Why the Distinction Matters
A Nelson–Aalen estimate for a transition is often drawn as a staircase. Its flat sections and jumps can look like a time-varying rate, but the plotted quantity is accumulated transition intensity. The jump at an event time is an increment in that accumulated quantity; it is not an instantaneous rate. Keeping those concepts separate matters when describing how quickly transitions occur, comparing time periods, or communicating what the estimate says.
The focus here is one specified transition, from state \(h\) to state \(j\), where \(h\ne j\). The risk set at time \(t\) consists of individuals still under observation and in state \(h\) immediately before \(t\). The earlier material on transition-specific risk sets and Nelson–Aalen estimation supplies the basic counting setup; the key task here is interpreting its units and time behavior.
Informally, an intensity describes a local rate: over a short interval of length \(\Delta\), the conditional probability of the specified transition is approximately \(\alpha_{hj}(t)\Delta\), when the intensity is regular over that interval. By contrast, \(A_{hj}(t)\) accumulates such local contributions over time. The cumulative quantity does not retain a unit of time in its measurement.
This local-probability approximation presumes that the stated transition intensity is the relevant one for the individual’s history and that the interval is short. If histories matter beyond current state and time, a single pooled \(\alpha_{hj}(t)\) need not describe everyone in state \(h\); the conditioning or model must account for that heterogeneity.
The Nelson–Aalen Step Function
Let \(N_{hj}(t)\) count observed transitions from \(h\) to \(j\) by time \(t\), and let \(Y_h(t)\) be the number at risk in state \(h\) immediately before \(t\). At an observed transition time \(u\), write \(dN_{hj}(u)\) for the number of those transitions at \(u\). The Nelson–Aalen estimator is the sum of observed transition counts divided by the corresponding risk-set sizes.
Thus the increment at an observed event time \(u\) is $$ \Delta\widehat A_{hj}(u) = \widehat A_{hj}(u)-\widehat A_{hj}(u-) = \frac{dN_{hj}(u)}{Y_h(u)}. $$ The estimate stays constant between observed transition times and jumps by this amount at each event time. Both the cumulative estimate and its increments are dimensionless. Dividing an increment by a time interval would produce a quantity with rate units, but that calculation would be a rate summary over that chosen interval, not the Nelson–Aalen increment itself.
Why the Increment Has This Form
The justification comes from the local meaning of an intensity. Over a short interval \([t,t+\Delta)\), suppose individuals at risk have a common transition intensity that is approximately constant. Each person at risk then has an approximate conditional transition probability \(\alpha_{hj}(t)\Delta\). With \(Y_h(t)\) people at risk, the expected number of transitions is approximately \(Y_h(t)\alpha_{hj}(t)\Delta\). Dividing the observed count by the risk-set size gives a local estimate of the probability contribution, and hence an estimate of intensity multiplied by interval length.
The right side is dimensionless: an inverse-time rate multiplied by a time interval. Summing these contributions over time approximates the integral of the intensity. In observed event-time data, the Nelson–Aalen estimator records these contributions as jumps \(dN_{hj}(u)/Y_h(u)\), rather than dividing each jump by the spacing to the next event. This is why a step-function increment should not be read as a per-year or per-day rate.
The argument relies on correctly identifying who is at risk for the transition, observing transition times and states accurately, and having censoring that does not selectively remove people in a way that invalidates the risk-set comparison. It also relies on a meaningful intensity for the people being pooled. If censoring is informative or the pooled group has materially different transition behavior, the estimator may not estimate the intended cumulative intensity without further modeling.
Worked Example: Reading Two Nelson–Aalen Jumps
Suppose a hypothetical cohort is followed for a transition from state \(h\) to state \(j\). At time 2 years, 2 of the 20 people in state \(h\) make that transition. At time 5 years, 1 of the 10 people then at risk in state \(h\) makes it. Assume the risk sets are correctly formed and censoring is non-informative for this estimation.
At 2 years, the increment is $$ \Delta\widehat A_{hj}(2)=\frac{2}{20}=0.10. $$ After this event time, the cumulative estimate is \(0.10\). At 5 years, the increment is $$ \Delta\widehat A_{hj}(5)=\frac{1}{10}=0.10, $$ so the cumulative estimate through 5 years is \(0.10+0.10=0.20\).
The result \(0.20\) is accumulated estimated transition intensity, not a 20% transition probability and not a rate of 0.20 transitions per year. Each \(0.10\) jump is also dimensionless. The event times are measured in years, but their spacing does not change the units or definition of the jump.
Worked Example: A Jump Is Not a Local Rate
Consider a hypothetical estimated step function with a jump of \(0.08\) at year 1 and another jump of \(0.08\) at year 5, with no other observed transitions of this type between those times. Suppose the risk sets are positive at both event times and the transition data satisfy the estimation assumptions.
Just after year 1, \(\widehat A_{hj}=0.08\). It remains at \(0.08\) until the year-5 event, when it jumps to \(0.16\). The estimator’s average increase over the interval \((1,5]\) is $$ \frac{\widehat A_{hj}(5)-\widehat A_{hj}(1)}{5-1} = \frac{0.16-0.08}{4} = 0.02\ \text{per year}. $$
This \(0.02\) per year is a difference quotient across a four-year window. It is not the instantaneous intensity at year 5: the step function is flat between observed events and has a jump at year 5. Nor does this calculation establish that the underlying intensity was constant over the four years. It only summarizes the estimated cumulative increase across the chosen window.
Worked Example: Interpreting a Tied-Event Increment
In a hypothetical cohort, 12 people are at risk in state \(h\) immediately before year 3. At that time, 3 make the specified transition to \(j\). Assume the time and transition type are recorded correctly, and that the risk set contains everyone eligible just before the tied event time.
The Nelson–Aalen increment at year 3 is $$ \Delta\widehat A_{hj}(3)=\frac{3}{12}=0.25. $$ If the estimated cumulative intensity just before year 3 is \(0.15\), its value just after the event time is \(0.15+0.25=0.40\).
The \(0.25\) is the tied-event contribution to the estimated cumulative intensity. It also equals the observed fraction of this risk set making this transition at that time. That fraction is a data-based risk-set quantity, not an instantaneous rate of \(0.25\) per unit time. The time unit does not appear in the division \(3/12\).
Intensity, Accumulation, and Transition Probability
Three quantities are easy to confuse, but they answer different questions:
| Quantity | What it describes | Units |
|---|---|---|
| Transition intensity \(\alpha_{hj}(t)\) | Local rate of transition from \(h\) to \(j\) at time \(t\) | Inverse time |
| Cumulative transition intensity \(A_{hj}(t)\) | Accumulated intensity through time \(t\) | Dimensionless |
| Transition probability over a stated period | Probability of occupying or entering a specified state over that period, as defined by the question | Probability, from 0 to 1 |
A cumulative intensity is not constrained to be at most 1, whereas a probability is. Even when a particular model allows a relationship between an intensity and a transition probability, that relationship depends on the model and the transitions that can occur. In a multi-state setting, another outgoing transition can happen first and alter the chance of reaching \(j\). The cumulative intensity for \(h\to j\) alone therefore does not provide that probability.
Likewise, a rate cannot be inferred from a cumulative value without specifying the period over which it accumulated and how intensity varied within that period. Dividing \(A_{hj}(t)\) by \(t\) gives an average accumulated intensity per unit time from 0 to \(t\), not generally the instantaneous intensity at \(t\).
Diagnostics and Common Pitfalls
Several checks help prevent an appealing staircase from being overinterpreted:
- Check the scale and units. Each jump is \(dN/Y\), a dimensionless quantity. Labeling it “per year” without a defined time denominator is an error.
- Check the risk set and transition count. The denominator is the number at risk in the origin state immediately before the event time; the numerator counts only the specified transition. A wrong origin state or event type changes the estimator.
- Investigate large jumps. A large jump can occur when the risk set is small, even if the event count is modest. It may be highly variable and should not automatically be interpreted as a sharp increase in the underlying rate.
- Inspect time recording and ties. Rounding can create tied event times. Use the appropriate tied-event count and pre-event risk set rather than treating the recorded jump as a measured instantaneous spike in the underlying process.
- Check observation and censoring. If people leave observation in a way related to their unobserved transition behavior, the observed risk set may be unrepresentative. A mathematically computable step function does not by itself validate its interpretation.
A common mistake is to say, “The Nelson–Aalen curve jumps by 0.25, so the instantaneous transition rate at that time was 0.25 per year.” The reasoning confuses an accumulated-intensity increment with a rate. The correction is to report the increment as \(0.25\), dimensionless, and to report a rate only when a model or a clearly specified time-window calculation supports one.
The Nelson–Aalen step estimate is useful because it directly displays accumulated transition intensity without imposing a smooth time pattern. Its limitation is equally important: its jumps occur at observed event times, and their sizes depend on the risk sets. The step plot alone does not reveal a smooth instantaneous intensity curve. Estimating a smooth rate requires additional choices and assumptions; it is not achieved merely by relabeling the Nelson–Aalen jumps.
Check Your Understanding
Answer each question, then compare with the solutions below.
- At a particular event time, \(Y_h=25\) and \(dN_{hj}=2\). What is the Nelson–Aalen increment, and what are its units?
- Why is a cumulative estimate of \(1.3\) not automatically impossible, even though a probability cannot exceed 1?
- A step estimate rises by \(0.12\) over a two-year interval. What is the difference quotient, and why need it not equal the instantaneous intensity at the interval’s endpoint?
- Name two conditions or data checks relevant to interpreting a Nelson–Aalen estimate for a transition.
Solutions. (1) The increment is \(2/25=0.08\), dimensionless; it is not \(0.08\) per unit time. (2) The estimate is accumulated intensity, not a probability, and cumulative intensity is not restricted to the interval from 0 to 1. (3) The difference quotient is \(0.12/2=0.06\) per year. It averages the estimated cumulative increase over the two-year window and does not specify the local rate at its endpoint. (4) Examples include verifying that the denominator is the pre-event risk set in state \(h\), counting only \(h\to j\) events, checking that transitions are timed and classified correctly, and assessing whether censoring undermines the risk-set interpretation.