From Transition Counts to a Cumulative Estimate
A multi-state data set records not only whether an event occurred, but also which state change occurred and when. For a specified transition, the Nelson–Aalen estimator combines the observed transition counts with the number of individuals at risk for that transition just before each event time. Its value accumulates across time: each observed transition contributes an amount determined by the corresponding risk set.
The calculation uses the transition-specific risk sets and event counts defined in the earlier material on transition-specific risk sets and event counts. The crucial bookkeeping rule is to use the risk set for the transition being estimated, at the time of that transition. A count or denominator for another transition is not substituted.
In the usual intensity notation, \(\alpha_{rs}(u)\) is the transition intensity for \(r\to s\), and its cumulative intensity is \(A_{rs}(t)=\int_0^t\alpha_{rs}(u)\,du\). Intensity is measured per unit time; cumulative intensity is dimensionless. The estimator is a step function: it changes at observed transition times for \(r\to s\), and stays constant between those times.
This calculation does not itself require that the process be Markov. The interpretation of \(\alpha_{rs}(u)\), however, depends on how the transition intensity is defined. If it is modeled as depending only on the current state and time, that is a Markov-style specification; with history-dependent intensities, the relevant history must be accounted for in the risk-set definition or analysis. The sum always gives the stated empirical calculation, but the target it estimates depends on the model and data-collection assumptions.
How to Calculate the Estimator
At each event time for the transition of interest, divide the number of observed transitions by the number at risk immediately beforehand. Add those fractions through the time point being reported. If there is no event for that transition at a time, that time contributes nothing to its sum. Keep the transition label attached throughout: \(\widehat A_{rs}\) and \(\widehat A_{rq}\) are separate estimates, even when both transitions start in state \(r\).
Worked Example: Two Event Times for One Transition
Hypothetical data. In a cohort followed for a transition from state A to state B, one A-to-B transition occurs on day 2 and another on day 5. The risk sets immediately before those times are 10 and 7 individuals, respectively. Assume these are distinct event times and the risk sets correctly count individuals eligible for A-to-B.
At day 2, the event count is 1 and the risk set is 10, so the contribution is \(1/10=0.1000\). At day 5, the event count is 1 and the risk set is 7, so the contribution is \(1/7\approx0.1429\).
Thus the estimated cumulative intensity for A-to-B through day 5 is 0.2429, a dimensionless quantity. The estimate remains at 0.1000 between day 2 and day 5, then increases to 0.2429 at day 5. It is not a 24.29% transition probability.
A useful way to express the same computation is as a sum over jumps of the counting process. For the observed step function \(N_{rs}\), integration against \(dN_{rs}(u)\) adds the integrand once for each event counted. Therefore,
This identity justifies the equivalent integral notation for the estimator. It also explains the arithmetic: at an event time with two observed transitions, the contribution is \(2/Y_{rs}(u)\), not \(1/Y_{rs}(u)\). The justification is a property of summing over jumps; it does not by itself establish that the estimator targets a particular population cumulative intensity.
Worked Calculations for Multiple Transitions and Time Points
Worked Example: Keeping Two Outgoing Transitions Separate
Hypothetical data. Patients in remission can transition to relapse or death. The following event times and risk sets are supplied by the study records; all listed event times are distinct. There is one event of each specified type at each listed time.
| Transition | Event time (months) | Event count | Risk set immediately before time |
|---|---|---|---|
| Remission to relapse | 1 | 1 | 20 |
| Remission to death | 2 | 1 | 18 |
| Remission to relapse | 4 | 1 | 13 |
| Remission to death | 4.5 | 1 | 11 |
For remission to relapse through month 4, use only its two event times:
For remission to death through month 4.5, use its own event times and denominators:
These are two transition-specific cumulative-intensity estimates. The table supplies the risk sets for the calculation; one should not replace them with a total cohort size or combine the two transition counts into a single numerator.
Worked Example: Estimating the Increase Over a Time Interval
Hypothetical data. For a specified transition, events occur at times 1, 2, 3.5, 6, and 8 days. Each event count is 1. The corresponding risk sets are 15, 12, 9, 6, and 4. There are no other events for this transition through day 8.
First calculate the cumulative estimate at day 3:
At day 8, all five contributions are included:
The estimated increase after day 3 through day 8 is the sum of the later contributions, equivalently \(\widehat A(8)-\widehat A(3)=0.6778-0.1500=0.5278\). This subtraction works because the cumulative estimate adds the contributions over time. Its value describes accumulated intensity over that interval, not the probability of making the transition during it.
Conditions, Checks, and Interpretation
The arithmetic is simple, but its scientific interpretation requires sound data and an appropriate target. The state definitions must make the transition \(r\to s\) unambiguous, event times and event types must be ascertained correctly, and each \(Y_{rs}(u)\) must count the individuals actually eligible and observed immediately before the event time. If the scientific target is a common transition intensity for individuals in the risk set, the analysis also needs a defensible assumption that the relevant intensity is shared or appropriately modeled across those individuals. If follow-up is censored, interpreting the estimate for the target population additionally requires suitable assumptions about how censoring relates to transition risk.
Before reporting a result, check that every denominator is positive and that the event count does not exceed the corresponding risk set. Verify that the risk set is measured just before the event time, and that only events of the transition being estimated enter the sum. For an estimator starting at zero, every contribution is nonnegative, so its values cannot decrease over time. A negative contribution, a decrease in the cumulative estimate, or a division by zero signals an error or a boundary that cannot be estimated from the available risk set.
Small risk sets can produce large jumps: one event among four at risk contributes 0.25, while one event among 100 contributes 0.01. Such jumps are not automatically computational mistakes, but they indicate that the estimate can be sensitive to a small number of observations. The step function also does not provide a smooth estimate of the instantaneous intensity between event times.
| Approach | What it calculates | Key distinction |
|---|---|---|
| Nelson–Aalen estimator | A cumulative transition-intensity estimate from event counts and risk sets | Nonparametric and stepwise; uses each transition’s own event contributions |
| Parametric intensity model | A cumulative intensity by integrating a specified fitted intensity function | Requires a chosen model form; can yield a smoother curve if the model is suitable |
| Smoothed intensity estimate | A locally smoothed estimate of intensity over time | Targets a time-varying rate directly and depends on smoothing choices |
The Nelson–Aalen estimator is useful when a transparent, nonparametric summary of accumulated transition intensity is wanted. It is limited by sparse risk sets, poor event-time or state classification, and assumptions needed to interpret the estimate beyond the observed risk sets. It is not a transition probability: converting cumulative transition intensities into transition probabilities requires additional multi-state calculations.
Common Mistakes
- Using the wrong denominator. Dividing by the original cohort size ignores who remained eligible and observed immediately before the event. Use the corresponding risk set at that time.
- Mixing transition types. Adding relapse events to the death-transition numerator changes the estimand. Calculate each specified transition separately.
- Calling the estimate a probability. A value such as 0.2429 is a cumulative-intensity estimate, not a 24.29% chance of transition.
- Forgetting earlier contributions. The cumulative estimate at time \(t\) includes all contributions for that transition up to \(t\), not only the latest event.
- Ignoring a small denominator. A large jump may be arithmetically correct but unstable. Report the risk-set context and avoid implying greater precision than the data support.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For one transition, there is one event at time 2 with 8 at risk and one at time 5 with 4 at risk. Calculate the cumulative estimate at time 5.
- Why should a transition-specific estimate not include an event count for a different transition?
- In a calculation, the risk set is 10 and the event count at a time is 2. What is that time’s contribution?
- An estimated cumulative intensity is 0.30 at time 4. Is this a 30% transition probability? Explain briefly.
Solutions. (1) \(\widehat A(5)=1/8+1/4=0.1250+0.2500=0.3750\). (2) Each estimator summarizes a specified transition; including another transition changes the event count and no longer estimates that transition’s cumulative intensity. (3) The contribution is \(2/10=0.2000\). (4) No. It is a dimensionless cumulative-intensity estimate; it is not itself a transition probability.