Numbers at risk in Kaplan–Meier plots, explained

Updated 2026-09-25 · TrialCurve methods team

Why it matters

• Reading the tail: a curve supported by 3 patients at 48 months is far less certain than one supported by 300.
• Separating censoring from events: the drop in numbers at risk between two times equals events plus censorings. The curve tells you the events, so the table reveals the censoring.
• Reconstruction: the Guyot algorithm uses these counts to pin down the censoring in each interval. Without them, reconstructed tails are much less reliable.

A simple consistency check

Censoring only removes patients, so the fraction still at risk can never exceed the survival probability: n(t)/n(0) ≤ S(t). If a digitised table breaks this, a cell was probably misread or rows were swapped between arms. TrialCurve runs this check automatically and flags any violation.

Adding the table to your own plots

• R: survminer::ggsurvplot(fit, risk.table = TRUE), or ggsurvfit::add_risktable().
• SAS: PROC LIFETEST ... plots=survival(atrisk=0 to 60 by 12).
• Stata: sts graph, by(arm) risktable.
• GraphPad Prism: add a "number at risk" table below the survival graph from the analysis results.

Extracting it from a published figure

Copy every cell exactly, together with the time each column refers to, which is usually the x-axis tick above it. Common pitfalls are columns that skip ticks, rows listed in a different order from the legend, and small digits misread in low-resolution images. TrialCurve's vision model reads every cell, zooming into the table where the digits are small, and the reconstruction reports how closely it reproduces the curve, so a misread count shows up as a mismatch.

FAQ