Questions: Kaplan-Meier Survival Analysis and Curves

Question 1

In a clinical trial with 100 patients, 60 experienced the event and 40 were censored (lost to follow-up or study ended). How does the Kaplan-Meier estimator handle the 40 censored patients?

• A. They are excluded because their outcomes are unknown, leaving 60 patients for analysis
• B. They are counted as having experienced the event at their censoring time, to be conservative
• C. They contribute survival time up to their censoring date, then are removed from the risk set for all subsequent event times
• D. Their outcomes are imputed based on the average time-to-event of similar patients who experienced the event

Answer: (C)

A censored participant is not missing data — their censoring time is real information: this person survived at least until that point. The KM estimator uses this by including them in the risk set up to their censoring time (contributing to the denominator in conditional survival calculations) and then removing them from the risk set going forward. Option A would discard real survival information and overestimate event rates. Option B would introduce false events and underestimate survival. The KM estimator's handling of censoring is its central innovation.

Question 2

Two groups are shown on a Kaplan-Meier plot. The curves cross at month 18: Group A has better survival for the first 18 months, but Group B has better survival thereafter. What is the correct interpretation?

• A. The analysis contains an error — properly constructed KM curves cannot cross
• B. The two treatments have identical overall survival and any apparent difference is random noise
• C. The relative survival benefit changes over time — Group A's treatment may have early benefit but late harm, or the groups have time-varying hazard differences
• D. The log-rank test result is automatically invalid whenever curves cross

Answer: (C)

Crossing KM curves are clinically meaningful and occur when the relative hazard between groups is not constant over time. An aggressive treatment might reduce early mortality but carry late toxicity that reverses the advantage. Curves that cross violate the proportional hazards assumption — which affects the log-rank test's sensitivity — but crossing is not an analysis error. It reflects real biology and is important clinical information about when and for whom a treatment is beneficial.

Question 3

A censored observation in survival analysis contains real information: it establishes that the participant survived at least until the time of censoring.

• T. True
• F. False

Answer: True

Censoring is not the same as a missing outcome. If a participant was followed for 3 years without experiencing the event before dropping out, we know they survived for at least 3 years. The KM estimator uses this information: the participant remains in the risk set for all event times up to their censoring point, contributing to the numerator and denominator of conditional survival estimates throughout that window. Treating such an observation as missing would waste real data.

Question 4

If a Kaplan-Meier curve seldom drops below 0.5, it means most participants in the study survived to the end of follow-up.

• T. True
• F. False

Answer: False

A KM curve that never reaches 0.5 means the median survival time cannot be estimated — not that all participants survived. This occurs when fewer than half the cohort experienced the event during follow-up, which can happen because of a high censoring rate, a short follow-up period, or genuinely excellent survival. The curve staying above 0.5 could reflect either a truly favorable outcome or heavy censoring, and distinguishing these interpretations requires examining the data carefully.

Question 5

Explain why simply excluding censored observations from a survival analysis would produce biased results, and describe how the Kaplan-Meier estimator avoids this problem.

Model answer: Excluding censored observations creates survivorship bias: only participants who experienced the event would contribute to the analysis. Since censored participants are typically people who survived longer (otherwise they would have experienced the event), excluding them overestimates the event rate and underestimates survival time. Counting them as events does the reverse. The KM estimator avoids both errors by using censored participants for the time they were observed — including them in the risk set up to their censoring date — then removing them from subsequent calculations without assigning an outcome.

The product-limit formula achieves unbiased estimates by computing conditional survival probabilities at each event time. The denominator (number at risk) reflects all participants still under observation at that moment. When a censoring occurs between event times, the risk set decreases silently — no drop in the survival curve. This approach is unbiased under the assumption that censoring is independent of the event, which is the key underlying assumption of the KM method.