Why Smooth Transition-Intensity Estimates?
The Nelson–Aalen estimator records transition-hazard accumulation as jumps at observed event times. Its increments retain the data’s time resolution, but a transition observed only a few times can produce a jagged estimate that is difficult to interpret as an instantaneous rate. Kernel smoothing replaces each increment by a weighted contribution over nearby times. This produces a curve that can reveal broad temporal patterns, while making the choice of smoothing bandwidth consequential.
The goal is not to change which transition is being estimated. For a specified transition from state \(i\) to state \(j\), smoothing uses that transition’s own risk set and event increments. The resulting curve is an estimate of the transition-specific instantaneous intensity, not a cumulative hazard and not a transition probability.
Conditions and Construction
This construction smooths the Nelson–Aalen increments; it does not remove the assumptions needed for those increments to represent the transition process. The states and transition of interest must be defined consistently, event times and state membership must be recorded adequately, and the transition-specific risk set must be correct. Censoring must not selectively hide transitions in a way that invalidates risk-set estimation; in particular, the censoring assumptions used for the Nelson–Aalen estimate must be appropriate. When several transitions leave state \(i\), the increment for \(i\to j\) uses the risk set in state \(i\), while events on other outgoing transitions are not counted as \(i\to j\) events.
For a conventional kernel estimate, take \(K\) to be nonnegative, symmetric, and integrable with total integral 1. A compactly supported kernel uses only increments within a specified distance of \(t\); a Gaussian kernel gives smaller but nonzero weight to more distant increments. In either case, \(h\) controls the time scale over which increments contribute. The basic estimate is most interpretable where the transition is observable and its intensity changes smoothly enough for nearby times to be informative about one another.
The discrete sum is often the simplest way to calculate the estimate. At each observed transition time \(u\), first calculate its Nelson–Aalen increment \(\Delta N_{ij}(u)/Y_i(u)\). Then multiply that increment by the kernel weight \(K_h(t-u)\) and add the contributions. If there are tied transitions at \(u\), use the increment at that time as defined by the Nelson–Aalen calculation; do not treat a tied set as separate event times merely to create more smoothing points.
Choose \(i\to j\), check its risk set and event increments, and state the time units to be used for \(t\) and \(h\).
Use a normalized kernel \(K\) and select \(h\) to determine how far in time increments will influence the estimate.
At each evaluation time \(t\), calculate \(K_h(t-u)\Delta N_{ij}(u)/Y_i(u)\) for each observed transition time \(u\), then add the terms.
Check whether \(t\) is near the observed time boundaries, whether the risk set has become empty, and whether the curve changes substantially under plausible bandwidth choices.
Worked Example: A Box-Kernel Estimate at an Interior Time
Suppose a hypothetical transition \(i\to j\) has three observed event times, 2, 3, and 5 days. Their Nelson–Aalen increments are respectively 0.10, 0.20, and 0.10. Estimate the intensity at \(t=3\) days using the box kernel \(K(v)=\tfrac12\) for \(|v|\leq1\), and \(K(v)=0\) otherwise, with bandwidth \(h=2\) days.
For this kernel, \(K_h(t-u)=1/(2h)=1/4\) per day when \(|t-u|\leq2\) days, and zero otherwise. All three event times are within two days of \(t=3\). Therefore, $$ \widehat\alpha_{ij,2}(3) =\frac14(0.10+0.20+0.10) =\frac14(0.40) =0.1000\ \text{per day}. $$
This estimate averages the nearby cumulative-hazard increments over a four-day window. It is an estimated instantaneous transition intensity at day 3, expressed per day; it is not the probability of transitioning during the next day. The calculation assumes that the recorded increments are valid for this transition and that pooling information across this local time window is a reasonable smoothing choice.
Why the Kernel Formula Smooths Increments
The justification follows from the meaning of a cumulative transition hazard. If the underlying intensity \(\alpha_{ij}(u)\) exists, then over a short interval the expected cumulative-hazard increase is approximately the intensity at that time multiplied by the interval length. Smoothing the increments with a normalized kernel therefore forms a local weighted average of the intensity, rather than simply plotting the jump sizes.
In more detail, write \(dA_{ij}(u)=\alpha_{ij}(u)\,du\) for the underlying cumulative-hazard increment in a continuous-time formulation. The corresponding smoothed target is $$ \int K_h(t-u)\,dA_{ij}(u) =\int K_h(t-u)\alpha_{ij}(u)\,du. $$ The kernel weights integrate to 1 over the full time line. If \(\alpha_{ij}(u)\) is nearly constant over the kernel’s effective window around \(t\), then it can be approximated by \(\alpha_{ij}(t)\) inside the integral, giving $$ \int K_h(t-u)\alpha_{ij}(u)\,du \approx \alpha_{ij}(t)\int K_h(t-u)\,du =\alpha_{ij}(t). $$ Replacing \(dA_{ij}\) with the estimated increments \(d\widehat A_{ij}\) yields the stated estimator. This argument explains the local-averaging interpretation, but does not guarantee that a finite-sample estimate is close to the true intensity. In particular, rapid changes in intensity and sparse events can make the local approximation poor.
Worked Example: A Sparse Transition and Two Bandwidths
Consider a hypothetical transition with only three Nelson–Aalen increments: 0.20 at day 2, 0.25 at day 5, and 0.50 at day 9. Estimate the intensity at day 5 using the same box kernel as above, first with \(h=2\) days and then with \(h=5\) days.
With \(h=2\), only the event at day 5 lies within the window \([3,7]\). The kernel weight is \(1/(2h)=1/4\) per day, so $$ \widehat\alpha_{ij,2}(5) =\frac14(0.25) =0.0625\ \text{per day}. $$ With \(h=5\), all three increments lie within the window \([0,10]\), and each receives weight \(1/(2h)=1/10\) per day. Thus $$ \widehat\alpha_{ij,5}(5) =\frac1{10}(0.20+0.25+0.50) =\frac1{10}(0.95) =0.0950\ \text{per day}. $$
The wider bandwidth produces a higher estimate at day 5 because it includes the relatively large increment at day 9. The narrow bandwidth is more local, but depends on just one observed increment; the wide bandwidth uses more events but blends information over a broader period. Neither number is automatically preferable. Their difference is a warning to examine whether the data support a stable local intensity estimate and whether the scientific question concerns short-term changes or a broader trend.
Bandwidth, Boundaries, and Diagnostics
Bandwidth controls a bias–variability trade-off. A narrow bandwidth preserves local features, but in a sparse transition it can produce a curve dominated by one or two increments and can leave many evaluation times with little or no nearby information. A wider bandwidth generally produces a smoother curve, but can blur a real change in intensity or combine periods that should not be treated as locally similar. These are qualitative consequences of local averaging; smoothing does not create additional observed transitions.
Boundary times require particular care. Near the beginning or end of follow-up, a symmetric kernel window extends beyond the time range where increments are observable. The full kernel weight of 1 is then not represented by observed increments. This can produce boundary distortion, and the direction and size of the distortion depend on the pattern of the underlying intensity and the chosen kernel. One-sided or boundary-adjusted smoothing can be considered, but it changes the construction and should be stated explicitly; it does not recover information that was never observed.
A zero-risk-set boundary is more serious than an ordinary sparse region. If nobody remains at risk in state \(i\), there are no later \(i\to j\) increments that can identify the transition intensity from these data. Smoothing earlier increments across that boundary may draw a visually continuous curve, but it does not make the intensity estimable beyond the supported observation period. Restrict interpretation to times with relevant risk-set support, and do not present a smoothed tail as data-supported merely because the kernel formula returns a number.
- Check risk-set support. Look for abrupt decreases in \(Y_i(t)\), long periods with very few people at risk, and times after the risk set becomes zero.
- Check event support. Mark how many increments actually contribute near each \(t\). A smooth curve can conceal that a local estimate is based on one transition.
- Check bandwidth sensitivity. Compare a small set of scientifically plausible bandwidths. Large changes signal that the apparent pattern is not well determined by local data.
- Check the time boundaries. Be especially cautious where the kernel window is truncated by the start or end of observation.
- Check the scale. The estimate should be nonnegative for a nonnegative kernel and nonnegative Nelson–Aalen increments, and its units should be inverse time.
Worked Example: Interpreting an Estimate Near the End of Follow-Up
In a hypothetical study, the last observed transition \(i\to j\) occurs at day 8, with increment 0.30. The transition-specific risk set becomes zero immediately after day 9. A box kernel with \(h=2\) days is used to estimate the intensity at day 9.
The observed increment at day 8 is one day from the evaluation time and is inside the kernel window. Its weight is \(1/(2h)=1/4\) per day. There are no other contributing increments in the stated data, so the calculation gives $$ \widehat\alpha_{ij,2}(9) =\frac14(0.30) =0.0750\ \text{per day}. $$
The arithmetic is valid as a kernel-weighted sum, but interpretation is limited: the estimate at day 9 is based on one increment and is at the boundary of the period with a positive risk set. It does not support claims about the intensity after day 9, when the risk set is zero. A curve extending beyond that point would be a smoothing extrapolation, not an estimate identified by later risk-set data. The example therefore passes the nonnegativity and unit checks, but fails to provide strong local event support.
How Smoothing Relates to Alternatives
| Approach | What it shows | Main consideration |
|---|---|---|
| Unsmoothed Nelson–Aalen increments | Individual estimated cumulative-hazard jumps at observed transition times | Faithful to event-time resolution, but not a continuous intensity curve |
| Kernel-smoothed increments | A local average of transition-hazard increments on the time scale set by \(h\) | Readable as a curve, but depends on kernel choice, bandwidth, and boundary support |
| Parametric intensity model | An intensity curve constrained to a chosen functional form | Can summarize sparse data efficiently if the form is credible, but may misrepresent patterns when the form is wrong |
Kernel smoothing is useful when a descriptive, time-varying curve is wanted without imposing a specific parametric shape. It is not a substitute for the unsmoothed cumulative-hazard estimate when the event-time jumps themselves are of interest, and it is not automatically a better choice than a model with explicit structure. Its main limitation is that apparent detail depends on the bandwidth and the observed support. In sparse transitions, a curve should be treated as a summary of limited evidence, not as proof of fine-scale temporal variation.
Common Mistake
Check Your Understanding
Answer each question using the definitions and calculations above. Brief solutions follow the questions.
- A transition has increments 0.12 at day 4 and 0.08 at day 6. Using a box kernel and \(h=2\) days, calculate the estimate at day 5.
- For the same two increments, what is the estimate at day 5 with \(h=1\) day?
- Why is a kernel-smoothed estimate not a transition probability?
- What should be concluded if the risk set for the origin state becomes zero after day 10, but a smoothed curve is drawn through day 12?
Solutions. (1) Both increments are within two days of day 5. The weight is \(1/(2h)=1/4\) per day, so the estimate is \((0.12+0.08)/4=0.0500\) per day. (2) Both increments are exactly one day from day 5 and are included by the stated box-kernel boundary rule \(|t-u|\leq h\). The weight is \(1/2\) per day, giving \((0.12+0.08)/2=0.1000\) per day. (3) It is a smoothed estimate of instantaneous rate, with units of inverse time; it is not a probability over a specified interval. (4) The curve after day 10 is not supported by a positive risk set for that transition. Smoothing cannot supply information about an intensity after the data cease to identify it.