Greenwood variance
Greenwood's formula estimates the sampling variance of a Kaplan-Meier survival estimate from event counts and risk-set sizes.
For ordinary unweighted, independent-unit, right-censored first-event data, the estimate is S_hat(t)^2 times the sum of d_i / (n_i * (n_i - d_i)) over event times through t. Here d_i counts events and n_i counts units at risk just beforehand; the displayed expression requires n_i > d_i. [1]
Its square root is a standard error of estimated survival, not a device lifetime spread. A confidence interval also needs a confidence level and construction: a log-log transformation can keep interior limits within the probability range, but its coverage remains approximate. [2]
These are pointwise intervals for a fixed time, not simultaneous coverage of a complete survival curve. S_hat equal to zero or one requires explicit boundary handling; undefined limits or a degenerate interval are not evidence of certainty. [2] [3]
An error bar does not correct informative censoring or justify counting repeated telemetry rows as independent devices. Clustered, weighted and repeated-event analyses need a variance method matching their sampling structure; the R survival documentation distinguishes these from the ordinary calculation. [4]
Relations
Citations
- Kaplan and Meier (1958): Nonparametric Estimation from Incomplete Observations, sections 1.4 and 2.2, equation (2f)
- Sawyer (2003): The Greenwood and Exponential Greenwood Confidence Intervals in Survival Analysis, equations 1.2-1.3
- SciPy 1.16.2: EmpiricalDistributionFunction.confidence_interval, clipping and undefined limits
- R survival package: Methods, sections 2.2.1-2.2.2, confidence intervals and robust variance