Deriving The Nelson-Aalen Increment

From a transition rate to an observed increment

A transition intensity is a rate: it describes how quickly transitions from state $j$ to state $k$ occur among individuals currently in state $j$. The data, however, arrive as event counts at particular times. The Nelson–Aalen increment connects these two scales by dividing the observed transition count by the number at risk just before the time in question. Understanding that division clarifies both what the estimator measures and what assumptions its interpretation requires.

Write $N_{jk}(t)$ for the number of observed $j$-to-$k$ transitions by time $t$, and $Y_j(t)$ for the number of individuals in state $j$ and under observation just before time $t$. Thus $dN_{jk}(t)$ is the transition count at $t$, and $Y_j(t)$ is the corresponding transition-specific risk set. Let $\lambda_{jk}(t)$ be the transition intensity for an individual in state $j$, and define its cumulative intensity by $A_{jk}(t)=\int_0^t\lambda_{jk}(u)\,du$. The intensity has units of inverse time; the cumulative intensity is dimensionless.

Definition: At a time $t$ with $Y_j(t)>0$, the Nelson–Aalen increment for the $j$-to-$k$ transition is $dN_{jk}(t)/Y_j(t)$. The corresponding cumulative estimator over an interval from $0$ to $t$ is the sum, or counting-process integral, of these increments: $$ \widehat A_{jk}(t)=\int_0^t\frac{dN_{jk}(u)}{Y_j(u)}. $$ When $Y_j(u)=0$, this integrand is not defined without an explicit convention. The full cumulative intensity is the target only when the risk set is positive throughout the interval, apart from sets of time of measure zero.

The denominator is not the total sample size. It is the number currently contributing exposure to this particular transition intensity. If one transition from $j$ to $k$ occurs while 10 individuals are at risk, the observed increment is $1/10$. If two transitions are counted at a time when 10 are at risk, the count-based increment is $2/10$. The increment is an observed change in the cumulative estimate, not itself an instantaneous rate.

Why the increment has this form

The intensity model gives the local interpretation. Conditional on the observed history just before $t$, a short interval of length $dt$ contributes approximately $Y_j(t)\lambda_{jk}(t)\,dt$ expected transitions. Dividing the observed count in that interval by $Y_j(t)$ therefore puts the count on a per-at-risk-individual scale. In the continuous-time limit, the accumulated per-risk-set counts correspond to integrating the individual transition rate over time.

That local-rate argument motivates the formula, but a counting-process decomposition makes the justification precise. Under the intensity model, the compensator of $N_{jk}$ is $\int_0^tY_j(u)\lambda_{jk}(u)\,du$. Define the compensated process $M_{jk}(t)=N_{jk}(t)-\int_0^tY_j(u)\lambda_{jk}(u)\,du$. In general, $M_{jk}$ is a local martingale, not necessarily a true martingale. A sufficient condition for it to be a true martingale through time $t$ is $E\int_0^tY_j(u)\lambda_{jk}(u)\,du<\infty$.

Assume $Y_j(u)>0$ throughout the interval of interest, up to sets of measure zero. Dividing the compensated-process decomposition by the risk set inside the stochastic integral gives:

\begin{aligned} \widehat A_{jk}(t) &=\int_0^t\frac{dN_{jk}(u)}{Y_j(u)}\\ &=\int_0^t\lambda_{jk}(u)\,du +\int_0^t\frac{1}{Y_j(u)}\,dM_{jk}(u)\\ &=A_{jk}(t)+\int_0^t\frac{1}{Y_j(u)}\,dM_{jk}(u). \end{aligned}

The second term is the estimation error represented by a risk-set-weighted martingale integral. It is a local martingale under the usual predictable-integrand conditions; it is a true martingale if an appropriate integrability condition holds. For a simple counting process with predictable intensity $Y_j\lambda_{jk}$, one sufficient condition is $E\int_0^t\mathbf{1}\{Y_j(u)>0\}\lambda_{jk}(u)/Y_j(u)\,du<\infty$. The decomposition does not say that the estimator equals the target in a particular sample. It says that the estimator is the cumulative intensity plus a stochastic error term.

If the risk set becomes zero during part of the interval, define the estimator's integrand to be zero there only if that convention is intended. The resulting compensator term is then $\int_0^t\mathbf{1}\{Y_j(u)>0\}\lambda_{jk}(u)\,du$, not necessarily the full $A_{jk}(t)$. The data provide no at-risk exposure for estimating the transition intensity during times when nobody is in state $j$ and under observation.

Worked Example: Summing Three Event-Time Increments

In a hypothetical study, suppose three distinct observed $j$-to-$k$ transitions occur at times $t_1<t_2<t_3$. The risk sets just before those events are 12, 9, and 6, respectively, and one such transition occurs at each time. Calculate the cumulative Nelson–Aalen estimate through $t_3$.

At each event time, $dN_{jk}(t_r)=1$, so the increments are $1/12$, $1/9$, and $1/6$. There are no contributions at times with no observed transition, since $dN_{jk}=0$ there.

\begin{aligned} \widehat A_{jk}(t_3) &=\frac{1}{12}+\frac{1}{9}+\frac{1}{6}\\ &=\frac{3+4+6}{36}\\ &=\frac{13}{36}\approx 0.3611. \end{aligned}

The estimate is 0.3611, a dimensionless estimate of cumulative transition intensity through $t_3$. It is not a probability that an individual has made the transition: cumulative intensity and transition probability are different quantities.

Interpreting and checking the calculation

Each increment is a count scaled by the available risk set. A larger risk set makes one event a smaller contribution; a smaller risk set makes it larger. This is why replacing $Y_j(t)$ with the original sample size generally changes the estimator: individuals who are no longer at risk do not contribute exposure at later times.

Worked Example: How the Risk Set Changes an Increment

Consider two hypothetical, distinct event times. At the first, one $j$-to-$k$ transition occurs among 20 individuals at risk. At the second, one such transition occurs among 5 individuals at risk. Calculate both increments and their combined contribution.

The first increment is $1/20=0.05$. The second is $1/5=0.20$. Thus these two events contribute $0.05+0.20=0.25$ to the cumulative estimate. The later event contributes four times as much because its risk set is one quarter as large. This scaling reflects the per-risk-set construction; it does not establish that the underlying intensity was four times larger at the second time.

A useful diagnostic is to check the ingredients at every event time: the count must refer to the $j$-to-$k$ transition, and the risk set must be the number eligible for that transition immediately before the event. Confirm that the denominator is positive wherever an increment is calculated. A zero denominator with a positive event count signals inconsistent data or an incorrectly defined risk set; a zero denominator with no event does not supply information about the intensity at that time.

Also check that the cumulative estimate is nondecreasing: counts and risk sets are nonnegative, so every defined increment is nonnegative. A negative increment indicates a calculation or data-processing error. The estimate can nevertheless be unstable when the risk set is small, because a single event then adds a relatively large amount. The decomposition identifies the stochastic error but does not, by itself, quantify its uncertainty.

What the derivation does—and does not—establish

The Nelson–Aalen construction is a nonparametric estimator of cumulative transition intensity under the specified transition-intensity model. Its event-time increments use the observed count and the contemporaneous risk set; summing them is the discrete counting-process counterpart of integrating an intensity over time. The martingale decomposition explains this correspondence and makes clear which cumulative intensity is being targeted.

The argument depends on a meaningful risk set and a valid intensity specification relative to the observed history. It does not guarantee that the intensity is constant over time, that the risk set is large enough for stable estimation, or that an estimate from a finite sample is close to its target. Nor does the cumulative-intensity estimate alone give transition probabilities. Those quantities require additional relationships among transition intensities.

This construction is distinct from estimating a rate by dividing a total count by the original cohort size and follow-up duration: that crude ratio does not update the denominator as individuals enter or leave the risk set. It is also distinct from a smoothed estimate of the instantaneous intensity. The Nelson–Aalen estimator accumulates observed increments without smoothing them over time.

Common pitfall: “One event occurred, so the increment is $1/n$, where $n$ is the original sample size.” This is wrong unless the original sample size happens to equal the relevant risk set at that event time. Use $dN_{jk}(t)/Y_j(t)$, with $Y_j(t)$ measured just before $t$. A second pitfall is calling the resulting sum a transition probability; it estimates cumulative intensity, which is not generally a probability.

The key reasoning is therefore: the intensity is a rate per individual at risk; the observed event count must be scaled by the risk set to put it on that scale; and summing the scaled counts estimates the accumulated rate, subject to the positivity and model conditions stated above.

Check Your Understanding

Answer each question, then compare your work with the solutions.

At two distinct event times, one transition occurs at each time. The risk sets are 12 and 8. What is their combined contribution to the cumulative estimate?
At a time when $Y_j(t)=7$, no $j$-to-$k$ transition occurs. What is the increment at that time?
Why does setting the integrand to zero when $Y_j(u)=0$ not automatically estimate the full cumulative intensity?
State the difference between the Nelson–Aalen cumulative estimate and a transition probability.

Solutions. (1) The contribution is $1/12+1/8=2/24+3/24=5/24\approx0.2083$. (2) Since $dN_{jk}(t)=0$, the increment is $0/7=0$. (3) Where the risk set is zero, the estimator has no at-risk exposure; setting its integrand to zero makes the compensator target $\int_0^t\mathbf{1}\{Y_j(u)>0\}\lambda_{jk}(u)\,du$, which can omit parts of the full cumulative intensity. (4) The cumulative estimate sums estimated intensity increments; a transition probability is a probability of occupying or entering a state and is not generally equal to cumulative intensity.