and life tables, which permit regression effects relative to an unspecified baseline hazard. These methods solve different inferential problems. A product-limit survival curve describes the observed cohort under censoring assumptions. A Cox regression describes conditional covariate associations under a hazard structure. Neither automatically reveals an unobserved frailty distribution; even correct adjustment for observed covariates leaves possible latent heterogeneity and selection.
The econometric treatment of unobserved duration heterogeneity provides a complementary warning. Elbers and Ridder (1982) examined true and spurious duration dependence and established identification results under particular proportional-hazard and covariate conditions. Their results should not be transposed into a claim that arbitrary insurance frailty mixtures are identifiable from any single aggregate life table. Identification depends on the family of models, covariate support, observation scheme, and restrictions placed on the baseline and latent distribution.
The insured risk population
An insurer's in-force cohort is the set of policies or insured individuals meeting a stated exposure definition at a valuation or experience date. Policy counts, insured lives, face-amount exposure, and annuity income exposure are different denominators. A death claim may terminate one policy or affect several coverages; a single individual may appear across multiple contracts. A study must define whether its unit is person, contract, life-year, amount at risk, or benefit-payment exposure. Aggregating unlike units can generate misleading patterns even when every data record is accurately coded.
Insurance entry introduces at least four observable dates: birth, application, underwriting decision, and policy issue or coverage commencement. Calendar year adds another axis. For a cohort of newly insured persons, the model may need attained age (x), issue duration (d), issue year (y), calendar period (c), and product class (p). Since (c=y+d) for annual-duration conventions, model terms based on these clocks can be linearly dependent. A latent frailty term absorbs their omitted effects if these dimensions are poorly specified.
National mortality data are essential contextual evidence, but they measure a different population. The Human Mortality Database (HMD) explains that its national mortality rates and life tables derive from vital death registration, population counts, births, and exposure estimation. It does not purport to identify underwriting selection in proprietary life-insurance cohorts. The HMD Methods Protocol, Version 6 documents reconstruction methods and the importance of exposure timing, including refinements to monthly-birth information. A consistent national baseline is informative for external comparison; it must be mapped to insurance age conventions, geography, observation periods, and selection structure before use as an insurer assumption.
Research boundaries and levels of evidence
This dossier separates four levels of evidence. First, analytic identities follow mathematically from stated hazard and distribution assumptions. Second, historical and methodological claims are grounded in identifiable original articles or methodological books; where full text is restricted, the record is limited to publisher metadata and abstract. Third, insured mortality patterns reported by professional bodies are evidence about the specified contributed populations, underwriting vintages, and observation windows, rather than universal population laws. Fourth, applications to any single insurer require its actual exposure, claim, lapse, and underwriting data, which are not present here. A conclusion on one level cannot be promoted to the next merely because the same equation or mortality label appears.
Frailty Model Structure and Mathematical Derivations
Conditional survival and the mixture identity
Let (T) denote remaining lifetime measured from a defined entry age or date, and let (Z\geq 0) be a time-invariant frailty with mean (E[Z]=1) at that origin. For a given observed covariate vector (X), assume the conditional hazard
[ \lambda(t\mid Z=z,X)=z,\lambda_0(t)\exp(\beta^\top X). ]
At fixed (X), write (H_X(t)=\int_0^t\lambda_0(u)\exp(\beta^\top X),du). Conditional survival is (S(t\mid z,X)=\exp[-zH_X(t)]). Integrating the unobserved frailty yields
[ S_M(t\mid X)=E_Z{\exp[-ZH_X(t)]}=L_Z{H_X(t)}, ]
where (L_Z) is the Laplace transform of the frailty distribution. Differentiation gives
[ \lambda_M(t\mid X)=\lambda_0(t)\exp(\beta^\top X)\frac{E[Z\exp(-ZH_X(t))]}{E[\exp(-ZH_X(t))]}. ]
The fraction is exactly the conditional mean (E[Z\mid T>t,X]). Survivors are weighted in proportion to (\exp[-zH_X(t)]), so larger (z) values lose representation faster. This weighting is the precise survivor-selection mechanism. It requires persistent relative frailty and a specified mortality time origin; if frailty changes over time, the conditional process and interpretation must be modified.
For nondegenerate fixed frailty with finite second moments, differentiation with respect to cumulative baseline hazard gives
[ \frac{d,E[Z\mid T>t,X]}{dH_X(t)}=-\operatorname{Var}(Z\mid T>t,X)\leq 0. ]
The result follows by differentiating the ratio of the first and zeroth exponentially weighted moments. As cumulative hazard increases, mean frailty among survivors weakly declines. The marginal mortality hazard itself can still increase with attained age: it is the product of a potentially rising baseline hazard and a declining composition factor. Consequently, an observed deceleration of aggregate hazard is consistent with selection, but does not alone establish that selection caused it.
The gamma model
A common tractable assumption is (Z\sim\operatorname{Gamma}(1/\theta,\theta)), where the second argument denotes scale and (\theta>0) is the variance; (E[Z]=1). The Laplace transform and marginal survival are
[ L_Z(s)=(1+\theta s)^{-1/\theta},\qquad S_M(t\mid X)=(1+\theta H_X(t))^{-1/\theta}. ]
The marginal hazard and conditional mean frailty among survivors become
[ \lambda_M(t\mid X)=\frac{\lambda_0(t)\exp(\beta^\top X)}{1+\theta H_X(t)},\qquad E[Z\mid T>t,X]=\frac{1}{1+\theta H_X(t)}. ]
The limit as (\theta\to0) is the homogeneous model (S_M(t)=\exp[-H_X(t)]). The parameter (\theta) measures latent dispersion relative to the declared covariate information and entry population. A different recorded risk classification can change its estimated value without any biological change in insured lives. A gamma law is therefore a modeling restriction chosen for interpretability and algebraic convenience, not an empirically universal distribution of susceptibility.
Baseline hazards, mortality improvements, and observational equivalence
Under a Gompertz-type baseline (\lambda_0(t)=a\exp(bt)), (a,b>0), cumulative hazard is (H_0(t)=a(\exp(bt)-1)/b). The gamma-mixture observed hazard is then
[ \lambda_M(t)=\frac{a\exp(bt)}{1+\theta a(\exp(bt)-1)/b}. ]
For constant (\theta>0), the observed hazard approaches (b/\theta) as (t\to\infty), while the conditional Gompertz hazard continues to rise. This limiting result concerns an idealized mathematical curve projected indefinitely; it does not establish a human longevity limit, the existence of an empirical mortality plateau, or the suitability of that asymptote at extreme insurance ages. Choice of tail model, recording error, censoring, and cohort depletion remain decisive.
A stronger identification problem is constructive. Suppose one observes only a valid, smooth marginal survival function (S_M(t)). A homogeneous model reproduces it exactly with cumulative hazard (H_{\mathrm{hom}}(t)=-\log S_M(t)). For every selected gamma variance (\theta>0), an alternative baseline cumulative hazard
[ H_{\theta}(t)=\frac{S_M(t)^{-\theta}-1}{\theta} ]
also reproduces that same observed curve because ((1+\theta H_{\theta}(t))^{-1/\theta}=S_M(t)). If (S_M) is decreasing, (H_\theta) is nondecreasing. Thus a single marginal survival curve is consistent with a continuum of baseline–frailty combinations. This algebraic observation illustrates precisely why selecting a gamma distribution, specifying the baseline, or adding credible cross-sectional/longitudinal information is necessary for identification. It reinforces the substantive message of equivalent mortality models, while remaining an elementary consequence of established Laplace-transform methods rather than a new doctrine.
Time-dependent frailty and alternative distributions
Time-invariant multiplicative frailty is one assumption, not a general law of individual aging. In lognormal frailty models, the log multiplier is Gaussian, producing a different distributional tail and a Laplace transform generally evaluated numerically. In inverse-Gaussian and positive-stable specifications, the implied survivor weighting differs. Shared, correlated, nested, and time-dependent frailty models permit clusters or evolving health states, but add new parameters and identification constraints. Hougaard's distributional construction and Wienke's comparative treatment of frailty families support comparing model classes rather than taking a single fitted gamma variance as a directly measured characteristic.
A frailty term may capture unmeasured causes, measurement error, omitted risk classification, contextual exposures, or misspecified time dependence. These sources imply different causal interpretations. If an exposure such as smoking changes after policy issue, a fixed frailty may incorrectly attribute evolving individual risk to a permanent latent effect. If health shocks are correlated within a cohort, an independent frailty model understates dependence. If death-certificate quality differs at extreme ages, a smooth fitted tail may absorb ascertainment artifacts. Model adequacy therefore requires tests directed at the observation mechanism as well as fit to survival totals.
Conditional and marginal proportionality
Consider two covariate classes, each with proportional conditional hazards (z\lambda_0(t)\exp(\beta X)) and the same entry frailty distribution. Their marginal gamma hazards have ratio
[ \frac{\lambda_M(t\mid X=1)}{\lambda_M(t\mid X=0)}=\exp(\beta)\frac{1+\theta H_0(t)}{1+\theta\exp(\beta)H_0(t)}. ]
Even though the conditional hazard ratio (\exp(\beta)) is constant, the marginal ratio generally moves toward one as cumulative exposure increases. Inference that equates such attenuation with biological convergence of risk classes confuses within-individual hazard behavior with changing class composition. The derivation also clarifies why a Cox regression without latent frailty may exhibit time-varying aggregate contrasts. Model selection must distinguish genuine time-varying covariate effects from survival selection and unobserved baseline differences.
Survivor Selection, Insurance Entry, and Portfolio Persistence
Four selection gates
In life insurance, selection arises successively at the application, underwriting, survival, and contract-persistence stages. Application selection depends on preferences, private information, affordability, distribution access, and coverage needs. Underwriting selection depends on medical evidence, occupational or behavioral classifications, product eligibility, and acceptance. Survival selection removes individuals according to their death hazard. Persistence selection conditions the experience portfolio on policies remaining in force. A fifth gate may be added when the reporting system excludes records because of data linkage, reinsurance reporting, incomplete claim development, or delayed notification. These mechanisms are analytically distinct even when they act on the same lives.
The first two gates determine the distribution of (Z) at policy issue. If the general-population frailty distribution is (f_{\mathrm{pop}}(z)) and a risk with frailty (z) applies and is accepted with probability (a(z)), then the issued portfolio has density proportional to (a(z)f_{\mathrm{pop}}(z)). Application and underwriting cannot generally be reconstructed from insured death counts alone because rejected, never-applying, and unobserved lives are absent. Aggregate comparison with population death rates is therefore a comparison between distributions conditioned on different admission mechanisms.
Survival selection reweights the issue distribution by (\exp[-zH(t)]). Persistence further changes the distribution. Let (L) be lapse or other contract exit time, and let (G_L(t\mid z,X)=P(L>t\mid Z=z,X)), with death and exit conditionally independent given the modeled latent and observed covariates for this expression. The frailty density among surviving, in-force lives is then
[ f(z\mid T>t,L>t,X)\propto f(z\mid\mathrm{issued},X)\exp[-zH_X(t)]G_L(t\mid z,X). ]
The mortality hazard of the in-force population is consequently the baseline hazard multiplied by the mean frailty under both weights. The usual survivor-only formula is recovered when persistence is independent of frailty, conditional on measured characteristics. If healthier insureds surrender at a different rate from less healthy insureds, a single competing-event adjustment need not restore independence. Selection on price, expected benefits, current health, and private information may all produce informative lapses.
Event histories and risk sets
Let (N_i(t)) record an observed death event and (Y_i(t)) indicate whether insured life (i) is alive, in force, and at risk just before (t). For a contract-based death study, a natural estimator of the central death rate in an age-duration-calendar cell (g) is
[ \widehat m_g=\frac{D_g}{E_g},\quad D_g=\sum_i\int_g dN_i(t),\quad E_g=\sum_i\int_gY_i(t),dt. ]
This ratio estimates a hazard averaged over the observed risk set, with further qualifications for time variation and exposure measurement. It is not automatically a one-year death probability (q_x); translating a central rate into a death probability requires an assumption about deaths within the interval. The distinction is material when converting insurer experience into pricing or valuation tables. The HMD exposure and rate definitions make an analogous denominator distinction for general populations.
A person who enters an experience study after a covered age or time is left truncated: only persons surviving to entry can contribute observations. Under independent left truncation, the survival contribution is conditional on survival to entry, (S(t_i)/S(a_i)), not unconditional survival from birth. A contract terminating for reasons other than death may be right censored for a lifetime study, but ordinary product-limit estimation assumes censoring independent of the unobserved failure time at the relevant conditioning level. Kaplan and Meier's incomplete-observation framework remains foundational precisely because the censoring process must be examined. Under health-dependent lapse, a censoring assumption that ignores latent health can produce selection bias.
Administrative observation boundaries, deferred claim reporting, beneficiary delays, migration between administration systems, and reinstatements further complicate risk sets. Policies surrendered and later reinstated cannot be represented by a single uninterrupted exposure. Underinsured lives receiving multiple overlapping contracts require deduplication or explicit correlation accounting. Face-amount-weighted death ratios answer a different financial question from life-count ratios; a change in average face amount can move the former without changing the latter. All such decisions must be specified before a latent frailty distribution is estimated.
Duration of insurance and underwritten selection
Early policy durations commonly differ from later durations because an underwritten issue cohort is filtered through health-selection rules. A declining advantage of newly selected lives over ultimate insured mortality can reflect the aging of the underwritten cohort, the disappearance of initial selection effects, differential lapses, and shifts in product distribution. Modeling all duration effects as unobserved frailty is a classification error. The Society of Actuaries' 2015 Valuation Basic Tables distinguish relevant mortality dimensions including sex, smoking status, and table conventions. The actuarial table is a calibrated reference rather than a claim that underwriting risk classes are biologically homogeneous.
The SOA 2019 Individual Life Mortality Experience Report, published in October 2024 and covering observation years 2012–2019, describes actual-to-expected analyses against the 2015 VBT and identifies shifts in age, duration, product, underwriting, and company mix as interpretation problems. Its source material explicitly describes data-validation effort after a change in the statistical data agent. The report provides documented insurer-contributed experience; it does not by itself estimate a universal gamma frailty variance for every life portfolio.
The SOA's U.S. Individual Life Mortality Improvement Analysis similarly addresses non-biometric changes in business composition. Its public indexed material identifies face amount band, risk class, plan type, issue era, and duration as covariates considered when attempting to normalize observed improvement. A practical deduction follows: when underlying segment weights change, an aggregate experience ratio can change even if the segment-specific mortality mechanisms remain constant. Normalization may reduce this compositional influence, while residual variation still requires sensitivity to omitted health or socioeconomic factors.
Annuity selection and the direction of cash-flow exposure
Annuity beneficiaries receive survival-contingent payments. People expecting longer lives may value such products more strongly, and product designs can attract different combinations of mortality prospects. The UK research of Finkelstein and Poterba found selection differences between voluntary and compulsory annuity markets and links between some product features and subsequent longevity. Their 2004 contract-characteristic evidence also illustrates an inferential discipline: selection can appear along one contract feature and be weak or absent along another. A positive result about one product characteristic does not establish every plausible selection channel.
For a life-insurance death benefit, earlier mortality generally accelerates claims, subject to contract terms, lapses, and reserves. For a lifetime annuity, lower mortality generally lengthens benefit payment streams. For contracts combining death and survival benefits, the direction depends on the benefit design. The same modeled frailty distribution can therefore generate opposite financial sensitivities across two products. Pooling mortality experience across those products without reconditioning the distribution at issue or purchase would confound underlying longevity with contractual selection.
The 2026 SOA–LIMRA payout annuity experience study reports 3.1 million contract-years of exposure and 143,190 deaths across the years 2020–2024; its detailed data package is subscription-based. The 2026 deferred annuity experience study reports 80 million contract-years and more than 1.4 million deaths over 2015–2024, again with detailed analysis behind a paid data package. These are studies of different products and observation populations. Their reported aggregate sizes do not permit a direct frailty estimate or a contract-level comparison without the underlying data, inclusion rules, and weighting definitions.
Selection is not one parameter
Survivor selection, underwriting selection, adverse selection, lapse selection, and observational exclusion have distinct time directions and counterfactual populations. An observed attenuation of mortality among survivors may arise from within-cohort depletion of high-risk lives, while an increasing death rate among remaining policyholders may arise from unhealthy lives keeping coverage when healthy lives lapse. Product discontinuation can abruptly alter the mix by issue year; an adverse public-health event can shift baseline hazards across every frailty class. A defensible model tracks these processes separately where the data permit and includes residual uncertainty where they do not.
Portfolio Tail Implications and Actuarial Valuation
A reproducible synthetic example
Consider a newly constituted cohort with a constant baseline force of mortality of (\lambda_0=0.02) per year, independent gamma frailty with (E[Z]=1), and variance (\theta=0.5). The choice is illustrative; no observed insurance data are used. Over twenty years the baseline cumulative hazard is (H_0(20)=0.4). A homogeneous cohort has survival (\exp(-0.4)=0.670320). The gamma-frailty cohort has survival ((1+0.5\times0.4)^{-2}=0.694444). Conditional mean frailty among its survivors is (1/1.2=0.833333), and its marginal hazard at year 20 is (0.02/1.2=0.016667) per year.
For orientation, the model's year-20-to-year-21 conditional death probability is
[ q_{20,21}^{M}=1-\frac{S_M(21)}{S_M(20)}=1-\left(\frac{1.21}{1.20}\right)^{-2}=0.016461. ]
The homogeneous constant-hazard one-year death probability is (1-\exp(-0.02)=0.019801). The difference follows solely from the initial latent mixture and survivor weighting; it is not evidence of mortality improvement, changes in medical care, or an observed insurer advantage.
| Years after entry | Homogeneous survival | Gamma-mixture survival | Gamma-mixture marginal hazard per year |
|---|---|---|---|
| 0 | 1.000000 | 1.000000 | 0.020000 |
| 10 | 0.818731 | 0.826446 | 0.018182 |
| 20 | 0.670320 | 0.694444 | 0.016667 |
| 40 | 0.449329 | 0.510204 | 0.014286 |
This stylized comparison must be read with its assumptions: a positive frailty variance is imposed, baseline mortality is held constant, and no one lapses or enters after inception. It isolates the mathematical selection effect, rather than representing real human mortality at those ages.
The sensitivity at twenty years with the same cumulative baseline hazard is:
| Initial frailty variance (\theta) | Survival at year 20 | Marginal hazard at year 20 |
|---|---|---|
| 0.00 | 0.670320 | 0.020000 |
| 0.25 | 0.683013 | 0.018182 |
| 0.50 | 0.694444 | 0.016667 |
| 1.00 | 0.714286 | 0.014286 |
| 2.00 | 0.745356 | 0.011111 |
Holding the baseline fixed while changing frailty variance produces large tail differences. In a fitted portfolio the baseline would often be re-estimated when variance changes; the table is a conditional sensitivity test, not an empirical ranking of model fits. The distinction matters especially when a tail extrapolation is produced from a short or narrow experience window.
Death benefits and life-contingent present values
For a death benefit (B(t)) paid at death within horizon (n), with deterministic discount factor (v(t)) and model mortality (S(t)), the expected present value in a single-life representation is
[ \mathrm{EPV}_{\mathrm{death}}=\int_0^n v(t)B(t)S(t)\lambda_M(t),dt. ]
For a continuous lifetime annuity paying (b(t)) while the insured remains alive, the corresponding expected present value is
[ \mathrm{EPV}_{\mathrm{annuity}}=\int_0^n v(t)b(t)S(t),dt. ]
These identities require the benefit and coverage conditions to match the modeled event process. Insurance calculations with mortality-linked surrender, policy loans, rider conditions, premium holidays, contingent beneficiaries, taxes, and stochastic financial variables require a fuller cash-flow specification. Frailty alone does not determine premium rates, reserves, or capital; it alters one mortality component in the valuation system.
The synthetic survival table illustrates the directional issue. With a fixed annuity payment and fixed discounting, larger (S(t)) increases expected survival-contingent payments at each positive (t). For a pure death benefit payable within a fixed twenty-year window with no other features, larger twenty-year survival means a lower probability of death by that horizon. For lifetime death cover the timing and discounting of later payments remain relevant. The same mixture can raise a longevity-tail measure and reduce near-term death claims relative to a homogeneous model calibrated to the same entry hazard. The financial effect must be stated contract by contract.
Tail uncertainty and dependence
Frailty heterogeneity is frequently mistaken for systemic mortality dependence. The distinction can be demonstrated without fitting data. Suppose two different insured persons share the same gamma frailty (Z) conditionally on which their survival times are independent, with cumulative individual baseline hazard (H). The probability both survive is (E[\exp(-2ZH)]=(1+2\theta H)^{-1/\theta}). Their individual marginal survival probabilities are each (S=(1+\theta H)^{-1/\theta}), so the survival-indicator covariance is
[ \operatorname{Cov}(I(T_1>t),I(T_2>t))=(1+2\theta H)^{-1/\theta}-(1+\theta H)^{-2/\theta}. ]
With (\theta=0.5) and (H=0.4), joint survival is (0.510204), while the square of marginal survival is (0.482253); covariance is approximately (0.027951), corresponding to a survival-indicator correlation of about (0.131725). By contrast, with two independent gamma frailties, joint survival equals the product of marginal survivals, and that covariance is zero. Both models have identical single-life marginal survival. Only the shared-frailty model produces this particular cross-life dependence. Hence a one-life mortality table cannot identify the diversification properties of the portfolio.
For (n) equal benefit exposures with survival or death indicators (I_i), the variance of aggregate claims is the sum of individual variances plus twice the sum of pairwise covariances. When individual frailties and events are independent, pooling reduces relative random variation as exposure grows under appropriate regularity conditions. When a common frailty or systematic mortality shock creates positive pairwise covariance, a component remains after diversification. An insurer's capital calculation must distinguish persistent common susceptibility, temporary epidemic or catastrophe effects, uncertain future mortality improvements, and estimation uncertainty. These factors are not interchangeable or recoverable from the variance of the marginal frailty distribution alone.
A common gamma frailty shared by every policyholder is an extreme illustrative structure and often implausible as a literal population model. More realistic risk studies may include family or geographic clusters, common age-specific shocks, country-level trends, asset-mortality dependence, and catastrophe scenarios. The covariance example is a proof of the importance of dependence assumptions, not a calibrated estimate of insurer tail losses.
Measurement bases and regulatory boundaries
Financial-reporting and solvency applications use different definitions and objectives. The official IFRS 17 Insurance Contracts overview describes recognition and measurement of groups of insurance contracts using current fulfillment cash-flow estimates, an explicit risk adjustment for nonfinancial risk, and a contractual service margin in the applicable circumstances; IFRS 17 applies to annual reporting periods beginning on or after January 1, 2023. A mortality frailty model may inform cash-flow probabilities, but the model's gamma variance is not itself an IFRS 17 risk adjustment or a contractual service margin. The latter quantities are determined under the standard's separate requirements.
For US principle-based life reserving, ASOP No. 52 concerns professional services under the NAIC Valuation Manual. Its discussions of relevant insurer experience, credibility, and sensitivity must be applied within the specified legal and product scope. The standard is a professional practice source, not a universal mortality parameter or rule binding foreign markets. ASOP No. 56, Modeling, addresses model selection, use, governance, and evaluation for actuarial services within its scope. A frailty extension should therefore have a documented purpose, validation strategy, relevant data, known limitations, and change controls.
For England and Wales projections, the Continuous Mortality Investigation publishes a distinct CMI Mortality Projections Model based on user-specified assumptions and population mortality experience. CMI Working Paper 211 describes CMI_2025, released in March 2026 and calibrated to mortality through December 2025. Working Paper 216 is an October 2026 consultation on prospective CMI_2026 methods; it is not an already issued 2026 projection model. A population improvement model and a latent insured-selection model answer different questions and require coherent combination before long-duration insurer use.
Portfolio segmentation and financial relevance
A practical portfolio tail analysis must identify the financial unit, horizon, and benefit schedule before selecting a frailty structure. For term death cover, concentrated short-horizon mortality, lapses, and face amounts may dominate. For whole-life or universal life contracts, continuing coverage, cash values, guarantees, and policyholder behavior alter future at-risk exposures. For annuities, survival at advanced ages and the interaction of payment size with longevity are central. For reinsurance, treaty attachment, reporting, and aggregation govern the transfer of adverse results. Cross-contract and common-cause dependence is crucial whenever portfolio capital depends on the far tail of total claims.
An insurer can observe acceptable aggregate mortality fit while materially misestimating uncertainty at the extremes. The construction is straightforward: two models can share the same marginal (S(t)) and differ in their joint frailty distribution, hence their claim count variance. Even if marginal survival is perfectly reproduced, stress tests and dependence assumptions are necessary for capital-oriented questions. Conversely, a complex correlated frailty model with little data support may create false precision. Complexity is justified by identifiable evidence and its influence on the specific decision.
Empirical Evidence, Comparative Research, and Validation
Evidence classes and the limits of available data
A defensible empirical study starts with data provenance. Insurer administrative records contain policy and coverage histories, exposure timing, product definitions, issue attributes, claims, and recorded causes of exit. Public population mortality is constructed from registration and population data. Neither source alone encodes all unobserved underwriting characteristics. The HMD population documentation and methods should be used to assess general-mortality comparability, completeness, source adjustments, and cohort coverage. Its standardized rates cannot identify the distribution of (Z) among insureds without an external mapping or additional data.
Actuarial experience studies vary in their units and data access. The SOA's individual-life mortality research index catalogs successive US insured-life studies and table reports; the 2024 mortality-improvement analysis examines compositional factors and normalization. Holland (2018), in Beyond Actual to Table, explicitly addresses distortion when heterogeneous experience segments are pooled for traditional actual-to-expected comparisons. Such sources substantiate the relevance of segmentation but do not provide a public estimate of a transferable latent frailty law for a particular company.
The 2026 payout and deferred annuity studies record substantial industry-contributed exposures. Some detailed materials are commercially restricted. This dossier uses only their publicly documented scope and summary totals; it does not infer unreported mortality rates or model coefficients from paywalled outputs. For country-to-country comparisons, changes in the observation regime, age conventions, underwriting rules, annuitization markets, benefit design, and coverage definitions can be at least as important as changes in the biological component of mortality risk.
A staged empirical design
A rigorous insured-life study would begin by creating a person-level or policy-level event history, identifying the unit of exposure and retaining all issue, coverage, lapse, reinstatement, and death dates. Record the underwriting vintage, product form, face amount, risk class, smoking status where legally and empirically appropriate, attained age, sex or other relevant classifications, and calendar period. Resolve duplicate insured persons if the estimand is individual mortality. Test for deaths occurring outside policy dates and for inconsistencies in exposures. A documented and versioned data dictionary is a prerequisite to interpreting the results.
Next, construct nonparametric death and exit summaries, with age, duration, and issue-year stratification. Compare count-weighted and amount-weighted experience, since changes in face mix can imitate mortality trends. Estimate a baseline hazard with observed covariates and check residual age-duration-calendar patterns. Examine proportional hazards assumptions and entry truncation before adding frailty. The purpose of these stages is to determine which observable sources of heterogeneity can be measured and explained directly.
A frailty model should then be fitted only within an explicit identification design: for example, a prespecified Gompertz or other parametric baseline plus gamma frailty; a cluster-structured frailty model with repeated related lives; or a carefully regularized multivariate event-history model. Estimators may involve integrated likelihood, penalized methods, simulation, or Bayesian computation; each requires convergence diagnostics and sensitivity to latent distribution choices. Comparisons should include a no-frailty baseline and alternative frailty distributions. Model complexity should be evaluated by out-of-sample predictive performance and decision-relevant calibration, not simply better in-sample likelihood.
Potential diagnostic comparisons include observed versus predicted deaths by age-duration-vintage cells, death and exposure totals, survival residuals, mortality ratios by original underwriting class, and stability under leave-one-vintage-out or leave-one-company-out evaluations. Tail extrapolations need stress ranges for frailty variance, baseline functional form, mortality trend, extreme age measurement, and lapse dependence. Bootstrap, posterior predictive, or other uncertainty intervals should account for the actual data structure, including relevant clustering. If model estimates vary strongly between a gamma and lognormal specification, the difference belongs in the disclosed valuation uncertainty rather than being eliminated by selecting the best-fitting point estimate alone.
Identifiability, validation, and falsifiability
An observed aggregate hazard shape cannot establish the existence or quantitative size of permanent individual frailty because of the algebraic equivalence demonstrated above. Identification becomes more plausible when externally supported baseline restrictions, observed covariates with sufficient variation, repeated events or shared groups, independent risk markers, or longitudinal outcomes add information. Yet every additional instrument or measurement requires justification: an apparent health predictor correlated with insurance lapse may affect observability as well as mortality. Selection and causal pathways must be examined together.
Validation should test hypotheses that could fail. If a fitted model predicts marked risk-class convergence, does convergence remain after equalizing issue-year and product mix? If an age-tail plateau is attributed to frailty depletion, does it persist under plausible age-misstatement adjustments or alternative baseline hazards? If an estimated frailty variance is stable in one insurer, is it stable by underwriting vintage, currency, country, and claim reporting system? Does the model reproduce conditional survival and deaths, or only the aggregate table? Are policy exits independent of unobserved mortality after adjusting observed characteristics? A positive answer must arise from actual tests, not the frailty equation itself.
Elbers and Ridder's identification analysis, alongside Yashin, Vaupel, and Iachine's model-equivalence study, demonstrates why a highly accurate empirical curve fit is not a unique scientific explanation. Testing identification assumptions is an indispensable part of the scientific review, particularly when frailty results are intended to inform economic or regulatory decisions.
Comparative geographic and institutional interpretation
The United States provides public Society of Actuaries valuation tables and insured mortality experience summaries, subject to the particular contributing carriers and product definitions. The United Kingdom supplies distinctive evidence about annuitant selection and CMI mortality improvement models. International HMD populations provide harmonized demographic reference series. These evidence systems are complementary rather than directly exchangeable. The comparability task is to map exposure definitions, period, age, reporting, selection, coverage, contract benefit, and risk-class composition. A universal cross-jurisdiction frailty variance cannot be deduced from their coexistence.
Actual law and professional practice must be attributed to the applicable jurisdiction. US ASOP 52 concerns specified US life reserving work. IFRS 17 concerns in-scope insurance contracts under the applicable reporting framework, as adopted or required in the relevant jurisdiction. CMI parameters are projection assumptions under a UK-centered professional model, not universal regulatory mortality rates. An insurer in another jurisdiction would need to identify its own legal, accounting, and supervisory requirements before relying on the same analysis.
Information quality and ethical use
A latent frailty estimate is a residual statistical measure. It may be correlated with omitted socioeconomic, demographic, behavioral, or medical attributes, but a model coefficient cannot certify why an individual has an elevated hazard. Use in underwriting or consumer decisions implicates local anti-discrimination rules, privacy, data permissions, access constraints, and model transparency. Statistical residuals should not be treated as individually verified medical facts. Risk classifications need relevance, accuracy, applicable legal authority, and a documented decision purpose.
ASOP No. 23, Data Quality, distinguishes review of data from formal audit and sets professional expectations about appropriateness and data limitations for work in its scope. ASOP No. 56, Modeling, supports purposeful model governance. These professional references reinforce the importance of recording data provenance, limitations, and the decision for which an estimated frailty term is being used, without substituting for law applicable in a particular market.
Research limitations
The research reviewed public scholarly abstracts or metadata for several subscription-restricted original papers and books. Public SOA documents were examined through available descriptions and indexed text; the paid 2026 SOA–LIMRA experience packages were not obtained, and the PDF copies could not all be independently visually read back in this research environment. Consequently, the dossier does not claim a replication of reported intercompany numerical results. No proprietary insurer microdata, family-linked mortality records, national-law-by-national-law regulatory inventory, or fitted empirical frailty coefficients were produced.
The numerical tables and covariance illustration are exact under their stated models and rounded only for display. They cannot determine which frailty distribution best describes insured populations. The available verified material supports methodological claims, published study scope, and selected contemporary practice descriptions. It does not support a new empirical claim about global insured-life heterogeneity, a new actuarial model standard, or an assertion of a newly discovered biological mortality mechanism.
Scientific Synthesis and Research Agenda
The analytic structure of mortality heterogeneity
The evidence and derivations support a clear scientific synthesis. A mortality portfolio is a selected, observed, evolving population. An entry distribution of risk is modified by survival and, for active contracts, by persistence. Conditional hazards, marginal hazards, aggregate mortality trends, and joint mortality dependence are therefore separate objects. A latent frailty distribution is a useful representation of residual heterogeneity when linked to an identifiable model and an explicit observational population. Interpreting its fitted variance without reference to baseline restrictions, underwriting mix, or exits is statistically unstable.
This synthesis also clarifies the hierarchy of conclusions. The Laplace-transform representation establishes how survivor selection works under a multiplicative model. The gamma example quantifies a model-specific survival change. The observational-equivalence construction demonstrates a structural identification limit. The shared-frailty example proves that matching single-life mortality cannot by itself validate portfolio-tail covariance. Insured-life studies document relevant heterogeneity in underwriting and product data but do not supply the missing joint distribution of latent risks for an arbitrary portfolio.
An operational research-to-valuation crosswalk
For an insurer evaluating heterogeneous mortality, the analytical work should proceed through a reproducible chain. Define the decision and benefit contingent on death or survival. Specify the risk population, entry clock, coverage status, and financial exposure. Verify person and policy identifiers, death and lapse histories, and observation gaps. Estimate and validate the observable risk structure before attempting latent frailty. Evaluate no-frailty and multiple-frailty specifications under explicit identification restrictions. Test sensitivity of survival and cash flows to baseline, variance, lapse, and trends. Separately analyze dependence and adverse common shocks. Transfer the resulting distribution, with disclosed uncertainty, into the appropriate valuation or capital framework. Maintain model provenance and re-estimation criteria.
This is an operational synthesis of existing survival, demographic, and actuarial methods. Its value is in keeping the inferential object aligned with the insurer decision throughout the process. Each stage prevents a specific category error: comparing general and insured populations as if selected alike, interpreting amount-weighted claims as individual death rates, inferring frailty from a fitted aggregate curve alone, treating latent dispersion as common shock correlation, or translating actuarial cash-flow uncertainty directly into an accounting or regulatory capital measure.
Future research questions
A primary empirical priority is joint estimation of mortality and lapse using insurer-linked records in which coverage transitions and deaths after termination can be investigated, subject to privacy and permission constraints. A second is the stability of frailty estimates across underwriting vintages and product structures after explicit control of business composition. A third is investigation of shared frailty in linked lives and its effect on insurer-tail claims under realistic clustering. A fourth is assessing high-age extrapolations against age-validation and mortality-rate reporting quality. A fifth is quantifying whether the improvements in out-of-sample cash-flow prediction from frailty modeling justify additional parameters, data collection costs, and governance burdens.
Additional research should compare conditional and marginal projections under alternate mortality-improvement assumptions, and distinguish persistent latent susceptibility from age-varying individual states. Where detailed empirical data remain restricted, a transparent reporting framework should publish exposure definitions, observed-versus-predicted performance, model sensitivities, and limitations without disclosing personal health information. Such a practice would improve scientific reproducibility and the quality of risk communication while respecting portfolio confidentiality.
Conclusion
Mortality heterogeneity is fundamental to the interpretation of life-insurance experience because insured cohorts enter through selection processes and then evolve through death, survival, and contractual persistence. Multiplicative frailty models offer a mathematically coherent description of one dimension of this process. Their marginal survival curves are Laplace transforms of conditional survival; the distribution of frailty among survivors becomes progressively weighted toward lower latent risk. Under gamma frailty, the model admits transparent formulas for marginal mortality and the survivor-risk distribution.
The central inferential result is that marginal mortality data alone do not uniquely identify frailty variance or baseline mortality. Alternative homogeneous and heterogeneous hazard structures can generate identical observed survival, while alternative dependence structures with identical single-life survival can produce different aggregate claim tails. Accurate fit to life tables is therefore a necessary but insufficient test for actuarial valuation and capital analysis. Underwriting selection, issue duration, lapse behavior, face-amount weighting, cohort composition, data provenance, and projection assumptions must be evaluated explicitly.
Evidence from historical demographic research, survival-analysis methodology, US insured mortality studies, UK annuity-selection studies, and current actuarial and accounting sources supports this interpretation within their documented scopes. The actuarial consequence is a requirement for model-purpose discipline: evaluate death benefits, annuity payments, long-duration liabilities, and correlated loss tails according to their distinct contingent cash flows and legal measurement frameworks. The original synthesis developed here establishes a replicable interpretive and verification structure for these questions without creating a new theoretical doctrine or claiming an empirically universal frailty distribution. Empirical calibration remains specific to the insured population and evidence actually observed.
References
Aalen, O. O., Borgan, Ø., & Gjessing, H. K. (2008). Survival and Event History Analysis: A Process Point of View. Springer, Statistics for Biology and Health. https://doi.org/10.1007/978-0-387-68560-1
Actuarial Standards Board. (2016). *Actuarial Standard of Practice No. 23: Data Quality (December 2016 revision). Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2017). Actuarial Standard of Practice No. 52: Principle-Based Reserves for Life Products under the NAIC Valuation Manual. Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/principle-based-reserves-life-products-naic-valuation-manual/
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/modeling-3/
Continuous Mortality Investigation. (2026). CMI Mortality Projections Model: CMI_2025 (Working Paper 211, March 10, 2026). Institute and Faculty of Actuaries. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-working-papers/mortality-projections/cmi-working-paper-211
Continuous Mortality Investigation. (2026). CMI_2026 Consultation (Working Paper 216, October 15, 2026). Institute and Faculty of Actuaries. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-working-papers/mortality-projections/cmi-working-paper-216
Continuous Mortality Investigation. (2026). Mortality Projections (official model overview). Institute and Faculty of Actuaries. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-investigations/mortality-projections
Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202 (article; discussion follows). https://doi.org/10.1111/j.2517-6161.1972.tb00899.x
Elbers, C., & Ridder, G. (1982). True and spurious duration dependence: The identifiability of the proportional hazard model. The Review of Economic Studies, 49(3), 403–409. https://doi.org/10.2307/2297364
Finkelstein, A., & Poterba, J. (2002). Selection effects in the United Kingdom individual annuities market. The Economic Journal, 112(476), 28–50. https://doi.org/10.1111/1468-0297.0j672
Finkelstein, A., & Poterba, J. (2004). Adverse selection in insurance markets: Policyholder evidence from the U.K. annuity market. Journal of Political Economy, 112(1), 183–208. https://doi.org/10.1086/379936
Holland, B. D. (2018). Beyond Actual to Table: Models in Experience Studies. Society of Actuaries, Individual Life Experience Committee. https://www.soa.org/globalassets/assets/files/resources/research-report/2018/beyond-actual-table.pdf
Hougaard, P. (1986). Survival models for heterogeneous populations derived from stable distributions. Biometrika, 73(2), 387–396. https://doi.org/10.1093/biomet/73.2.387
Human Mortality Database. (2026, accessed November 14). Overview: Scope, Data, and Computation of Mortality Rates and Life Tables. University of California, Berkeley; Max Planck Institute for Demographic Research; French Institute for Demographic Studies (international collaborative database). https://www.mortality.org/Project/Overview
IFRS Foundation / International Accounting Standards Board. (2017, amended subsequently). IFRS 17 Insurance Contracts (official standards overview; effective annual reporting periods beginning January 1, 2023). IFRS Foundation. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. https://doi.org/10.1080/01621459.1958.10501452
Society of Actuaries. (2015; associated report updated 2018). 2015 Valuation Basic Report and Tables. Society of Actuaries. https://www.soa.org/resources/experience-studies/2015/2015-valuation-basic-tables
Society of Actuaries. (2026). Individual Life—Mortality Experience Studies (research catalog). Society of Actuaries Research Institute. https://www.soa.org/research/topics/indiv-mort-exp-study-list/
Society of Actuaries. (2026). Predictive Analytics Experience Studies (research catalog listing Holland 2018). Society of Actuaries. https://www.soa.org/research/topics/pred-analytics-exp-study-list/
Society of Actuaries Individual Life Experience Committee. (2024a). 2019 Individual Life Insurance Mortality Experience Report. Society of Actuaries Research Institute, October 2024; observation period 2012–2019. https://www.soa.org/globalassets/assets/files/resources/research-report/2024/ilec-mort-main.pdf
Society of Actuaries Individual Life Experience Committee, Mortality Improvement Analysis Subgroup. (2024b). U.S. Individual Life Mortality Improvement Analysis. Society of Actuaries Research Institute, May 2024. https://www.soa.org/49b043/globalassets/assets/files/resources/research-report/2024/ind-life-mort-analysis.pdf
Society of Actuaries Research Institute and LIMRA. (2026a). 2020–2024 Individual Payout Annuity Mortality Experience Study. Society of Actuaries. https://www.soa.org/resources/experience-studies/2026/2020-24-individual-payout/
Society of Actuaries Research Institute and LIMRA. (2026b). 2015–2024 Deferred Annuity Mortality Experience Study. Society of Actuaries; released September 17, 2026. https://www.soa.org/resources/experience-studies/2026/2015-24-deferred-annuity-mort/
Vaupel, J. W., Manton, K. G., & Stallard, E. (1979). The impact of heterogeneity in individual frailty on the dynamics of mortality. Demography, 16(3), 439–454. https://doi.org/10.2307/2061224
Wienke, A. (2011). Frailty Models in Survival Analysis. Chapman & Hall/CRC Biostatistics Series, CRC Press. https://www.routledge.com/Frailty-Models-in-Survival-Analysis-1st-Edition/Wienke/p/book/9780429139604
Wilmoth, J. R., Andreev, K., Jdanov, D., Glei, D. A., & Riffe, T., with contributors. (2025). Methods Protocol for the Human Mortality Database, Version 6, revised August 5, 2025. Human Mortality Database. https://www.mortality.org/File/GetDocument/Public/Docs/MethodsProtocolV6.pdf
Yashin, A. I., Vaupel, J. W., & Iachine, I. A. (1994). A duality in aging: The equivalence of mortality models based on radically different concepts. Mechanisms of Ageing and Development, 74(1–2), 1–14. DOI: 10.1016/0047-6374(94)90094-9. https://www.sciencedirect.com/science/article/pii/0047637494900949