Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Biostatistics / Survival analysis

Time-to-event data, censoring and the Kaplan-Meier curve

15 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Define time-to-event outcomes and the three kinds of censoring

  • Construct a Kaplan-Meier survival curve by hand

  • Read a survival curve: survival at a time, median survival and numbers at risk

Many trial outcomes are not 'did it happen?' but 'when did it happen?': time to death, to relapse, to viral failure, to first malaria episode, to hospital discharge. Such data are called time-to-event or survival data (whatever the event is). Two features make them special.

  1. Times are positive and usually skewed.
  2. Censoring: by the time the study ends, some participants have not had the event, or are lost to follow-up. We know only that they were event-free up to a certain time. Discarding them throws away information; treating their last follow-up as the event time underestimates survival.

Types of censoring

  • Right censoring (the usual): the event has not happened by the end of observation (study ends, patient moves away, dies from an unrelated cause).
  • Left censoring: the event happened before observation began (time of HIV infection in a person first tested positive).
  • Interval censoring: the event occurred between two visits (HIV seroconversion between 6-monthly tests).
Key assumption: non-informative censoring. Participants who are censored must have the same future risk as those who stay in. If sicker patients drop out, survival is overestimated.
Month 0Patient enrolledMonth 9Event (died): completeobservationMonth 14Moved away: censoredMonth 24Study ends: censored
Each participant contributes follow-up time until the event or until they are censored.

The Kaplan-Meier estimator

The Kaplan-Meier (KM) estimator estimates the survival function S(t), the probability of remaining event-free beyond time t, using every participant for as long as they were observed. At each event time, multiply the previous survival by the proportion who survive that moment among those still at risk:

S(t) = Π_(t_i ≤ t) (1 - (d_i)/(n_i))

d_i = events at time t_i, n_i = number still at risk just before t_i. Censored participants are in the denominator until they leave, and then silently drop out.

Worked example by hand: 10 patients

Follow-up (months): 2, 3, 3\*, 5, 6, 8\*, 9, 12, 14\*, 18 (an asterisk marks a censored time; the others are deaths).

TimeAt risk (n)Deaths (d)Survival \hat S(t)
21010.900
3910.900 × 8/9 = 0.800
5710.800 × 6/7 = 0.686
6610.686 × 5/6 = 0.571
9410.571 × 3/4 = 0.429
12310.429 × 2/3 = 0.286
18110.286 × 0/1 = 0

Note that the censored patient at month 3 (and those at 8 and 14) reduce the number at risk afterwards but do not cause the curve to drop. The median survival is the first time the curve falls to 0.5 or below: 9 months.

Reading a KM plot

  • The curve is a step function: flat between events, dropping at each event.
  • Survival at a time: read up from the time on the x-axis to the curve, then across (e.g. 1-year survival).
  • Median survival: where the curve crosses 0.5. If the curve never reaches 0.5, the median is 'not reached' (and you should report survival at a fixed time instead).
  • Tick marks show censored times. A long run of ticks at the right end signals few people left, so the tail is unreliable.
  • Number at risk table beneath the plot shows how many remain. Always include it: curves with very few at risk are wobbly and wide in uncertainty.
  • Confidence bands widen as numbers at risk fall.
00.250.50.751051015MonthsProportion survivingSingle arm (illustrative)Median = 9 months
Kaplan-Meier curve for the ten patients. Median survival is where the curve crosses 50%.

Do not use simple proportions when follow-up differs

'20 of 100 died' sounds like 20% mortality, but if half the patients were followed for only 6 months and half for 5 years, that figure is not comparable between studies. KM estimates avoid this by accounting for who was still being observed. Likewise, an average of survival times that includes censored times is biased low.

What to report

Median survival (with 95% CI) or survival at a prespecified time, the number of events and the median follow-up (estimated by the reverse Kaplan-Meier method).

Summary

  • Survival data record time to an event; censoring means the event had not occurred when observation stopped.

  • Censoring must be non-informative for the analysis to be valid.

  • Kaplan-Meier multiplies conditional survival probabilities at each event time using those at risk.

  • Median survival is where the curve crosses 0.5; show the number at risk beneath the plot.

Check your understanding

Question 1
1

A patient moves away and is last seen alive at month 14. This patient is:

Question 2
2

In the table, why does survival not drop at a censoring time?

Question 3
3

If 6 patients are at risk and 1 dies, and the previous survival was 0.60, the new survival is:

Question 4
4

The median survival time is defined as the time when the KM curve first reaches:

Question 5
5

Which assumption makes censoring acceptable?

Question 6
6

What does a long cluster of tick marks at the end of a KM plot suggest?

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Poisson regression: counts, rates and person-timeComparing survival curves: the log-rank test and hazard ratios