Why Censoring Independence Matters
A Kaplan-Meier estimate uses people who are still under observation and have not yet experienced the event to learn about survival. That interpretation depends on a key question: do the people who remain observed have the same future event-time prospects as comparable people who were censored? If censoring preferentially removes people with shorter or longer event times, the observed risk sets can become systematically unrepresentative.
Censoring is not itself an event. A right-censored record tells us that an event time exceeds the last time the person was known to be event-free; it does not tell us when the event eventually occurs. The earlier material on right-censored records distinguishes this partial information from an exact event time. Here the concern is how the reason or timing of censoring affects interpretation of the people who remain at risk.
A common example is a study that ends on a fixed calendar date. If people enter the study at different times for reasons unrelated to their event prospects, some simply have less potential follow-up. This can produce administrative censoring that is independent of event time. By contrast, if a person leaves observation because a deteriorating condition makes an event more likely, censoring may carry information about \(T\).
Independence is a statement about the data-generating process, not a feature that can usually be confirmed from the observed records alone. For a censored person, the later event time is unknown, so it is impossible to compare it directly with that person's censoring time. Study design, knowledge of the reasons for censoring, and recorded covariates therefore matter.
What a Risk Set Represents
At time \(t\), the risk set consists of individuals who have not yet experienced the event and have not yet been censored. In event-time notation, they satisfy \(T\ge t\) and \(C\ge t\), subject to the study's convention for records occurring exactly at \(t\). The people in this set are the ones who can contribute information about what happens next.
Under independent censoring, membership in the risk set because of continued observation does not, by itself, select people with different event-time prospects. Among people who have survived event-free to \(t\), those still observed have the same distribution of future event times as all people who survived event-free to \(t\). This is the population interpretation that allows risk-set information to represent survival in the source population.
The equality follows from independence: once we condition on having remained event-free to \(t\), the additional condition that observation has continued to \(t\) supplies no further information about \(T\). This is a statement about the conditional event-time distribution, not a claim that censoring never occurs or that every person is followed indefinitely.
The Kaplan-Meier product-limit estimate is built from conditional survival information at successive event times. This tutorial focuses on the assumption that makes its risk sets interpretable, rather than on the estimator's calculation. If censoring changes the composition of a risk set according to event risk, the observed conditional survival may describe a selected group instead of the source population.
Worked Example: Independent Administrative Follow-Up
A hypothetical study follows 1,000 policyholders for a specified event. By time 1, the probability of remaining event-free is 0.80. Among those event-free at time 1, the probability of an event during the next interval, from time 1 to time 2, is 0.25. Follow-up availability is assigned independently of event time: half the cohort has observation through time 2, while the other half is censored at time 1. Find the expected number at risk at time 1 among those observed through time 2, the expected number of observed events in the next interval, and the event proportion in that risk set.
Method. Use independence to calculate the expected size of the observed risk set, then apply the event probability among people who are event-free at time 1.
The expected number event-free at time 1 in the full cohort is \(1{,}000(0.80)=800\). Because half have follow-up through time 2, independence gives \(800(0.50)=400\) expected people who are both event-free at time 1 and observed through time 2.
The expected number of events in the next interval among those 400 is \(400(0.25)=100\). The corresponding expected number remaining event-free through time 2 is \(400-100=300\).
The expected event proportion in this observed risk set is \(100/400=0.25\). It matches the specified event probability for all people event-free at time 1 because follow-up availability did not select on event time.
Check. The expected 800 people event-free at time 1 split by follow-up availability into 400 observed through time 2 and 400 censored at time 1. Among the observed 400, 100 events plus 300 event-free people account for all 400. The common mistake this avoids is treating the 400 people censored at time 1 as if their event times were known; they contribute no observed event status for the next interval.
When Censoring Is Informative
Censoring is informative for the event time when the censoring mechanism is associated with event prospects. For instance, people at higher event risk may be more likely to withdraw, or a financial limit may cause high-loss cases to stop being observed sooner. The remaining risk set can then have a different risk profile from the original population, even if every recorded censoring time has been coded correctly.
A useful way to reason about this is to ask what the censoring process removes. If censoring removes people in a way unrelated to their event times among those still event-free, it reduces available information but does not systematically change the event-time distribution represented by the risk set. If it removes people with different future event prospects, the risk set is selected.
Worked Example: High-Risk People Leave Observation More Often
A hypothetical cohort contains 100 people in a high-risk category and 100 in a low-risk category. During the next interval, the event probability is 0.60 for each high-risk person and 0.10 for each low-risk person. Before that interval starts, 80 high-risk people and 20 low-risk people are censored. Calculate the expected event proportion for the original cohort and for those who remain observed. Explain the interpretation.
Method. Calculate expected events by category before censoring to obtain the original-cohort rate, then recalculate using only the retained people.
In the original cohort, expected high-risk events are \(100(0.60)=60\); expected low-risk events are \(100(0.10)=10\). Thus the expected total is \(60+10=70\), and the expected event proportion is \(70/200=0.35\).
After censoring, 20 high-risk people and 80 low-risk people remain. Their expected event counts are \(20(0.60)=12\) and \(80(0.10)=8\), respectively. The total is 20 expected events among 100 retained people, giving \(20/100=0.20\).
The retained group's expected event proportion, 0.20, differs from the original cohort's 0.35 because censoring disproportionately removed the high-risk category. This is informative censoring in the pooled cohort: risk category is related to event prospects and to censoring.
Check. The retained counts are \(20+80=100\), and their expected events sum to \(12+8=20\). The error to avoid is interpreting 0.20 as the original population's event probability merely because it was calculated from people still observed. Under this selection mechanism, the risk set does not represent the original risk mix.
Conditional Independence and the Target Population
Censoring need not be independent of event time marginally if relevant covariates explain their association. Let \(X\) denote a measured characteristic such as risk category. A conditional independence assumption says that \(T\) and \(C\) are independent within each value of \(X\). In notation, \(T\perp C\mid X\). This can support valid survival comparisons within covariate groups, provided the covariate is measured and the conditional assumption is appropriate.
However, conditional independence does not automatically make a pooled risk set representative of the overall population. If censoring rates differ by \(X\), the mix of covariates in the pooled risk set may change over time. When event prospects also differ by \(X\), that changing mix can alter the pooled event experience. A population-level interpretation may require group-specific estimation followed by standardization to a clearly specified target population.
Worked Example: Conditional Independence but a Changing Risk Mix
A hypothetical target population has 100 people in group A and 100 in group B. In the next year, event probabilities are 0.10 in A and 0.50 in B. Before follow-up for that year, 10% of A and 80% of B are censored. Assume censoring is random within each group, so event time and censoring time are independent conditional on group. Find the expected pooled event proportion among those retained, compare it with the target population's event proportion, and state an appropriate interpretation.
Method. Calculate the target population rate using its equal group sizes. Then calculate expected retained counts and events by group. Random censoring within each group preserves that group's event probability among those retained.
In the target population, expected events are \(100(0.10)=10\) in A and \(100(0.50)=50\) in B. The target total is 60 events among 200 people, for an expected event proportion of \(60/200=0.30\).
The expected retained counts are \(100(1-0.10)=90\) in A and \(100(1-0.80)=20\) in B. Expected events among those retained are \(90(0.10)=9\) in A and \(20(0.50)=10\) in B. The pooled retained proportion is \((9+10)/(90+20)=19/110\approx0.1727\).
Within group A the expected rate remains 0.10, and within group B it remains 0.50. The pooled retained rate is lower than the target rate of 0.30 because the retained group contains many more people from low-risk A than high-risk B. Group-specific estimates could be combined using the target's equal group weights: \(0.5(0.10)+0.5(0.50)=0.30\).
Check. The retained counts total \(90+20=110\), and the expected event counts total \(9+10=19\). The mistake this avoids is assuming conditional independence guarantees that a pooled risk set retains the target population's covariate mix.
Assumptions, Scope, and Common Traps
Independent censoring is often a plausible design assumption, but it should be stated with its scope. The independence condition must apply over the time range for which survival is being interpreted. If no one can remain under observation beyond a particular time, the data contain no direct risk-set information about survival later than that point. More generally, there must be a positive chance of remaining uncensored long enough to observe the event-time behavior of interest; otherwise, that behavior is not identified by the observed follow-up.
The relevant population also needs to be explicit. A Kaplan-Meier estimate based on a selected sample does not automatically describe every person who might be of interest. It describes survival for the population represented by the sampled risk sets, subject to the censoring assumptions and study design. If entry into the study, eligibility, or follow-up rules select a subgroup, conclusions should be framed for that subgroup unless additional methods justify broader generalization.
A common exam trap is to claim that censoring is independent simply because censored records are not counted as events. Correct coding is necessary, but it does not establish independence. Another trap is to call every loss to follow-up informative. Loss to follow-up may be independent, informative, or conditionally independent given measured covariates; its reason and relationship to event prospects determine the issue.
Check Your Understanding
For each problem, distinguish the event-time information from the censoring mechanism and identify the population represented by the resulting risk set.
- A hypothetical cohort has 500 people. At time 2, 300 are event-free. Of these, 40% are independently censored before the next interval; among event-free people, the next-interval event probability is 0.20. Find the expected number available at the start of the interval and expected events among them. (Answer: 180 available; 36 expected events.)
- A population has 60 people with event probability 0.40 and 140 people with event probability 0.10 in the next interval. Before the interval, 30 of the first group and 20 of the second group are censored. Calculate the original-cohort and retained-group expected event proportions. (Answer: Original \(38/200=0.1900\); retained \((12+12)/150=0.1600\).)
- Two equally sized groups have next-interval event probabilities 0.20 and 0.60. Censoring is random within each group, but leaves 80 people from the first group and 20 from the second. Find the target population rate and the pooled retained rate. (Answer: Target \(0.4000\); pooled retained \((16+12)/100=0.2800\).)
- A study reports that the probability of remaining under observation is 0.70 for event-free people with a measured risk marker and 0.90 for those without it. Within each marker group, the censoring process is independent of event time. What additional feature must be considered before interpreting a pooled estimate as survival for the original population? (Answer: Whether event prospects differ by marker and the pooled risk-set marker mix differs from the target mix; conditional independence alone does not ensure pooled representativeness.)
- In a hypothetical study, every person is administratively censored by time 5, and no one remains observed after time 5. What can the data support about survival after time 5, even if censoring is independent? (Answer: They provide no direct risk-set information about survival beyond time 5; independent censoring does not overcome the lack of follow-up support.)