Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Multistate Models for Life and Health Insurance

Abstract. Life and health insurance obligations depend on sequences of contingent states: a policyholder may remain active, experience disability, recover, relapse, die, or discontinue coverage, with the amount and timing of cash flows changing along each path. A multistate actuarial model represents those paths through an admissible transition structure and values contractually specified cash flows using occupancy probabilities and transition intensities or their semi-Markov analogues. This monograph establishes the mathematical correspondence between event history, stochastic state processes, transition-dependent payments, benefit present values, prospective reserves, and empirical experience calibration. It derives continuous-time Markov valuation and Thiele-type reserve equations, distinguishes them from duration-dependent semi-Markov calculations, analyzes identification under exact and panel observation, and develops a fully specified four-state numerical contract example with sensitivity analysis. The evidence integrates foundational actuarial texts and transition-matrix estimation, recent disability research, practical statistical software, and international financial-reporting and actuarial-practice materials. The principal analytical conclusion is that the relevant model must preserve both the path dependence of insured health and the distinct legal process governing continued coverage. Reliable estimates require coherent state definitions, transition-specific exposure, censoring and selection control, contract-level reward specification, and validation of predictions at the horizons at which benefits and reserves are measured.

Keywords: Multistate models; life insurance; health insurance; disability insurance; recovery; relapse; semi-Markov processes; Markov chains; transition intensities; Aalen–Johansen estimator; state occupancy; Thiele differential equations; life contingencies; lapse; actuarial reserves; competing risks; experience calibration; model validation.

Introduction

Insurance benefits arise through events occurring within a particular contractual relationship. Mortality and sickness are biological and demographic processes; enrollment, premium payment, waiting periods, exclusions, renewals, and lapse are contractual processes. The actuarial task is to translate the interaction of these processes into probabilities of liability and dated cash flows. Treating a single incident, recovery, or death as an isolated decrement is adequate only when the contractual benefit under study genuinely depends on that limited event structure. Income protection, disability annuities, critical-illness riders, long-term care, waiver-of-premium guarantees, and some life insurance arrangements require the joint history of multiple events.

The discipline has a mature technical lineage. The multi-state chapters in Haberman and Pitacco (1998) situate disability, long-term care, critical illness, and transition-intensity graduation within one actuarial framework; the third edition of Actuarial Mathematics for Life Contingent Risks provides contemporary life-contingent valuation foundations. The central methodological bridge into biostatistics is the competing-risks and multi-state tutorial of Putter, Fiocco, and Geskus (2007), which connects event-history data to transition and occupancy probabilities.

This Research Dossier examines state-transition models for insurance contingent benefits, with attention to disability incidence, recovery, repeat episodes, death, and lapse. The research method combines explicit stochastic assumptions, derivation, numerical verification, and an audit of primary and scholarly sources. Its interpretations are reasoned syntheses of established mathematics and observed research; it introduces no separate named theory, doctrine, or proprietary model. Every numerical insurance example below is hypothetical, with monetary units and timing specified. A literature result established in a clinical population is evidence of a statistical mechanism and not automatically a calibrated estimate for an insured portfolio.

The analytical organization is consequential. The state structure comes first because a probability model cannot repair an incomplete definition of coverage. A Markov or semi-Markov specification then determines which history is retained. Occupancy is derived from transitions and distinguished from a claim count. Valuation attaches monetary rewards to paths. Calibration finally tests whether the data can identify the parameters that the contract requires. These steps form the existing mathematical workflow of life-contingent risk analysis, not an additional theoretical system.

State Spaces and Transition Intensities

An admissible contractual state space

Let X(t)X(t) take values in a finite state space EE at duration tt from policy inception. The simplest disability structure has active AA, disabled DD, and deceased MM. A recoverable disability product permits A→DA\to D, D→AD\to A, A→MA\to M, and D→MD\to M. The state MM is absorbing for the life history. A life insurer may also require a terminated-coverage state LL, because life after lapse is still possible even though future covered benefits may have ceased. In a coverage-limited model, LL is an absorbing bookkeeping state. In a combined health-and-coverage model it is more accurate to represent an ordered pair (H(t),C(t))(H(t),C(t)) where HH is health and CC is coverage status; health transitions can continue after contractual termination.

This product-state distinction matters for data interpretation. A claims system that stops receiving medical records at lapse records the end of observation rather than the end of disability. If lapse is treated as death or recovery, the estimated recovery intensity is distorted. A valid four-state contract model can collapse post-lapse health only when the sole estimand is benefit liability after lapse and when all relevant post-lapse contractual obligations, including any surrender values or continuation rights, are represented. State aggregation is thus an assertion about the valuation target as well as an assertion about biology.

For each admissible transition i→ji\to j, let the instantaneous rate at time tt be

qij(t)=lim⁡h↓0Pr⁡(X(t+h)=j∣X(t)=i)h,i≠j,q_{ij}(t)=\lim_{h\downarrow0}\frac{\Pr(X(t+h)=j\mid X(t)=i)}{h}, \quad i\neq j,

when the limit exists and the Markov conditioning is appropriate. Forbidden direct transitions have zero rate. Define qii(t)=−∑j≠iqij(t)q_{ii}(t)=-\sum_{j\ne i}q_{ij}(t). The generator matrix Q(t)Q(t) consequently has nonnegative off-diagonal elements and row sums zero. An intensity has units of inverse time; a one-year transition probability is dimensionless. Substituting a one-year probability for a continuous rate without conversion is a potentially material actuarial error.

For a four-state model ordered A,D,M,LA,D,M,L, with active disability incidence α\alpha, recovery ρ\rho, active mortality μA\mu_A, disabled mortality μD\mu_D, and active and disabled lapse rates ℓA,ℓD\ell_A,\ell_D,

Q(t)=(−(α+μA+ℓA)αμAℓAρ−(ρ+μD+ℓD)μDℓD00000000).Q(t)=\begin{pmatrix} -(\alpha+\mu_A+\ell_A)&\alpha&\mu_A&\ell_A\\ \rho&-(\rho+\mu_D+\ell_D)&\mu_D&\ell_D\\ 0&0&0&0\\ 0&0&0&0 \end{pmatrix}.

Each entry can depend on attained age x+tx+t, calendar time, sex where relevant and legally permissible, morbidity class, occupation, underwriting selection, contractual duration, and other justified covariates. The model must explicitly state which influences are observed and which are assumed. Disability incidence and recovery concern different risk sets. The numerator of a recovery rate counts recoveries from disability; its denominator is disabled person-time, not all insured person-time.

Multiple decrements and path recurrence

From AA, incidence, death, and lapse compete. With constant rates the total exit intensity is a=α+μA+ℓAa=\alpha+\mu_A+\ell_A, so the time to the first departure from AA, conditional on starting there, has survivor probability exp⁡(−at)\exp(-at). The probability the first departure is to DD by horizon TT equals α(1−exp⁡(−aT))/a\alpha(1-\exp(-aT))/a. This is a first-exit probability; it is not the probability of occupying DD at TT, because recovery and further exits are possible.

The distinction becomes especially important when A→D→A→DA\to D\to A\to D can repeat. Disability episode incidence counts entrances into DD. A person can contribute more than one episode. Point-in-time disability prevalence counts occupants of DD once each. Contractual disability-days reflect time spent in DD and can differ from either measure. Accurately measuring them requires the model to carry the appropriate indicator or counting process. In the notation of Putter, Fiocco, and Geskus (2007), transition-specific event histories permit such measures to be analyzed in the same stochastic process, while preserving their distinct meanings.

Benefits, exclusions, and administrative states

Medical illness is not automatically covered disability. A policy may require inability to perform an occupation, a minimum waiting period, diagnosis validation, or ongoing eligibility. One may represent a waiting period by tracking disability duration, or by splitting a health state into contractual substates such as reported, qualifying, in-payment, and suspended. A claim notice date and a disability onset date require separate fields. Similarly, recovery from a medical condition may not imply satisfaction of a policy's return-to-work definition. Where a benefit terminates at an age limit, lapse, death, or maximum payment duration, the exposure window for that benefit must be encoded.

The disability actuarial framework is useful precisely because the policy's rules alter the state system and its cash flows. In an actual valuation, the transition diagram should be auditable against contract wording, administrative practice, and any material variations by cohort or market. State parsimony improves estimation only when it preserves contractually relevant information.

Markov and Semi-Markov Assumptions

The Markov representation

A finite-state time-inhomogeneous Markov model assumes that conditional future state probabilities depend on the current state and current time, given the covariates included in the state specification. Formally, for s<ts<t, the conditional law of X(t)X(t), given the past through ss, is determined by X(s)X(s) and the specified time-varying explanatory information. This assumption does not require constant rates. It does require that remaining duration and prior episodes convey no additional transition information after conditioning on included variables.

Let P(s,t)P(s,t) be the matrix with entry pij(s,t)=Pr⁡(X(t)=j∣X(s)=i)p_{ij}(s,t)=\Pr(X(t)=j\mid X(s)=i). The Chapman–Kolmogorov identity requires P(s,t)=P(s,u)P(u,t)P(s,t)=P(s,u)P(u,t) for s≤u≤ts\le u\le t. Its forward equation is

∂P(s,t)∂t=P(s,t)Q(t),P(s,s)=I.\frac{\partial P(s,t)}{\partial t}=P(s,t)Q(t),\qquad P(s,s)=I.

The backward equation is ∂P(s,t)/∂s=−Q(s)P(s,t)\partial P(s,t)/\partial s=-Q(s)P(s,t). If QQ is constant, P(s,t)=exp⁡[(t−s)Q]P(s,t)=\exp[(t-s)Q]. With time-varying, noncommuting generators, the correct solution is a time-ordered product integral; replacing it with an ordinary exponential of ∫Q(u) du\int Q(u)\,du generally gives the wrong answer. The original Aalen–Johansen transition-matrix paper (1978) provides the nonparametric product-integral foundation for Markov transition probability estimation from censored histories.

If rates are piecewise constant on intervals, the transition matrix is an ordered product of interval-specific matrix exponentials. A valid numerical implementation should preserve row sums close to one, nonnegative probabilities within tolerance, and the specified zero transitions for direct hazards. An annual stochastic transition matrix is not always embeddable in a valid continuous-time generator: simply applying a matrix logarithm can produce negative off-diagonal entries or an ambiguous generator. Calibrating a continuous-time model to annual transitions requires an explicit identification and fit assessment.

Semi-Markov duration dependence

For disability recovery, time already spent disabled often has substantive predictive content. Let τ(t)\tau(t) be the most recent entry time into the current state, and let U(t)=t−τ(t)U(t)=t-\tau(t) denote sojourn duration. A semi-Markov-style model writes the exit intensity as qij(t,U(t),z)q_{ij}(t,U(t),z), where zz represents observed covariates. Conditional on (X(t),U(t))(X(t),U(t)) and suitable covariate history, the augmented process can be represented as Markov; the observed health state X(t)X(t) alone generally cannot.

The existing paper by D’Amico, Guillen, and Manca (2009) explicitly augments disability dynamics with backward recurrence time. Its empirical application motivates duration-dependent modeling but does not establish a universal recovery function. An open-access study by De Giovanni and colleagues (2025) uses Health and Retirement Study records to investigate non-Markov features in functional disability. These studies support testing duration dependence when a contract's liability is sensitive to disabled spell length.

A model may distinguish calendar time from attained age, time since policy inception from time since the disability episode, and number of prior episodes from current spell duration. Their effects should not be conflated. One parsimonious intensity might be ρ(t,u)=ρ0(x+t)exp⁡{βg(u)}\rho(t,u)=\rho_0(x+t)\exp\{\beta g(u)\}, where ρ0\rho_0 is an age-dependent baseline and g(u)g(u) measures spell duration. The sign and shape of βg\beta g must be estimated or scenario-tested. It is scientifically incorrect to assume recovery always decreases or always increases with duration across conditions and policy populations.

For a disabled spell entered at time ss, the conditional survival of that spell without exit over duration uu is

SD(u∣s,z)=exp⁡[−∫0u(ρ(s+v,v,z)+μD(s+v,v,z)+ℓD(s+v,v,z)) dv].S_D(u\mid s,z)=\exp\left[-\int_0^u \left(\rho(s+v,v,z)+\mu_D(s+v,v,z)+\ell_D(s+v,v,z)\right)\,dv\right].

The chance of recovery before duration uu, before competing death or lapse, is the integral of SD(v∣s,z)ρ(s+v,v,z)S_D(v\mid s,z)\rho(s+v,v,z) over vv from zero to uu. An exit probability computed from recovery alone overstates actual recovery when competing exits occur. This is directly related to the treatment of competing events in Putter, Fiocco, and Geskus (2007).

Inference, observation and latent states

When exact dates of onset, recovery, and death are observed, transition-specific counting-process methods can estimate intensities. A medical record observed only at periodic visits instead yields panel observations and interval uncertainty. The likelihood uses conditional transition probabilities across observation intervals, which requires care about a health change followed by a reversal between visits. Jackson's msm methodology (2011) supplies continuous-time models for panel observations, including hidden Markov variants for misclassified or noisy states. A panel record listing a person as active at consecutive visits does not prove that no disability episode occurred between them.

Semi-Markov state augmentation resolves a probabilistic representation; it does not automatically resolve statistical identifiability. Sparse panel visits may fail to locate entry dates accurately, while claim reports can be delayed or contingent on a benefit threshold. Identification of a duration-specific rate requires enough observed event-times and appropriately measured duration risk sets. Flexible parametric forms, including tools described in Jackson's flexsurv article (2016), can help model hazards, but flexibility should be limited by observed information and validated against held-out periods.

The choice of model class is ultimately empirical and contractual. A simpler Markov approximation may be sufficient for a short-horizon product whose benefits are insensitive to spell duration. A duration-dependent process may be essential for a multi-year disability income stream. Comparing them requires common contract inputs and a target metric such as prediction calibration, expected cash flow error, or risk capital sensitivity, not merely relative likelihood in a different population.

Occupancy Probabilities

State occupancy versus transition history

Fix a policy initially active at time zero, represented by row vector eA=(1,0,0,0)e_A=(1,0,0,0). Under a Markov model, the occupancy row is π(t)=eAP(0,t)\pi(t)=e_AP(0,t). Its component πD(t)\pi_D(t) is the probability that the insured is disabled and covered at tt. The expected discounted disability time over [0,T][0,T] at force of interest δ(t)\delta(t) is

∫0Texp⁡[−∫0tδ(v) dv]πD(t) dt.\int_0^T \exp\left[-\int_0^t\delta(v)\,dv\right]\pi_D(t)\,dt.

This quantity is measured in discounted years per starting policy. Multiplied by an annual payment rate, it becomes the expected present value of an in-state benefit, subject to contract eligibility. Total disability episodes over the interval have expectation ∫0TπA(t)α(t) dt\int_0^T\pi_A(t)\alpha(t)\,dt when only active-to-disabled entrances are counted. The difference between the two integrals is not a contradiction; one counts transitions, the other measures state occupation.

The forward equation for occupancy is π′(t)=π(t)Q(t)\pi'(t)=\pi(t)Q(t). In the active–disabled–death–lapse system,

πD′(t)=α(t)πA(t)−(ρ(t)+μD(t)+ℓD(t))πD(t).\pi_D'(t)=\alpha(t)\pi_A(t) -\big(\rho(t)+\mu_D(t)+\ell_D(t)\big)\pi_D(t).

Incidence supplies entrants and recovery, mortality, and lapse remove occupants. The corresponding active equation includes the recovery inflow. Consequently, a recovery improvement can raise the active occupancy at future horizons while lowering disabled occupancy. It can also alter future incidence because recovered lives again enter the active risk set. This endogenous feedback makes single-decrement shortcuts unreliable for products with recurrent claims.

Nonparametric estimation and the Aalen–Johansen relation

For transitions i→ji\to j, let Nij(t)N_{ij}(t) count observed events and Yi(t)Y_i(t) be the observed risk set in state ii. Under suitable event observation and censoring conditions, the integrated hazard increment is estimated by dA^ij(t)=dNij(t)/Yi(t)d\widehat A_{ij}(t)=dN_{ij}(t)/Y_i(t). Off-diagonal increments are positive and diagonal increments are chosen for zero row sums. The Aalen–Johansen product integral then constructs an estimated transition matrix from successive I+dA^(t)I+d\widehat A(t) factors. The multiplicative structure respects sequential transitions and can be used in multistate survival analysis.

A crucial identification distinction is that ordinary Aalen–Johansen transition estimates conditional on a state at a later landmark can fail under general non-Markov dependence. Titman (2015) develops alternatives for non-Markov transition probabilities, and the landmark Aalen–Johansen approach of Putter and Spitoni (2018) uses a landmark subset to estimate conditional transitions under specified assumptions. One must therefore state whether the desired estimand is marginal occupancy from policy inception, conditional transition from a state occupied at ss, or conditional transition further stratified by spell history. These are different probability statements.

Censoring can depend on measured health or administrative conditions. Datta and Satten (2002) propose weighted estimation of transition hazards and state occupation under certain covariate-dependent censoring mechanisms in non-Markov systems. In an insurance portfolio, termination of observation through data extraction must be separated from genuine contract lapse, death, transfer, policy-year expiry, and migration across claims systems. Each has a different effect on the risk set and contractual liability.

Continuous and discrete observation

For discretely reported state observations, empirical proportions at an arbitrary date need adjustment for delayed reporting, selection, and loss to follow-up. If the state is accurately observed annually and a continuous-time generator is assumed, a one-year transition matrix P(t,t+1)P(t,t+1) gives probabilities of endpoint states, including paths with multiple intervening transitions. The count of observed endpoints alone cannot identify each intermediate episode without more information or additional assumptions. In particular, a transition from active to active in annual records may include an active–disabled–active path.

Estimation uncertainty attaches to transitions and occupancies and must be propagated into benefit values. Variance calculations based on independent complete observations can underestimate uncertainty when there is within-policy dependence, heterogeneous observation intensity, or parameters shared across cohorts. Bootstrap units should follow the independent sampling unit: for a policy-level panel, resample policies or appropriately defined clusters, not individual transitions as if they were unrelated. A transparent empirical study should report exposure volumes and confidence or prediction intervals alongside point estimates.

Recovery and Relapse

Recovery as a competing transition

The economic meaning of recovery depends on contract terms. A return to activity may stop a disability annuity, restore premiums, restart an elimination period, or reactivate a rider. Recovery rates are often conditional on time since onset, severity, diagnosis, treatment, and occupation. The recovery transition itself can be observed medically or administratively; the two event definitions may yield different dates. For a benefits-in-payment model, a recovery effective date should align with the date contractual payments cease. For a population-health model, the target date should align with the clinical or functional status definition.

Suppose a disabled life has constant recovery intensity ρ\rho, mortality intensity μD\mu_D, and lapse intensity ℓD\ell_D, and consider the next exit only. The probability that recovery occurs before the two other competing exits is ρ/(ρ+μD+ℓD)\rho/(\rho+\mu_D+\ell_D). This is a first-next-exit probability and allows no distinction between short and long past disability spells. A duration-specific model changes both the time pattern and cumulative recovery incidence. Applying the constant-rate expression to a heavily duration-dependent cohort can misstate the length of benefits and reserves even when the aggregate annual recovery fraction appears similar.

A chronic condition may permit partial recovery, relapse, or progression in impairment. Modeling a binary active/disabled variable merges partial states; that is defensible only if the contract treats those states equivalently. Where benefit percentages depend on severity, multiple disabled states or a continuous severity covariate may be necessary. An increase in state resolution improves contractual fidelity but multiplies transition parameters. The practical choice requires enough observation exposure for each transition, a coherent medical definition, and stability of classification through time.

Relapse and prior episode history

After recovery, disability can recur. A Markov model can represent recurrence through another A→DA\to D transition, but if second or later episodes have different incidence or recovery hazards, simply returning to the same undifferentiated active state discards information. One solution expands active and disabled states by episode order; another uses a history covariate, perhaps with state augmentation. The choice should be justified by contractual sensitivity and estimability. Event-history data should preserve original onset, recovery, relapses, and contract amendments even if a final production model later pools some categories.

The empirical evidence from De Giovanni et al. (2025) gives a concrete reason to test non-Markov patterns in functional-disability transitions. Their Health and Retirement Study setting has specific survey measurements, age distributions, and event definitions. It does not supply an unmodified insured disability-income recovery assumption. Likewise, D’Amico et al. (2009) shows the practical attraction of backward recurrence time in a specific disability-data application, subject to the limitations of the underlying portfolio and available reporting.

Waiting periods and benefit survival

For a disability payment beginning only after a waiting period ww, one requires the probability that a spell survives in a qualifying state to ww, followed by subsequent in-payment occupancy. For a newly disabled person entering at ss, the probability of surviving the waiting period without exit is SD(w∣s,z)S_D(w\mid s,z), with the relevant competing transitions included. If the contract permits a short recovery followed by recurrent disability within an aggregation window, the waiting-period logic becomes more complex and needs an explicit contractual rule. The model should implement the actual definition rather than assume uninterrupted disability is always required.

A disability reserve at duration uu after onset should account for conditional future cash flow from that duration, including recovery, recurrence, mortality, expiry, offsets and any waiver of premiums. The relevant prospect is not the unconditional average disabled spell. When intensity depends on time disabled, reserve amounts for otherwise similar policyholders may differ materially by duration even at equal attained age. In a Markov approximation without duration states, this variation is absent by assumption.

Selection and experience bias

Recovery statistics drawn from accepted claims may exclude minor disabilities not satisfying waiting periods, claims not yet adjudicated, and unreported illnesses. Incidence calibrated on all policyholders and recovery calibrated on accepted long-duration claims can be combined only when the implied state and observation definitions are compatible. Similarly, successful return-to-work records may be more complete than records of ongoing disabled claims. Administrative delay can create apparent duration effects. Observed claim-closure hazards also mix clinical recovery, contractual settlement, expiry, death, and fraud or compliance decisions. A model should retain the precise event type, model genuinely competing exits, and document the provenance of each duration variable.

Multistate Benefit Valuation

Pathwise cash flows and expected present value

Let bi(t)b_i(t) be the monetary rate of net benefit while the contract occupies covered state ii, positive when the insurer pays and negative when it receives premiums. Let Bij(t)B_{ij}(t) be a lump sum paid upon transition i→ji\to j, including death benefit or contractual surrender payment where applicable. Let the force of interest be δ(t)\delta(t) and the deterministic discount factor from inception v(0,t)=exp⁡[−∫0tδ(s) ds]v(0,t)=\exp[-\int_0^t\delta(s)\,ds]. For a fixed expiry TT, the expected present value from initial state distribution π(0)\pi(0) is

EPV=∫0Tv(0,t)∑iπi(t)[bi(t)+∑j≠iqij(t)Bij(t)]dt,\mathrm{EPV}=\int_0^T v(0,t) \sum_i\pi_i(t)\left[b_i(t)+\sum_{j\ne i}q_{ij}(t)B_{ij}(t)\right]dt,

with separately specified amounts at fixed dates or the maturity date added as discounted occupancy-weighted terminal benefits. This follows from integrating expected continuous state reward and compensator-based expected transition rewards. The formula is a Markov expression; in a semi-Markov model one integrates over the joint distribution of state and sojourn duration, with duration-dependent rewards and transition rates.

A premium fixed for all active times is attached to bA(t)b_A(t). Premiums payable in disabled states, waived during disability, or guaranteed while a policy remains active require distinct cash-flow definitions. Death benefits dependent on health state are represented by BAMB_{AM} and BDMB_{DM}, perhaps with different coverage amounts or exclusion effects. Surrender or lapse benefits require BALB_{AL} and BDLB_{DL}, even if the lapse state is absorbing for later covered benefit purposes. The set of modeled cash flows must match the contract's cash-flow perimeter. Net expected present value, gross premium valuation, and financial-reporting fulfilment cash flows are related but have different expense, risk adjustment, and accounting conventions.

Prospective reserves and the differential equation

Let Vi(t)V_i(t) be the expected present value at time tt of remaining net insurer payments conditional on the contract occupying state ii, at a valuation basis with deterministic force δ(t)\delta(t). Assume no payment exactly at tt and a terminal reserve condition Vi(T)=0V_i(T)=0 after all expiry payments have been separately settled. Applying first-step conditioning over [t,t+dt][t,t+dt] gives

Vi′(t)=δ(t)Vi(t)−bi(t)−∑j≠iqij(t)[Bij(t)+Vj(t)−Vi(t)].V_i'(t)=\delta(t)V_i(t)-b_i(t) -\sum_{j\ne i}q_{ij}(t)\big[B_{ij}(t)+V_j(t)-V_i(t)\big].

These are the multistate Thiele differential equations with the stated sign convention: positive insurer outflows and negative premium receipts. A death or lapse state has reserve zero only if the contract has no further obligations after its transition benefit. If an in-force contract has maturity benefit mim_i, set the terminal condition Vi(T)=miV_i(T)=m_i for eligible states instead of zero. Backward numerical integration from TT derives the reserve trajectory. Forward occupancy integration and backward reserves should reconcile at inception when based on the same QQ, rewards, discounting, and terminal terms.

This mathematics is consistent with the life-contingent valuation exposition in Dickson, Hardy, and Waters (2019). There is no assumption that actuarial discounted expected value by itself equals the balance-sheet liability under every accounting regime. The IFRS 17 standard overview identifies fulfilment cash flows, risk adjustment for non-financial risk, and contractual service margin where applicable. Multi-state projections may supply expected cash flows to such measurement, but the discounting, risk adjustment, group aggregation, expense perimeter, and recognition rules follow the applicable accounting framework.

Fully reproducible four-state example

Consider one hypothetical five-year contract beginning active at time zero. States are AA active, DD disabled, MM deceased, and LL lapsed; MM and LL are absorbing for insurance liability. Constant transition intensities, per year, are α=0.080\alpha=0.080, ρ=0.200\rho=0.200, μA=0.015\mu_A=0.015, μD=0.040\mu_D=0.040, ℓA=0.025\ell_A=0.025, and ℓD=0.015\ell_D=0.015. The generator rows are A=(−0.120,0.080,0.015,0.025)A=(-0.120,0.080,0.015,0.025) and D=(0.200,−0.255,0.040,0.015)D=(0.200,-0.255,0.040,0.015), followed by two zero rows. These numbers are deliberately illustrative and are not calibrated to any actual country, age, medical condition, or insurer.

The contract collects 1,500 monetary units per year while active, pays 12,000 per year while disabled, pays a 50,000 death benefit upon active-to-death transition, and 40,000 upon disabled-to-death transition. It also pays 1,000 at the instant of lapse from either covered living state. Disability premiums are waived; no other administrative expenses, benefit offsets, claim waiting period, maturity benefit, tax, inflation or risk margin is included. Discounting uses constant continuously compounded force δ=0.03\delta=0.03 per year. Benefits are paid continuously in the occupied state or at a modeled transition, as applicable; expiry is exactly five years.

Matrix exponentiation gives the following probability distributions for a single contract starting active. Rows sum to one within rounding.

Years from inception Active Disabled Deceased Lapsed
1 0.893716 0.066550 0.015586 0.024148
3 0.742279 0.141004 0.048676 0.068042
5 0.640395 0.170408 0.081989 0.107208

Numerical quadrature of the discounted occupancy functions over five years yields 3.685070 active years and 0.519769 disabled years per policy at inception. The monetary decomposition is:

Expected discounted component Monetary units
Active-state premiums received −5,527.61
Disability annuity benefits +6,237.23
Active-to-death lump sums +2,763.80
Disabled-to-death lump sums +831.63
Lapse lump sums from active +92.13
Lapse lump sums from disabled +7.80
Net insurer expected present value +4,404.98

A positive result here is expected net insurer outflow under the stated premium, not a recommended insurance price and not an IFRS 17 liability. The discount-integrated occupancy values can be independently reconstructed as the first row of ∫05exp⁡[(Q−δI)t] dt\int_0^5 \exp[(Q-\delta I)t]\,dt, restricted to active and disabled columns. For invertible Q−δIQ-\delta I, the matrix expression equals (Q−δI)−1[exp⁡((Q−δI)5)−I](Q-\delta I)^{-1}[\exp((Q-\delta I)5)-I]. Numerical integration avoids matrix-inverse instability and works under piecewise generators.

The sensitivity calculation replaces the annual disabled-to-active recovery intensity while keeping all other rates, rewards, and discounting unchanged. The disabled row diagonal is recomputed as minus the sum of exits for each case. At ρ=0.100\rho=0.100, the net insurer expected present value is 5,453.58 units; at the baseline ρ=0.200\rho=0.200, it is 4,404.98; and at ρ=0.300\rho=0.300, it is 3,554.63. The increase from baseline cost when recovery falls to 0.100 is 1,048.60 units, about 23.8% of the baseline net cost. This effect follows from longer discounted disability occupancy and the simultaneous changes in future premium exposure and death/lapse transitions. It is a scenario comparison, not a causal empirical estimate of medical rehabilitation benefits.

The model's expected counts and benefits require further distinctions. If the full contract starts active and recovery permits renewed disability, the expected number of active-to-disabled entries over five years is α∫05πA(t)dt\alpha\int_0^5\pi_A(t)dt, without discounting; the expected discounted disability annuity uses the discounted disabled occupancy integral. A death benefit's expected present value is based on the rate of covered death transitions, not the final cumulative death probability multiplied by a generic discount at year five. Conflating these quantities changes the premium and reserve even when the final occupancy table is correct.

Payment timing, stochastic interest and competing risk metrics

Discrete monthly benefits require sums over actual payment dates, with a definition of whether disability eligibility is tested at the start or end of each period and how deaths during the month affect entitlement. Continuous-annuity formulas approximate discrete payments only with a stated timing convention. A policy with a 90-day waiting period cannot be faithfully valued simply by deleting the first 90 days of payments from all paths; eligibility depends on spell history and repeated entries. A duration state or pathwise contractual algorithm is needed.

Deterministic discount rates make the expected value calculation linear in state probabilities. When mortality, morbidity, lapse, and interest are stochastic and dependent, replacing their joint expectation by a product of marginal expectations generally loses covariance effects. A consistent asset–liability model may therefore require a joint economic and biometric state system or a properly justified conditional simulation. Diversification, capital and tail quantiles are further questions about a distribution of values across policies, not automatic consequences of a single-policy expectation.

Experience Calibration

Data architecture and state assignment

A credible empirical exercise begins with exposure and event construction. At minimum, policy identifiers, coverage effective dates, coverage end reasons, birth dates or attained ages, enrollment/selection variables, observed disability onset and recovery dates, death dates, and the benefit definition should be available with documented consistency. For duration modeling, each spell entry date must be retained; a final aggregated disability status cannot recover elapsed spell duration. Payments and adjudication records are connected to episodes but should not replace event-time history without an explicit reporting-delay model.

Data may come from policy administration, claims, medical assessments, mortality registries and population surveys. Agreement across sources requires linkage quality, privacy controls, de-duplication and definitions of who is at risk of each transition. The HMD methods overview documents national death rates derived from population deaths and estimated exposures. These can support mortality benchmarking or a demographic starting point. The database does not, by itself, provide an insured population's disability incidence, recovery, lapse or health-state joint transitions. Importing national mortality into an insurance model without selection, age, contract and socioeconomic adjustment requires explicit justification.

The US ASOP 23 data-quality standard provides a practical institutional reference for evaluating completeness, relevance, questionable data and disclosure, within its jurisdictional and professional scope. It does not imply that a statistical record has been independently audited. Each data extract should be accompanied by definition version, observation window, permissible lag, date precision, curation actions, and a reconciliation of starting and ending policy counts. Across countries, policy waiting periods, disability definitions, health-system reporting and mortality-coding changes can make superficially comparable rates substantively different.

Exact-path likelihood

When complete paths are observed for independent policies conditional on covariates, the usual continuous-time Markov likelihood factorizes into event rates at observed transitions and survival of current-state exposure against its total exit rate. For an observed path occupying states i(t)i(t) over [0,T][0,T] and jumping at times tkt_k from iki_k to jkj_k,

log⁡L=∑klog⁡qikjk(tk)−∫0T∑j≠i(t)qi(t)j(t) dt,\log L=\sum_k\log q_{i_kj_k}(t_k) -\int_0^T\sum_{j\ne i(t)}q_{i(t)j}(t)\,dt,

with additional terms where state misclassification or incomplete observation must be modeled. This formula assumes exact state and event times, correctly constructed observation start and right-censoring, and specified rates. It cannot be applied directly to annual endpoints by inventing jump times. A semi-Markov likelihood conditions rates on observed or latent sojourn durations and handles recurrent entries; histories are necessary to reconstruct the relevant risk sets.

A constant-rate estimate with fully observed exposure is the number of qualifying transitions divided by person-time in the origin state, conditional on the event and censoring structure. For time-varying rates, Poisson exposure approximations, piecewise-constant hazards, parametric regression, or Cox-type transition-specific hazards are feasible. The mstate implementation guide describes transition-specific covariates and semiparametric estimation, and msm's panel-data account makes clear that intermittently observed states require a likelihood over transition matrices. These are alternative observation models, not interchangeable estimators.

Identifiability and observation bias

A disability prevalence series alone generally does not identify incidence and recovery separately: many combinations of entry and exit rates yield similar occupancy. Death and lapse further complicate the balance. A complete parameterization needs either transition observations or sufficiently restrictive, justified assumptions. Even with recorded claims, the cause of a claim closure may be ambiguous. When death or lapse is coded as generic termination, reclassifying it as recovery will mechanically overstate recovery hazard. Time-dependent underwriting selection and cessation of premiums can also create informative dropout.

Correct use of censored data requires a target estimand and a mechanism for observation. Independent administrative cutoff may justify conventional independent censoring; missing visits linked to worsening health generally call for sensitivity analysis or an identified observation model. Datta and Satten (2002) address particular covariate-dependent censoring conditions, not all forms of informative missingness. Hidden Markov models can adjust modeled misclassification if emission probabilities are identified or constrained, but they cannot recover information that the data do not contain.

Validation at contractual horizons

Calibration fit should be assessed by transition and by resulting contract value. Tests include state occupancy at predetermined horizons, competing cumulative incidence, spell-duration distributions, recurrent episode counts, claim-in-payment prevalence, and the timing of premiums, death benefits and disability payments. Where feasible, holdout validation should separate calendar periods and cohorts, and pay attention to underwriting changes, new diseases, occupational trends, medical practice, claims administration, and lapse behavior. Predictive intervals should include material parameter and model uncertainty rather than only Monte Carlo noise.

A reserve comparison should hold contract provisions and interest assumptions constant when testing biometric model alternatives. Conversely, an integrated actuarial assessment should propagate all justified simultaneous changes. A sensitivity grid for recovery duration, incidence, lapse and disabled mortality can reveal cancellation effects in aggregate reserves even when individual rates are poorly calibrated. A seemingly accurate total expected present value may be achieved by overstating one benefit while understating another; benefit-level reconciliation is required.

The United States ASOP 56 Modeling standard addresses model risk mitigation, output checking, intended application and communication for actuaries subject to that standard. The IAA international practice catalogue identifies ISAP 1 on general actuarial practice and ISAP 4 on IFRS 17 among other model-actuarial materials. They are model standards whose application depends on professional and national adoption; their existence is not a declaration of universally identical legal obligations. The relevant rule for any actual valuation must be checked for local authority and effective date.

Parameter uncertainty, model uncertainty and stress

Experience data exhibit several distinct forms of uncertainty. Process variation is the randomness of transitions conditional on an assumed intensity surface. Parameter estimation uncertainty concerns the surface itself, often larger in thin recovery or disabled-mortality cells. Model-form uncertainty arises from Markov versus semi-Markov assumptions, duration smoothing, age dependence and heterogeneity. Economic assumptions and future changes in medical care add scenario uncertainty. A report that provides only conditional process variance leaves these other dimensions undocumented.

A disciplined validation sequence begins with a state-and-benefit dictionary and exposure reconciliation, then checks event coding, estimates alternative hazard structures, predicts out-of-time state occupancies, attaches contractual payments, compares reserve decomposition, and documents stress results. For semi-Markov models it also checks calibration by current spell duration rather than merely calendar-year age. Internal consistency constraints include absorbing-state monotonicity for death and terminal coverage, nonnegative intensities, normalized transition probabilities, and terminal reserve behavior. Under the defined target, each discrepancy should be traceable to data, parameters, discounting, or benefit logic.

Research Methods and Reproducibility

Source verification and limits of inference

The investigation used primary bibliographic and methodology sources, publisher records, open-access academic articles, and primary professional institutional materials available as of November 14, 2026. The Aalen–Johansen foundational record was checked at its journal archive, while Putter et al. (2007), the Journal of Statistical Software materials, and official standards were checked on publisher or institution pages. Where a book or subscription article was available primarily through its publisher description or abstract, its verified contribution is limited to the stated subject, publication identity and described method; detailed inaccessible derivations are not attributed to it.

The analytic derivations above follow from the explicitly displayed generator, conditioning, and reward assumptions, rather than relying on bibliographic authority. The numerical example was evaluated by continuous-time transition matrix exponentials and independent numerical quadrature of discounted occupancy. All figures are rounded for presentation; the formulas define the reproducible calculation. The sensitivity cases modify only the disabled-to-active transition intensity and the associated row diagonal. No insurer records, confidential databases, empirical fitted parameter surfaces or external peer review are represented as having been used.

Independent numerical reproduction protocol

A numerical reproducer can construct the four-by-four generator QQ shown above, evaluate π(t)=eAexp⁡(Qt)\pi(t)=e_A\exp(Qt), integrate the two discounted occupancy components over [0,5][0,5], and multiply them by their contract-specific net flow terms. The active-state expected net rate equals −1,500+0.015(50,000)+0.025(1,000)=−725-1{,}500+0.015(50{,}000)+0.025(1{,}000)=-725 units per active year. The disabled-state expected net rate equals 12,000+0.040(40,000)+0.015(1,000)=13,61512{,}000+0.040(40{,}000)+0.015(1{,}000)=13{,}615 per disabled year. Thus the stated inception value also equals

−725∫05e−0.03tπA(t)dt+13,615∫05e−0.03tπD(t)dt.-725\int_0^5e^{-0.03t}\pi_A(t)dt +13{,}615\int_0^5e^{-0.03t}\pi_D(t)dt.

Using the displayed discounted occupancies gives 4,404.98 monetary units after rounding. This compact reconciliation checks the sum of separate premium, disability, death, and lapse components against the reward-vector method. A second independent valuation method solves the Thiele equations backward, under terminal zero conditions, and should return the same inception value within numerical tolerance. A second numerical verification solved the four-state Thiele differential equations backward from zero terminal reserves using an adaptive ODE solver; the active-state initial reserve was 4,404.98282732 units, agreeing with the forward matrix-exponential/quadrature value to within 7 × 10⁻¹⁰ monetary units. Changing only the recovery intensity to 0.100 and 0.300 reproduced the corresponding stated net present values within 10⁻⁸ monetary units.

Scope boundaries and cross-disciplinary interpretation

The results are actuarial and statistical statements. The clinical appropriateness of health-state categories, jurisdictional recognition of disability, and economic interpretation of lapses require their own evidence. Some private disability contracts pay to a maximum age; long-term care may involve changes in dependency grade and care setting; term life contracts may be dominated by death transitions; and critical-illness products may pay once on diagnosis. Each uses a related transition apparatus with a different state and reward map. The general equations remain reusable when their assumptions are satisfied; the model inputs, observational design and benefit logic must be changed.

A model's scientific sufficiency is determined by whether its state and history variables carry the information required to estimate the relevant contractual future cash flows at the required precision. More states always create more parameters, so claims of improved realism must be balanced against increased variance, data fragmentation and instability. The operational criterion is out-of-sample fit to contract-relevant quantities, with transparent source and model uncertainty. This is an original editorial synthesis of the established methods discussed, without a new named conceptual system.

Conclusion

Multistate models link event histories to life and health insurance liabilities by assigning transition intensities or duration-dependent transition laws to a precisely defined contractual state space. They distinguish the number of claim episodes, the probability of being disabled at a date, duration of disability, death contingencies, coverage lapse, and dated benefit cash flows. Each quantity is a different functional of the underlying stochastic process, and each may require a different observation and contractual treatment.

The Markov representation gives transition matrices through forward and backward equations and supports prospective reserves through multistate Thiele differential equations. Semi-Markov specifications retain spell duration when recovery, death, or other exits depend on time already spent in a state. Empirical estimation must distinguish exact event histories from panel observations, independent censoring from missingness related to health, and coverage termination from mortality or recovery. Foundational Aalen–Johansen estimation, non-Markov landmark methods and modern statistical software give a mature collection of methods whose validity depends on these precise conditions.

The hypothetical four-state calculation shows why the choice of recovery assumptions is financially material: varying an annual recovery intensity between 0.100 and 0.300, with all other specified assumptions fixed, changes a five-year net expected insurer outflow from approximately 5,453.58 to 3,554.63 monetary units. That result is transparent, reproducible and bounded to its explicitly hypothetical contract. It substantiates the general methodological point that transition rates, contract rewards and discounting must be checked together.

The research's principal synthesis is that multistate insurance valuation is reliable when health history, contractual eligibility, event observation, exposure construction, discounting and accounting perimeter remain separately identified and then coherently connected. A model can fit observed state frequencies while misvaluing benefits if its durations or payments are wrong; a correct reward equation can misvalue coverage if its state model is misidentified. Scientific and editorial quality therefore require both probabilistic validation and contract-level cash-flow reconciliation. These findings elaborate established actuarial science and do not constitute a new proprietary theory.

References

  • Aalen, Odd O., and Søren Johansen. 1978. “An Empirical Transition Matrix for Non-Homogeneous Markov Chains Based on Censored Observations.” Scandinavian Journal of Statistics 5 (3): 141–150. Source record

  • Putter, Hein, Marta Fiocco, and Ronald B. Geskus. 2007. “Tutorial in Biostatistics: Competing Risks and Multi-State Models.” Statistics in Medicine 26 (11): 2389–2430. doi:10.1002/sim.2712. Source record

  • de Wreede, Liesbeth C., Marta Fiocco, and Hein Putter. 2011. “mstate: An R Package for the Analysis of Competing Risks and Multi-State Models.” Journal of Statistical Software 38 (7): 1–30. doi:10.18637/jss.v038.i07. Source record

  • Jackson, Christopher H. 2011. “Multi-State Models for Panel Data: The msm Package for R.” Journal of Statistical Software 38 (8): 1–29. doi:10.18637/jss.v038.i08. Source record

  • Jackson, Christopher H. 2016. “flexsurv: A Platform for Parametric Survival Modeling in R.” Journal of Statistical Software 70 (8): 1–33. doi:10.18637/jss.v070.i08. Source record

  • Titman, Andrew C. 2015. “Transition Probability Estimates for Non-Markov Multi-State Models.” Biometrics 71 (4): 1034–1041. doi:10.1111/biom.12349. Source record

  • Putter, Hein, and Cristian Spitoni. 2018. “Non-parametric Estimation of Transition Probabilities in Non-Markov Multi-State Models: The Landmark Aalen–Johansen Estimator.” Statistical Methods in Medical Research 27 (7): 2081–2092. doi:10.1177/0962280216674497. Source record

  • D’Amico, Guglielmo, Montserrat Guillen, and Raimondo Manca. 2009. “Full Backward Non-Homogeneous Semi-Markov Processes for Disability Insurance Models: A Catalunya Real Data Application.” Insurance: Mathematics and Economics 45 (2): 173–179. doi:10.1016/j.insmatheco.2009.05.010. Source record

  • Haberman, Steven, and Ermanno Pitacco. 1998. Actuarial Models for Disability Insurance. 1st ed. London: Chapman & Hall. Publisher reissue listing ISBN 9780849303890. Source record

  • Dickson, David C. M., Mary R. Hardy, and Howard R. Waters. 2019. Actuarial Mathematics for Life Contingent Risks. 3rd ed. Cambridge: Cambridge University Press. Source record

  • Actuarial Standards Board. 2019. Actuarial Standard of Practice No. 56: Modeling. Revised December 2019; effective October 1, 2020. Source record

  • Actuarial Standards Board. 2016. Actuarial Standard of Practice No. 23: Data Quality. Revised December 2016; effective 2017. Source record

  • International Actuarial Association. 2026. International Standards of Actuarial Practice: Official Standards Catalogue (including ISAP 1, ISAP 4 and model governance materials). Source record

  • IFRS Foundation. 2026. IFRS 17 Insurance Contracts: Standard Overview and Key Principles. IASB; standard first issued 2017, effective 2023. Source record

  • Human Mortality Database. 2026. Project Overview: Scope, Data Sources, Exposure-to-Risk and Mortality Rate Methods. University of California, Berkeley, Max Planck Institute for Demographic Research, and collaborators. Source record

  • De Giovanni, Domenico, Massimiliano Menzietti, Marco Pirra, and Fabio Viviano. 2025 (online publication; 2026 issue). “Functional Disability in the Health and Retirement Study: A Semi-Markov Multi-State Analysis.” Annals of Operations Research 363: 1133–1160. doi:10.1007/s10479-025-06526-7. Source record

  • Datta, Somnath, and Glen A. Satten. 2002. “Estimation of Integrated Transition Hazards and Stage Occupation Probabilities for Non-Markov Systems under Dependent Censoring.” Biometrics 58 (4): 792–802. doi:10.1111/j.0006-341X.2002.00792.x. Source record