Kaplan–Meier curves
In one sentence
A Kaplan–Meier curve estimates the probability of remaining free of a defined event over time while accounting for incomplete follow-up.
The intuition
The curve steps down when events occur. People whose follow-up ends without an observed event contribute information until that time. They are not counted as permanently event-free; their later outcome is unknown to the analysis.
How it works
The horizontal axis is time from a specified origin. The vertical axis is the estimated probability of remaining free of the endpoint. A death curve and a recurrence curve can have different steps because they count different events.
Censoring means follow-up ends before an event is observed. It may occur at a data cutoff, loss to follow-up or another prespecified circumstance. Standard interpretation assumes censoring is sufficiently non-informative; patients dropping out because they are worsening can threaten that assumption.
The number-at-risk table shows how many people remain observed and event-free at each interval. Estimates near the tail can rest on very few people. Confidence bands often widen there, even when the line looks flat. A five-year point is an estimate under that follow-up, rather than a count showing every enrolled person reached five years.
Median survival is the time the estimated curve reaches 50%. If it never reaches 50% by the cutoff, the median is not reached. That does not mean no events occurred or that the outcome is infinite. A hazard ratio summarizes another aspect of the comparison; medians and a ratio cannot be substituted for each other.
Why it matters in cancer
It helps readers inspect time, event definitions and follow-up before quoting a survival percentage.
Worked example
A study enrolls 100 people but only ten remain at risk near six years. A flat late segment does not demonstrate a precise six-year result for all 100. Inspect censoring, the risk table and confidence interval before quoting the tail.
Common confusions
- Censored does not mean cured or known to have had an event.
- “Median not reached” does not mean all patients are event-free.
- Apparent curve separation is not alone proof of a meaningful or causal treatment effect.
Related concepts
Hazard ratios and endpoints.
Sources and scope
General teaching, source-checked October 8, 2026. The examples are hypothetical unless a study is explicitly named. Expert and learner review remain pending.