Kaplan-Meier Survival

A survival figure carries three things a reviewer looks for: the step curves, the log-rank p-value, and the number still at risk under the x-axis. Drop the risk table and the curves at the right edge become impossible to trust, because you cannot tell whether a flat tail rests on forty patients or four.

This page draws the full figure for a two-arm trial with right-censored follow-up. R uses survminer, which produces the curves, the table, and the p-value in one call. Python uses lifelines. Both run in the pinned pubplot container and return the same log-rank p.

The data

survival_trial.csv has one row per patient: time in months, event coded 1 for the event and 0 for censored, and arm. Sixty patients per arm. The treated arm survives longer.

Curves, p-value, risk table

The tick marks are censored patients. event = 0 means the patient left the study before the event, and marking them is what separates an honest survival curve from a line that pretends everyone was followed to the end.

R

library(survival)
library(survminer)
library(ggpubr)

df$arm <- factor(df$arm, levels = c("control", "treated"))
fit <- surv_fit(Surv(time, event) ~ arm, data = df)

g <- ggsurvplot(fit, data = df, pval = TRUE, risk.table = TRUE, conf.int = FALSE,
                censor = TRUE, palette = c("#3B6DB3", "#C1432B"),
                legend.title = "Arm", legend.labs = c("control", "treated"),
                xlab = "Months", ylab = "Survival probability",
                ggtheme = theme_classic(base_size = 13),
                tables.theme = theme_cleantable())

# ggsurvplot returns the curve and the table as separate ggplots; stack them.
ggarrange(g$plot, g$table, ncol = 1, nrow = 2, heights = c(0.74, 0.26), align = "v")

Python

import matplotlib.pyplot as plt
import pandas as pd
from lifelines import KaplanMeierFitter
from lifelines.plotting import add_at_risk_counts
from lifelines.statistics import multivariate_logrank_test

order, pal = ["control", "treated"], ["#3B6DB3", "#C1432B"]
fig, ax = plt.subplots(figsize=(5.6, 5.8))

fitters = []
for color, grp in zip(pal, order):
    sub = df[df["arm"] == grp]
    kmf = KaplanMeierFitter().fit(sub["time"], sub["event"], label=grp)
    kmf.plot_survival_function(ax=ax, ci_show=False, color=color,
                               show_censors=True, censor_styles={"marker": "|", "ms": 6})
    fitters.append(kmf)

p = multivariate_logrank_test(df["time"], df["arm"], df["event"]).p_value
ax.text(0.04, 0.08, f"log-rank p = {p:.2g}", transform=ax.transAxes, fontsize=11)
add_at_risk_counts(*fitters, ax=ax)   # the table under the axis

Both engines return log-rank p = 0.007. The one visible difference is the table. survminer prints number at risk only; lifelines adds censored and event counts by default, which is more information in the same space. Neither is wrong; pick the one your journal expects.

Making it publication-ready

• Always show the risk table. A survival curve without it is not publishable in most clinical journals, and for good reason.
• Mark the censored patients. The tick marks tell the reader the tail is real, not extrapolated.
• Put the log-rank p on the plot, not only in the caption. It is the headline result.
• Do not extend the x-axis past the last informative time. When the smaller arm drops to one or two at risk, the curve past that point is noise.

For an adjusted comparison, a hazard ratio from a Cox model rather than a raw log-rank test, see the Cox forest plot. The runnable scripts and fixture are in the companion code repo under guides/figures/kaplan-meier/.