The most common misreading of a Kaplan-Meier curve is treating the tick marks on the line as events. They are the opposite: a tick mark is a censored observation — a participant who left the study, was lost to follow-up, or reached the end of the study window without the event happening. The curve only drops at an event; it never drops at a censoring mark. Get that one distinction right and most of the rest of the plot falls into place.
A second common error is reading the tail of the curve with the same confidence as the early portion. A Kaplan-Meier curve is a step function estimated from shrinking numbers of people at risk, and the further right you look, the fewer people that estimate is built on. A flat-looking tail at month 36 built on four remaining participants is not the same kind of evidence as the steep early drop built on 200. The numbers-at-risk table underneath the plot exists specifically to let you check this, and skipping it is the single biggest reason KM curves get over-interpreted.
What a Kaplan-Meier Curve Actually Plots
A Kaplan-Meier curve estimates the survival function S(t) — the probability that the event of interest (death, relapse, equipment failure, grant termination, whatever “event” means in the specific study) has not yet occurred by time t. The y-axis runs from 1.0 (100% event-free) down to 0, and the x-axis is time since some defined origin (enrollment, diagnosis, randomization). “Survival” is the generic statistical term for the method; it is used identically for non-mortality time-to-event outcomes like time-to-relapse, time-to-device-failure, or time-to-manuscript-acceptance, and it is the standard descriptive method for time-to-event outcomes in a cohort study design.
The method, developed by Edward Kaplan and Paul Meier and published jointly in 1958 after the two submitted near-identical papers to the same journal, is a nonparametric estimator: it makes no assumption about the shape of the underlying survival distribution. That is precisely why it tolerates censored data, which is the condition under which most real time-to-event datasets are collected — not everyone in a study reaches the event before the study ends.
Why the Curve Is a Step Function
The line is flat everywhere except at the exact times an event occurred, where it drops. At each event time ti, the estimator multiplies the running survival probability by (1 − di/ni), where di is the number of events at that time and ni is the number still at risk immediately before it. The drop’s height is proportional to how many events happened relative to how many people were still being tracked — a single event among 10 remaining people produces a much bigger visible drop than a single event among 200. This is why early, well-populated portions of the curve tend to look smoother and later portions look choppier: the same single event has a bigger multiplicative effect on a small remaining risk set.
Censoring Marks: What the Tick Marks Mean
The vertical tick marks (sometimes plus signs, sometimes small crosses) on the line mark censoring times — points where a participant’s follow-up ended without the event being observed. Three things cause censoring in most designs: the study ended while the participant was still event-free (administrative censoring), the participant withdrew or was lost to follow-up, or the participant experienced something other than the event of interest (competing-risk censoring, which needs its own caveat below). A censoring mark does not lower the curve. It does, however, reduce the number at risk for every subsequent event-time calculation, which is why a cluster of censoring marks followed shortly by an event produces a visibly larger drop than the same event would have produced without the preceding censoring.
The counter-example worth knowing: if a participant is lost to follow-up and the event is later discovered to have happened before the censoring date (a corrected event, not a censoring), that is not plotted as a tick mark — it’s re-entered as an event at the true event time and the curve is recalculated. A tick mark specifically represents “event status unknown after this point,” not “event known not to have happened.”
Reading Off Median Survival
Median survival time is the time at which the curve crosses S(t) = 0.5 — the point where half the studied population has had the event and half has not (yet). To read it: draw a horizontal line at y = 0.5 across to where it meets the step function, then drop straight down to the x-axis. Because the curve is a step function, the crossing point is usually the x-value of whichever event caused the curve to drop to or below 0.5, not an interpolated point.
Two things trip people up here. First, if the curve never drops to 0.5 by the end of follow-up (common in trials with good outcomes, or short follow-up), median survival is not estimable from that curve — reported instead as “not reached,” which is a real, correctly reported result, not a missing one. Second, median survival is not the same as mean survival, and mean survival generally cannot be estimated reliably from a censored KM curve at all, because the tail beyond the last event is unknown. Papers that report a KM-derived mean survival are almost always restricting it to a fixed time horizon (restricted mean survival time, RMST) rather than computing a true unrestricted mean.
The Numbers-at-Risk Table: Why the Tail Can Mislead
Underneath the plot, most published KM figures include a “number at risk” row (or one row per group) showing how many participants remained under observation at each labeled time point on the x-axis. This table is not decorative — it’s the honesty check on the plot above it. A curve that looks stable or that shows a late, dramatic separation between two groups deserves much less weight if the risk table shows the remaining sample has dropped to single digits by that point. A wide, confident-looking gap between two curves built on 3 people per arm is a statistical accident waiting to be reported, not a finding.
As a working rule: treat any portion of a KM curve past the point where the risk table drops below roughly 10–15% of the starting sample size as descriptive only, not as a basis for a strong claim, and check whether the paper’s own confidence bands (if plotted) widen accordingly in that region — they should.
Worked Example: Interpreting a Two-Group KM Plot
Illustrative example — not data from a real trial, constructed to show the reading process.
| Time (months) | Arm A at risk | Arm A events | Arm B at risk | Arm B events |
|---|---|---|---|---|
| 0 | 120 | 0 | 118 | 0 |
| 6 | 101 | 8 | 95 | 14 |
| 12 | 84 | 6 | 70 | 11 |
| 18 | 62 | 5 | 41 | 7 |
| 24 | 39 | 3 | 19 | 4 |
Reading this: both curves start at 1.0 at month 0. By month 12, Arm A’s estimated survival is noticeably higher than Arm B’s — the drops in Arm B have been both more frequent and occurring against a smaller risk set, so each event pulls the curve down further. By month 24, Arm B’s risk set (19) is thin enough that the last segment of its curve should be read cautiously; Arm A’s risk set (39) supports somewhat more confidence but is also thinning. The sentence a researcher would write from this: “Kaplan-Meier estimated survival at 12 months was higher in Arm A than Arm B; the separation persisted to 24 months, though risk-set sizes beyond 18 months were small in both arms and later estimates should be interpreted with caution.” Note what that sentence does not claim: it does not put a precise number on “how much better,” because that comparison belongs to a formal test (below), not to eyeballing the plot.
Comparing Curves: Where the Log-Rank Test and Hazard Ratio Come In
A KM plot showing two curves that look separated is a description, not a statistical test. The formal comparison is the log-rank test — a nonparametric, chi-square-based test of whether the two survival distributions differ, which returns a p-value — or a Cox proportional-hazards model, which yields a hazard ratio with its own confidence interval. A hazard ratio is frequently misread as a straightforward risk ratio — “Arm B had 2x the risk of the event” — but it is an instantaneous-hazard ratio averaged (under the model’s proportional-hazards assumption) across the whole follow-up period, not a statement about cumulative risk or about any single time point. It functions as a time-to-event effect size measure, but reading it correctly, and reading the KM curve it came from, are related but separate skills — the curve shows you when separation happens; the hazard ratio (under its own assumption of proportional hazards, checkable via a log-log plot or a Schoenfeld residuals test) summarizes the relative rate across the whole curve into one number.
Assumptions and When Not to Trust the Curve
- Non-informative (independent) censoring. The Kaplan-Meier estimator assumes that being censored is unrelated to your underlying risk of the event — a participant censored at month 6 is assumed to have the same future event risk as anyone else still at risk at month 6. If sicker participants are systematically more likely to drop out before their event is recorded, the curve is optimistically biased — in effect, follow-up time itself becomes a confounding variable. This assumption is rarely directly testable and should be argued from study design, not assumed by default.
- Competing risks. If participants can leave the risk set via a different, competing event (e.g., death from another cause when the outcome of interest is disease relapse), standard KM treats that competing event as ordinary censoring, which can overestimate the cumulative incidence of the event of interest. A cumulative incidence function or competing-risks regression is the more defensible method whenever a competing event is common in the population studied.
- Small late-follow-up risk sets. As covered above, the tail of any KM curve is only as trustworthy as the number of people still contributing to it — always cross-check against the risk table before drawing conclusions from the right-hand side of the plot.
- Proportional hazards is a separate assumption from anything required to draw the KM curve itself — you can plot and correctly read a KM curve with no proportional-hazards assumption at all. It only becomes relevant the moment you want to summarize the comparison with a single hazard ratio from a Cox model.
Related Interpretation Guides
Reading a KM curve correctly is one instance of a broader skill — knowing what a given statistical output can and can’t support. Related interpretation guides on CASRAI: Cronbach’s Alpha, Intraclass Correlation Coefficient (ICC), Normality of Distribution, Skewness: Left-Skewed vs. Right-Skewed, Sensitivity vs. Specificity, and Statistical vs. Clinical Significance.
Frequently Asked Questions
How do you read a Kaplan-Meier curve?
Read the y-axis as the estimated probability of remaining event-free at each point in time on the x-axis. The line only drops at an actual event; tick marks show censoring, not events. Compare curves by their vertical separation over time, read median survival off the point where a curve crosses y = 0.5, and check the numbers-at-risk table before drawing conclusions from the far-right portion of the plot.
What is censoring in survival analysis?
Censoring means a participant’s exact event time is unknown because their follow-up ended before the event occurred or was observed — commonly because the study ended, they withdrew, or they were lost to follow-up. Censored participants still contribute information (they are known to be event-free up to their censoring time) and are represented by tick marks on the KM curve, not by a drop in the line.
What is median survival time?
Median survival time is the time point at which the Kaplan-Meier curve crosses a 50% survival probability — when half of the studied population has experienced the event. If the curve never reaches 50% within the follow-up period, median survival is reported as “not reached” rather than estimated by extrapolation.
What does it mean if a Kaplan-Meier curve never reaches 0%?
It means that, by the end of observed follow-up, not everyone remaining in the risk set has experienced the event — some were censored before the study ended. A curve that flattens above zero and stays there through the last censoring mark is showing exactly that: no further events were observed, but the final at-risk participants left the study without one either.
Is a Kaplan-Meier curve the same as a hazard ratio?
No. The KM curve is a descriptive estimate of survival probability over time for one or more groups. A hazard ratio is a single summary number from a separate model (typically Cox proportional-hazards) comparing the instantaneous event rate between groups across the whole follow-up period. You can produce a KM curve without ever fitting a hazard ratio, and a hazard ratio does not tell you when in time a difference between groups emerged — only the curve does that.
