Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Multiple Decrement and Competing Risks Models

Abstract. Multiple-decrement models represent the first exit of a covered person or insured contract from an initial risk state through several mutually exclusive causes, including mortality, lapse, surrender, disability, retirement, withdrawal, and other contractually specified terminations. Competing-risks statistics supplies the corresponding framework of cause-specific hazards, event-free survival, cumulative incidence, event-time likelihoods, and censoring. This monograph derives the central probability identities and their actuarial transformations; distinguishes observed crude incidence from associated single-decrement survival and latent-time counterfactuals; examines nonparametric estimation, regression, identifiability, uncertainty, and policyholder selection; and maps first-exit probabilities to insurance and pension benefits. Exact hypothetical calculations provide probability-conservation tests, cohort expectations, discounted benefit valuations, and sensitivity analysis under changes in lapse intensity. Historical statistical literature is integrated with modern actuarial textbooks, insurer experience studies, and jurisdiction-specific professional standards. The resulting original analytic synthesis explains how an unchanged cause-specific mortality intensity can coexist with a changed probability of a payable death claim when competing lapses vary. The work advances a reproducible interpretation of established methods and contains no new theoretical doctrine or original portfolio-derived empirical claims.

Keywords: Multiple decrement; competing risks; cause-specific hazard; cumulative incidence; event-free survival; associated single-decrement tables; net and crude decrement; mortality; lapse; disability; withdrawal; retirement; Aalen–Johansen estimator; Gray's test; Fine–Gray regression; censoring; identifiability; policyholder selection; life insurance; pension valuation; actuarial models.

Introduction: Scientific Scope and Research Design

Insurance and pension liabilities depend on the event that first removes a policyholder, member, or covered exposure from a defined contractual state. Death, surrender, premium nonpayment, disability, retirement, and withdrawal can occur at different times and produce different payment obligations. One exit changes the population exposed to later exits. A death-benefit valuation that uses mortality without accounting for coverage lapses consequently describes a different exposure process from the actual one, unless lapse has already been incorporated through an appropriate joint model.

The scientific object is a time-to-first-event system. The actuarial multiple-decrement method and the statistical competing-risks method describe the same core probability architecture. The actuarial multiple-decrement chapters of Dickson, Hardy, and Waters (2020) and demographic decrement processes in Preston, Heuveline, and Guillot (2001) organize its traditional life-contingency and population applications. The competing-risks and multistate tutorial of Putter, Fiocco, and Geskus (2007) places first exits within the wider family of possible state transitions.

The investigation asks how competing first exits determine actual event probabilities, what can be recovered from observed data, how hypothetical cause-deleted projections should be interpreted, and how these quantities enter insurance benefits, reserves, and pension obligations. The method combines direct mathematical derivation, original statistical source appraisal, exact hypothetical examples, and a separately scoped assessment of professional actuarial practice. Textbook and subscription publication claims were checked against original publisher records and accessible abstracts; publicly available original research articles were examined where available. No insurer's confidential dataset was modeled and no external reviewer is credited.

A complete specification requires a population, a time origin, an exposure state, legal event definitions, event hierarchy, observation window, and endpoint. Policy age, attained age, calendar time, duration since underwriting, service tenure, and disability onset can produce different conditional hazards. A calendar-year experience study mixing policy durations is not equivalent to a cohort of policies issued at the same date. The SOA–LIMRA lapse study (2024), which reports the contributing insurers' experience, illustrates the practical importance of measured lapse exposure; the Human Mortality Database overview documents population mortality statistics that address a different sampling population from an insurer's policy-termination records.

The manuscript maintains one boundary throughout: a first exit from a state can be absorbing for the chosen model even when the person subsequently lives, recovers, retires, or enters a new insurance contract. Repeated disability episodes, post-retirement death, renewals, and reentries require explicit additional transition states. That boundary permits exact modeling of first decrements without silently assuming that all events after a first exit have ceased to exist.

Decrement Events and Rates

The initial state and its event partition

Let T be time from a declared origin to the first exit from an active state and J∈{1,...,K} identify the cause. Every first event has one mutually exclusive cause within the model's taxonomy. At time t, Y(t)=1(T≥t) indicates continued eligibility immediately before an event. Contract termination by death, lapse, disability, surrender, or withdrawal must follow the actual product terms, including grace periods, reinstatement rules, benefit eligibility, and simultaneous-event adjudication.

This taxonomy is an indispensable scientific decision. A pension member may retire and begin an annuity, making retirement the end of active service and the beginning of another valuable state. A disability claim may terminate an employment risk state while creating years of benefit payments. A policyholder may miss a premium before a reported death, but the contract may continue through a grace period. Selecting the earliest administrative timestamp regardless of contractual validity would misclassify the insured benefit event.

An administrative data cutoff creates right censoring, not an observed lapse. A missing file due to a system migration can create potentially informative loss of follow-up. A predetermined contract maturity is often a known termination date, not a randomly estimated cause in the same sense as incidental death. Researchers must distinguish these observation arrangements from real exit events before fitting hazards or generating cash flows.

Central occurrence rates and probabilities

For cause j over interval [a,b], a simple event-count rate is m^j=Dj/E\widehat m_j=D_j/E, with E=∑i∫abYi(t) dtE=\sum_i\int_a^b Y_i(t)\,dt in exposed policy-years or person-years. D_j counts the qualifying first events and E includes time until any qualifying competing event. The rate has dimensions of inverse time and cannot be inserted unchanged into a probability table. Death and lapse rates estimated on different denominators do not form an additive total hazard without a verified common exposure convention.

For a fully followed cohort l_x active at exact age x, let d_x^(j) be those leaving during one year through cause j. The actual crude probability is qx(j)=dx(j)/lxq_x^{(j)}=d_x^{(j)}/l_x. Its total is qx(∙)=∑jqx(j)q_x^{(\bullet)}=\sum_jq_x^{(j)}, and continuation is px(∙)=1−qx(∙)p_x^{(\bullet)}=1-q_x^{(\bullet)}. These are cohort accounting identities. When censoring, truncation, or staggered entry occurs, observed fractions are not automatically consistent population probabilities; event-history estimation is then required.

Observed event proportions among exits, Pr⁡(J=j∣T≤t)=Fj(t)/(1−S(t))\Pr(J=j\mid T\le t)=F_j(t)/(1-S(t)), answer a further question: the composition of those who exit by t. They should not be interpreted as the probability that an initially active policy will exit by that cause. Distinguishing cause composition, conditional hazard, annual crude probability, and associated single-decrement probability prevents mixing incompatible actuarial inputs.

Selection through competing exits

Even if individual hazards were stable conditional on unobserved characteristics, the risk composition among survivors changes with time. High-lapse policyholders can be selected away faster than low-lapse policyholders. If lapse propensity and mortality share health or financial predictors, the mean observed mortality among those remaining active may change through selection alone. A constant pooled portfolio mortality hazard is consequently a modeling assumption about a changing risk set, not a general empirical law.

The term and whole life experience study (SOA Research Institute & LIMRA, 2024) documents one public lapse/surrender experience collection and its exposure basis. It is not evidence that lapses occur independently of mortality, nor does it justify transferring one portfolio's annual lapse rates to every jurisdiction or product. Mortality rates from the HMD project overview likewise do not directly yield insured-life death benefit probabilities when lapse, selection, underwriting, and coverage rules alter who remains insured.

The first-event boundary in insurance products

For term insurance the first event triggering payment may be covered death, while lapse closes coverage without a death claim. For cash-value life insurance, surrender carries a contractually defined payment; it is a competing exit with financial value. For long-term care and disability insurance, first admission or disability onset may be the relevant incidence endpoint, with recovery and death modeled in a subsequent state structure. For a pension, active employment exit does not necessarily terminate the accrued liability.

Thus a decrement category is specified relative to an exposure state and an obligation. Equating an event's biological or social label with an actuarial cash-flow trigger obscures the actual model. The life-contingent insurance textbook (Dickson et al., 2020) treats these contractual applications with multiple-decrement and multistate tools.

Cause-Specific Hazards

Definitions and basic identities

For a continuous first-event distribution, define the cause-specific hazard:

λj(t)=lim⁡h↓0Pr⁡(t≤T<t+h, J=j∣T≥t)h.\lambda_j(t)=\lim_{h\downarrow0}\frac{\Pr(t\le T<t+h,\ J=j\mid T\ge t)}{h}.

The total exit hazard is λ∙(t)=∑j=1Kλj(t)\lambda_\bullet(t)=\sum_{j=1}^K\lambda_j(t). If the cumulative hazard is absolutely continuous, event-free survival is

S(t)=Pr⁡(T>t)=exp⁡{−∫0tλ∙(u) du}.S(t)=\Pr(T>t)=\exp\left\{-\int_0^t\lambda_\bullet(u)\,du\right\}.

The cumulative incidence for first exit by cause j is

Fj(t)=Pr⁡(T≤t,J=j)=∫0tS(u)λj(u) du.F_j(t)=\Pr(T\le t,J=j)=\int_0^tS(u)\lambda_j(u)\,du.

Consequently S(t)+∑jFj(t)=1S(t)+\sum_jF_j(t)=1, and where differentiable, fj(t)=Fj′(t)=S(t)λj(t)f_j(t)=F_j'(t)=S(t)\lambda_j(t). These equations define a probability-conserving system with a single initial risk state. They do not require statistical independence among unobserved hypothetical latent cause-specific failure times. The distinction between an instantaneous cause-specific rate and a cumulative marginal event probability is explained in competing-risks probability analysis (Andersen et al., 2012).

The identity shows how cause j's incidence depends on the entire vector of exit intensities: its density includes S, and S depends on every cause. A lower death cumulative incidence among initially insured policies can arise from a higher competing lapse intensity while the conditional death hazard among those still in force remains unchanged.

Time-specific and integrated hazards

For hazards with jumps or atoms—for example mandatory retirement at an exact date—the continuous exponential formula should be replaced by the appropriate product-integral or discrete transition formulation. The original Aalen–Johansen estimator (Aalen & Johansen, 1978) and counting-process statistical foundation (Andersen et al., 1993) provide general transition-probability machinery, including product-integral construction and counting-process uncertainty. Scheduled events can be incorporated as deterministic cash-flow dates or explicit discontinuities, depending on the chosen state definition.

For subject i, N_ij(t)=1(T_i≤t,J_i=j) counts its first cause-j exit, while Y_i(t) records at-risk status under observation. In a correctly specified multiplicative-intensity model, the event intensity is Y_i(t)λ_ij(t). The aggregate risk set falls after each cause-specific exit, ensuring that future denominators represent only contracts still eligible for their first transition.

The failure-time statistical treatment of Kalbfleisch and Prentice (2002) and the counting-process monograph (Andersen et al., 1993) underpin event-time inference with right censoring, left truncation, and event-specific hazard estimation. Event time must be measured consistently: a policy anniversary, calendar-day event, or age-based decrement table requires an explicit temporal map.

Cause-specific hazard versus event probability

A cause-specific hazard answers the question how quickly a currently exposed member experiences a particular exit. Cumulative incidence answers how many initially exposed members have experienced that exit by a horizon. The derivative of the latter uses the full survival process. These objects are therefore related but distinct regression and valuation targets.

In a cause-specific proportional-hazards model one may write λj(t∣z)=λj0(t)exp⁡(z⊤βj)\lambda_j(t|z)=\lambda_{j0}(t)\exp(z^\top\beta_j). This uses the general structure of Cox's original regression model (1972), applied to each first-event cause. People experiencing other exits leave the corresponding at-risk set. For prediction, the complete set of modeled λ_j must be combined to reconstruct S and F_j. A coefficient on mortality hazard does not uniquely determine mortality CIF when lapse and other hazards also change.

For a constant death hazard λ_d, lapse λ_l, and disability λ_i, put H=λ_d+λ_l+λ_i. Then

Fd(t)=λdH(1−e−Ht)F_d(t)=\frac{\lambda_d}{H}(1-e^{-Ht})

when H>0. Holding λ_d and λ_i fixed yields

∂Fd(t)∂λl=−∫0tuλde−Hu du<0\frac{\partial F_d(t)}{\partial\lambda_l} =-\int_0^t u\lambda_d e^{-Hu}\,du<0

for t>0 and λ_d>0. This is an exact sensitivity of the defined contractual model. It is not evidence that increasing lapse prevents the person's biological death. The sign concerns exposure to the death benefit while the policy remains in force.

Conditional covariates, heterogeneity, and changing risk sets

A practical model can include age, duration, underwriting class, premium adequacy, policy type, market returns, and calendar effects. A covariate's effect on one hazard may be offset or amplified in the cumulative incidence by its effect on other hazards. The rate-and-risk interpretation (Andersen et al., 2012) expressly warns against confusing these relationships.

Hazards estimated within heterogeneous policy groups are conditional on the composition still under risk. A mixture of different mortality and surrender propensities can produce time dependence even if each constituent's conditional hazard is constant. The researcher should therefore distinguish a hypothetical fixed risk class from the dynamically selected in-force population. Dynamic prediction and aggregate financial projection may require the joint distribution of risk classes among survivors.

Associated Single-Decrement Tables

Construction from a cause-specific hazard

The associated single-decrement survivor function for cause j is

Sj∗(t)=exp⁡[−∫0tλj(u) du],qj∗(t)=1−Sj∗(t).S_j^*(t)=\exp\left[-\int_0^t\lambda_j(u)\,du\right], \qquad q_j^*(t)=1-S_j^*(t).

The construction holds the selected cause-specific intensity as specified and mathematically suppresses the other exit mechanisms. The actual crude cause-j probability is Fj(t)=∫0tS(u)λj(u) duF_j(t)=\int_0^t S(u)\lambda_j(u)\,du. Since S(u)≤Sj∗(u)S(u)\le S_j^*(u) for nonnegative competing intensities, Fj(t)≤qj∗(t)F_j(t)\le q_j^*(t) under the same continuous-hazard assumptions.

The relationship is part of the traditional multiple-decrement demographic calculus (Preston et al., 2001) and the actuarial associated-table framework (Dickson et al., 2020). Its interpretation is quantitative: a cause-only table describes a hypothetical exposure calculation under an unchanged selected intensity, while a crude table describes the observed first-cause incidence with all exits operating.

This distinction is especially important for mortality and lapse. Suppose a policy terminates before the life dies. The person remains biologically exposed to death, but ceases to be exposed to a death claim under that policy. A cause-only death table may be useful for isolating mortality intensity, but its death probability is not the policy death-benefit frequency over the same interval.

Constant-force and constant-cause-share conversions

If cause intensities are constant throughout a year, let H=Σ_jλ_j. The probability of any exit is q(∙)=1−e−Hq^{(\bullet)}=1-e^{-H}, and crude cause-j probability is q(j)=(λj/H)q(∙)q^{(j)}=(\lambda_j/H)q^{(\bullet)}. The associated single-decrement probability is qj∗=1−e−λjq_j^*=1-e^{-\lambda_j}.

Rearranging yields

qj∗=1−(1−q(∙))q(j)/q(∙).q_j^*=1-\left(1-q^{(\bullet)}\right)^{q^{(j)}/q^{(\bullet)}}.

This conversion remains valid if cause shares λ_j(t)/H(t) remain constant through the year while H(t) varies, provided the underlying no-exit process is continuous and its integrated total hazard is inferred from q(∙)q^{(\bullet)}. It therefore uses a fractional-period structural assumption; an annual crude q by itself does not reveal the exact shape of its within-year hazard.

For small rates and short intervals, cause-j probability is approximately λ_jΔt, but the approximation error grows with total intensity and exposure duration. Within-year uniform distribution of exits and constant force are different assumptions. Both can match an annual year-end q but produce different discount-weighted payments because benefit timing differs. A data dictionary must specify whether an input field stores a rate, crude annual probability, associated net probability, or cumulative incidence.

Why associated tables do not identify interventions

A formal cause-deleted survival function is determined by a specified observed cause-specific hazard, but interpreting it as a causal prediction under elimination of a competing cause requires additional structural restrictions. The Tsiatis nonidentifiability result (1975) shows that the joint distribution of latent potential failure times generally cannot be recovered from first-event competing-risk observations. Shared health, economic, or behavioral factors can affect both death and lapse; removing lapse may change the composition and behavior of the continuously insured group.

Independence of hypothetical latent failure times is a possible assumption, but an observed (T,J) sample does not itself verify it. For real contractual valuation under existing lapse and mortality hazards, the crude CIF is directly relevant; a causal intervention on retention, withdrawal, or treatment demands a defined counterfactual mechanism and further identification assumptions. It is scientifically legitimate to calculate a cause-only projection while explicitly describing it as a projection rather than a demonstrated causal outcome.

The Kaplan–Meier comparison

If competing exits are treated as censored events in a single-cause Kaplan–Meier analysis and one minus the resulting survivor function is interpreted as actual cause-j cumulative incidence, the quantity is generally overestimated in the presence of competing events. The original estimator analysis by Gooley et al. (1999) and epidemiological competing-risk review by Andersen et al. (2012) explain this distinction. In a constant-hazard model, the hypothetical cause-only probability 1−e−λjt1-e^{-\lambda_jt} exceeds the actual Fj(t)F_j(t) whenever other positive competing intensities operate.

This does not prohibit modeling the cause-specific hazard with appropriately constructed risk sets. It distinguishes a hazard estimation procedure from an estimator of actual marginal claim probability. When the actuarial estimand is a payable first-event claim under existing coverage, the crude CIF or its equivalent hazard-derived probability is the appropriate frequency target.

Multiple-Decrement Probabilities

Exact hypothetical cohort model

Consider 100,000 identical new policies in force at t=0, followed for one year. Assume three constant cause-specific intensities, in inverse policy-years: death λ_d=0.015, lapse λ_l=0.080, and disability λ_i=0.025. Thus H=0.120, S(1)=e^{-0.120}=0.8869204367, and total crude exit probability is 0.1130795633.

First exit Annual hazard Actual crude probability Expected events Associated cause-only probability
Death 0.015 0.0141349454 1,413.49 0.0148880604
Lapse 0.080 0.0753863755 7,538.64 0.0768836536
Disability 0.025 0.0235582424 2,355.82 0.0246900880
All first exits 0.120 0.1130795633 11,307.96 —
Continued exposure — 0.8869204367 88,692.04 —

Every number follows from the disclosed input intensities. The three crude probabilities sum to the total exit probability; adding S(1) yields one to rounding precision. Expected counts can be fractional because they are mathematical expectations, not empirical counts. The associated cause-only probabilities cannot be summed to produce the joint exit probability, because each belongs to a different hypothetical exposure system.

Using cause-only death probability 0.0148880604 instead of actual death incidence 0.0141349454 would overstate the one-year contract death incidence by approximately 5.33% relative to its correct value in this example. The error exists before any statistical uncertainty, and it is created solely by confusing net exposure with competing exit.

Multi-year recurrence and dynamic exposure

Let qr(j)q_r^{(j)} be cause-j probability conditional on remaining active at the beginning of year r. Let pr=1−∑jqr(j)p_r=1-\sum_jq_r^{(j)}. Event-free continuation to year n is

np=∏r=0n−1pr.{}_np=\prod_{r=0}^{n-1}p_r.

The unconditional probability of cause-j first exit by year n is

Fj(n)=∑r=0n−1(∏s=0r−1ps)qr(j).F_j(n)=\sum_{r=0}^{n-1}\left(\prod_{s=0}^{r-1}p_s\right)q_r^{(j)}.

The empty product equals one. This recurrence generalizes the constant-force example to annual mortality, lapse, withdrawal, disability, and retirement rates that depend on age, policy duration, service, product, or economic environment. Simply summing the yearly conditional q values would count previously exited members as though they remained at risk. An actuarial projection must update the active cohort before adding the next year's exits.

The same recursion can be conditioned on demographic or product cells. Aggregation should use projected surviving exposures rather than retaining fixed initial weights for all years. Selective lapses can materially alter the risk mix; ignoring them may misstate later death and surrender cash flows even when each initially fitted hazard appears reasonable.

Fractional-year assumptions and benefit timing

A known year-end q does not uniquely determine event timing. Under a constant-force assumption, S(t)=[1−q(∙)]tS(t)=[1-q^{(\bullet)}]^t for 0≤t≤1 and the cause CIF is Fj(t)=[q(j)/q(∙)][1−S(t)]F_j(t)=[q^{(j)}/q^{(\bullet)}][1-S(t)]. Under a uniform distribution of exits, S(t)=1−tq(∙)S(t)=1-tq^{(\bullet)} and Fj(t)=tq(j)F_j(t)=tq^{(j)}. Both match annual probabilities, while a payment immediately on exit is discounted differently under each. Year-end death claims and at-event death claims cannot be valued with the same timing convention without a documented approximation.

If an exit j pays B_j(t) at time t, its expected present value is ∫0τv(t)Bj(t) dFj(t)\int_0^\tau v(t)B_j(t)\,dF_j(t), where v(t) is the appropriate discount factor. This formula includes the possibility of other exits through F_j. It also shows why a year-end crude probability is insufficient when contractual benefit amounts or interest rates vary materially within the year.

Lapse stress under unchanged mortality hazard

Hold death intensity at 0.015 and disability at 0.025, while doubling lapse intensity from 0.080 to 0.160. Then H=0.200, the new one-year death incidence is 0.015[1−e−0.20]/0.20=0.01359519350.015[1-e^{-0.20}]/0.20=0.0135951935, and associated single-decrement death probability remains 1−e−0.015=0.01488806041-e^{-0.015}=0.0148880604.

The decline in actual insured death incidence follows from reduced exposure duration to covered death. It does not mean the person becomes less likely to die biologically or that a retention intervention would reproduce this exact scenario. In a real portfolio, financial distress or poor health may affect lapse and mortality together, requiring a stress design that allows other intensities to change as well.

Probability conservation and implementation tests

A computational implementation should confirm that each predicted F_j is nonnegative, S is nonincreasing, every F_j is nondecreasing, and S+∑jFj=1S+\sum_jF_j=1 at each projected horizon within numerical tolerance. A cause-only table should be labeled separately and should never appear in the same additive probability vector as the crude exits.

For discretized hazards the interval probability of each cause must respect conditional survival through the interval; taking λ_jΔt as exact can generate sums above one under high hazard. For a model with scheduled age or calendar events, explicit jump transitions and the timing of valuation cash flows must be documented. These are mathematical consistency tests that complement, but do not replace, empirical calibration and source-data review.

Competing Risks Estimation

Sampling frame, censoring, delayed entry

An event-history file records entry, exit or censoring date, cause, and relevant covariates for each policy or member. The risk indicator Y_i(t) must reflect delayed entry, the policy's eligibility, and removal after the first qualifying event. The competing-event statistical tutorial (Putter et al., 2007) and failure-time data monograph (Kalbfleisch & Prentice, 2002) describe the statistical machinery for censoring and truncation.

Right censoring occurs when an exposure is still active when observed follow-up ends, or when its further status cannot be ascertained. It differs from an observed lapse: a lapse is an event that competes with a death claim, while a censoring mechanism obscures the first-event outcome after its time. Statistical consistency requires suitable independent-censoring assumptions, potentially conditional on recorded covariates. A transfer to an unknown administrator, a missing system file, or attrition triggered by impending claims can create informative observation loss.

Left truncation arises when inclusion requires continued activity until the observation window begins. An experience study that starts with already in-force policies cannot automatically describe unconditional risks from issue date, because early failures were excluded. Model fitting must distinguish issue duration and attained age from calendar time and account for the conditional entry structure.

Aalen–Johansen cumulative incidence

At each observed event time u let Y(u) be the number at risk immediately before events, d_j(u) the observed number of cause-j events, and d_\bullet(u)=Σ_jd_j(u). Define

S^(t)=∏u≤t(1−d∙(u)Y(u)),F^j(t)=∑u≤tS^(u−)dj(u)Y(u).\widehat S(t)=\prod_{u\le t}\left(1-\frac{d_\bullet(u)}{Y(u)}\right), \quad \widehat F_j(t)=\sum_{u\le t}\widehat S(u-)\frac{d_j(u)}{Y(u)}.

This is the competing-risks form of the Aalen–Johansen estimator (1978). All exits enter the survivor estimate, while each cause's event increments enter its own incidence curve. The result reconstructs actual crude first-event incidence without an independent latent failure-times assumption. Correct censoring and risk-set construction remain necessary for inference.

As a transparent arithmetic example, let ten persons start active with no censoring. At time one, one death and two lapses occur, yielding S=0.7, death F_d=0.1, lapse F_l=0.2. At time two, two deaths and one disability occur among the seven remaining. Survival becomes 0.7(4/7)=0.40.7(4/7)=0.4; death CIF becomes 0.1+0.7(2/7)=0.30.1+0.7(2/7)=0.3; disability CIF becomes 0.7(1/7)=0.10.7(1/7)=0.1; lapse CIF remains 0.2. Together 0.4+0.3+0.2+0.1=1. This is a constructed identity demonstration, not a new estimator, data study, or sampling-variance result.

The competing-event CIF should be compared with one-minus Kaplan–Meier, which treats other exits as censored for a single endpoint. The Gooley estimator analysis (1999) and Austin–Lee–Fine introduction (2016) show why the latter generally fails as an estimator of actual marginal cause incidence when competing events occur.

Risk-set construction for cause-specific regression

A Cox model may be fitted for a cause-specific intensity with each other cause removed from the risk set at its event time. This is appropriate for estimating the conditional event hazard, with the model's risk-set, sampling and censoring assumptions. It should not be conflated with one-minus Kaplan–Meier estimation of the observed CIF. The first procedure targets λ_j; the second, if misinterpreted, substitutes a hypothetical single-cause quantity for F_j.

When predicted CIFs are required from cause-specific Cox models, the estimated hazards of every competing cause must be combined to compute total survival and then integrate each cause-specific density. Estimating only the death hazard cannot by itself determine actual insured death incidence when the lapse hazard is unspecified.

Subdistribution hazards and different estimands

The Fine–Gray subdistribution model (Fine & Gray, 1999) targets covariate associations with a particular cumulative incidence. For continuous distributions its subdistribution hazard can be written

λ~j(t)=fj(t)1−Fj(t),Fj(t)=1−exp⁡{−∫0tλ~j(u) du}.\widetilde\lambda_j(t)=\frac{f_j(t)}{1-F_j(t)}, \quad F_j(t)=1-\exp\left\{-\int_0^t\widetilde\lambda_j(u)\,du\right\}.

Its denominator is not S(t), and its special risk-set construction mathematically retains the effect of competing outcomes. It therefore differs from the cause-specific hazard fj(t)/S(t)f_j(t)/S(t). A Fine–Gray coefficient concerns a subdistribution-hazard relationship, not a directly interpretable increment in the physical death intensity among policies still active. Inserting λ~j\widetilde\lambda_j into the ordinary additive total hazard with other cause-specific rates would combine quantities with different denominators and violate the intended model.

Cause-specific regression helps describe the mechanisms of active-state exit, while subdistribution regression is often convenient for predicting or comparing marginal incidence. Both are conditional statistical associations requiring separate causal assumptions before being treated as effects of changes in premiums, retention policy, underwriting or treatment. The hazard-versus-incidence discussion (Andersen et al., 2012) and survival-methods introduction (Austin et al., 2016) emphasize the distinction.

Testing and uncertainty

The Gray cumulative-incidence tests (1988) compare cause-specific CIF curves between groups under their own test assumptions. A Gray-test result addresses a different endpoint from a log-rank comparison of overall event-free survival or a cause-specific hazard coefficient. Differences in lapse incidence can alter death CIF even when the death cause-specific hazard shows a different pattern.

Standard errors or confidence bands for CIF estimates must reflect uncertainty from the common risk set and all competing exits. Sparse cause counts, long extrapolation, informative censoring, and delayed entry can make simple normal approximations fragile. Resampling should preserve dependence structure where several policies share a policyholder, household, employer, economic environment, or branch. The counting-process statistical framework (Andersen et al., 1993) provides foundational asymptotic theory, while implementation still requires appropriate finite-sample diagnostics.

Validation, calibration, and scientific limitations

A good validation plan compares predicted versus observed or consistently nonparametrically estimated CIFs at multiple horizons, stratified by age, duration, contract type, benefit class, underwriting and calendar period. It also checks the full vector of exit types. A model can calibrate overall mortality but miscalibrate lapse, producing incorrect later exposure composition and erroneous projected death benefits.

Randomly dividing rows from the same policy's history between training and validation datasets may leak future information. Temporal validation with fixed observation cutoffs and coherent delayed labels better represents operational forecasting. Classification audits should inspect cancellation dates, reinstatements, grace periods, claim adjudications, duplicate insured entities, and changes in administrative coding. The ASOP 23 data standard (ASB, 2016) addresses data appropriateness and significant limitations; the ASOP 56 modeling standard (ASB, 2019) concerns model design, use, validation, and communication within its professional scope.

Observed first-event data identify the observed competing-risk distribution under the relevant observation assumptions. They generally do not identify the joint distribution of all hypothetical failure times. The Tsiatis nonidentifiability article (1975) limits any causal interpretation of eliminating one risk. Flexible models, high prediction accuracy, or extensive policy covariates cannot themselves overcome missing counterfactual information.

Insurance Benefit Applications

General actuarial valuation

Let B_j(t) be the amount payable at a first cause-j exit, v(t) the discount factor, c(t) the signed net cash-flow rate while the policy remains active, and G(τ) a terminal benefit payable if the active state survives to τ. Then

V=∑j=1K∫0τv(t)Bj(t)S(t)λj(t) dt+∫0τv(t)c(t)S(t) dt+v(τ)G(τ)S(τ).V=\sum_{j=1}^K\int_0^\tau v(t)B_j(t)S(t)\lambda_j(t)\,dt +\int_0^\tau v(t)c(t)S(t)\,dt+v(\tau)G(\tau)S(\tau).

The sign of c(t) must be declared consistently—for example, premiums may be negative liability outflows. The expression assumes a single active state and first-exit benefits. When exiting to disability, retirement, or a deferred benefit state creates continuing obligations, additional state-specific valuation terms are required. The actuarial life-contingency formulation (Dickson et al., 2020) places multiple decrements inside this broader valuation setting.

The factor S(t)λ_j(t) is the actual competing-exit density. Substitution of S_j^*(t)λ_j(t) would give a different cause-only product, retaining members in death exposure despite lapse or disability events that would have removed their coverage. Pricing should therefore follow the contract's actual coverage states and event order.

Explicit death and lapse cash-flow calculation

Return to the one-year hypothetical hazards λ_d=0.015, λ_l=0.080, λ_i=0.025 and total H=0.120. Assume death pays $100,000 at the instant of covered death; lapse pays $2,000 at exit; disability has no immediate lump-sum payment; and continuous discount force is δ=0.03 per year. All figures are stylized and exclude premiums, expense, taxes, capital, and follow-on benefits.

For constant hazards, the present value for an exit payment B_j is Vj=Bjλj[1−e−(H+δ)]/(H+δ)V_j=B_j\lambda_j[1-e^{-(H+\delta)}]/(H+\delta). Thus

Vd=100,000(0.015)1−e−0.150.15=$1,392.92,V_d=100{,}000(0.015)\frac{1-e^{-0.15}}{0.15}=\$1{,}392.92,

and

Vl=2,000(0.080)1−e−0.150.15=$148.58.V_l=2{,}000(0.080)\frac{1-e^{-0.15}}{0.15}=\$148.58.

The combined expected outflow for these two exit benefits is $1,541.50 per starting policy. It is not a policy premium or a complete reserve. A financially complete valuation must incorporate premiums, acquisition and renewal expenses, expenses after exit, claims administration, profit or risk margins where required, reinsurance, and any downstream benefit obligations.

Under lapse stress λ_l=0.160, with death and disability hazards fixed, H=0.200. The one-year death present value becomes 100,000(0.015)[1−e−0.23]/0.23=$1,340.00100{,}000(0.015)[1-e^{-0.23}]/0.23=\$1{,}340.00. The approximate $52.92 decline in death claims is only one component of the scenario. Higher lapses may remove premium income and increase lapse payments while changing reserve release and portfolio selection. It is consequently invalid to infer overall financial improvement from the decline in one benefit component.

Term life and cash-value insurance

For term life insurance the relevant cause-j exit can be covered death, while lapse terminates future premiums and death coverage. A policy still under a contractual grace period requires a specific rule for when coverage ends. A cash-value product adds surrender proceeds and possibly contractual nonforfeiture options. The timing and amount of surrender benefits can make competing-exit changes financially material even when death incidence shifts only modestly.

Policyholders respond to interest rates, income shocks, guarantees, health information, and premium changes. Therefore a lapse stress should specify whether cause-specific mortality also changes or whether the analysis deliberately holds it fixed to isolate contractual exposure. The SOA–LIMRA experience evidence provides a real-world measurement context but cannot be interpreted as proof of independence between policy surrender, health, and external economic conditions.

Pensions and employment decrements

Active pension membership may end through retirement, disability, resignation, death, or other scheme-specific exits. The immediate benefit can be a refund, retirement pension, disability pension, deferred vested benefit, or survivor entitlement. Exiting active status can trigger rather than erase the continuing scheme liability. A first-exit model is therefore often the initial layer of a pension projection, followed by benefit-state survival and payment models.

The correct time axis may jointly involve attained age, service duration, vesting status, eligibility age, and calendar conditions. Retirement can be concentrated at particular contractual ages, so a smooth continuous hazard over a broad interval may hide discontinuities. Competing retirement and withdrawal rates can lower observed death incidence among active employees without changing the age-specific biological mortality of a reference population.

Disability and long-term-care benefits

For a disability income policy, first disability incidence is an entry event to a claim-paying state; recovery, death, retirement, and recurrent episodes require a multistate or episode model. For long-term care, transitions between community care, assisted care, institutional care, recovery and death can carry substantially different benefit levels. The multistate models tutorial (Putter et al., 2007) provides the statistical architecture beyond first events. A first-decrement estimate can inform the probability of new claims but not the full payment-duration distribution.

In health research a competing death may make another endpoint unobservable; administrative loss to follow-up is conceptually separate. The Austin–Lee–Fine methodology (2016) clarifies these statistical differences. An insurance adaptation must additionally define the exact covered event, waiting periods, elimination periods, policy continuation rules, and follow-on benefits.

Validation, stress, and international governance

An insurer may need a vector of exit intensities under economic scenarios, together with dependence among policyholders, underwriting selection, and assets. At individual level a multiple-decrement model supplies first-exit probabilities. It does not alone deliver a portfolio-tail capital model, causal assessment of retention measures, or joint economy-wide shocks; these tasks require additional assumptions and structures.

The actuarial data quality standard and professional modeling standard express U.S.-specific expectations about suitable data, intended use, modeling assumptions, validation and limitations. The ISAP 1 international register identifies an international model professional standard; the enforceability and adoption of any model standard are jurisdiction-dependent. Applying professional governance requires precise identification of institution, effective date, scope, and local legal requirements. The standards support disciplined model work and transparent communication, not automatic certification of the mathematical model for a particular portfolio.

The model governance record should contain a contractual event map; documented exposure and censoring logic; rates and probabilities with units; source-lineage evidence; probability conservation tests; calibration and stress results; sensitivity to alternative fractional-year assumptions; and a list of the financial cash flows triggered by each transition. A validation report should specify whether it tests cause-specific intensity, cumulative incidence, or expected present value, since passing one criterion does not imply passing the others.

Original analytic synthesis

The evidence yields three complementary scientific interpretations of a first-exit system. The observational interpretation concerns actual incidence Fj(t)F_j(t) under the existing competing hazards. The conditional-intensity interpretation concerns λj(t)\lambda_j(t) among policies still active and therefore the evolution of the exposed risk set. The hypothetical interpretation concerns a model in which causes or behaviors are altered, such as associated cause-only survival or an intervention on lapse; the latter calls for additional identification assumptions and cannot be read directly from first-event data.

This distinction explains several otherwise confusing results in one analytic structure. Death hazard can remain constant while the contract's death incidence changes with lapse. Cause-specific and subdistribution regressions can give different covariate patterns because they describe different risk sets. A statistically accurate death CIF can coexist with a financially incomplete policy valuation when surrender cash flows, premium cessation or post-disability liabilities are omitted. The synthesis organizes existing mathematics and empirical responsibilities; it establishes no new named theory, proprietary causal principle, or unsupported scientific novelty.

Conclusion

Multiple-decrement and competing-risks methods form a coherent mathematical account of how several possible first events remove exposure from a defined insurance or pension state. Cause-specific hazards add to the total exit hazard, the total hazard determines event-free survival, and the product of survival and a selected cause-specific hazard determines that cause's observed incidence density. The resulting cumulative incidence probabilities and survival sum to one when computed consistently.

Associated single-decrement tables are valid mathematical transformations of a specified cause-specific intensity. They describe a different exposure system from the observed multiple-decrement portfolio and should not be summed as competing event probabilities. Counterfactual claims about eliminating a cause require identification assumptions beyond the observed distribution of first-event times and labels.

The hypothetical cohort exhibits the difference in exact numbers: death hazard 0.015, lapse 0.080 and disability 0.025 yield an actual one-year insured death probability of 0.0141349454 versus the associated cause-only death probability of 0.0148880604. Aalen–Johansen estimation reconstructs actual incidence from event histories under suitable observation assumptions; cause-specific Cox and Fine–Gray regression use different hazard definitions and serve distinct statistical targets. Competing exit timing then determines the present value of death, lapse, retirement, and disability benefits according to each contract's obligations.

The controlling scientific requirement is alignment among contract, event definitions, exposed population, observation process, probability estimand, valuation cash flows, and validation method. Explicit source limitations, mathematical identities, coherent timing assumptions, and professionally scoped model governance make the analysis auditable. The result is a complete interpretive and quantitative integration of established competing-risk methods for the international insurance context.

References

  • Aalen, O. O., & Johansen, S. (1978). An empirical transition matrix for non-homogeneous Markov chains based on censored observations. Scandinavian Journal of Statistics, 5(3), 141–150. Original source or publisher record.

  • Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality. Original source or publisher record.

  • Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Original source or publisher record.

  • Andersen, P. K., Borgan, Ø., Gill, R. D., & Keiding, N. (1993). Statistical Models Based on Counting Processes. Springer. Original source or publisher record.

  • Andersen, P. K., Geskus, R. B., de Witte, T., & Putter, H. (2012). Competing risks in epidemiology: Possibilities and pitfalls. International Journal of Epidemiology, 41(3), 861–870. Original source or publisher record.

  • Austin, P. C., Lee, D. S., & Fine, J. P. (2016). Introduction to the analysis of survival data in the presence of competing risks. Circulation, 133(6), 601–609. Original source or publisher record.

  • Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202. Original source or publisher record.

  • Dickson, D. C. M., Hardy, M. R., & Waters, H. R. (2020). Actuarial Mathematics for Life Contingent Risks (3rd ed.). Cambridge University Press. Original source or publisher record.

  • Fine, J. P., & Gray, R. J. (1999). A proportional hazards model for the subdistribution of a competing risk. Journal of the American Statistical Association, 94(446), 496–509. Original source or publisher record.

  • Gooley, T. A., Leisenring, W., Crowley, J., & Storer, B. E. (1999). Estimation of failure probabilities in the presence of competing risks: New representations of old estimators. Statistics in Medicine, 18(6), 695–706. Original source or publisher record.

  • Gray, R. J. (1988). A class of K-sample tests for comparing the cumulative incidence of a competing risk. The Annals of Statistics, 16(3), 1141–1154. Original source or publisher record.

  • Human Mortality Database. (2026, accessed November 14). Project overview. Original source or publisher record.

  • International Actuarial Association. (2018). International Standard of Actuarial Practice 1: General Actuarial Practice. Original source or publisher record.

  • Kalbfleisch, J. D., & Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data (2nd ed.). Wiley. Original source or publisher record.

  • Preston, S. H., Heuveline, P., & Guillot, M. (2001). Demography: Measuring and Modeling Population Processes. Blackwell Publishers. Original source or publisher record.

  • Putter, H., Fiocco, M., & Geskus, R. B. (2007). Tutorial in biostatistics: Competing risks and multi-state models. Statistics in Medicine, 26(11), 2389–2430. Original source or publisher record.

  • Society of Actuaries Research Institute & LIMRA. (2024). 2015–2022 Term/Whole Life Lapse/Surrender Experience Study. Original source or publisher record.

  • Tsiatis, A. (1975). A nonidentifiability aspect of the problem of competing risks. Proceedings of the National Academy of Sciences, 72(1), 20–22. Original source or publisher record.