Why Late Estimates Need Special Care
A cumulative incidence estimate can be mathematically well defined at a late follow-up time and still be supported by very little information. By then, many participants may have experienced an event or been censored, leaving only a small number at risk. One event among ten people at risk represents a much larger fraction of that risk set than one event among one hundred. Consequently, late steps in the estimated cumulative incidence function can be conspicuous, and the estimate may be sensitive to the outcomes of only a few people.
This is not a reason to discard all late estimates. It is a reason to interpret them alongside the risk set, event counts, censoring pattern, and follow-up horizon. In particular, a numerical estimate does not by itself show that the tail is precise, that the remaining participants represent everyone of interest, or that the estimated curve can be extended beyond observed follow-up.
The estimate at a chosen horizon is the sum of the type-\(k\) increments through that horizon. The factor \(d_{kj}/Y_j\) is the observed event-type-specific fraction of the risk set having that event at \(t_j\); the survival factor weights that fraction by the estimated probability of having remained event-free until then. These are the same quantities used in the Aalen–Johansen estimator. Here the emphasis is on what their sizes and data support imply near the end of follow-up.
What a Small Risk Set Does
A small \(Y_j\) makes each observed event a larger fraction of the risk set. For fixed \(\widehat S(t_j-)\), one type-\(k\) event contributes \(\widehat S(t_j-)/Y_j\). If the risk set has ten people rather than one hundred, that one event's increment is ten times as large. The estimate can therefore change noticeably in response to just one or two late events.
This is a statement about the scale of the observed step, not a complete calculation of uncertainty. A large step does not establish that the population risk is high, and a small step does not establish that the estimate is precise. Formal standard errors and confidence intervals require additional calculations covered in the material on variance and interval estimation. Even before those are calculated, the number at risk and the late event counts are essential context for interpreting the point estimate.
Worked Example: The Same Late Event Count With Different Risk Sets
Consider two hypothetical analyses at a late event time. In both, the estimated event-free probability just before the time is \(0.60\), and two of the people at risk experience the event type of interest. There are no other events at that time. In the first analysis, \(Y_j=10\); in the second, \(Y_j=80\). Suppose the estimated cumulative incidence before this time is \(0.25\) in both.
For the first analysis, the increment is \(\widehat S(t_j-)\,d_{kj}/Y_j=0.60(2/10)=0.12\). The cumulative incidence after the step is \(0.25+0.12=0.37\).
For the second analysis, the increment is \(0.60(2/80)=0.015\). The cumulative incidence after the step is \(0.25+0.015=0.265\).
Thus, the same count of two events produces different step sizes because it is divided by different risk-set sizes. These hypothetical calculations do not imply that either estimate is biased; they show why late counts must be read with their denominators. The estimate based on ten people is especially exposed to the outcomes of a few individuals, so it warrants a clear warning about sparse late follow-up.
Censoring: Information Loss and Assumptions
Right censoring occurs when a participant's event status is known only up to a time, after which follow-up ends without an observed event. A censored participant contributes to the risk set up to the censoring time and is no longer counted at risk afterward. Censoring can therefore shrink later risk sets even when few events have occurred. A high early event-free survival estimate does not prevent the tail from being sparse if many participants are later censored.
The usual Aalen–Johansen estimate with right censoring relies on independent censoring for the target analysis: conditional on the information used to define the analysis, censoring must not carry additional information about the participants' future event times and event types. Informally, among people still event-free and under observation at a given time, those who remain observed should adequately represent those who are censored with respect to subsequent event experience. If this condition holds, censoring reduces the available information and often leaves a small tail risk set, but does not by itself invalidate the estimator.
If censoring is informative—for example, if participants with a worse unmeasured prognosis are more likely to leave follow-up—those remaining under observation may not represent the target population. A small risk set makes this concern especially important, but the small count alone cannot diagnose whether censoring is informative. The reason for loss to follow-up, its timing, and observed differences between retained and censored participants matter. Adjusting for recorded predictors may help only if the analysis appropriately accounts for them; it does not guarantee that unmeasured informative censoring has been removed.
Worked Example: Calculating a Tail Estimate After Censoring
In a hypothetical cohort, 20 people are at risk just before time 1. At that time, two experience the event of interest and one experiences a competing event. Five of the remaining participants are censored before time 2. Just before time 2, 12 people are at risk; one experiences the event of interest and one experiences a competing event. Assume the independent censoring condition is plausible for this analysis.
At time 1, the event-free estimate falls by all three observed events, so \(\widehat S(1)=1-(2+1)/20=0.85\). The cumulative-incidence increment for the event of interest is \(1(2/20)=0.10\).
The five censored people are no longer at risk at time 2, but censoring does not itself count as an event in the survival-factor update. The risk set at time 2 is 12, as given. The event-free estimate immediately before time 2 remains \(0.85\), and the type-of-interest increment is \(0.85(1/12)=0.0708333\). The cumulative incidence after time 2 is therefore \(0.10+0.0708333=0.1708333\), or approximately \(0.171\).
The arithmetic gives a point estimate, not a guarantee that the tail is precise. The five censorings have reduced the later risk set, and only 12 people remain at risk at time 2. Interpreting the result requires reporting that limited support and considering whether independent censoring is credible. If the people censored before time 2 have different future risks from those still observed, the calculation may not estimate the intended population quantity reliably.
Check the Tail Against the Observed Data
Before describing a late estimate, inspect the event-time risk sets and the censoring history. A useful report pairs the estimated value at the selected horizon with the number still at risk near that horizon and the counts of relevant and competing events. If the last displayed portion of a curve is based on very few people, say so plainly. Avoid presenting a smooth-looking line or a precise-looking decimal as a substitute for that information.
A plateau also requires care. The estimated cumulative incidence is a step function: it changes at observed event times and stays constant between them. If no further event is observed, the estimate stays flat. After the last observed event, that flat segment is a property of the estimator and the available records; it does not demonstrate that no one in the target population would experience the event later. Beyond the end of observed follow-up, the data do not directly establish the cumulative incidence.
Identify the time at which the estimate will be interpreted and the last observed event and follow-up times relevant to that estimate.
Check the number at risk and the counts of events of interest, competing events, and censorings near the horizon. A small risk set signals that few observations support the tail.
Consider whether independent censoring is plausible for the target analysis and whether loss to follow-up may depend on future event experience.
Report the estimate at the supported horizon, with its risk-set context. Do not treat a flat tail as evidence about unobserved future time.
Choosing and Comparing Interpretations
The estimate at a late horizon can be useful if that horizon matters clinically and the data still provide adequate follow-up support. When support is weak, a shorter, prespecified horizon may yield a more defensible summary, provided it addresses the scientific question. Choosing a horizon after inspecting the results solely to obtain a preferred estimate is not a neutral remedy. A transparent report can give the estimate at the intended horizon while explicitly describing the sparse tail.
A crude proportion of observed events is not a replacement for the cumulative incidence estimate when follow-up is censored: it generally ignores both differing observation times and the competing-event structure. Likewise, one minus a Kaplan–Meier estimate that treats competing events as censoring does not estimate the cause-specific cumulative incidence. The Aalen–Johansen estimator accounts for the event-free probability and observed event-type increments, but it cannot create information where the risk set has become small or correct informative censoring merely by its formula.
A different horizon changes the time-specific quantity being reported, rather than repairing a weak tail estimate at the original horizon. A different model or adjustment may be appropriate for a specific censoring or prediction problem, but its assumptions must be stated and justified. For a descriptive cumulative incidence analysis, a practical approach is to show the curve and risk counts, identify the chosen horizon, and state the limitations imposed by sparse follow-up.
Common Mistakes and Limitations
- Reading a large step as strong evidence of high population risk. The step can be large because \(Y_j\) is small. Report the risk-set size and event count; do not interpret the step without them.
- Treating censoring as a competing event. Censoring removes a participant from later risk sets but is not an observed event type. The event-free survival update uses event counts, not censoring counts.
- Calling every tail estimate biased because few people remain. Sparse risk sets signal limited information and sensitivity, not automatic bias. Informative censoring, in contrast, can undermine the estimator's validity.
- Interpreting a flat tail as zero future risk. No observed event after a time means no additional estimated step there; it does not establish what would happen after observed follow-up ends.
- Reporting many decimal places without context. Numerical precision is not evidential precision. Round sensibly and provide the horizon, risk-set context, and limitations.
The estimator also has limits beyond sparse risk sets. It describes a population cumulative incidence under its censoring assumptions; the observed data alone cannot verify those assumptions. If the target population, follow-up process, or event classification changes, the interpretation may change as well. Tail interpretation should therefore be framed as an assessment of what the observed follow-up supports, not as a guarantee about future events.
Worked Example: Interpreting an Estimate With a Very Small Final Risk Set
In a hypothetical analysis, the estimated event-free probability just before time 3 is \(0.715625\), and the cumulative incidence for event type 1 through the preceding event time is \(0.190\). Ten people remain at risk just before time 3, and two experience event type 1 at that time. There are no other events at time 3.
The increment is \(\widehat S(3-)(d_{1j}/Y_j)=0.715625(2/10)=0.143125\). Thus the updated cumulative incidence is \(0.190+0.143125=0.333125\), which rounds to \(0.333\).
The calculation is internally consistent: the two observed events are 20% of the ten-person risk set, and the estimated event-free probability weights that fraction. But the step is based on two events among only ten people at risk. The estimate should therefore be reported as a late point estimate with sparse support, not as a highly precise risk assessment. It also should not be extended beyond the observed follow-up merely because the curve remains flat afterward.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(\widehat S(t-)=0.50\), \(d_{kj}=1\), and \(Y_j=5\), what is the increment? What would it be if \(Y_j=50\), with the other quantities unchanged?
- Does censoring itself count as an event in the event-free survival update? What does it change for later event times?
- A cumulative incidence curve is flat after the last observed event. What can and cannot be concluded from that flat segment?
- Why does a small late risk set call for caution even when the independent censoring condition is plausible?
- What should accompany a reported cumulative incidence estimate at a late horizon?
Solutions. (1) The first increment is \(0.50(1/5)=0.10\); with 50 at risk it is \(0.50(1/50)=0.010\). (2) No. Censoring removes a participant from later risk sets but is not an event in the event-free survival update. (3) The estimate has no observed event-driven increase during that segment; the flatness does not establish zero risk after observed follow-up. (4) Independent censoring supports validity under its assumptions, but a small risk set means few observations contribute and a few outcomes can strongly affect the estimate. (5) Report the horizon and relevant risk-set size and event counts, and describe important censoring and sparse-tail limitations.