Uncertainty For Kaplan-Meier Curves

Why a Kaplan–Meier Curve Needs Uncertainty Bands

A Kaplan–Meier curve is an estimate of the survival function, not a known population curve. Its steps depend on how many subjects remain at risk and how many events occur at each observed event time. In samples with few subjects at risk, the estimated survival probability can be especially uncertain. Reporting only the curve can therefore give a misleading impression of precision.

This tutorial develops pointwise uncertainty for the Kaplan–Meier estimator. The central calculation is Greenwood’s formula, which estimates the variance at a selected time. A normal interval can then be formed directly on the survival-probability scale. A second approach first transforms the estimate using the log-log function; this often behaves better near zero or one because the resulting interval, after inversion, stays within the probability range.

Reading Greenwood’s Formula

Each summand in Greenwood’s formula reflects the information at one event time. A larger number of events relative to the risk set generally adds more uncertainty. The terms accumulate across event times, while multiplication by \(\hat S(t)^2\) puts the result on the variance scale for the survival estimate itself. Censoring affects these calculations by changing who remains in each later risk set; a censored subject contributes to the risk set until censoring, but not as an event.

The usual large-sample justification assumes independent subjects and noninformative, independent right censoring: censoring must not selectively remove subjects according to their unobserved event times, at least under the conditioning structure of the analysis. Under these conditions, the Kaplan–Meier estimator has an approximately normal sampling distribution at a fixed time when the estimate is away from problematic boundaries and there is adequate information. Greenwood’s formula estimates its variance. Neither the formula nor the normal approximation makes the interval exact in a small sample.

Pointwise Normal Intervals

A pointwise confidence interval concerns the survival probability at one specified time \(t\). It does not, by itself, provide a confidence band that covers the entire curve simultaneously. A simple large-sample interval uses the approximate normal distribution of \(\hat S(t)\):

Here \(1-\alpha\) is the confidence level and \(z_{1-\alpha/2}\) is the corresponding standard normal quantile; for a 95% interval, it is approximately 1.96. This construction treats the estimate as approximately normal on the probability scale. Consequently, the lower endpoint can be below zero or the upper endpoint above one, particularly when the estimate is close to a boundary or its standard error is large. Truncating such endpoints to zero or one does not change the underlying approximation and does not generally solve its coverage problem.

Log-Log Confidence Intervals

A common alternative constructs an interval for a transformation of the survival estimate. Define \(g(S)=\log[-\log(S)]\) for \(0<S<1\). Its inverse is \(g^{-1}(u)=\exp[-\exp(u)]\), which is decreasing. A normal interval for \(g\{\hat S(t)\}\) can therefore be transformed back to the survival scale, reversing the endpoint order.

The standard error on the transformed scale follows from the delta method. Differentiating gives \(g'(S)=1/[S\log(S)]\). Multiplying the squared derivative by Greenwood’s variance estimate gives

The absolute value gives a positive standard error. If \(L_g\) and \(U_g\) are the lower and upper normal limits on the transformed scale, then \(L_g<U_g\), but the inverse is decreasing. The corresponding survival interval is therefore \([g^{-1}(U_g),g^{-1}(L_g)]\). Because \(g^{-1}(u)\) is strictly between zero and one for any finite \(u\), finite transformed limits yield survival limits inside that range.

Why the Transformation Helps Near a Boundary

On the original scale, a symmetric normal interval adds and subtracts the same amount from \(\hat S(t)\). Near one, this can produce an upper limit above one; near zero, it can produce a lower limit below zero. A log-log interval instead makes its normal approximation on a transformed scale and maps the limits back through a function whose range is \((0,1)\). The resulting interval is bounded and typically asymmetric.

Bounded endpoints are a practical advantage, not a guarantee of accurate coverage. If only a few subjects remain at risk, the large-sample approximation can still be poor. At \(\hat S(t)=1\), \(\log[-\log\hat S(t)]\) is not finite, and at \(\hat S(t)=0\) it is also undefined. Thus the ordinary log-log construction cannot be applied at these boundary estimates; inference there requires particular caution rather than blind use of the formula.

Common Pitfalls and Interpretation

• Using the Greenwood sum as the variance. The sum \(G(t)\) is not itself the variance of \(\hat S(t)\). Greenwood’s formula multiplies it by \(\hat S(t)^2\).
• Forgetting that the inverse transformation decreases. The transformed lower limit maps to the larger survival endpoint. Reverse the order after applying \(\exp[-\exp(u)]\).
• Treating a pointwise interval as a curve-wide guarantee. A 95% interval at one fixed time does not imply 95% coverage of the entire estimated curve at once.
• Calling the interval a probability statement about the parameter. The confidence level describes the repeated-sampling performance of the procedure under its assumptions; it is not the probability that the fixed survival probability lies in the observed interval.
• Assuming bounded limits ensure good inference. Log-log limits stay within \((0,1)\) when finite, but a small risk set or informative censoring can undermine the approximation or the estimator’s interpretation.