Comparing Accumulated Recurrent-Event Burden
A recurrent-event analysis can compare groups through an event rate or a model coefficient, but those summaries do not show the whole time course of accumulated events. The mean cumulative function (MCF) gives a direct group-level description: how many events, on average, have accumulated by each time point? Comparing group MCF curves makes it possible to see when a difference emerges, whether it persists, and whether the curves cross.
The curve comparison is descriptive unless the study design and assumptions support a causal interpretation. In a randomized trial, for example, a contrast can be interpreted as an effect of assignment under appropriate follow-up and analysis conditions. In an observational comparison, differences may also reflect group composition or other confounding.
An MCF is nondecreasing: subjects can accumulate events over time, but their cumulative event counts do not decrease. A difference between two MCFs need not be nondecreasing. It can grow, shrink, or change sign as the groups accumulate events at different times. The vertical gap at time \(t\) is in events per subject by time \(t\); it is not an instantaneous rate difference.
The group curves must refer to a comparable event definition, time origin, and follow-up horizon. The earlier material on estimating the mean cumulative function provides the estimator; here the emphasis is on comparing its group-specific curves and contrasts.
Assumptions Behind a Group Comparison
Before interpreting a comparison, specify the population and follow-up process to which it applies. The usual curve-based comparison relies on the following conditions, or on a suitable adjustment that addresses their failure:
- Comparable measurement: recurrent events are defined and ascertained in the same way across groups, and time is measured from a common or otherwise clearly specified origin.
- Independent observational units: subjects are independent within each group for ordinary subject-level variance calculations. If subjects are clustered, inference must account for that dependence.
- Noninformative censoring for the estimator: censoring must not selectively remove subjects with different future event histories in a way left unaccounted for by the estimation procedure. Administrative censoring at a common study end is often plausible; loss to follow-up can be more problematic.
- Comparable target populations: group membership and eligibility define the populations being summarized. A causal interpretation additionally requires an appropriate design or adequate control of confounding, along with consistency and positivity.
- Adequate follow-up support: estimates near the end of follow-up can be unstable when few subjects remain observed. Comparisons should not extend beyond the time range with useful information in both groups.
The independent-group variance formula below also assumes that the two group estimates are independent. That is natural for disjoint randomized arms with independent subjects, but it does not apply unchanged to matched or paired groups. The data structure and censoring should be checked before choosing the inference method.
Reading Curves and Contrasts
Suppose the group of interest is labelled 1 and the reference group 0. If \(\hat\Delta(t)<0\), the estimated group 1 mean cumulative event count is lower by \(|\hat\Delta(t)|\) events per subject by time \(t\). This does not say that every subject in group 1 had fewer events, nor that group 1 had a lower event rate at every instant.
At a prespecified time \(t\), let \(\widehat V_g(t)\) estimate the variance of \(\hat M_g(t)\). For independent groups, the variance of the estimated difference is the sum of the group variances. A large-sample pointwise confidence interval and test can then be formed as follows.
The interval is a pointwise 95% confidence interval: under the model and regularity conditions, the procedure has approximately 95% coverage for the contrast at that particular time across repeated samples. The corresponding two-sided test assesses \(H_0:\Delta(t)=0\). A small p-value is evidence against equality at that time under the test assumptions; it is not the probability that the null hypothesis is true.
When a conclusion concerns the entire trajectory, a collection of pointwise intervals is not a 95% simultaneous confidence band. The earlier material on uncertainty bands for the MCF addresses uncertainty over a range of times. In addition, bands for two separate group curves should not be treated as a direct test of their difference: the relevant uncertainty is the uncertainty of the contrast itself.
Worked Example: Describing a Pair of Group Curves
Consider a hypothetical randomized study with recurrent events counted over 12 months. The event definition, follow-up schedule, and time origin are the same in both groups. Assume independent subjects, and assume censoring is administrative at month 12. The estimated MCF values are:
| Time (months) | Control: \(\hat M_0(t)\) (events per subject) | Treatment: \(\hat M_1(t)\) (events per subject) | Difference: treatment minus control |
|---|---|---|---|
| 0 | 0.00 | 0.00 | 0.00 |
| 3 | 0.28 | 0.20 | -0.08 |
| 6 | 0.61 | 0.42 | -0.19 |
| 9 | 0.91 | 0.61 | -0.30 |
| 12 | 1.20 | 0.78 | -0.42 |
Both estimated curves are nondecreasing, as cumulative event counts should be. The treatment-minus-control gap becomes more negative over the displayed follow-up. At month 12, the estimated mean cumulative count is 0.42 events per subject lower in the treatment group. This summarizes accumulated burden through 12 months; it does not establish that treatment lowered the instantaneous rate throughout the study.
The assumptions support a straightforward comparison here: the groups are randomized and disjoint, event measurement and time origin match, and censoring occurs administratively at a common horizon. The table alone does not provide uncertainty, however. A visual gap is not enough to determine whether a difference is statistically distinguishable from zero.
Worked Example: A Pointwise Contrast at 12 Months
For the hypothetical study above, suppose the estimated standard errors at month 12 are 0.18 events per subject for control and 0.14 events per subject for treatment. Use treatment minus control, so \(\hat\Delta(12)=0.78-1.20=-0.42\).
The independent-group standard error is
The pointwise 95% confidence interval is
Thus, the estimated treatment-group mean cumulative count is 0.42 events per subject lower at 12 months, but the interval includes zero. For the two-sided test of \(H_0:\Delta(12)=0\), the statistic is
The two-sided normal-approximation p-value is approximately 0.0655. At the 0.05 significance level, this test does not reject equality of the 12-month MCFs. It would be incorrect to conclude that the groups are equal: the interval also allows a modest difference in either direction. This inference concerns one time point, not the complete curves.
Why Curves Can Cross
A difference in accumulated counts is affected by the timing of events. One group can accumulate more events early, while the other catches up later. A crossing therefore carries information that a single summary can hide. It is not, by itself, proof of a data or model error.
Worked Example: Interpreting Crossing MCF Curves
In a second hypothetical study, two groups have these estimated MCF values, with the same event definition and time origin:
| Time (months) | Group A (events per subject) | Group B (events per subject) | A minus B (events per subject) |
|---|---|---|---|
| 3 | 0.50 | 0.30 | 0.20 |
| 6 | 0.80 | 0.95 | -0.15 |
| 12 | 1.10 | 1.15 | -0.05 |
At month 3, group A has an estimated 0.20 more events per subject accumulated than group B. By month 6, group B has accumulated 0.15 more events per subject. At month 12, the difference has narrowed to 0.05 fewer events per subject in group A. Each MCF remains nondecreasing, while their difference changes sign.
The appropriate summary depends on the scientific question. If early accumulated burden matters, the month-3 contrast is relevant; if the target is total burden by month 12, the month-12 contrast is more directly relevant. Reporting only an overall statement such as “group A has fewer events” would conceal the time pattern. The table supplies no standard errors, so it cannot establish whether any of these differences exceed sampling uncertainty.
Diagnostics, Alternatives, and Limitations
Before interpreting group curves, inspect whether the estimates behave plausibly and whether the comparison is supported by the data. Useful checks include:
- Confirm that each estimated MCF is nondecreasing and starts at the value implied by the chosen time origin and counting convention.
- Check the number still under observation over time. A tail based on very few observed subjects can look dramatic while being highly uncertain.
- Look for unequal loss to follow-up, delayed entry, or different follow-up schedules across groups. These features may invalidate a simple comparison or require methods that account for them.
- Verify that repeated events are analyzed at the subject level. Treating each event record as an independent subject can understate uncertainty.
- Display uncertainty for the contrast or use an appropriate simultaneous band when making claims across many time points. Repeatedly testing pointwise differences inflates the chance of at least one false positive.
An MCF comparison is not interchangeable with a rate ratio from a marginal rate model or a count-model coefficient. The MCF difference is an absolute difference in expected accumulated counts by time; a rate ratio compares rates under its model. A GEE model can target mean count contrasts while accounting for within-subject dependence through its covariance structure, but its chosen mean model may impose a different time pattern. These approaches answer related but distinct questions, so the estimand should guide the choice.
Curve comparisons also have limitations. A fixed-time contrast depends on the chosen horizon, and later estimates may be imprecise. When event rates vary over time, a single relative effect can obscure the pattern; conversely, a curve contains many time-specific comparisons and should not be reduced to whichever point looks most favorable. In observational data, an unadjusted difference is generally an association, not a causal effect. If death or another terminal event ends the opportunity to experience recurrent events, the MCF describes accumulated events under the observed follow-up process; it does not isolate a recurrent-event effect from the terminal-event process.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- At 8 months, \(\hat M_1(8)=0.90\) and \(\hat M_0(8)=1.25\) events per subject. What is the group 1 minus group 0 contrast, and how should it be interpreted?
- At a fixed time, independent groups have estimated MCF standard errors of 0.10 and 0.15 events per subject. Find the standard error of their difference.
- Can two valid MCF curves cross? Explain what a crossing means and what it does not establish.
- Why is testing the difference at many times with separate pointwise 5% tests not equivalent to a single 5% test of the whole trajectory?
Solutions. (1) The contrast is \(0.90-1.25=-0.35\) events per subject. By 8 months, group 1 has an estimated mean cumulative count 0.35 events per subject lower than group 0; this is not a statement about each individual or an instantaneous rate. (2) Under independent-group estimation, \(SE=\sqrt{0.10^2+0.15^2}=\sqrt{0.0325}=0.1803\) events per subject. (3) Yes. Each cumulative mean must be nondecreasing, but their difference may change sign as one group accumulates events earlier and the other catches up. A crossing does not alone imply an error or a statistically meaningful difference. (4) Separate pointwise tests do not control the probability of at least one false positive across all tested times at 5%; inference for the whole trajectory needs an appropriate simultaneous procedure.