Why an Event Date Is Not an Analysis Date
A clinical event can occur before it is reported, entered into a database, cleaned, and confirmed through endpoint adjudication. The event may therefore have an occurrence date before a planned data cutoff but still not be usable in the analysis at that cutoff. For planning an interim look, it is the availability of confirmed, analysis-ready records—not merely the dates on which events occurred—that determines how many events can contribute.
The issue is distinct from the time needed to schedule or conduct an analysis after a dataset is ready. Here the delay is between an event’s occurrence and the point at which its record is sufficiently complete and confirmed to be counted. A projection that ignores this interval can place a data cutoff too early or overstate the number of available events.
The basic planning question is therefore: by a proposed cutoff, how many of the events expected to have occurred will also have completed the steps required for analysis? A delay adjustment answers that question without changing the underlying event-occurrence model.
Fixed Delays and Variable Delays
If every event has the same known delay \(d\), an event target projected to occur at time \(T\) becomes available at \(T+d\). This is a direct calendar adjustment. It is appropriate only when the delay is reasonably represented as fixed for the events in question. Calendar days and business days must not be confused: a delay specified as ten business days should be added using the applicable working calendar, not as ten calendar days.
In practice, delays vary. Let \(G(x\mid t)\) denote the conditional cumulative distribution function of the delay for an event occurring at time \(t\). Thus, for \(x\ge 0\), \(G(x\mid t)=P(D_i\le x\mid E_i=t)\). For an event occurring at \(t\le C\), the probability it is available by cutoff \(C\) is \(G(C-t\mid t)\). Events with \(t>C\) cannot yet be available at \(C\), since the delay is nonnegative.
The distribution matters because a typical or average delay does not tell us exactly how many records will be ready at a particular cutoff. For example, two delay distributions can have the same mean but different proportions of short delays, and therefore different numbers of events available by a given date.
Worked Example: Adding a Fixed Adjudication Delay
Suppose a hypothetical trial projects that its target of 240 events will occur on August 20. Assume all relevant event records take exactly 10 calendar days from occurrence to confirmation and analysis readiness. Find the projected date when the target is available for analysis.
Under the fixed-delay assumption, the target becomes available \(d=10\) days after its projected occurrence date:
Substituting the hypothetical dates gives August 20 plus 10 calendar days, or August 30. The projected occurrence target is still August 20; August 30 is the date by which those target events are expected to have completed the assumed delay.
The assumption must be checked against the planning context. If some records take longer than 10 days, the August 30 date does not ensure that all 240 will be ready. If the delay is in business days, the date also needs to be recalculated using the relevant calendar.
Expected Confirmed Counts Under Variable Delays
Suppose a projection supplies expected event counts by occurrence time. Each expected occurrence can be weighted by its probability of being available at the cutoff. For a set of projected events with occurrence times \(t_1,\ldots,t_m\), the expected number available at \(C\) is the sum of their availability probabilities. This uses the fact that the expected value of a sum of event indicators is the sum of their expectations; independence among the events is not required for this expectation.
Here the \(m\) projected events are indexed by \(i\), and their occurrence times are treated as specified for this calculation. If a projected event count in a time interval is an expectation rather than a list of individual events, its expected count can instead be weighted by the appropriate average availability probability for that interval.
For a continuous-time projection, let \(\lambda(t)\) be the expected event-occurrence rate at time \(t\), in events per unit time. Under the assumptions that this rate describes expected occurrences and that the delay distribution conditional on occurrence time is \(G(\cdot\mid t)\), the expected number confirmed by \(C\) is:
The lower limit can be replaced by the start of the period in which events are projected. To justify the expression, divide time before \(C\) into short intervals. An interval around \(t\) has expected event count approximately \(\lambda(t)\,dt\), and each event in it has probability approximately \(G(C-t\mid t)\) of being ready by \(C\). Its contribution to the expected available count is therefore approximately \(\lambda(t)G(C-t\mid t)\,dt\). Summing these contributions and taking the limit as the intervals become arbitrarily short gives the integral.
This calculation estimates an expected count, not a guarantee that the realized count will reach a target. Choosing a cutoff to make the expected count equal a target is a planning convention; uncertainty in the realized count remains. It is also possible to specify a more conservative planning criterion, but that requires an explicit probability or prediction-interval requirement rather than just an expected count.
Worked Example: Event-Level Availability Probabilities
In a hypothetical trial, a data cutoff is planned for day 20. Consider 40 projected events, all expected to occur on day 10. Assume each delay is independent of occurrence time and has this distribution: 40% of records are ready after 2 days, 35% after 7 days, and 25% after 12 days. Find the probability that a single event is ready by the cutoff and the expected number of the 40 events available.
An event occurring on day 10 has \(20-10=10\) days remaining until the cutoff. It is ready by day 20 if its delay is at most 10 days. The 2-day and 7-day delays qualify, while the 12-day delay does not. Therefore:
The expected available count is the sum of the 40 identical availability probabilities:
Thus, the expected number of confirmed, analysis-ready records is 30. This does not mean that exactly 30 will be available: the actual number can be lower or higher. The 12-day records are not counted at this cutoff even though their events occurred before it; they become available only after their delay has elapsed.
Worked Example: A Continuous Accrual Rate with Exponential Delays
Consider a hypothetical projection with events expected at a constant rate of 4 events per day from day 0 through day 30. Suppose the delay \(D_i\) is exponential with mean 5 days, independent of occurrence time. For an exponential delay with mean 5 days, \(G(x)=1-\exp(-x/5)\) for \(x\ge0\). Calculate the expected number available by day 30, and compare it with the expected number of events that have occurred by then.
For an event occurring on day \(t\), the remaining time until the cutoff is \(30-t\), so its availability probability is \(1-\exp(-(30-t)/5)\). Applying the continuous-time formula:
The expected number of events that have occurred by day 30, without requiring that their records be ready, is \(4(30)=120\). The delay-adjusted expectation is about 100.05 events, roughly 19.95 fewer. The difference is not an estimate of events that will never occur; it represents events expected to have occurred by the cutoff but not yet to have completed the assumed delay. Events near day 30 have little time to become available, which is why the delay adjustment is not equivalent to removing the same fraction of events from every date.
Planning a Cutoff for a Confirmed-Event Target
When the objective is a target of confirmed events, first distinguish the event target from its data-availability requirement. With a fixed delay \(d\), a projected date \(T\) for reaching an occurrence target translates to \(T+d\) for that target to be available, if the fixed-delay assumption is appropriate. With variable delays, there is generally no single exact date at which all events associated with a projected occurrence target will be ready.
One planning approach is to identify the earliest cutoff \(C\) at which the projected expected available count reaches the desired number \(K\):
This is an expectation-based planning rule. It does not imply that the actual available count at \(C\) will be at least \(K\) with certainty, or even with any particular probability. A probability-based requirement would instead specify a probability such as \(P(N_{\mathrm{available}}(C)\ge K)\) and use an appropriate predictive model for event counts and delays. The chosen rule should match the purpose of the planning decision.
The delay distribution should reflect the stage at which the event becomes eligible to count. If the analysis requires central adjudication, a report from a site is not yet availability. If only cleaned records are required, waiting for adjudication may be unnecessary. Define the endpoint of the delay consistently with the planned analysis dataset.
Diagnostics, Alternatives, and Limitations
A delay adjustment is only as credible as the assumptions behind it. Compare assumed delays with observed elapsed times from event occurrence to the chosen readiness milestone, using records that have completed the process. Check whether delays differ by calendar period, site, event type, or other relevant characteristics. A common delay distribution may be misleading if, for example, adjudication backlogs have increased or a subset of events needs substantially more documentation.
- Check the time origin and endpoint. Confirm that delay starts at the event date and ends at the precise data-ready milestone used for planning.
- Check units and calendars. Do not combine days with weeks or calendar days with business days without conversion.
- Check the delay model. Ensure delays are nonnegative and that the assumed distribution represents the current reporting and adjudication process.
- Check for incomplete records. Recent events may have had less time to complete adjudication. Treating their current elapsed delays as completed delays can distort estimates of the full delay distribution.
- Separate delay from event occurrence. A record not yet confirmed at cutoff is not automatically evidence that no event occurred. It may simply be pending; a separate adjudication outcome model is needed if some candidate events are ultimately rejected.
A fixed lag is simple, transparent, and useful when completion times are tightly controlled. A distribution-based adjustment uses more information when delays vary and provides expected counts by cutoff. A conservative percentile-based lag can also be used when the planning objective is to allow time for a specified fraction of records to complete; that fraction must be stated and the percentile estimated from relevant delay data. These are alternatives for representing delay, not substitutes for the underlying event-occurrence projection.
Important limitations remain. Delay estimates from a small or historical sample may not represent future workload; process changes can alter the distribution. A mean delay alone does not determine the probability of availability at a cutoff. Expected confirmed counts also do not describe their variability. Finally, if adjudication can conclude that a candidate is not a qualifying endpoint event, delay modeling alone cannot determine the final number of valid events; the probability of confirmation as a valid event must also be represented.
Common Mistake: Counting Events Before Their Records Are Ready
A frequent error is to count every event with occurrence date before the data cutoff as available. This silently assumes zero delay. For example, suppose an event occurred 5 days before cutoff and the delay is fixed at 7 days. Only 5 days remain, and the 7-day delay exceeds that time, so the event is not available by the cutoff. The correct condition is \(D_i\le C-E_i\), not merely \(E_i\le C\).
Another mistake is to add the mean delay to a projected event date and describe the result as a date when the target is certain to be ready. For variable delays, adding the mean gives neither a guarantee nor, in general, a stated probability of readiness. Use the full delay distribution for expected availability, or specify a percentile or predictive probability that matches the planning objective.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A hypothetical target of 180 events is projected to occur on November 4. All records have a fixed 6-calendar-day delay. What is the corresponding projected availability date?
- An event occurs 5 days before a cutoff. Its delay is 7 days. Is it available at the cutoff? State the condition that determines your answer.
- Twenty projected events occur on the same day, 9 days before cutoff. For each event, the delay is 3 days with probability 0.6 and 12 days with probability 0.4. Find the probability that an event is available and the expected available count.
- In your own words, why does an expected available count of 50 not guarantee that at least 50 events will be ready?
Solutions. (1) Add 6 calendar days to November 4: the projected availability date is November 10. (2) No. Only 5 days remain, and \(7>5\), so the delay has not elapsed by cutoff; availability requires \(D_i\le C-E_i\). (3) Nine days remain. The 3-day delay qualifies and the 12-day delay does not, so the availability probability is \(0.6\). The expected count is \(20(0.6)=12\) events. (4) An expectation is an average over possible outcomes under the model, not a promise about the realized count at a particular cutoff. The actual count varies because event occurrence and delay outcomes vary.