Why tied transition times need careful handling
In multi-state data, two or more people can be recorded as making the same transition at the same time. For example, several patients may move from an “admitted” state to a “discharged” state on the same recorded day. Nelson–Aalen estimation does not require assigning these events an invented order. Instead, it uses the number of events of the specified transition at that time and the risk set immediately before that time.
The key is to distinguish a tie within one transition type from events of different transition types that share a recorded time. Each transition has its own event count and cumulative transition-hazard estimate. Risk-set definitions and a consistent rule for same-time censoring are essential to both calculations.
The Nelson–Aalen increment for transition \(h\to j\) at time \(s\), when \(Y_h(s)>0\), is
The cumulative estimate through time \(t\) is the sum of these increments over recorded transition times:
The denominator is the risk set just before the time’s events are counted. People making the transition at \(s\) are therefore included in \(Y_h(s)\). They are not removed one by one while calculating the same-time increment. The estimate is transition-specific: \(d_{hj}(s)\) counts only transitions from \(h\) to \(j\), not every event leaving state \(h\).
Why the tied-event increment is a count divided by one risk set
The short justification comes from the usual transition-intensity setup. While a subject is in state \(h\), the transition intensity to \(j\) describes the instantaneous event rate for that transition, conditional on the modeled history. The Nelson–Aalen construction compares the observed number of \(h\to j\) events with the amount of time subjects were at risk for that transition. At a recorded event time, the risk set \(Y_h(s)\) is the relevant exposure immediately before the jump.
More formally, under the intensity model, the expected increment in the transition counting process, conditional on the history just before \(s\), is governed by the risk set and the transition-hazard increment. This motivates estimating that increment by solving the observed-count equation at the jump:
This is an estimating argument, not a claim that the observed count must equal its conditional expectation. Crucially, the whole count at a tied time is paired with the same pre-event risk set. Replacing that denominator with a sequence of shrinking risk sets would impose an ordering that the recorded data do not supply.
The usual interpretation also relies on valid state and transition definitions, correct event and censoring records, and censoring that is non-informative for the transition process given the modeled information. The risk set must be positive at any time for which an increment is calculated. If event times are recorded only to a day or week, events sharing a timestamp may have occurred in an unknown order within that interval; the calculation then estimates using the recorded-time grouping, not a recovered exact ordering.
Worked Example: Three Tied Events of One Transition
Suppose 12 hypothetical patients are in state \(h\) and under observation immediately before day 10. Three make the same transition, \(h\to j\), on recorded day 10. No other event of this transition occurs on that day. Calculate the Nelson–Aalen increment at day 10.
The tied count is \(d_{hj}(10)=3\), and the pre-event risk set is \(Y_h(10)=12\). Thus
The estimated cumulative transition hazard increases by \(0.25\) at day 10. This is a cumulative-hazard increment, not a statement that each individual has a 25% probability of transition on that day. In particular, the calculation does not reduce the denominator to 11 and then 10 for the second and third events.
Different transition types at the same recorded time
Suppose subjects in state \(h\) can move to either \(j\) or \(k\). At a shared recorded time \(s\), count the \(h\to j\) events and the \(h\to k\) events separately. Both transition-specific increments use the same pre-event \(Y_h(s)\), provided both transitions are possible from \(h\) and the risk set is defined the same way. An event of type \(h\to k\) contributes nothing to \(d_{hj}(s)\), and vice versa.
This calculation assumes that one subject cannot make two distinct outgoing transitions at precisely the same instant. If the data appear to show that, the state definitions, timestamp precision, and event-recording rules need checking. Multiple people can make different transitions at the same recorded time without any one person making two transitions.
Worked Example: Two Outgoing Transition Types Tied in Time
Immediately before recorded time \(s=6\), 10 hypothetical subjects are at risk in state \(h\). At that time, two make transition \(h\to j\), and one makes transition \(h\to k\). Calculate both increments and the combined amount of outgoing transition-hazard increments at \(s\).
For \(h\to j\), the count is \(d_{hj}(6)=2\). For \(h\to k\), it is \(d_{hk}(6)=1\). Each uses \(Y_h(6)=10\):
The sum of the two transition-specific increments is \(0.20+0.10=0.30\). The estimate for \(h\to j\) still uses only the two \(h\to j\) events; it is not calculated as \(3/10\). Under the stated data, three distinct people make one outgoing transition each, so the combined event count does not exceed the 10 people at risk.
Same-time censoring and risk-set conventions
An event and a censoring time can share a recorded timestamp. With exact times, their ordering may be known. With coarse recording, it may not be. A calculation must specify how the data are handled rather than silently changing the denominator. A common convention is to count events as occurring just before censoring at the same recorded time. Under this convention, people censored at \(s\) are included in the risk set for events at \(s\), since they are considered under observation immediately before that time.
That convention is not a way to discover the true within-day ordering. If the recorded times are too coarse for the assumed ordering to be credible, the analyst should report the limitation and consider whether the study’s recording process supports the desired analysis. The independent-censoring assumption does not by itself resolve unknown event-versus-censoring order.
Worked Example: A Censoring Time Tied With Events
At recorded day 4, eight hypothetical subjects are at risk in state \(h\) just before the day begins. Two make transition \(h\to j\), and one is censored on day 4. Use the event-before-censoring convention. One of the five remaining subjects makes transition \(h\to j\) on day 6, when the risk set is five. Find the cumulative estimate through day 6.
At day 4, include the subject censored on that day in the pre-event risk set. The increment is
After the two transitions and one censoring, five subjects remain under observation in state \(h\). At day 6, the increment is
Therefore, the cumulative estimate through day 6 is
This arithmetic follows the stated risk-set convention. If censoring were instead placed before events at day 4, the risk set and resulting increment could differ; without known ordering, that alternative would represent a different convention, not a correction that can be inferred from the timestamps alone.
Checks, alternatives, and limitations
Before calculating a tied-time increment, check the origin state, destination state, event count, risk set immediately before the timestamp, and rule for same-time censoring. Also verify that no subject has been counted as making two distinct outgoing transitions at the same instant. A useful data check is that the total number of distinct outgoing events at a shared time cannot exceed the number at risk in the origin state, under the one-transition-per-person-at-an-instant setup.
Common signs of a calculation going wrong include a denominator that changes within one tied group, an event of another type added to the transition-specific numerator, or a censored subject removed without an explicit timing convention. If recorded times are heavily grouped, inspect how many ties occur and how the event-time precision was defined. Many ties may signal coarse measurement rather than genuinely simultaneous events.
When exact ordering is available, use that information and calculate increments at the distinct event times in that ordering. When only grouped timestamps are available, the aggregate Nelson–Aalen increment is a transparent conventional calculation; arbitrarily jittering times or inventing an event order can create unsupported risk sets. For data that are intrinsically discrete-time, a model designed for discrete-time transition probabilities may be preferable if those probabilities, rather than a Nelson–Aalen cumulative-hazard estimate, are the target. Neither option recovers unobserved within-interval event order without additional information.
The Nelson–Aalen calculation is also limited by the data and assumptions supporting its risk sets. Informative loss to follow-up, misclassified states, inaccurate transition types, or an inconsistent censoring rule can distort the estimate. The increment is not itself a transition probability, and a large increment should not be read as a probability exceeding one or as a direct individual-level risk. Its interpretation concerns the estimated cumulative transition hazard for the specified transition.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- At a recorded time, 15 people are at risk in state \(h\), and four make transition \(h\to j\). What is the increment?
- With 10 people at risk, two make transition \(h\to j\) and three make transition \(h\to k\) at the same time. What is each transition-specific increment?
- Why should tied same-transition events not be given sequentially shrinking denominators unless their ordering is actually known?
- Under an event-before-censoring convention, a person censored at the same recorded time as an event is included in which risk set?
- At one timestamp, the transition-specific increments are \(2/8\) and \(1/8\) for two outgoing transitions. What is their sum, and what should it not be called?
Solutions. (1) \(4/15=0.2667\). (2) The increments are \(2/10=0.20\) for \(h\to j\) and \(3/10=0.30\) for \(h\to k\); each numerator counts only its own transition type. (3) A shared timestamp does not establish an order among events, so sequential denominators impose unsupported information. Use the common pre-event risk set. (4) The person is included in the risk set immediately before that time, under the stated convention. (5) The sum is \(2/8+1/8=3/8=0.375\). It is the sum of outgoing cumulative-hazard increments at that time, not a transition probability.