Why Project Event Accrual?
In an event-driven trial, the primary analysis is planned after a specified number of events has accrued, rather than solely at a fixed calendar date. This can align the analysis with the amount of outcome information available, but it makes the calendar timeline depend on how quickly participants enroll and experience the event. A model-based projection connects those design choices to an expected event count at a proposed analysis date.
The projection is a planning calculation, not a promise about when the target will be observed. It depends on assumptions about event rates, enrollment, follow-up, and censoring. A useful plan therefore compares plausible scenarios and makes clear which assumptions are driving the projected date.
Model and Assumptions
Consider two randomized arms indexed by \(j=0\) for control and \(j=1\) for treatment. Let \(T_j\) be time from randomization to the event of interest, and let \(S_j(t)=P(T_j>t)\) be the probability of remaining event-free beyond \(t\) months. For a simple projection, assume a constant event hazard \(\lambda_j\) within each arm. Under this exponential time-to-event model, \(S_j(t)=e^{-\lambda_j t}\). The hazard has units of events per person-month; it is an instantaneous event rate among participants who remain event-free.
The basic calculation assumes the event is defined consistently across arms; participants are assigned to the stated arms; follow-up begins at randomization; and the event hazard model is a reasonable description over the projected follow-up. The examples below also assume complete event ascertainment through the analysis date and no competing events that prevent the event of interest. Those assumptions should not be treated as facts merely because the calculation is convenient.
Suppose \(n_j\) participants are enrolled in arm \(j\) uniformly over an accrual period of \(A\) months. If the analysis occurs \(F\) months after the last participant is enrolled, then participant follow-up durations range from \(F\) to \(F+A\). Under the constant-hazard model, average event-free probability in that arm is the average of \(e^{-\lambda_j t}\) over this duration range.
The corresponding expected event count in arm \(j\) is \(n_j[1-\overline{S}_j(F,A)]\). Adding the arm-specific expectations gives the total expected count. This calculation uses linearity of expectation; it does not require event outcomes for different participants to be statistically independent. Independence and additional modeling details matter when estimating the distribution of the total count or the probability of reaching a target by a date.
Deriving the Projection and Reading It Correctly
The average-survival expression follows directly from the enrollment schedule. With uniform enrollment, the fraction of participants enrolled in any small time interval within the accrual period is proportional to the interval’s length. Their follow-up times are therefore averaged uniformly from \(F\) to \(F+A\). Integrating survival over those times gives the first expression below; evaluating the integral gives the closed form.
The result is an expected count. It is not the number that must occur in a particular trial, and it is not by itself the probability that the event target will be reached. Event times vary across trials, and estimates of the hazards and enrollment schedule are uncertain. If the decision depends on the chance of reaching the target by a deadline, a distributional calculation or simulation is needed in addition to the expected-count projection.
Worked Example: Projecting Events for a Proposed Analysis Date
Hypothetical plan. A trial enrolls 300 participants per arm uniformly over 14 months. The proposed analysis is 12 months after the last participant is enrolled. Assume constant hazards of \(\lambda_0=0.0400\) events per person-month in control and \(\lambda_1=0.0300\) events per person-month in treatment. Calculate the expected total number of events.
The model implies exponential event-free probabilities. At the analysis, participant follow-up ranges from 12 to 26 months. The assumed uniform enrollment schedule and complete ascertainment make the average-survival formula applicable for this calculation.
For control, the average event-free probability is
Thus the expected control event count is \(300(1-0.4738013)=157.8596\), or about 157.9 events. For treatment,
The expected treatment event count is \(300(1-0.5696912)=129.0926\), or about 129.1 events. The total is \(157.8596+129.0926=286.9522\), approximately 287.0 events. This plan’s expected count is below a target of 300 events. It does not prove the target cannot be reached at that date: actual accrual may be higher or lower than its expectation.
Compare Enrollment and Follow-Up Plans
An event projection lets a development team compare calendar plans using the same event model. A shorter accrual period can shorten the time to the analysis date, but it also changes the spread of participant follow-up times. More time after the last enrollment generally increases expected events, while a higher assumed event hazard also increases them. The next example illustrates a plan close to its target, where it is especially important not to confuse an expected count with a guaranteed count.
Worked Example: Checking a Target at an Alternative Date
Hypothetical plan. Keep 300 participants per arm and the hazards from the first example, but assume uniform enrollment over 12 months and assess the trial 12.3 months after the last participant is enrolled. Follow-up therefore ranges from 12.3 to 24.3 months. The constant-hazard assumptions are unchanged, and the enrollment interval is still assumed uniform.
The average event-free probabilities are
The expected control count is \(300(1-0.4855765)=154.3271\) events, and the expected treatment count is \(300(1-0.5806508)=125.8048\) events. Together, the arms contribute \(280.1319\), or about 280.1 expected events. A target of 280 is therefore approximately met in expectation at this date, but the small margin does not establish that the trial will have 280 events by then. A planning team should consider the variability of event accrual and uncertainty in the assumed hazards before treating the date as dependable.
Worked Example: Sensitivity to a Different Treatment Hazard
Hypothetical sensitivity scenario. Return to 300 participants per arm, 14 months of uniform enrollment, and 12 months of follow-up after the last enrollment. Keep the control hazard at 0.0400 per person-month, but suppose the treatment hazard is 0.0350 rather than 0.0300. This remains an exponential model with complete event ascertainment; the purpose is to see how a plausible change in an input shifts the projected count.
The control projection is unchanged at 157.8596 expected events. For treatment,
The treatment event count is approximately \(300(1-0.5194)=144.2\). The total is approximately \(157.9+144.2=302.0\) expected events, compared with 287.0 in the first example. Both scenarios retain a lower treatment hazard than control, but the higher treatment hazard yields more expected events by the same analysis date. Thus, an event-target timeline can be sensitive to the assumed treatment effect even when the control assumption and trial size are unchanged.
Diagnostics, Alternatives, and Limitations
Before relying on a projection, check whether the inputs and assumptions are credible for the intended use. Compare the assumed hazards with relevant data and assess whether a constant hazard is plausible over the projected time span. Check enrollment feasibility against the planned recruitment rate, and confirm that follow-up begins and ends as represented. In particular, distinguish withdrawal from follow-up from remaining event-free: if participants stop contributing event information, the complete-ascertainment assumption is not satisfied.
The exponential model is convenient, but it may misrepresent hazards that change over time, such as hazards that rise after randomization or differ early and late in treatment. A piecewise-exponential model can assign different constant hazards to prespecified time intervals. Other time-to-event models may be suitable when their assumptions match the available evidence. If competing events can preclude the event of interest, simply using \(1-S(t)\) from an event-free survival model may not estimate the probability of that event; the competing-event process must be represented appropriately.
Enrollment is also rarely perfectly uniform. If recruitment varies over calendar time, calculate follow-up for the planned enrollment cohorts or use the expected enrollment curve rather than applying the uniform-accrual formula. Dropout, missed visits, and incomplete event ascertainment can delay observed events; omitting them may make a projection optimistic. A useful diagnostic is to compare projected accrual with observed enrollment and events as the trial proceeds, then investigate material departures rather than assuming the original projection remains calibrated.
| Approach | Useful when | Main limitation |
|---|---|---|
| Closed-form exponential projection | Hazards are reasonably represented as constant and enrollment is close to uniform. | Can give misleading timing if hazards or enrollment change substantially. |
| Cohort-based calculation | Enrollment dates or follow-up differ across cohorts and can be represented explicitly. | Still depends on the specified event and censoring models. |
| Simulation across plausible scenarios | The decision needs uncertainty in timing or sensitivity to several uncertain inputs. | Results depend on the scenarios and model assumptions supplied to the simulation. |
A target event count is usually chosen as part of the trial’s inferential design. This tutorial addresses the complementary planning question: under specified event, enrollment, and follow-up assumptions, when might that count accrue? Meeting the projected count does not by itself establish that the analysis will answer the scientific question; that also depends on the design and the quality of the data.
Common Pitfall
Define precisely which event counts and when follow-up begins.
State arm-specific survival assumptions, numbers enrolled, accrual pattern, and censoring assumptions.
Average event probabilities over participant follow-up, then sum the expected counts.
Vary uncertain assumptions and distinguish expected accrual from the chance of meeting a deadline.
Check Your Understanding
Answer each question, then compare with the solutions.
- With a constant hazard of 0.0500 events per person-month and 10 months of follow-up for every participant, what is the expected event count among 80 participants?
- For 100 participants enrolled uniformly over 10 months, with 6 months of follow-up after the last enrollment and a hazard of 0.0400 per person-month, calculate the expected event count.
- If the number enrolled in each arm doubles while the hazard and follow-up plan remain unchanged, what happens to the expected event count?
- Why might the formula \(1-S(t)\) be inadequate for projecting the probability of a specific event when competing events can occur?
Solutions. (1) The event probability is \(1-e^{-0.0500(10)}=1-e^{-0.5}\approx0.3935\). Thus the expectation is \(80(0.3935)\approx31.48\) events. (2) Follow-up ranges from 6 to 16 months, so average survival is \([e^{-0.0400(6)}-e^{-0.0400(16)}]/[0.0400(10)]\approx0.6483\). The expected count is \(100(1-0.6483)\approx35.17\) events. (3) With unchanged individual event probabilities, doubling enrollment doubles the expected count. (4) A competing event can prevent the event of interest from occurring; the event-of-interest probability must account for that competing process rather than treating all non-occurrence as ordinary event-free survival.