Why Censoring Conditions Matter
Gray’s test compares the cumulative incidence function (CIF) for a specified cause across groups. Censoring affects which events are observed and which people remain under observation. If censoring selectively removes people with different future event risks, the observed data can give a misleading picture of the CIFs and their difference. A valid test therefore depends not only on how the test statistic is calculated, but also on how observation ends.
Censoring does not mean that an event did not occur. It means that the event time and cause are not observed after a particular time. A cause-specific event that occurs before censoring is still observed, even if that person would have been censored later. This distinction is essential when reasoning about the amount of observed event information.
The Censoring Condition for Gray’s Test
The standard condition is that censoring be independent of the event process within each comparison group. In a basic formulation, if \(G\) denotes group membership, then \(C\) is independent of \((T,J)\) conditional on \(G\). This says that, among people in the same group, censoring time does not carry additional information about when an event will occur or which cause will occur.
The condition is about independence within groups; it does not require the censoring-time distributions to be identical across groups. For example, one group may have more administrative censoring because its participants entered the study later. If censoring is independent of future event time and cause within each group, and there is adequate observation over the time range being compared, different censoring rates alone do not violate the standard condition.
The relevant assumption concerns the full competing-risks event process, not just the target cause. If loss to follow-up is related to the risk of a competing event, the observed risk set can also become unrepresentative, affecting the target CIF. Competing events that are observed are not censoring: they remove a person from the possibility of experiencing the target event first, and they must be recorded with their cause.
In some settings, censoring is independent only after conditioning on measured covariates, such as disease severity or calendar time. That conditional statement does not automatically validate an ordinary, unadjusted Gray test. The analysis must use a method that appropriately accounts for the covariates and censoring mechanism. Likewise, independent censoring is not enough if event causes are misclassified or if censoring is so heavy that important parts of the CIF are unsupported by observed follow-up.
Why Independent Censoring Enters the CIF Estimate
The CIF accumulates the probability of experiencing a particular cause over time. In continuous time, it can be written as the integral of the probability of remaining event-free just before time \(s\) multiplied by the cause-specific event hazard at \(s\). Under independent censoring, the individuals still observed at each time provide information about the event hazards in the group. This is the basis for estimating the CIF from observed event and risk-set data, and for constructing Gray’s test.
A useful way to see the issue is to calculate the probability that a target event is observed by a fixed time \(\tau\). Let \(F_1\) be the distribution of target-cause event times, and let \(G(s-)=P(C\geq s)\), using the convention that an event at \(s\) is observed if censoring has not occurred before \(s\). Then
If censoring is independent of event time and cause, the conditional probability in the integrand becomes \(G(s-)\). The resulting probability depends on the censoring survival at each event time, not simply on the probability of remaining uncensored through the end of the interval. For discrete event and censoring times, the calculation must also follow the study’s tie rule.
Worked Example: Calculating the Observed Event Probability
Suppose a hypothetical group has a target-cause CIF of 0.12 by \(\tau=1\). Half of these target events occur at time 0.25 and half at time 0.75, so each event-time mass is 0.06. Censoring is independent of event time and cause. Everyone is observed through 0.25, while 20% are censored at time 0.5; the others remain observed through 1. Assume there are no ties between censoring and events.
At time 0.25, the probability of remaining observed is 1. At time 0.75, it is 0.80. Thus the probability of an observed target event by time 1 is
So 10.8% of the original group are expected to have an observed target event by time 1. The underlying target-cause CIF is 12.0%. Multiplying 0.12 by 0.80 would incorrectly give 9.60%: it would treat everyone censored at 0.5 as if they had been unable to experience an observed event earlier at 0.25. The calculation must account for when events occur relative to censoring.
What Informative Censoring Can Do
Censoring is informative when, within a group, censoring time is associated with future event time or cause. For example, people with worsening health may be more likely to leave follow-up, while also being more likely to experience the event of interest soon. The censored individuals then differ systematically from those who remain under observation. Treating their censoring as if it were unrelated to future events can distort estimated CIFs.
The direction of distortion is not fixed. If people at especially high risk of the target event are censored early, the observed data may understate that cause’s incidence. If people at unusually low risk are censored early, the observed data may overstate it. If the relationship between censoring and event risk differs across groups, the apparent group contrast can be exaggerated, attenuated, or even reversed.
Worked Example: High-Risk Participants Lost Before an Event
Consider a hypothetical group of 100 people followed to time 1. In the underlying population, 20 would experience the target event at time 0.5, 20 would experience a competing event at time 0.75, and 60 would have no event by time 1. Suppose 10 of the 20 people destined for the target event are instead censored at time 0.25 because their worsening symptoms make them more likely to leave the study. The other 10 target events are observed at 0.5. Assume the 20 competing events at 0.75 are observed.
The true target-cause CIF by time 1 is \(20/100=0.20\). An Aalen–Johansen-style calculation that treats censoring as independent uses the observed risk sets. At 0.25, 10 people are censored, leaving 90 at risk. At 0.5, 10 of those 90 have the target event. The estimated target incidence contributed at that time is
At 0.75, 20 of the 80 remaining people experience the competing event. This reduces the estimated event-free probability but adds no target-cause incidence. The estimated target CIF therefore remains about 0.1111 through time 1, rather than the true 0.20. The discrepancy arises because the early censoring selectively removed half of the people who would have had the target event. A Gray test based on the same unrepresentative follow-up can likewise give a distorted comparison.
Different Censoring Rates Do Not Necessarily Mean Informative Censoring
It is important not to confuse differential censoring with informative censoring. The former means that censoring patterns differ between groups; the latter means that censoring is associated with future event experience within a group. Differential censoring can make raw observed event proportions incomparable, even when the independence condition holds and the CIF estimator can account for censoring.
Worked Example: Independent but Unequal Censoring
Suppose two hypothetical groups each have a target-cause CIF of 0.20 by time 1: 0.10 of the group experiences the target event at 0.25, and another 0.10 at 0.75. Assume censoring is independent of event time and cause in each group. In Group A, 20% are censored at 0.6 and the rest remain observed through 1. In Group B, 10% are censored at 0.2, another 40% at 0.6, and the rest remain observed through 1.
In Group A, all target events at 0.25 are observed, and 80% of those at 0.75 are observed. The expected observed target-event fraction is
In Group B, 90% remain observed at 0.25 and 50% at 0.75. Its expected observed target-event fraction is
The raw fractions differ even though both groups have the same underlying target CIF of 0.20. The difference is caused by unequal observation, not by a difference in event incidence. Because censoring was stipulated to be independent within each group, standard CIF estimation and Gray-test procedures can account for these patterns in principle; the censoring distributions need not be equal. In finite samples, estimates can still be imprecise, especially where few people remain under observation.
Assessing the Assumption and Choosing a Response
Independent censoring cannot generally be proven from the observed data alone: after censoring, the person’s event time and cause are unknown. Assessment therefore combines study design, knowledge of why observation ended, and available data. Useful questions include whether censoring was administrative or due to loss to follow-up, whether reasons for loss are associated with prognosis, and whether relevant predictors of both censoring and events were measured.
- Review the censoring mechanism. Prespecified study-end censoring is often more plausible as independent than withdrawal prompted by deteriorating health, but it is not automatically independent. For example, staggered entry or calendar time can matter if prognosis also changes over calendar time.
- Examine measured predictors. If variables such as severity or treatment adherence predict both censoring and event risk, an unadjusted analysis may not satisfy the needed within-group independence condition. Conditional modeling or weighting may be appropriate only if the required variables are measured and the method is justified.
- Check support over time. If almost nobody in a group remains observable late in follow-up, estimates there rely on little information. No method can recover a well-supported comparison at times where observation probability is essentially zero.
- Check event ascertainment and coding. A competing event must be distinguished from censoring, and the event-of-interest cause must be recorded consistently. Misclassifying an observed competing event as censoring changes the analysis.
- Consider sensitivity analysis. When informative censoring is plausible, compare results under defensible assumptions about the unobserved outcomes of censored people. A single standard Gray-test p-value does not quantify sensitivity to departures from independent censoring.
A model for the censoring process, such as inverse-probability-of-censoring weighting, can address censoring that is conditionally independent given measured variables, provided the model is adequate and the probability of remaining observed is positive over the relevant range. Such approaches do not solve dependence on unmeasured prognosis without additional assumptions. A complete-case comparison or a crude proportion of observed target events is not a general substitute for CIF analysis, because both discard the timing structure and can be biased by unequal follow-up.
Common Mistakes and Alternatives
| Incorrect reasoning | Correction |
|---|---|
| “The groups have different censoring rates, so Gray’s test is invalid.” | Different rates alone do not violate independent censoring. The key question is whether censoring is independent of future event time and cause within each group, with adequate follow-up support. |
| “Anyone censored before the end of follow-up could not have had an observed event.” | An event before the censoring time is observed. The probability of an observed event depends on censoring survival at each event time, not just at the interval endpoint. |
| “A person censored at time \(C\) had no event.” | The event status is unknown after \(C\). Censoring is not evidence that an event did not occur later. |
| “If censoring is independent within groups, every analysis is automatically valid.” | The condition supports standard CIF and Gray-test inference only alongside correct event classification, suitable sampling assumptions, adequate support, and a correctly implemented procedure. |
The main alternative when the condition is doubtful is not to reinterpret Gray’s test as though it automatically adjusts for informative loss to follow-up. Instead, clarify the censoring mechanism, use a method that addresses measured predictors when justified, and assess sensitivity to plausible unmeasured dependence. If those assumptions are not credible, the data alone may not identify the desired CIF comparison.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Does the standard censoring condition require the censoring-time distributions to be identical across groups? State the condition that matters instead.
- A group has target-event probability 0.15 by time 1: 0.05 of events occur at time 0.2 and 0.10 at time 0.8. Under independent censoring, \(P(C\geq0.2)=1\) and \(P(C\geq0.8)=0.70\). What is the probability of observing a target event by time 1?
- Why can censoring high-risk participants before their likely event time underestimate a target CIF?
- If censoring is independent only after conditioning on disease severity, is an ordinary unadjusted Gray test automatically justified? Explain.
- What is the difference between a competing event and a censoring time in the analysis?
Solutions. (1) No. Censoring distributions may differ across groups. The relevant standard condition is independence of censoring time from event time and cause within each group, together with adequate observation support. (2) The observed probability is \((0.05)(1.00)+(0.10)(0.70)=0.05+0.07=0.12\), or 12.0%. It is not \(0.15\) times the censoring survival at time 1, because events at 0.2 can be observed before later censoring. (3) The people remaining under observation then have lower target-event risk than the original group. Their observed event rate does not represent the risk among those selectively censored, so the estimated CIF can be too low. (4) No. The unadjusted test does not automatically account for severity-related censoring. A method that appropriately handles the conditioning variable is needed, along with adequate measurement and modeling assumptions. (5) A competing event is an observed event of another cause that prevents the target event from being the first event. Censoring ends observation while the event status after that time is unknown.