Why quantify uncertainty in a cumulative transition hazard?
A Nelson–Aalen estimate summarizes the accumulated transition intensity from one state to another. Its value alone does not show how much it may vary across repeated samples. A transition observed only a few times, or in a small risk set, will generally produce a less precise estimate than the same pattern of events observed in much larger risk sets. A variance estimate makes that difference quantifiable.
This tutorial focuses on uncertainty for one transition at a time. It uses the transition-specific risk sets and event counts introduced in the earlier material on transition-specific risk sets and Nelson–Aalen estimation. It does not calculate transition probabilities: a cumulative transition hazard is not itself the probability of being in a destination state.
The denominator is the risk set for the origin state \(i\), not the number of people who will ever make transition \(i\to j\). At a tied event time, use the risk set just before that time for all \(d_{ij}(t_k)\) events in the count. The contribution to the estimated variance is therefore \(d_{ij}(t_k)/Y_i(t_k-)^2\), not \(d_{ij}(t_k)/Y_i(t_k-)\) and not \(d_{ij}(t_k)^2/Y_i(t_k-)^2\).
The estimate \(\widehat A_{ij}\), its variance, and its standard error are dimensionless. A hazard rate has units of inverse time, but its integral over time is dimensionless. These quantities describe a cumulative intensity scale; they do not make the estimate a probability.
Assumptions and a brief justification
The usual interpretation of this variance estimate relies on a correctly specified state space and correctly constructed risk sets, with transition times and censoring times observed as intended. Individuals are assumed independent of one another. Censoring should be non-informative for the transition process, at least conditional on the observed history used to define the analysis. In counting-process terms, the transition count has an intensity governed by the risk set and transition intensity, and the risk-set process is determined just before the time at which events are counted.
These conditions allow the usual counting-process variance calculation for the Nelson–Aalen estimator. Informally, each event adds an increment \(d_{ij}/Y_i\) to the estimate. The estimated variance adds a corresponding event-count contribution scaled by the squared risk-set size. This gives the formula above. The formal martingale derivation is a separate step; here the key operational point is that each event-time contribution depends on the number of events for the transition and the risk set immediately before those events.
For a single transition estimate, the formula gives a marginal variance estimate. It does not provide covariance between estimates for different transitions. In particular, outgoing transitions from the same state share a risk set and may be statistically related. Their covariance requires additional calculations, rather than being inferred from the separate variance formulas.
Worked Example: A Single Transition With Untied Event Times
Suppose a hypothetical study follows individuals in state 0 and estimates the cumulative hazard for transition \(0\to1\) through year 4. The data at the three event times are:
| Time (years) | At risk in state 0 just before time | Events \(0\to1\) |
|---|---|---|
| 1 | 20 | 1 |
| 2 | 15 | 2 |
| 4 | 10 | 1 |
Assume the state assignments and event times are recorded correctly, individuals are independent, and censoring does not carry additional information about the transition after conditioning on the observed history. Then the standard variance calculation applies. The cumulative transition-hazard estimate is $$ \widehat A_{01}(4)=\frac{1}{20}+\frac{2}{15}+\frac{1}{10} =0.05000+0.13333+0.10000=0.28333. $$ The estimated variance is $$ \widehat{\operatorname{Var}}\{\widehat A_{01}(4)\} =\frac{1}{20^2}+\frac{2}{15^2}+\frac{1}{10^2} =0.002500+0.008889+0.01000 =0.02139. $$ Therefore, the estimated standard error is $$ SE\{\widehat A_{01}(4)\}=\sqrt{0.0213889}=0.14625. $$
The estimate is 0.28333 on the cumulative-hazard scale, with an estimated standard error of 0.14625. That standard error quantifies sampling uncertainty under the stated assumptions; it is not the standard deviation of individual event times or a probability of transition.
Competing outgoing transitions and tied events
If several transitions leave state \(i\), the risk set \(Y_i(t-)\) includes everyone still in state \(i\) just before time \(t\). The numerator for estimating \(A_{ij}\), however, counts only events of type \(i\to j\). An event from \(i\) to another destination does not add an increment to \(\widehat A_{ij}\), but the individual making that transition leaves state \(i\) and is therefore no longer in later risk sets for transitions out of \(i\).
Worked Example: Competing Exits From the Origin State
In a hypothetical multi-state follow-up, state 0 has two possible outgoing transitions: \(0\to1\) and \(0\to2\). The target is the cumulative hazard for \(0\to1\). The event counts and risk sets are:
| Time (years) | At risk in state 0 just before time | Events \(0\to1\) | Events \(0\to2\) |
|---|---|---|---|
| 1 | 30 | 2 | 0 |
| 2 | 23 | 1 | 5 |
| 3 | 12 | 2 | 0 |
Suppose the risk sets properly remove individuals after either outgoing transition, and the independence and non-informative-censoring conditions described above are reasonable. For the \(0\to1\) estimate, count only that transition in the numerator: $$ \widehat A_{01}(3)=\frac{2}{30}+\frac{1}{23}+\frac{2}{12} =0.06667+0.04348+0.16667 =0.27681. $$ The variance estimate is $$ \widehat{\operatorname{Var}}\{\widehat A_{01}(3)\} =\frac{2}{30^2}+\frac{1}{23^2}+\frac{2}{12^2} =0.002222+0.001890+0.013889 =0.01800. $$ Thus, $$ SE\{\widehat A_{01}(3)\}=\sqrt{0.018001}=0.13417. $$
The five transitions \(0\to2\) at year 2 do not appear in the numerator or variance sum for \(0\to1\). They do help explain why the risk set falls from 30 to 23 and then to 12. Using a denominator that excluded only \(0\to1\) events would incorrectly leave people at risk after they had left state 0.
For tied events, \(d_{ij}(t_k)\) is the number of transitions of the target type at that event time. The risk set is measured immediately before the tied events; it is not reduced one event at a time within the tie. This convention makes the calculation depend on the recorded event-time grouping. If the event times have been rounded or grouped, that grouping may affect the estimate and its variance.
Worked Example: A Tied Event Time and a Log-Scale Interval
Consider a second hypothetical analysis of transition \(A\to B\). Through year 3, there are two event times: at year 1, \(Y_A=25\) and \(d_{AB}=1\); at year 3, \(Y_A=10\) and \(d_{AB}=2\). The two transitions at year 3 are tied. Assume the risk sets and counts are correct and the working independence and censoring assumptions are appropriate.
The estimate and its variance are $$ \widehat A_{AB}(3)=\frac{1}{25}+\frac{2}{10}=0.2400, $$ $$ \widehat{\operatorname{Var}}\{\widehat A_{AB}(3)\} =\frac{1}{25^2}+\frac{2}{10^2} =0.001600+0.02000=0.02160. $$ Consequently, \(SE=\sqrt{0.02160}=0.14697\). The tied events contribute \(2/10^2=0.02000\), using the risk set of 10 for both events.
A normal interval formed directly as estimate plus or minus 1.96 standard errors has a lower endpoint below zero, so it is unsuitable if the interval is intended to respect the nonnegative scale of a cumulative hazard. One approximate alternative uses the log scale and the delta method. For a positive estimate, the approximate standard error of its logarithm is \(SE/\widehat A\). Here, $$ SE\{\log(\widehat A_{AB})\}\approx\frac{0.14697}{0.2400}=0.6124. $$ The approximate 95% log-scale limits are \(\log(0.2400)\pm1.96(0.6124)\), or approximately \((-2.627,-0.227)\). Exponentiating gives approximately \((0.072,\ 0.797)\). This is an approximate interval for the cumulative transition hazard, not a 95% probability statement about the fixed parameter. Its coverage depends on the large-sample approximation being adequate.
Reading the variance and checking the calculation
A large estimated variance can result from a small risk set, many events relative to that risk set, or both. For a fixed event count \(d\), the contribution \(d/Y^2\) decreases as \(Y\) increases. This explains why an event late in follow-up can contribute substantially to uncertainty when few individuals remain at risk. The estimate itself and its variance should always be reported on their respective scales, with the standard error obtained by taking the square root of the variance.
- Check the transition label. Confirm that each numerator count is for \(i\to j\), not for all events leaving \(i\).
- Check the time ordering. Each denominator must be the risk set immediately before that event time.
- Check tied counts. A tied count \(d\) contributes \(d/Y\) to the estimate and \(d/Y^2\) to the variance.
- Check units and arithmetic. The estimate, variance, and standard error are dimensionless.
- Inspect late sparse times. A very small risk set can make the estimate sensitive to one or two events and make normal approximations poor.
If there are no observed \(i\to j\) events through time \(t\), both the Nelson–Aalen estimate and this plug-in variance estimate are zero. That does not establish that the underlying transition intensity is zero. It means only that this data-based estimate has no event contributions; the formula does not express uncertainty about unobserved possibilities when the transition is absent from the observed sample.
How this estimate compares with alternatives
The formula here is the standard plug-in variance estimate paired with the Nelson–Aalen cumulative transition-hazard estimator. It is direct to compute from event counts and risk sets, and it is useful for standard errors and approximate intervals when the data support the large-sample approximation. It is not interchangeable with every variance formula used for other estimators or data structures. For example, a model-based variance may follow from a specified parametric transition-intensity model, while a resampling variance is based on repeated resampling of individuals. Those approaches involve additional modeling or resampling choices; they do not change the definition of the Nelson–Aalen estimate itself.
This calculation is marginal for one transition. It does not determine uncertainty in a state-occupation probability or transition probability, nor does it give the covariance between two transition-hazard estimates. Those questions require additional methods. Also, if observations are clustered or individuals have dependent histories across clusters, treating individuals as independent can understate or otherwise misrepresent uncertainty. Informative censoring, misspecified state definitions, and very sparse risk sets can likewise make the standard estimate unreliable.
Common mistake
Check Your Understanding
Work through each question, then compare with the solutions below.
- At one event time, \(Y_i=40\) and \(d_{ij}=3\). Find the estimate increment and variance contribution, and state whether they are dimensionless.
- At two times, the risk sets are 20 and 10, and the target event counts are 1 and 2. Calculate the cumulative estimate, variance estimate, and standard error.
- At a tied event time, there are two \(i\to j\) events and four \(i\to k\) events. Which count enters the variance formula for \(i\to j\), and which risk set is used?
- Why does a transition \(i\to k\) affect a later risk set for \(i\to j\), even though it is not counted in the \(i\to j\) variance numerator?
Solutions. (1) The estimate increment is \(3/40=0.0750\), and the variance contribution is \(3/40^2=0.001875\); both are dimensionless. (2) The estimate is \(1/20+2/10=0.2500\); the variance is \(1/20^2+2/10^2=0.02250\); the standard error is \(\sqrt{0.02250}=0.1500\). These quantities are dimensionless. (3) Use \(d_{ij}=2\), not the four \(i\to k\) events, and use the state-\(i\) risk set immediately before the tied event time. (4) A person making \(i\to k\) leaves state \(i\), so is no longer at risk for a later \(i\to j\) transition, even though that event is not an \(i\to j\) event.