Greenwood Standard Errors and Survival Intervals

Why quantify uncertainty in a Kaplan–Meier estimate?

A Kaplan–Meier curve gives an estimated survival probability over time, but an estimate alone does not show how precisely it has been determined. Uncertainty is especially important late in follow-up, when fewer individuals may remain at risk. Greenwood’s formula provides a standard estimate of the variance of the Kaplan–Meier estimator, which can then be used to construct pointwise confidence intervals.

These intervals describe uncertainty in survival at a particular time. They are not, by themselves, a simultaneous confidence band for the entire curve, and they do not compare two survival groups. The focus here is the uncertainty calculation for one Kaplan–Meier estimate at a time.

The event count \(d_j\) includes all events tied at \(t_j\); the risk set \(n_j\) is counted immediately before that time. Censored individuals contribute to the risk set up to their censoring time, but do not count as events. The introductory material on the Kaplan–Meier estimator establishes this product-limit construction.

Greenwood’s formula and its justification

Greenwood’s expression is a variance estimate, not an additional factor in the Kaplan–Meier product. To see its structure, write the survival factor at an event time as \(\hat q_j=1-d_j/n_j\). Under a binomial approximation for the event count within a risk set, the estimated variance of this factor is approximately \(d_j(n_j-d_j)/n_j^3\). Dividing by \(\hat q_j^2=\{(n_j-d_j)/n_j\}^2\) gives the relative variance

For a product of estimated factors, the first-order variance approximation adds their relative variances. Multiplying that sum by the squared product \(\hat S(t)^2\) gives Greenwood’s formula. This justification relies on the usual large-sample approximation for the product-limit estimator; it does not make a small-sample interval exact.

The standard setting is independent individuals with right-censored event times and censoring that is independent of the event time, or independent conditional on relevant groups or covariates. The risk sets must correctly reflect who is under observation and at risk; with delayed entry, they must include only individuals who have entered by the event time. The formula can accommodate tied event times by using the total \(d_j\) at each distinct time. It does not correct bias from informative censoring or an incorrectly constructed risk set.

Pointwise confidence intervals

A pointwise interval concerns \(S(t)\) at one fixed time \(t\). The simplest construction uses a normal approximation for \(\hat S(t)\):

For a 95% interval, \(z_{1-\alpha/2}\) is approximately 1.96. The Wald interval is easy to calculate, but it is formed on the probability scale. It can extend below 0 or above 1, even though survival probabilities must lie between 0 and 1. It can also behave poorly near either boundary or when the normal approximation is weak.

A common alternative is to form an interval on the log-minus-log scale. For \(0<\hat S(t)<1\), define \(g(S)=\log\{-\log(S)\}\). Its derivative is \(g'(S)=1/\{S\log(S)\}\). Applying the delta method to the Greenwood variance gives

Construct a normal interval for \(g(S(t))\), then transform the endpoints back. Since \(S=\exp\{-\exp(g)\}\) decreases as \(g\) increases, the endpoint order reverses:

This transformation keeps the resulting limits between 0 and 1 and often gives a more plausible interval when the estimated survival probability is near a boundary. It does not make the interval exact; it still relies on an approximate normal distribution on the transformed scale.

Reading the uncertainty and spotting problems

Greenwood’s sum is cumulative over event times. A large value increases the estimated standard error relative to \(\hat S(t)\), but an individual contribution \(d_j/\{n_j(n_j-d_j)\}\) depends on both \(n_j\) and \(d_j\). Holding \(d_j\) fixed, the contribution increases as \(n_j\) decreases over its valid range. But a smaller risk set does not guarantee a larger contribution if the event count changes too. For example, \(n=10,d=5\) contributes \(5/\{10(5)\}=0.1000\), whereas \(n=5,d=1\) contributes \(1/\{5(4)\}=0.05000\).

Useful checks include verifying the risk sets and event counts, confirming that censoring is not being counted as an event, and comparing the width of the interval with the number at risk at the time of interest. A sharp change in the curve with a small risk set can leave late survival estimates imprecise. Also check whether the chosen interval method is producing limits in the allowable range. When \(\hat S(t)=1\), the log-minus-log transformation is undefined; when \(\hat S(t)=0\), it is also undefined, and Greenwood’s expression may encounter a zero denominator if \(d_j=n_j\). These boundary cases require special care rather than mechanical use of the displayed formulas.

Choosing an interval and understanding its limits

The Wald interval and the log-minus-log interval use the same Greenwood standard error but handle the probability scale differently. The Wald interval is transparent and may be adequate when the estimate is well away from 0 and 1 and the approximation is reasonable. The transformed interval respects the range of survival probabilities and is often preferred near the boundaries. Neither repairs sparse data, informative censoring, or a misspecified risk set. Resampling methods can be another option for assessing uncertainty, but they also require a suitable resampling scheme and enough information in the observed data.

Both types of interval here are pointwise: their nominal coverage applies separately at a fixed time under the approximation and assumptions, not to the whole survival curve simultaneously. A collection of pointwise 95% intervals is not automatically a 95% band for every time on the curve. Report the time point, interval method, and units of time, and interpret the interval in terms of the survival probability for the population represented by the data.