Multi-State Life Table

Multi-State Life Table: Concept Architecture

A multi-state life table follows a hypothetical population as its members move among living states and death. It can estimate remaining years lived in each state, such as years without and with disability, while allowing onset, recovery, and state-specific mortality. This page explains the transition inputs, the projection and person-time calculation, a worked example, and the interpretation limits that matter in population health and health economic models.

Define states, entry, and time before calculating expectancy

States must be mutually exclusive at each time point and sufficiently clear to classify a person, while together covering the modeled possibilities. A simple table might use healthy, disabled, and dead; a disease-specific application may use several severity states and remission. Death is usually absorbing, and nonfatal states may be reversible when the clinical process allows recovery.

A life table begins with a specified starting age, living state or mixture of states, and period or cohort perspective. Its transitions can depend on age, sex, calendar period, treatment, and other modeled characteristics. A period multi-state life table applies age-specific transition conditions observed in a defined period to a hypothetical cohort as it ages; its projected remaining years are not a direct forecast of the actual future of a birth cohort when transition patterns change. Specify the width of each age or time interval and whether a transition probability means destination state at the next boundary or occurrence of an event at any time during the interval.

Follow people using age-specific transition probabilities

For states indexed by $i$ and $j$, let $p_{ij}(a)$ be the probability of occupying state $j$ at age $a+\Delta$ conditional on occupying state $i$ at age $a$, under the modeled population and interval width $\Delta$. Collect the probabilities in a row-stochastic matrix $P_a$: every entry is nonnegative and every row sums to one. Let $m_a$ be a row vector of probabilities of occupancy at the start of the interval. Then:

$$m_{a+\Delta}=m_aP_a,\qquad m_{a+n\Delta}=m_aP_aP_{a+\Delta}\cdots P_{a+(n-1)\Delta}.$$

When the probabilities depend on age, do not replace the ordered product by a power of one fixed matrix. Conditional on a living initial state $s$, the initial vector $m_a$ has a one in position $s$ and zero elsewhere. This yields state-specific remaining expectancies; a mixture of initial states produces a weighted expectancy for that specified mixture. In a population with transitions from healthy to disabled and back, the two directions need distinct estimates, and death probabilities can differ by starting living state.

Transition probabilities are not interchangeable with instantaneous transition rates. With a continuous-time intensity matrix $Q(a)$, conversion to interval probabilities needs an appropriate multi-state calculation; for a constant $Q$ over an interval, $P(\Delta)=\exp(Q\Delta)$. Applying a separate single-event rate-to-probability conversion to each of several competing exits may create a row that does not sum to one or misrepresent their competition. The time scale and estimation assumptions must match the life table.

Convert occupancies into years lived in each state

Expected time in state $j$ is the sum of its occupancy probability over successive intervals, with a justified rule for within-interval occupancy. In continuous time, the expected years in living state $j$ from age $a$ to horizon $\tau$ are $E[T_j(a,\tau)]=\int_a^{\tau}P(X_u=j\mid X_a=s)\,du$. In an annual discrete table, a trapezoidal or half-cycle approximation uses average occupancy at adjacent boundaries:

$$E[T_j(a,\tau)]\approx\sum_{k=0}^{K-1}\frac{\Delta}{2}\left[m_{a+k\Delta,j}+m_{a+(k+1)\Delta,j}\right],\qquad \tau=a+K\Delta.$$

These are years per person from the specified starting condition; multiplying by an initial life-table radix gives aggregate person-years for that hypothetical cohort. The sum of expected years over all living states is remaining life expectancy over the same horizon. Counting time in the absorbing death state as lived years would be an error. The trapezoidal rule is an approximation to within-interval occupancy, not a repair for incorrect transitions or coarse state definitions. For complete remaining life expectancy, extend the table far enough or justify an explicit terminal-age treatment; a short finite horizon yields restricted state expectancies.

Worked example with recovery and unequal death risks

Suppose the following illustrative transition matrix applies for each of the first two one-year intervals. Rows identify the starting state and columns the state at the next boundary. These are invented probabilities for teaching, not empirical age-specific estimates.

Every row sums to one. Recovery from disabled to healthy has probability 0.10, and the one-year probability of death differs between healthy (0.05) and disabled (0.20). Starting healthy gives $m_0=(1,0,0)$, $m_1=(0.80,0.15,0.05)$, and $m_2=(0.655,0.225,0.120)$. For example, healthy occupancy at year two is $0.80(0.80)+0.15(0.10)=0.655$; disabled occupancy is $0.80(0.15)+0.15(0.70)=0.225$; and death occupancy is $0.80(0.05)+0.15(0.20)+0.05(1)=0.120$.

Using average endpoint occupancy within each one-year interval, expected healthy years through year two are $(1+0.80)/2+(0.80+0.655)/2=1.6275$. Expected disabled years are $(0+0.15)/2+(0.15+0.225)/2=0.2625$. Their sum is 1.8900 years alive within the two-year horizon. These figures depend on the stated within-cycle approximation; they are neither unrestricted lifetime expectancies nor direct counts of people experiencing disability. An exact continuous-time model could produce different person-time even if its boundary occupancies agree.

Distinguish transitions from prevalence weighting

The multi-state approach ordinarily uses transition information, such as incidence of disability, recovery, and mortality conditional on the current state. It can explicitly represent reversible movement and the different survival of people in different living states. Prevalence alone reports a cross-sectional proportion in a state and generally does not identify all of those transitions.

The Sullivan method instead combines a conventional life table's person-years with age-specific prevalence of a health state to partition life expectancy. It is valuable when longitudinal transition data are unavailable, and it is the method used for Eurostat's Healthy Life Years indicator. Its output may be similar to a transition-based estimate in some settings, but the two methods use different inputs and may respond differently when incidence, recovery, and prevalence are changing. Do not label a prevalence-weighted calculation a transition-based multi-state life table, or infer recovery probabilities from one cross-section alone.

Estimate transitions and check assumptions

Longitudinal observations can inform transitions among living states and death, but observed visits may miss movement that occurs between them. If a person is healthy at both visits, they may have remained healthy or passed through disability and recovered; the observation scheme must be considered in the estimation method. Death ascertainment, health-state measurement, informative dropout, left truncation, and competing events also affect estimated probabilities.

The model's memory assumption deserves particular attention. If the next transition depends on duration in disability, earlier episodes, or treatment history even after the current state and age are known, a simple first-order state definition may be inadequate. Duration-specific tunnel states, expanded histories, semi-Markov methods, or individual-level simulation may be justified. Calendar-time, age, and time-in-state are different clocks; age-dependent transition matrices do not automatically capture duration since an individual entered disability. Validate probabilities, row sums, survival, recovery patterns, and state-specific mortality against available evidence.

Interpret health and economic outputs carefully

State years can reveal whether longer life is spent in healthier or less healthy states. If state-specific utilities $u_j$ are applied consistently, an illustrative health-weighted time calculation is $\sum_{j\in L}u_jE[T_j]$, where $L$ is the set of living states; additional within-state variation, event disutilities, discounting, and the evaluation's QALY conventions may require a richer calculation. Healthy life expectancy as a count of years in a defined healthy state is not automatically a QALY total or the World Health Organization's health-adjusted life expectancy measure.

Similarly, a state occupancy table can support projected resource use and costs when states have defensible cost assignments. Clinical treatment effects must be translated into supported transition changes, not inferred merely from a difference in observed prevalence. For comparative economic evaluation, keep starting populations, state definitions, horizons, transition evidence, and discounting consistent across alternatives; report uncertainty in both transition estimates and model structure.

Checks before reporting a multi-state life table

These checks connect the mathematical table to the health question. They also reveal whether a result is a period estimate, a cohort forecast, or a restricted-horizon teaching example. Record each choice alongside the reported years.

1. Define states and entry. Specify health-state criteria, reversibility, absorbing death, starting age, initial state, and population.
2. Specify time. Distinguish age, calendar period, cohort, time since disease onset, and time since state entry; state interval widths.
3. Verify input estimates. Check transition denominators, row sums, mortality by state, recovery, observation intervals, missingness, and evidence sources.
4. Project occupancy. Multiply matrices in age order and confirm each occupancy vector remains nonnegative and sums to one.
5. Integrate person-time. State the within-interval approximation or continuous-time method; sum only living states for remaining life expectancy.
6. Validate interpretation. Compare outputs with observed health-state patterns and mortality, and explain sensitivity to state definitions, transition assumptions, and terminal age.

Sources and further study

The methodological literature distinguishes transition-based life tables from prevalence-weighted health expectancy. The official population-health sources clarify how public indicators are calculated, so similarly named measures are not silently treated as identical. The numerical example above is original and calculated independently.

1. Saito Y, Robine JM, Crimmins EM. The methods and materials of health expectancy. Research review of incidence-based transitions, recovery, and state-specific mortality in health expectancy estimation.
2. Eurostat. Methodology for Healthy Life Years and health methodology. Official explanation of the transition approach and Eurostat's Sullivan-based indicator.
3. WHO. Healthy life expectancy (HALE) metadata. Official description of a related but differently constructed health-adjusted measure.
4. The Multistate Life Table Method: An Application to Contraceptive Switching Behavior. Original applied methods paper on transition matrices, occupancy, state time, and visits.