Mathematics
Survival Analysis and Kaplan-Meier Curves
Quick fact
The Kaplan-Meier curve gracefully handles the missing information from censored patients by assuming that censored patients have the same future survival prospects as those still under observation—a key assumption that shapes all survival estimates.
Why this is interesting
Imagine a clinical trial where some patients are still alive at the end of the study—their survival times are incomplete. How can we still estimate survival probabilities accurately when we don't know when those patients will eventually die?
Read the full explanation
Understanding Survival Analysis and Kaplan-Meier Curves
Survival analysis is a set of statistical methods for analyzing data where the outcome is the time until an event occurs. The event could be death, disease recurrence, machine failure, or even a customer churning. Often, we cannot observe the event for all subjects because the study ends, or a subject drops out. This is called censoring. If we simply ignore censored patients, we would underestimate survival. The Kaplan-Meier estimator provides a way to estimate the survival probability over time, step by step, using the information from all subjects, including those censored. The idea is to break the study period into intervals defined by the times when events actually occur. At each event time, we calculate the proportion of subjects at risk who experienced the event. Then we multiply these proportions cumulatively to get the survival probability. The result is a step function that drops at each event time. Visualize it: at time zero, everyone is alive, so survival is 1. As events happen, the curve steps down. Censored subjects are carried along until they drop out, contributing to the 'at risk' population until that point. The curve is easy to read: the height at any time gives the probability of surviving beyond that time.
A deeper explanation
The Kaplan-Meier estimator is a nonparametric maximum likelihood estimate of the survival function S(t) = P(T t). It works by considering each event time ti. At that moment, let ni be the number of subjects still at risk (not yet event or censored), and di be the number who experience the event. The probability of surviving beyond ti is the product of (1 - di/ni) for all event times up to ti. This formula arises from the product of conditional probabilities: the chance of surviving past ti given that you survived just before ti. Censoring is handled by removing the censored subject from the risk set at the time of censoring, without counting them as an event. This is valid under the assumption that censoring is independent of the event time (i.e., censoring does not provide information about the future survival of the subject). The Kaplan-Meier curve is a step function that remains flat between event times, dropping only at observed events. The variance of the estimator can be calculated using Greenwood's formula, allowing confidence intervals. The method is widely used in medical research to compare survival between treatment groups, often using the log-rank test for formal comparison. It also appears in reliability engineering to estimate the probability that a component survives a certain duration. While Kaplan-Meier is descriptive and nonparametric, it assumes that censoring is non-informative and that the risk of the event is the same for censored and uncensored subjects. For analyzing the impact of multiple covariates, proportional hazards models (like Cox regression) extend this framework.