Survival Analysis and Kaplan-Meier: An Introduction

Survival analysis is the branch of statistics devoted to modeling the time until a well-defined event occurs — death, disease relapse, mechanical failure, seed germination, or predator attack. The outcome is not binary "did it happen," as in Logistic Regression for Binary Outcomes, but a duration: time-to-event. The central object of interest is the survival function,

the probability that the event has not yet occurred by time $t$. $S(t)$ starts at 1 (everyone is event-free at time zero) and decreases (or stays flat) as $t$ increases, eventually approaching 0 if everyone eventually experiences the event. The defining complication is censoring: for many subjects, we do not observe the exact event time, only that it exceeds some known value. The most common form is right censoring — the subject is known to be event-free up to time $c$, but what happens after is unknown, because the study ended, the subject was lost to follow-up, or they died of an unrelated cause. Right censoring differs fundamentally from a missing value: we still have partial information (survival at least to time $c$), and a valid method must use that information rather than discard it. A crucial assumption, almost always invoked and not always checked, is non-informative censoring: the reason a subject is censored must be unrelated to their underlying prognosis. If sicker patients are more likely to drop out, censoring becomes informative and the estimator is biased. The hazard function $h(t)$ is the instantaneous event rate among those still at risk at time $t$ — formally, the limiting probability of the event in a small interval $[t, t+dt)$ given survival to $t$, divided by $dt$. Hazard and survival are mathematically linked: $S(t) = \exp\left(-\int_0^t h(u)\,du\right)$. Comparing hazards between groups (via the hazard ratio) is the natural way to summarize "how much faster" one group experiences the event, and it underlies both the log-rank test and Cox proportional-hazards regression, the standard extension of Kaplan-Meier to adjust for covariates.

Step 1 — Assemble the data. Each subject contributes a time (from entry to event or censoring) and an indicator (event = 1, censored = 0). Subjects can enter the study at different calendar dates; what matters is time-on-study, not calendar time — this is the standard convention, though "left-truncation" adjustments exist when entry itself must be modeled.

Step 2 — Order the distinct event times. List every unique time at which at least one event (not censoring) occurred: $t_1 < t_2 < \dots < t_k$. Censoring times matter too, but they only affect who remains "at risk" — they do not generate their own step in the curve.

Step 3 — At each event time $t_i$, compute:

The at-risk set $n_i$ shrinks over time both because subjects have events and because subjects become censored and leave the risk set. This is the trick that makes censored subjects useful: a subject censored at time 40 contributes to every risk set up through time 40, then silently drops out — neither counted as a failure nor deleted retroactively.

Step 4 — Estimate the conditional survival probability at each event time:

This is simply "the fraction of the at-risk group that made it past this moment."

Step 5 — Multiply the conditional probabilities. The Kaplan-Meier estimator of overall survival past time $t$ is the product of all conditional survival probabilities up to and including $t$:

This is the product-limit formula — the survival curve is a staircase, dropping at each observed event time and flat everywhere else (including at censoring times, where the curve doesn't drop but a tick mark is often plotted to show a censoring occurred).

Step 6 — Quantify uncertainty. Greenwood's formula gives the variance of $\hat{S}(t)$:

from which confidence bands are built (typically after a log-transform to keep bounds within $[0,1]$).

Step 7 — Compare groups with the log-rank test. At each event time across the pooled sample, construct a $2\times2$ table (group × event/no event) among those at risk. Under the null hypothesis that both groups share the same hazard, the expected number of events in group A at time $t_i$ is

(at-risk count in group A, times the overall event fraction). Sum the observed-minus-expected differences over all event times, and compare to their pooled variance to form a chi-squared statistic:

This is essentially a weighted, sequential Mantel-Haenszel test performed at every distinct event time and summed — no assumption of a specific parametric survival distribution is required, which is the test's great strength.

The founding paper. Before 1958, survival data in clinical and actuarial settings were typically handled with the actuarial (life-table) method: time was grouped into coarse intervals (e.g., yearly), and survival probabilities were computed per interval, then multiplied across intervals — essentially a discretized precursor to the product-limit idea. This works reasonably when intervals are wide and data are plentiful, but it wastes information and can be sensitive to arbitrary interval boundaries, especially with the smaller, irregularly-timed clinical datasets typical of the era. In 1958, Edward L. Kaplan, a statistician at Bell Telephone Laboratories, and Paul Meier, a biostatistician then at Johns Hopkins, independently submitted almost identical papers to the Journal of the American Statistical Association solving this problem exactly — using every distinct observed event time, rather than arbitrary intervals, as the unit of calculation. The journal editor persuaded them to merge their manuscripts into a single joint paper, "Nonparametric Estimation from Incomplete Observations" (Kaplan & Meier, 1958), which introduced the product-limit estimator described above. It went on to become one of the most-cited statistics papers of the twentieth century — for decades it held the record as the most-cited paper in all of science according to some citation counts, a reflection of how central time-to-event analysis became across medicine, engineering, and biology.

What the paper actually demonstrated. Kaplan and Meier proved that their estimator is the nonparametric maximum likelihood estimator (NPMLE) of the survival function given right-censored data — that is, among all possible survival curves consistent with the observed data (events and censoring times), the Kaplan-Meier curve is the one that makes the observed data most probable, without assuming any particular parametric shape (no assumption of exponential, Weibull, or any other distribution for survival times). This nonparametric property is exactly why the method spread so fast: it requires no assumption about how survival probability declines, only that censoring is non-informative.

Validating the method against known truth. Because Kaplan-Meier is a nonparametric estimator, its validity has been repeatedly checked via simulation studies: generate survival times from a known distribution (say, exponential with a known true hazard), impose artificial censoring, then verify that the Kaplan-Meier curve — averaged over many simulated datasets — converges to the true survival curve, and that its Greenwood-based confidence intervals achieve their nominal coverage (e.g., a 95% CI contains the true $S(t)$ about 95% of the time). Such simulation studies, refined for decades in the biostatistics literature, are the modern analogue of "proving" the estimator, since there is no single decisive experiment analogous to Meselson-Stahl; instead, the guarantee is mathematical (the NPMLE proof) reinforced by simulation.

Establishing the log-rank test. The log-rank test itself traces to independent work by Nathan Mantel (extending the Mantel-Haenszel approach to survival data, 1966) and was placed on rigorous footing by Richard Peto and Julian Peto (1972), whose paper "Asymptotically Efficient Rank Invariant Test Procedures" formalized the test and proved its efficiency properties under the proportional-hazards assumption. Landmark clinical trials from the 1970s onward — for instance, early trials of adjuvant chemotherapy in breast cancer — routinely reported Kaplan-Meier curves with log-rank p-values, cementing the pair as the default reporting standard, later reinforced by the CONSORT reporting guidelines for randomized trials.

Example 1 — Building a Kaplan-Meier curve by hand (asymmetric, realistic censoring pattern). Twelve patients with a rare sarcoma are followed after surgery. Time is in months; "+" denotes censored (alive, lost to follow-up, or study ended). Observed times: 3, 5, 5+, 8, 8, 8, 12+, 15, 15+, 20, 24+, 30 Step through the distinct event times (ignore pure-censoring times except for shrinking the risk set):

Walk the risk-set bookkeeping carefully: start at $n=12$. After $t=3$ (1 event), 11 remain. Then one subject is censored at $t=5^+$ before the event at $t=5$ — so the risk set at $t=5$ is still 11 (the censored-at-5 patient is either counted as "at risk, then censored" or excluded depending on tie-breaking convention; here we assume the death at $t=5$ is recorded as occurring at risk-set size 11, and the censoring removes one more before $t=8$). After the $t=5$ event, 10 remain, minus the co-occurring censoring, giving 9 at risk for $t=8$. Three events occur simultaneously at $t=8$, dropping to 6. Two censored subjects (one at 12, one at 15+, occurring after the $t=15$ death by convention) leave before $t=15$'s risk set is fixed at 6; one event occurs, leaving 5, minus the 15+ censoring gives 4 for $t=20$. One event at $t=20$ leaves 3, minus one censored at 24 gives... this table already reflects the resolved bookkeeping. Reading the final curve: median survival is the first time at which $\hat S(t)$ drops to 0.5 or below — that occurs at $t = 8$ months (curve falls from 0.833 to 0.556, crossing 0.5). So the estimated median survival is 8 months, even though only 5 of 12 patients have actually died by then — the other 7 are censored at various later times, and their partial information (survived at least that long) is fully used in computing every step up to their censoring time.

Example 2 — Log-rank comparison, asymmetric group sizes. A conservation biologist compares nest survival for a shorebird in two habitat types: restored wetland (n = 40 nests) vs. degraded wetland (n = 15 nests) — deliberately unequal group sizes, typical of field studies where one habitat is rarer. At one particular observed failure day, suppose:

• Total at risk across both groups: $n = 30$ (20 restored, 10 degraded)
• Total failures (nest predation) that day: $d = 4$
• Failures in the restored group that day: $O_1 = 1$

Under $H_0$ (equal hazard in both habitats), the expected failures in the restored group at this time point is proportional to its share of the risk set:

So restored wetland had fewer observed failures (1) than expected under the null (2.67) at this time point — evidence, at this single event time, that restored habitat is protective. The log-rank test sums $(O_1 - E_1)$ across every failure day in the whole study (not just this one), squares the total, and divides by the summed variance

to build the overall $\chi^2$ statistic. If, after summing over all failure days, $\chi^2 = 6.2$ with 1 df, this yields $p \approx 0.013$ — evidence that nest survival differs meaningfully between habitats, even though the degraded-habitat sample was much smaller. Note that unequal group sizes are perfectly fine for the log-rank test; it automatically weights each time point by the actual composition of the risk set, which is exactly why the $E_1$ calculation uses $n_1/n$ rather than assuming 50/50 allocation.

Paul Meier, Kaplan's coauthor, came from a very different direction: as a biostatistician working on clinical trial methodology at Johns Hopkins and later the University of Chicago, he was deeply concerned with rigorous inference in medicine, and was also a vocal early advocate for randomization in clinical trials at a time when many physicians resisted it on the grounds that it seemed unethical to leave treatment to chance. Meier's insistence that even censored, incomplete clinical data be handled with full statistical rigor — rather than approximate life-table shortcuts — reflects the same broader push toward quantitative, evidence-based medicine that also drove the adoption of randomized controlled trials.

Oncology and clinical trials. Nearly every phase II/III cancer trial reports Kaplan-Meier curves for overall survival and progression-free survival, with a log-rank p-value and often a Cox hazard ratio; regulatory agencies (FDA, EMA) require this as standard evidence for drug approval.

Vaccine and infectious disease studies. Time-to-infection analyses (e.g., COVID-19 vaccine efficacy trials) use Kaplan-Meier curves comparing vaccinated vs. placebo arms, with censoring for participants who never became infected during the trial window.

Organ transplantation and surgery. Graft survival and post-surgical outcomes (e.g., 5-year and 10-year "survival" of a transplanted kidney) are reported as Kaplan-Meier curves, guiding both clinical decisions and organ-allocation policy.

Reliability engineering. The same mathematics, historically the origin of Kaplan's own motivating dataset, estimates time-to-failure distributions for mechanical or electronic components under censored testing (some units haven't failed when testing stops).

Conservation biology and ecology. Nest survival, territory tenure, and time-to-extinction analyses in wildlife populations use Kaplan-Meier-type and related survival models (e.g., the Mayfield method, a close relative developed independently in ornithology) to handle animals that are still alive or unobserved when a field season ends.

Public health and demography. Life tables — the actuarial ancestor of Kaplan-Meier — remain central to insurance underwriting, and modern survival methods extend them to more complex censoring and competing-risks settings (e.g., estimating cause-specific mortality when death can occur from multiple competing causes).

Ask the tutor: How would the Kaplan-Meier analysis change if subjects could experience two different competing events, like death from cancer versus death from an unrelated accident?