Quick answer
Kaplan-Meier analysis estimates survival probability over time while handling censored observations. It helps compare treatment groups visually and statistically through log-rank tests. Reliable interpretation requires checking censoring patterns, reporting median survival, and complementing curves with hazard modeling when covariate effects matter.
Key takeaways
- Define the time origin, event, and censoring rules before creating the curve.
- A Kaplan-Meier curve estimates event-free probability while the risk set changes over time.
- Always display numbers at risk and avoid interpreting unstable curve tails.
- Use regression methods when adjusted effects or covariate relationships are required.
What Kaplan-Meier Estimates
The curve estimates cumulative survival through observed event times, adjusting risk sets as participants are censored or experience events.
Comparing Groups
Log-rank tests assess survival differences across groups under proportional hazards assumptions. Complement with Cox models when adjusting for covariates.
Reporting Standards
Include median survival, confidence bands, number at risk, and event counts. Explain censoring reasons and follow-up completeness.
Practical method
Step-by-Step Workflow
- 1
Define time consistently
Choose one clinically meaningful origin, one event definition, and a transparent rule for participants without an observed event.
- 2
Audit follow-up
Review event counts, censoring reasons, follow-up duration, delayed entry, and whether losses differ between groups.
- 3
Estimate and compare curves
Plot confidence bands and numbers at risk. Use a log-rank test for an unadjusted comparison when its weighting is appropriate.
- 4
Add an adjusted model
Use Cox regression or an alternative survival model when confounding, effect modification, or non-proportional hazards matter.
Worked example
Median survival is not reached
- Scenario
- At 24 months, more than half of the intervention group remains event-free, so its Kaplan-Meier curve never falls below 0.50.
- Approach
- Do not extrapolate a median. Report that the median was not reached and present survival estimates at clinically selected times, each with a confidence interval and number at risk.
- Interpretation
- A longer-looking curve does not by itself establish a constant treatment benefit. Examine the entire follow-up pattern and assess proportional hazards before summarizing it with one hazard ratio.
Common Mistakes to Avoid
- Treating censored participants as event-free forever
- Omitting the number-at-risk table
- Over-interpreting the sparsely observed curve tail
- Describing a hazard ratio as a risk ratio
Frequently Asked Questions
What does censoring mean in survival analysis?
Censoring means outcome timing is partially known, such as participants lost to follow-up or event-free at study end.
When should I use a Cox model with Kaplan-Meier?
Use Cox regression when you need adjusted hazard estimates for multiple predictors or confounders.