Why Uncertainty Must Be Considered Over Time
A mean cumulative function (MCF) describes how the average number of recurrent events accumulates as follow-up proceeds. An estimated curve is therefore more than one estimate: it is a sequence of estimates, one for each time \(t\). Uncertainty at a prespecified time and uncertainty about the whole curve are different questions. A pointwise confidence interval addresses the first; a simultaneous uncertainty band addresses the second.
This distinction matters in practice. A pointwise interval can be appropriate when the scientific question concerns the mean cumulative number of events at one fixed horizon. It does not, by itself, provide a stated confidence level for the entire curve. If the curve will be examined across many times, or if the claim concerns its behavior throughout follow-up, simultaneous coverage is the relevant target.
Estimator and Assumptions Behind Its Uncertainty
For clarity, consider independent subjects followed from a common time origin. Let \(Y_i(u)\) equal 1 when subject \(i\) is under observation and at risk immediately before time \(u\), and 0 otherwise. Write \(Y(u)=\sum_{i=1}^n Y_i(u)\), and let \(dN(u)=\sum_{i=1}^n dN_i(u)\) be the total number of events at \(u\). The Nelson–Aalen-type MCF estimator is
At each event time, this estimator adds the number of events at that time divided by the number of subjects still under observation just before it. A subject who experiences an event remains at risk for later recurrences while still observed. This is not a first-event analysis.
The usual large-sample uncertainty calculations rely on independent subjects, a suitable independent or noninformative censoring mechanism, and enough subjects remaining at risk over the time range being analyzed. Informally, after accounting for the observation scheme, censoring must not selectively remove subjects according to their unobserved future event experience. The target is the population MCF under that observation assumption. Repeated events within one subject need not be independent: the variance calculation must retain each subject's event history as a unit.
One useful robust variance estimate is built from subject-level contributions. Define
The contributions represent each subject's event increments relative to the increments expected from their observed time at risk, scaled by the total risk set. Because the estimator satisfies its pooled estimating equation, their sum is zero. Their squared sum estimates sampling variability while retaining within-subject dependence. Set \(\widehat{SE}\{\hat M(t)\}=\sqrt{\sum_i\hat q_i(t)^2}\). The formula is a large-sample approximation; it cannot repair informative censoring, dependence between subjects, or a risk set that has nearly disappeared.
Pointwise Confidence Intervals
At a fixed time chosen for the analysis, an asymptotic normal interval with nominal confidence level \(1-\alpha\) is
where \(z_{1-\alpha/2}\) is the corresponding standard normal quantile. The interval is an approximate repeated-sampling procedure: under its assumptions, intervals constructed this way would contain the fixed \(M(t)\) in about \(100(1-\alpha)\%\) of repeated samples. The endpoints are in units of mean event count per subject. A normal interval can have a negative lower endpoint even though an MCF is nonnegative; this flags the limits of the approximation, especially when the estimate is near zero.
Worked Example: A Pointwise Interval With Complete Follow-Up
Suppose, hypothetically, six independent subjects are all followed to a fixed horizon \(T\), and their event counts by \(T\) are \(0,1,1,2,2,3\). With complete follow-up, the MCF estimate at \(T\) is the sample mean count.
The estimate is \(\hat M(T)=(0+1+1+2+2+3)/6=9/6=1.50\) events per subject. The squared deviations from 1.50 sum to \(5.50\), so the sample variance is \(s^2=5.50/(6-1)=1.10\) event-count-squared units. Thus \(SE=s/\sqrt{6}=\sqrt{1.10/6}=0.4282\) events per subject.
Using \(z_{0.975}=1.96\), the approximate 95% pointwise interval is
At this horizon, the data estimate an average of 1.50 events per subject, and the approximate pointwise interval extends from 0.661 to 2.34 events per subject. Because there are only six subjects, the normal approximation may be poor; the calculation illustrates the method, not a guarantee of reliable small-sample coverage.
Worked Example: A Robust Pointwise Standard Error
In a hypothetical censored cohort, suppose the risk-set estimator at 12 months is \(\hat M(12)=2.00\) events per subject. The subject-level contributions calculated from the observed event and risk histories are \(0.20,-0.10,0.05,-0.25,0.15,-0.05\) events per subject. Their sum is zero, as expected from the pooled estimating equation.
The estimated variance is the sum of their squares:
Therefore \(\widehat{SE}=\sqrt{0.1400}=0.3742\) events per subject. The approximate 95% pointwise interval is \(2.00\pm1.96(0.3742)=2.00\pm0.7334\), or \([1.267,2.733]\) events per subject. This interval refers only to the mean cumulative number at 12 months; it does not assert 95% coverage for the curve at all times.
Simultaneous Bands Across Follow-Up
A simultaneous band widens the critical value to account for examining the curve over a range of times. One approach uses subject-level multiplier resampling of the estimated contributions. For replicate \(b\), draw independent mean-zero, variance-one multipliers \(\xi_i^{(b)}\), one for each subject, and form the perturbed curve
On a prespecified time grid where the standard error is positive, calculate the largest absolute standardized deviation in that replicate:
The \(1-\alpha\) quantile \(c_{1-\alpha}\) of the replicate values gives a simultaneous band \(\hat M(t)\pm c_{1-\alpha}\widehat{SE}\{\hat M(t)\}\). In implementation, the supremum is taken over the chosen grid; the grid should represent the time range adequately. At time zero the MCF is known to be zero, so a standardized statistic with zero standard error is not used there. An alternative is to resample whole subjects, including each subject’s complete event and censoring history, and use the maximum standardized deviation from each bootstrap curve. Resampling event records individually would break the within-subject structure and is not appropriate.
Worked Example: Reading a Simultaneous Band
Suppose a hypothetical analysis uses subject-level multipliers and 2,000 replicates over a prespecified follow-up grid. The 95th percentile of the replicate maximum standardized deviation is \(c_{0.95}=2.45\). At three times, the estimates and standard errors are as follows.
| Time (months) | \(\hat M(t)\) (events per subject) | Standard error (events per subject) |
|---|---|---|
| 6 | 1.20 | 0.18 |
| 12 | 2.40 | 0.35 |
| 18 | 3.10 | 0.52 |
At 12 months, for example, the band endpoints are \(2.40\pm2.45(0.35)=2.40\pm0.8575\), giving \([1.543,3.258]\) events per subject. At 6 months they are \(1.20\pm2.45(0.18)=[0.759,1.641]\); at 18 months they are \(3.10\pm2.45(0.52)=[1.826,4.374]\).
The simultaneous interpretation is that the procedure is designed to cover the true MCF at all times on the analyzed grid together in 95% of repeated samples, subject to the model and resampling assumptions. It is not a 95% probability statement about the fixed curve after these data have been observed. The band is wider than a 1.96-based pointwise interval because its critical value accounts for the maximum deviation across times.
Justification, Diagnostics, and Alternatives
The distinction between pointwise and simultaneous coverage follows from the event being controlled. A pointwise interval controls the standardized error at one fixed \(t\). For a band, the relevant error is the largest standardized error anywhere in the range. If \(Z(t)\) denotes the standardized estimation error, then
The quantile of this supremum is generally larger than the pointwise normal quantile. This is why substituting 1.96 at every time does not produce a 95% simultaneous band. The multiplier procedure approximates the distribution of the standardized estimation process while preserving the shared subject-level contributions that make estimates at different times dependent.
Check the risk-set size and the estimated standard error across follow-up before interpreting a band. A sharp increase in uncertainty near the end of follow-up is expected when few subjects remain observed. A band that becomes extremely wide there is a warning that the data provide little precision, not a reason to ignore the late times. Also inspect whether the event and censoring histories, time origin, and analysis horizon support the assumed target population and independent censoring. Severe imbalance in follow-up or informative dropout can undermine the estimator and its uncertainty calculation.
The main alternatives answer different practical needs. A normal pointwise interval is straightforward for a prespecified time and is often adequate when inference is limited to that time. A simultaneous multiplier or subject-bootstrap band is appropriate for claims about a whole curve, but depends on a credible resampling approximation and adequate sample size. A pointwise bootstrap interval at many times still does not become simultaneous merely because it is bootstrapped. The earlier material on survival analysis provides related large-sample interval ideas, but the resampling unit here must be the subject because each subject contributes a recurrent-event history.
These bands summarize uncertainty about the population mean cumulative number of events, not the range of event counts individual subjects may experience. They also do not adjust automatically for choosing the time range or analysis after looking at the results. If the endpoint is selected after inspecting many candidate times, nominal coverage can be misleading. In small samples, near zero counts, or sparse late risk sets, asymptotic normal and bootstrap procedures may have inaccurate coverage; report those limitations rather than treating the displayed endpoints as exact.
Common Mistake
- Wrong reasoning: “I drew a 95% pointwise interval at every observed time, so there is a 95% chance the entire MCF is inside all of them.”
- Correction: Each pointwise interval targets coverage at its own fixed time. The probability that all intervals cover simultaneously is generally lower than 95%. Use a critical value calibrated to the maximum deviation over the time range when the claim concerns the whole curve.
- Wrong reasoning: “The repeated events are separate observations, so I can resample event records.”
- Correction: Resample subjects, retaining each subject’s full event and censoring history. Events within a subject may be dependent, and treating them as independent can understate uncertainty.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A 95% pointwise interval is reported at each of 20 follow-up times. Does this imply 95% coverage of the MCF at all 20 times together? Why?
- At a fixed time, an estimate is 1.80 events per subject and its standard error is 0.25. Calculate the approximate 95% pointwise interval using 1.96.
- Why should a subject-level bootstrap resample entire event histories rather than individual events?
- A simultaneous band becomes much wider near the end of follow-up. Give one data-based explanation and one interpretation of that widening.
- What is the correct interpretation of a 95% simultaneous band over \([0,\tau]\)?
Solutions. (1) No. Pointwise coverage applies separately at each fixed time; simultaneous coverage requires calibrating the maximum deviation over all 20 times. (2) \(1.80\pm1.96(0.25)=1.80\pm0.49\), giving \([1.31,2.29]\) events per subject. (3) Events from one subject can be dependent, so resampling them separately would destroy within-subject dependence and can misstate uncertainty. (4) Fewer subjects may remain under observation late in follow-up; the wider band indicates less precision there. (5) Under the assumptions, the band-producing procedure covers the entire true MCF over the stated range in about 95% of repeated samples.