Cumulative Exposure As A History Predictor

Why accumulated state occupancy can matter

Two individuals can be in the same state at the same follow-up time, and can have spent the same amount of time in that state on their current visit, yet have different earlier histories. If earlier time in a particular state has lasting relevance, current state and current-visit duration alone may not summarize the history needed to predict a transition. A cumulative-exposure predictor records that earlier occupancy as a running total.

For example, in an illness-death process, time spent in an illness state might be associated with a later death intensity even after a person leaves and re-enters that state. The model below expresses that association; it does not, by itself, establish that time in the state causes a change in the transition rate.

Definition and intensity model

Let \(X(t)\) denote the state occupied at time \(t\), with time measured from a common origin. Consider a specified state \(r\). Its cumulative occupancy exposure by time \(t\) is the total time spent in state \(r\) from the origin through \(t\):

A_r(t)=\int_0^t I\{X(u)=r\}\,du.

Here \(I\{\cdot\}\) is 1 when its condition holds and 0 otherwise. Thus \(A_r(t)\) has units of time, is nondecreasing, and increases at rate 1 while the process occupies \(r\), and at rate 0 otherwise. It adds together all visits to \(r\); it is not just the duration of the current visit. For a transition intensity at time \(t\), only occupancy history up to \(t\) is used. The value at a single time point does not affect the integral.

Definition: For distinct states \(j\) and \(k\), the transition intensity \(\lambda_{jk}(t\mid\mathcal H(t-))\) is the instantaneous rate of a transition from \(j\) to \(k\) at time \(t\), conditional on the observed history \(\mathcal H(t-)\) immediately before \(t\). In an intensity model with cumulative exposure, \(A_r(t)\) is a predictor summarizing prior time in state \(r\). The model is applied to the \(j\)-to-\(k\) transition while the process is at risk in state \(j\).

A common proportional-intensity specification is

\lambda_{jk}(t\mid\mathcal H(t-)) = I\{X(t-)=j\}\lambda_{0,jk}(t) \exp\{\beta_{jk}A_r(t)\}.

The indicator makes the transition intensity zero when the process is not in the origin state \(j\). Equivalently, among individuals currently in \(j\), the at-risk transition rate is \(\lambda_{0,jk}(t)\exp\{\beta_{jk}A_r(t)\}\). The baseline intensity \(\lambda_{0,jk}(t)\) may vary with calendar time \(t\). It has units of inverse time; because \(A_r(t)\) is measured in time, \(\beta_{jk}\) has units of inverse time, making the exponent dimensionless.

At a fixed calendar time and for individuals currently in \(j\), an additional \(\Delta\) units of cumulative exposure multiplies the modeled transition intensity by \(\exp(\beta_{jk}\Delta)\), holding the other model terms fixed. A positive coefficient means higher accumulated exposure is associated with a higher transition intensity under this model; a negative coefficient means a lower one. The interpretation is conditional on the specified model and is not automatically causal.

Worked Example: Accumulating Exposure Across Visits

Suppose a hypothetical process alternates between states 0 and 1 before a possible transition from state 1 to state 2. Its path is state 0 from time 0 to 2 years, state 1 from 2 to 5, state 0 from 5 to 6, and state 1 from 6 to 8. Calculate cumulative exposure in state 1 at time 8 and the modeled 1-to-2 intensity there, using \(\lambda_{0,12}(8)=0.0800\) per year and \(\beta_{12}=0.300\) per year.

The two visits to state 1 last 3 years and 2 years. Therefore,

A_1(8)=(5-2)+(8-6)=3+2=5\text{ years}.

The process is in state 1 at time 8, so it is at risk for the 1-to-2 transition. Substituting the values into the model gives

\lambda_{12}(8\mid\mathcal H(8-)) = 0.0800\exp\{0.300(5)\} = 0.0800\exp(1.50) \approx 0.359\text{ per year}.

The current visit to state 1 has lasted only 2 years, but the cumulative exposure is 5 years because it includes the earlier visit. The computed value is an instantaneous transition intensity at time 8, not the probability of transitioning within a year.

How the model uses the path

A state path that is constant between jumps makes the exposure calculation direct. If the process occupies states \(x_m\) on successive intervals \([t_m,t_{m+1})\), then the exposure accumulated by time \(t\) is the sum of the lengths of the portions of those intervals before \(t\) for which \(x_m=r\):

A_r(t) = \sum_m I\{x_m=r\} \left[\min(t,t_{m+1})-t_m\right]_+,

where \([v]_+=\max(v,0)\), and only intervals beginning before \(t\) contribute. This form emphasizes that exposure is calculated from the observed path, not inferred just from the current state.

The intensity is conditional on the history available just before the instant being considered. In a short interval \((t,t+\Delta t]\), for a process currently in state \(j\), the model implies

P\{\text{a }j\to k\text{ transition in }(t,t+\Delta t]\mid\mathcal H(t-)\} = \lambda_{0,jk}(t)\exp\{\beta_{jk}A_r(t)\}\Delta t +o(\Delta t).

This is the defining small-interval interpretation of an intensity. It does not say that the probability over a long interval equals intensity times interval length: the intensity and the state at risk can change during that interval. The cumulative exposure is predictable from the path immediately before \(t\), so the predictor uses past occupancy rather than future information.

Worked Example: Comparing Two Histories at the Same Time

Consider two hypothetical individuals who are both in state 1 at calendar time 10 years, and suppose the baseline 1-to-2 intensity at that time is \(0.0500\) per year. Their current visits to state 1 are each 2 years long. However, one has accumulated \(A_1(10)=3\) years in state 1 and the other \(A_1(10)=6\) years. Use \(\beta_{12}=0.400\) per year to calculate their modeled intensities and compare them.

For the first history,

\lambda_{12}^{(1)}(10) = 0.0500\exp\{0.400(3)\} = 0.0500\exp(1.20) \approx 0.166\text{ per year}.

For the second history,

\lambda_{12}^{(2)}(10) = 0.0500\exp\{0.400(6)\} = 0.0500\exp(2.40) \approx 0.551\text{ per year}.

Their intensity ratio is

\frac{\lambda_{12}^{(2)}(10)}{\lambda_{12}^{(1)}(10)} = \exp\{0.400(6-3)\} = \exp(1.20) \approx 3.32.

Thus, under this model, the history with 3 additional years of accumulated state-1 exposure has an estimated 1-to-2 intensity 3.32 times as large at year 10, conditional on being in state 1. The calculation compares modeled intensities, not probabilities of eventual transition.

Assumptions, justification, and checks

The model is meaningful only when its time scale, exposure definition, and risk set match the scientific question. In particular, the following conditions should be made explicit:

Well-defined path and origin. The state process and time origin must be consistently defined, and state occupancy must be known or appropriately observed over the intervals used to calculate exposure.
Correct risk set and transition. The \(j\)-to-\(k\) intensity applies only while the process occupies \(j\); the predictor cannot make an impossible transition available.
Predictable history. Exposure at time \(t\) is computed from occupancy before \(t\). Future state occupancy must not enter the predictor.
Specified functional form. The exponential model assumes that each additional unit of exposure has the same multiplicative effect on the intensity, conditional on the baseline and other included predictors.
Appropriate baseline and observation process. Calendar-time variation, censoring, and the observation of transitions must be handled in a way compatible with the chosen estimation method. In particular, missing or intermittently observed state changes can make a path-based exposure inaccurate.

The exponential form follows directly from a log-linear intensity model. If, among those currently in \(j\), the log intensity is specified as a baseline log intensity plus a linear exposure term, then exponentiating gives the proportional-intensity expression:

\log\{\lambda_{jk}(t\mid\mathcal H(t-))\} = \log\{\lambda_{0,jk}(t)\}+\beta_{jk}A_r(t) \quad\Longrightarrow\quad \lambda_{jk}(t\mid\mathcal H(t-)) = \lambda_{0,jk}(t)\exp\{\beta_{jk}A_r(t)\}.

For two histories with exposures \(a_1\) and \(a_2\) at the same time, both currently in \(j\), division cancels the shared baseline:

\frac{\lambda_{jk}(t\mid A_r(t)=a_2)} {\lambda_{jk}(t\mid A_r(t)=a_1)} = \exp\{\beta_{jk}(a_2-a_1)\}.

This derivation justifies the rate-ratio interpretation, provided the histories share the other modeled conditions. It also shows why the coefficient's units matter: if exposure is recorded in years, a one-unit increase means one additional year, not one additional visit.

Useful checks include confirming that calculated exposure is nondecreasing, never exceeds elapsed follow-up time, and increases only during occupancy of the specified state. Compare hand calculations with the reconstructed state path, especially around gaps, censoring, and repeated visits. In model assessment, inspect whether the assumed log-linear exposure effect is plausible over the observed exposure range and whether the baseline specification leaves systematic time patterns. Sparse transitions among individuals with large exposure can make estimates unstable; extrapolation beyond observed exposure is especially uncertain.

Worked Example: Checking a Candidate Exposure Calculation

A hypothetical path is in state 2 from time 0 to 1 year, state 1 from 1 to 4, state 2 from 4 to 6, and state 1 from 6 to 7. A proposed data record says \(A_1(7)=5\) years. Check it and calculate the correct state-1 exposure.

The process occupies state 1 for \(4-1=3\) years on its first visit and \(7-6=1\) year on its second. Thus,

A_1(7)=(4-1)+(7-6)=3+1=4\text{ years}.

The proposed value of 5 years is impossible: it exceeds the total elapsed follow-up of 7 years only if counted alongside another 3 years in state 2? More directly, the intervals in state 1 total 4 years, so the record overstates exposure by 1 year. A reliable path-based audit finds this by summing only the intervals assigned to state 1; the total time in states 1 and 2 is \(3+4=7\) years, matching follow-up.

What cumulative exposure does and does not capture

Cumulative occupancy is distinct from both current-visit duration and calendar time. Current-visit duration resets when a new visit to a state begins; cumulative occupancy does not. Calendar time advances regardless of the state occupied. A model can include a calendar-time baseline and a cumulative exposure predictor together, but their effects need not be easy to distinguish if exposure changes in a highly predictable way with calendar time.

This is also different from a semi-Markov model whose transition intensity depends on time since entry into the current state. That duration-based model addresses the current sojourn; cumulative exposure addresses total past occupancy in a chosen state, potentially across several visits. A model could include both predictors when the data and scientific question support it, but then the coefficients describe their conditional associations and may be difficult to estimate if the predictors are strongly related.

A linear cumulative total treats all past exposure as equally retained: one year spent in the state long ago contributes as much as one year spent recently. It also treats each additional year as having the same log-intensity effect. These assumptions may be poor if exposure is temporary, saturates, or depends on how exposure is distributed across episodes. Possible alternatives include a nonlinear function of \(A_r(t)\), separate predictors for different occupancy states, or a weighted history that gives different importance to recent and distant exposure. Each alternative changes the scientific question and requires adequate data.

History omitted: If accumulated occupancy is relevant but omitted, predictions may differ for histories that share the same current state and current-visit duration.
Wrong exposure summary: Total time alone cannot distinguish one long visit from several short visits of the same total length.
Functional-form error: A constant log-linear effect may misrepresent thresholds, saturation, or reversibility.
Overinterpretation: State occupancy is part of the evolving process. An estimated association with a later intensity is not automatically the effect of intervening to change that occupancy.

The main advantage of cumulative exposure is a compact predictor with a clear unit and a direct path-based calculation. Its limitation is the same compactness: different histories can have the same total, so information about timing and sequence is discarded. Whether that loss is acceptable depends on the intended prediction or scientific interpretation.

Key takeaway: Define cumulative exposure from the state path, calculate it using only prior occupancy, and interpret its coefficient per unit of accumulated time. It captures total past occupancy, not the order, timing, or number of visits that produced that total.

Check Your Understanding

Answer each question, then compare with the brief solutions.

A process occupies state 3 from year 0 to 2, state 1 from year 2 to 4, state 3 from year 4 to 5, and state 1 from year 5 to 8. What is \(A_1(8)\)?
For a transition intensity with \(\beta=0.25\) per year, what is the modeled intensity ratio for two otherwise comparable histories that differ by 2 years of cumulative exposure?
Why is cumulative occupancy a different predictor from duration of the current visit?
State one assumption made by the log-linear cumulative-exposure specification and one diagnostic check for calculating exposure from a state path.

Solutions. (1) State 1 is occupied for \(4-2=2\) years and \(8-5=3\) years, so \(A_1(8)=2+3=5\) years. (2) The ratio is \(\exp\{0.25(2)\}=\exp(0.50)\approx1.65\); the history with 2 additional years has about 1.65 times the modeled intensity. (3) Current-visit duration starts over on re-entry to a state, whereas cumulative occupancy adds time across all visits. (4) The specification assumes a constant linear effect of exposure on log intensity; a path check is to verify that exposure increases only in the specified state and never exceeds elapsed follow-up time.