Greenwood's Formula

Greenwood's formula is the standard biostatistical estimator used to compute the asymptotic variance and standard error of non-parametric survival probability curves, including actuarial life tables and the Kaplan–Meier estimator. Derived in 1926 by British physician and epidemiologist Major Greenwood in a British Ministry of Health report evaluating the "natural duration" of untreated cancer, the formula precedes the 1958 Kaplan–Meier product-limit estimator by more than three decades. By applying the first-order delta method to a product of conditional binomial survival probabilities, Greenwood's formula quantifies how estimation uncertainty accumulates across sequential failure times in right-censored epidemiological cohorts.

Product-Limit Variance Expression

Consider a survival cohort with ordered observed event times t₁ < t₂ < ··· < t_k. At each event time t_i, let n_i represent the number of individuals at risk just before t_i, and let d_i represent the number of observed deaths or failures at t_i. The Kaplan–Meier estimator estimates the probability of surviving beyond time t as the cumulative product of conditional survival proportions p̂_i = 1 − (d_i / n_i):

Ŝ(t) = ∏_{ti ≤ t} (1 − d_i / n_i) = ∏_{ti ≤ t} p̂_i

Greenwood's formula estimates the variance of Ŝ(t) at any follow-up time t as:

Var_G(Ŝ(t)) = [Ŝ(t)]² · ∑_{ti ≤ t} [d_i / (n_i (n_i − d_i))]

Taking the square root of Var_G(Ŝ(t)) yields the Greenwood standard error SE(Ŝ(t)), which enters directly into pointwise confidence intervals for S(t), median survival time confidence bounds (via the Brookmeyer–Crowley method), and the variance of restricted mean survival time.

Step-by-Step Derivation via the Delta Method

Conditioned on the risk set size n_i at time t_i, the number of deaths d_i follows a binomial distribution Binomial(n_i, q_i), where q_i = 1 − p_i is the conditional failure probability. The conditional variance of the survival proportion p̂_i = (n_i − d_i) / n_i is therefore:

Var(p̂_i | n_i) = [p_i (1 − p_i)] / n_i ≈ [p̂_i (1 − p̂_i)] / n_i = [d_i (n_i − d_i)] / n_i³

To convert the product Ŝ(t) = ∏ p̂_i into an additive sum, take the natural logarithm:

ln Ŝ(t) = ∑_{ti ≤ t} ln(p̂_i)

Applying the first-order Taylor expansion (the delta method), where Var(ln X) ≈ [1 / E(X)²] · Var(X), the variance of each logarithmic increment is:

Var(ln p̂_i) ≈ (1 / p̂_i²) · Var(p̂_i) = [n_i / (n_i − d_i)]² · [d_i (n_i − d_i) / n_i³] = d_i / [n_i (n_i − d_i)]

Because martingale increments across distinct non-overlapping time intervals are uncorrelated, the variance of the sum equals the sum of the variances:

Var(ln Ŝ(t)) ≈ ∑_{ti ≤ t} [d_i / (n_i (n_i − d_i))]

Finally, applying the delta method a second time to invert the logarithm via Ŝ(t) = exp(ln Ŝ(t)), where Var(exp(Y)) ≈ [exp(Y)]² · Var(Y), multiplies the sum by [Ŝ(t)]² and yields Greenwood's formula. When single untied deaths occur in continuous time (d_i = 1) and n_i is large, the summation term d_i / [n_i(n_i − d_i)] closely approximates d_i / n_i², which is Aalen's martingale variance estimator for the Nelson–Aalen estimator of cumulative hazard.

Linear versus Complementary Log-Log (cloglog) Intervals

The naive linear (Wald-type) 95% confidence interval for survival at time t is Ŝ(t) ± 1.96 · SE_G(Ŝ(t)). However, near Ŝ(t) = 1 or Ŝ(t) = 0, the sampling distribution of Ŝ(t) is skewed, and linear intervals frequently produce impossible survival probabilities above 1.0 or below 0.0.

To ensure confidence bounds always respect the probability range (0, 1) and accelerate convergence to asymptotic normality, John D. Kalbfleisch and Ross L. Prentice (1980) introduced the complementary log-log (cloglog) transformation, g(S) = ln(−ln S(t)) = ln H(t). By the delta method, the standard error on the cloglog scale is:

SE(ln(−ln Ŝ(t))) = √[∑_{ti ≤ t} (d_i / [n_i (n_i − d_i)])] / |ln Ŝ(t)|

If θ̂ = SE(ln(−ln Ŝ(t))), the back-transformed (1 − α) confidence interval on the survival probability scale is:

CI_1−α(S(t)) = [Ŝ(t)^{exp(+z_1−α/2 · θ̂)} , Ŝ(t)^{exp(−z_1−α/2 · θ̂)}]

Tail Instability and Peto's Effective Sample Size Adjustment

Between distinct event times t_i and t_i+1, even if numerous participants are right-censored and the risk set n(t) shrinks substantially, Greenwood's variance estimate Var_G(Ŝ(t)) remains completely flat because it updates only when a new death d_i > 0 occurs. Consequently, in the far right tail of a survival curve where few patients remain at risk, Greenwood's formula can underestimate true sampling variability. To guard against overconfident tail intervals, Richard Peto and colleagues (1977) proposed the conservative effective-sample-size variance approximation Var_Peto(Ŝ(t)) = [Ŝ(t)]² [1 − Ŝ(t)] / n(t), which widens the confidence band dynamically as censoring reduces the current risk set n(t). In generalized competing risks analysis, Greenwood's variance extends via the Aalen–Johansen covariance matrix to provide standard errors for cumulative incidence functions.

How could Greenwood's formula be published in 1926 if the Kaplan–Meier estimator was invented in 1958?

Major Greenwood derived his variance formula in 1926 for actuarial life tables divided into fixed annual intervals. When Edward L. Kaplan and Paul Meier took the continuous-time limit of the life-table estimator in 1958 by shrinking the intervals around each individual death time, Greenwood's discrete product variance formula carried over directly without structural modification.

Why are complementary log-log confidence intervals preferred over linear Greenwood intervals?

Linear confidence intervals of the form Ŝ(t) ± 1.96 · SE can exceed 100 percent survival at early follow-up times or drop below 0 percent survival at late follow-up times. The complementary log-log transformation maps the bounded interval (0, 1) onto the entire real line (−∞, +∞) before applying the normal approximation, guaranteeing back-transformed confidence limits that stay inside (0, 1).

What happens to Greenwood's formula if the last remaining person in the cohort dies?

If the final individual at risk experiences an event (d_k = n_k), then both Ŝ(t_k) = 0 and the denominator n_k − d_k = 0, making the standard Greenwood fraction 0 / 0 indeterminate. Statistical software packages either carry forward the variance from the immediately preceding time point or apply the Nelson–Aalen estimator variance d_i / n_i², which remains finite when d_k = n_k.

1. Greenwood, M. (1926). A Report on the Natural Duration of Cancer (Reports on Public Health and Medical Subjects, No. 33, pp. 1–26). His Majesty's Stationery Office, London. https://wellcomecollection.org/works/w69pqcc2
2. Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. https://doi.org/10.1080/01621459.1958.10501452
3. Kalbfleisch, J. D., & Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data (2nd ed.). John Wiley & Sons. https://doi.org/10.1002/9781118032985
4. Sawyer, S. (2003). The Greenwood and Exponential Greenwood Confidence Intervals in Survival Analysis. Department of Mathematics, Washington University in St. Louis. https://www.math.wustl.edu/~sawyer/handouts/greenwood.pdf