Why Censoring Matters to the Logrank Test
The logrank test compares event experience among participants who remain at risk over time. A participant who is censored contributes information up to the censoring time, but after that time their unobserved event time is not available to the comparison. The test therefore relies on the composition of the risk sets: if censoring preferentially removes people who would otherwise have earlier events, the remaining participants may no longer represent the event experience the test is intended to compare.
Let \(T\) be a participant’s event time and \(C\) their censoring time. The observed follow-up time is \(Y=\min(T,C)\), and the event indicator is \(\Delta=I(T\leq C)\). A participant with \(\Delta=0\) is censored at \(C\); their event time is not observed. Here the event may be death, recurrence, or another specified endpoint, and the meaning of censoring depends on that endpoint.
The condition concerns the relationship between censoring and event prognosis, not whether the arms have equal censoring rates. It is also a statement about the event times that would be observed under continued follow-up, including the unobserved event times after censoring. That makes it a substantive assumption, not a fact that can usually be confirmed by inspecting the observed data alone.
How Independent Censoring Supports the Risk-Set Comparison
At an event time \(t_j\), let \(n_{1j}\) and \(n_{0j}\) be the numbers at risk in arms 1 and 0 immediately before the event, with \(n_j=n_{1j}+n_{0j}\). Let \(d_{1j}\) and \(d_{0j}\) be the numbers of events at that time in the respective arms, and \(d_j=d_{1j}+d_{0j}\). The logrank score for arm 1 is the sum of observed events minus expected events, where the expectation is calculated using the risk-set allocation under the null hypothesis.
To see the role of censoring, first consider an instant with one event and no tied event times. Under the null hypothesis \(H_0\) of equal event hazards, suppose each participant at risk has the same instantaneous event hazard \(\lambda(t)\). If censoring is independent of event time within each arm, being present in the risk set does not select participants in that arm according to their latent event time. Conditional on an event occurring at \(t\), each at-risk participant is therefore equally likely to be the one with the event. The probability that the event is in arm 1 is \(n_{1j}/n_j\), so the expected arm 1 event contribution is \(n_{1j}/n_j\). Summing these observed-minus-expected contributions gives \(U\), whose expectation under \(H_0\) is zero.
With multiple events, the usual logrank calculation uses the corresponding risk-set expectation and variance, with the standard handling of tied event times. The key point is unchanged: expected events are allocated according to the participants actually at risk, not according to the numbers originally randomized. Independent censoring is what makes that risk-set comparison a valid basis for testing the event-hazard null.
Worked Example: Different Censoring Rates, but Independent Within Arm
Consider hypothetical risk sets at one event time. Because of different follow-up patterns, 30 participants in arm 1 and 20 in arm 0 remain at risk. Suppose five events occur at this time, four in arm 1 and one in arm 0. Assume that censoring is independent of event time within each arm, although the censoring distributions differ between arms.
Under the null hypothesis, the expected number of arm 1 events in this tied event block is
The observed-minus-expected score contribution is \(4-3=1\). Using the standard hypergeometric variance for this tied block,
The standardized contribution is \(1/\sqrt{1.102}=0.9535\). This single block is not a full trial analysis or a basis for a trial-level conclusion. It illustrates why the risk-set expectation uses 30 and 20, rather than the original allocation, and why unequal censoring rates do not by themselves invalidate the calculation. The conclusion still depends on the stated within-arm independence assumption.
How Informative Censoring Can Distort the Comparison
Suppose censoring is associated with prognosis. If participants with a high latent risk of an event are more likely to leave follow-up, the observed risk set becomes depleted of high-risk participants. The observed event hazard among those still followed may then differ from the hazard that would have been seen with complete follow-up. If this selection differs between arms, the logrank score can shift even when treatment has no effect on event times.
Worked Example: Prognostic Selection Creates a Risk-Set Imbalance
Imagine a hypothetical randomized trial with 100 participants per arm. In each arm, 50 have a high-risk baseline prognosis and 50 have a low-risk prognosis. In the next time interval, the event probability is 0.40 for high-risk participants and 0.10 for low-risk participants, identically in both arms. Thus, within each prognosis group, treatment has no effect.
Before the interval, however, half of the high-risk participants in arm 1 are censored, while no participants in arm 0 are censored. The expected arm 1 risk set is then 25 high-risk and 50 low-risk participants, or 75 in total. The expected number of events in arm 1 is
Arm 0 retains 50 high-risk and 50 low-risk participants, so its expected event count is
Using these expected counts to illustrate the risk-set calculation, the total is 40 events and the arm 1 risk-set fraction is \(75/175\). The logrank expected count for arm 1 is therefore \(40(75/175)=17.14\), whereas its expected observed count under the prognosis-dependent censoring pattern is 15. The resulting expected score contribution is \(15-17.14=-2.143\), despite there being no treatment effect within either prognosis group.
Here censoring changes the prognostic composition of the arms’ risk sets. The crude logrank comparison may consequently reflect selection as well as event experience. If prognosis is measured, an analysis that accounts for it may be appropriate; if relevant prognosis is unmeasured, the distortion may be difficult to remove.
Diagnostics: What the Data Can and Cannot Tell You
Diagnostics are useful for finding warning signs, but they do not prove independent censoring. A censoring indicator and time can be compared with recorded baseline prognostic variables, treatment arm, and observed history before censoring. Investigators can also review reasons for withdrawal and compare the timing and frequency of censoring across arms. Strong links between censoring and known predictors of the event raise concern, particularly if those links differ by arm.
Worked Example: Interpreting Unequal Censoring Patterns
In a hypothetical trial, 30 of 100 participants in the active-treatment arm and 10 of 100 in the control arm are censored before an event. The overall censoring percentages are 30% and 10%. A review by baseline prognosis finds that 20 of 40 high-risk active-arm participants and 8 of 40 high-risk control participants are censored. Among low-risk participants, 10 of 60 active-arm participants and 2 of 60 control participants are censored.
The high-risk censoring percentages are \(20/40=50\%\) and \(8/40=20\%\); the low-risk percentages are \(10/60=16.67\%\) and \(2/60=3.333\%\). The pattern is a warning: censoring differs by arm and is more frequent among high-risk participants in both arms. If high-risk status predicts event time, the unadjusted independence condition is questionable.
These observed percentages do not establish whether censoring is independent of the unobserved event times. Nor do the overall percentages alone prove a violation: arm-specific rates may differ while censoring remains independent of event time within each arm. The useful conclusion is that the prognostic pattern warrants scrutiny and may motivate adjustment or sensitivity analysis, not that the assumption has been definitively tested.
Common warning signs include withdrawals after worsening health, loss to follow-up associated with treatment response, and censoring that depends on a post-randomization variable related to future events. Reasons recorded at withdrawal may be informative, but they are not a complete substitute for the missing event outcomes. A lack of observed association with measured predictors also cannot rule out dependence on unmeasured prognosis.
Alternatives, Interpretation, and Limitations
| Approach | When it may help | Important qualification |
|---|---|---|
| Ordinary logrank test | Censoring is independent of event time within the comparison groups. | Different arm-specific censoring distributions are allowed under that condition. |
| Adjustment for recorded prognostic factors | Censoring and event prognosis are related through measured factors. | The analysis must condition on relevant factors appropriately; adjustment cannot repair dependence on unmeasured factors by itself. |
| Inverse-probability weighting or related methods | Censoring is explainable by recorded variables and the censoring process can be modeled. | Validity depends on adequate measurement, a suitable model, and sufficient probability of remaining uncensored. |
| Sensitivity analysis | Important dependence may remain unmeasured or cannot be credibly modeled away. | Results describe how conclusions change under specified departures from the primary assumptions. |
The logrank test is a test of equality of event hazards using observed risk sets. It is not a test that censoring is independent, and a statistically significant result does not establish that the censoring assumption held. Conversely, failure to reject \(H_0\) does not demonstrate independent censoring. Under non-proportional hazards, positive and negative score contributions over time can also offset; that is a separate limitation from informative censoring.
The event definition matters as well. A competing event may be treated as censoring for a cause-specific hazard comparison, but it is not generally independent censoring for the question of a participant’s probability of experiencing the event of interest before competing events. The logrank test must be interpreted for the endpoint and estimand its risk-set construction supports.
In a group sequential trial, repeated-look boundaries address the effect of examining the test repeatedly; they do not correct informative censoring. At every interim analysis, the same concern applies: if observed risk sets are selectively depleted according to event prognosis, the score and its interpretation can be compromised. The analysis plan should specify how censoring is defined, how reasons and prognostic predictors will be reviewed, and what analyses will assess departures from the primary assumption.
Check Your Understanding
Answer each question, then compare with the solutions below.
- Can censoring be more common in one treatment arm than the other without violating the independent-censoring condition? State the condition that matters.
- Under the null hypothesis, a risk set has 24 participants in arm 1 and 16 in arm 0, and three events occur. What is the logrank expected number of arm 1 events?
- Why might censoring high-risk participants more often in one arm bias an unadjusted logrank comparison?
- Can observed censoring percentages prove that censoring is independent of latent event time? Explain.
- Does a group sequential boundary correct a problem caused by informative censoring?
Solutions: (1) Yes. The censoring distributions may differ by arm; a sufficient condition is \(T\perp C\mid A\), so censoring is independent of event time within each arm. (2) \(3(24/40)=1.8\) expected arm 1 events. (3) It changes the prognostic composition of the risk sets, potentially changing event experience among those still observed even when treatment has no effect within prognosis groups. (4) No. Event times after censoring are unobserved, so the dependence cannot generally be verified from observed percentages alone. (5) No. Sequential boundaries account for repeated testing, not selective loss of participants according to event prognosis.