Nelson–Aalen Estimator

Nelson–Aalen estimator is a non-parametric estimator of the cumulative hazard function derived from right-censored, left-truncated, or recurrent time-to-event data in survival analysis, biostatistics, and reliability engineering. Introduced independently by Wayne Nelson in 1969 and 1972 for industrial life-testing and generalized by Odd Aalen in 1975 and 1978 using martingale and counting-process theory, the estimator accumulates the observed empirical hazard ratios at each distinct event time. While the Kaplan–Meier estimator directly estimates the survival probability curve, the Nelson–Aalen estimator targets the cumulative hazard H(t), offering superior small-sample statistical performance and extending naturally to recurrent clinical events, competing risks analysis, and Markov multi-state transition models.

Discrete Step-Function Definition

Consider a cohort of individuals followed over time where failure or disease onset occurs at ordered distinct event times t₁ < t₂ < ··· < t_k. At each event time t_i, let d_i denote the number of individuals experiencing the event of interest, and let n_i denote the number of individuals remaining at risk (alive, uncensored, and event-free) just prior to time t_i. The cumulative hazard function H(t) = ∫₀^t h(u) du is estimated by the right-continuous step function Ĥ_NA(t):

Ĥ_NA(t) = ∑_{ti ≤ t} (d_i / n_i)

At each observed event time t_i, the ratio d_i / n_i represents the empirical conditional probability of experiencing the event at time t_i given survival up to t_i⁻. Between distinct event times, Ĥ_NA(t) remains constant, jumping upward by d_i / n_i at each t_i. Visually, the local slope of the Nelson–Aalen curve over any time window reflects the instantaneous hazard rate h(t): a linear trajectory indicates a constant hazard (exponential survival), a concave trajectory indicates a decreasing hazard over time, and a convex trajectory indicates an accelerating hazard.

Martingale Variance and Confidence Intervals

Under Aalen's multiplicative intensity model, the counting process N(t) recording cumulative events up to time t decomposes into a predictable compensator ∫₀^t Y(u)h(u) du (where Y(u) is the size of the risk set at time u) plus a zero-mean local square-integrable martingale M(t). Using stochastic integration with respect to this martingale, Aalen derived the unbiased asymptotic variance estimator for Ĥ_NA(t):

Var_Aalen(Ĥ_NA(t)) = ∑_{ti ≤ t} (d_i / n_i²)

When multiple tied events occur at the same discrete recording time (d_i > 1), an alternative hypergeometric variance adjustment analogous to Greenwood's formula is frequently employed:

Var_{Greenwood-type}(Ĥ_NA(t)) = ∑_{ti ≤ t} [d_i (n_i − d_i) / (n_i² (n_i − 1))]

To prevent confidence limits from falling below zero at early follow-up times, pointwise (1 − α) confidence intervals are constructed on the logarithmic scale and exponentiated back to the hazard scale:

CI_1−α(H(t)) = Ĥ_NA(t) · exp(± z_1−α/2 · [SE(Ĥ_NA(t)) / Ĥ_NA(t)])

The Fleming–Harrington Survival Estimator

In continuous time, the survival function S(t) and the cumulative hazard function H(t) are linked by the identity S(t) = exp(−H(t)). Plugging the Nelson–Aalen cumulative hazard directly into this identity yields the Fleming–Harrington (or Breslow) survival estimator:

Ŝ_FH(t) = exp(−Ĥ_NA(t)) = ∏_{ti ≤ t} exp(−d_i / n_i)

By Taylor series expansion, exp(−x) = 1 − x + x²/2 − ··· ≈ 1 − x for small increments x = d_i / n_i. Consequently, when the risk set n_i is large relative to d_i, the Fleming–Harrington survival curve and the Kaplan–Meier estimator Ŝ_KM(t) = ∏(1 − d_i / n_i) are asymptotically equivalent. Because exp(−x) > 1 − x for all x > 0, Ŝ_FH(t) is always slightly larger than Ŝ_KM(t), and unlike the Kaplan–Meier curve, it never reaches zero even if the final remaining participant in the risk set experiences an event.

Baseline Hazard in Cox Proportional Hazards Models

In the semi-parametric Cox proportional hazards model, the partial likelihood estimates the regression coefficients β without specifying the baseline hazard h₀(t). Once the parameter vector β̂ is obtained, the Breslow estimator generalizes the Nelson–Aalen formula to recover the cumulative baseline hazard Ĥ₀(t) by weighting each individual in the risk set R(t_i) by their covariate risk score:

Ĥ₀(t) = ∑_{ti ≤ t} [d_i / ∑_{j ∈ R(ti)} exp(X_j^Tβ̂)]

When all covariate coefficients equal zero (β̂ = 0), the denominator reduces to n_i, and the Breslow estimator collapses identically to the Nelson–Aalen estimator.

Left Truncation, Delayed Entry, and Multi-State Transitions

Because the risk set n_i in the Nelson–Aalen formula is evaluated dynamically at each event time t_i, individuals do not need to be under observation at time zero. In observational cohort studies using chronological age as the time scale, participants enter the risk set at their age at enrollment (left truncation or delayed entry) and exit at their age at event or censoring.

Furthermore, in multi-state models with state space {1, 2, ..., S}, each directed transition from state g to state h possesses its own transition intensity α_gh(t). The matrix of component-wise Nelson–Aalen cumulative intensity estimators Â(t) drives the non-parametric Aalen–Johansen product-integral estimator of the state transition probability matrix P̂(s, t), which generalizes survival estimation to illness-death models, nosocomial infection tracking, and competing risks analysis.

Recurrent Clinical Events and Mean Cumulative Function

Many chronic conditions—such as asthma exacerbations, epileptic seizures, bladder tumor recurrences, or repeated malaria episodes—involve recurrent events within the same individual. When individuals remain in the risk set after experiencing an event, the Nelson–Aalen formula estimates the Mean Cumulative Function (MCF), Λ̂(t), which quantifies the expected cumulative number of recurrent events experienced per subject up to time t.

Why use the Nelson–Aalen estimator instead of the Kaplan–Meier estimator?

While the Kaplan–Meier estimator is preferred for displaying probability of survival on a 0-to-1 scale, the Nelson–Aalen estimator directly estimates cumulative hazard, which is additive across independent competing causes and directly reflects changes in the underlying hazard rate via its slope. It also exhibits smaller mean squared error in small risk sets and naturally handles recurrent events where a survival probability is undefined.

How do you interpret the slope of a Nelson–Aalen cumulative hazard plot?

Because the cumulative hazard H(t) is the integral of the instantaneous hazard rate h(t), the derivative (slope) of the Nelson–Aalen curve represents the current risk rate. A straight line indicates a constant hazard rate over time, an upward-curving (convex) slope indicates increasing risk with age or duration, and a flattening (concave) slope indicates diminishing risk among survivors.

How are Instantaneous Hazard curves derived from the Nelson–Aalen step function?

Because the raw step function Ĥ_NA(t) is non-differentiable at its jump points and flat elsewhere, biostatisticians convolve the discrete jumps d_i / n_i with a smooth symmetric kernel function (such as an Epanechnikov or Gaussian kernel) over a smoothing bandwidth b to produce a continuous estimate of the instantaneous hazard rate ĥ(t).

1. Nelson, W. (1972). Theory and applications of hazard plotting for censored failure data. Technometrics, 14(4), 945–966. https://doi.org/10.1080/00401706.1972.10488991
2. Aalen, O. (1978). Nonparametric inference for a family of counting processes. The Annals of Statistics, 6(4), 701–726. https://doi.org/10.1214/aos/1176344247
3. Andersen, P. K., Borgan, Ø., Gill, R. D., & Keiding, N. (1993). Statistical Models Based on Counting Processes. Springer. https://doi.org/10.1007/978-1-4612-4348-9
4. Fleming, T. R., & Harrington, D. P. (1991). Counting Processes and Survival Analysis. John Wiley & Sons. https://doi.org/10.1002/9781118150672