Common Errors In Event-Date Forecasts

Why Event-Date Forecasts Go Wrong

A forecast of when a trial will reach an event target is a forecast about when events occur, not simply a calendar-date calculation based on the number of participants enrolled. Participants enter at different times, have different amounts of follow-up, and may leave observation before an event is recorded. In addition, an event may occur before it is cleaned, adjudicated, and available for an analysis. Mixing these processes can make a forecast look precise while targeting the wrong date.

The errors considered here are especially important when an event target determines analysis planning. A useful forecast makes its assumptions explicit: the enrollment pattern, event-time distribution, follow-up cutoff, censoring or dropout, and any delay between event occurrence and availability of a verified record. It should also state whether the target is expected events, observed events, or events available in the analysis data.

Definition: An event date is the calendar date on which the endpoint event occurs. A report date is the date on which the event is available in the data used for monitoring or analysis. An event-date forecast estimates when a specified number of endpoint events will have occurred. It does not, by itself, forecast when those events will be reported or when an analysis can be conducted.

Staggered Entry and Unequal Follow-Up

Let \(u_i\) be participant \(i\)'s enrollment date, \(T\) the calendar date of interest, and \(F(s)\) the probability that an enrolled participant experiences the event within \(s\) units of follow-up. If every participant is followed through \(T\), the expected number of events by \(T\) is

E\{D(T)\}=\sum_{i=1}^{n} F(T-u_i),

where \(D(T)\) is the event count by \(T\), and participants with \(u_i>T\) contribute zero. This formula assumes the event-time distribution applies to the participants being projected and that events through \(T\) are completely ascertained. If follow-up is still ongoing, the event probability depends on each participant's available follow-up, not just the total number enrolled.

For a constant event hazard \(\lambda\), with time measured from enrollment, the event-time survival probability is \(S(s)=\exp(-\lambda s)\), so \(F(s)=1-\exp(-\lambda s)\). This model assumes a constant event hazard over the projected follow-up. It can be a useful illustration or planning approximation, but it is not automatically appropriate for a changing hazard or a population with substantial heterogeneity.

Worked Example: The Error From Treating Staggered Entrants as Fully Followed

Suppose 100 participants enroll uniformly over 10 months, at an average rate of 10 per month. Assume a constant event hazard of 0.10 per month, no censoring, and complete event recording through month 10. Compare the cohort-based expected count with the shortcut that gives all 100 participants 10 months of follow-up.

Under continuous uniform enrollment, participants have follow-up durations ranging from approximately 0 to 10 months. The expected event count is the enrollment rate multiplied by the integrated event probability over those durations:

E\{D(10)\} =10\int_0^{10}\left(1-e^{-0.10s}\right)ds =10\left[10-\frac{1-e^{-1}}{0.10}\right].

Since \(e^{-1}\approx0.3679\), the calculation is \(10[10-6.321]=36.79\) expected events. By contrast, treating all 100 participants as enrolled at month 0 gives \(100(1-e^{-1})=63.21\) expected events. That shortcut overstates the expected count by about 26.42 events because it assigns the later entrants follow-up they have not had. The assumptions are internally consistent with this comparison: both calculations use the same hazard, enrollment total, and month-10 cutoff; only the assumed entry pattern differs.

Uniform Accrual Is Not a Default Event Model

A common shortcut assumes events accrue at a constant calendar-time rate. That can be a reasonable local approximation if the event rate is stable over the forecast interval and enrollment and follow-up conditions are not changing substantially. It is not a general consequence of having a fixed endpoint target. Early in a study, few participants may have had time to experience an event; later, more participants are at risk and have accumulated follow-up.

The distinction can be seen even when enrollment proceeds at a constant rate. Suppose participants enroll continuously at rate \(r\), have a constant event hazard \(\lambda\), and follow-up begins at enrollment. With no censoring, the expected event count by calendar time \(t\) is

E\{D(t)\} =r\int_0^t \left(1-e^{-\lambda(t-u)}\right)du =r\left[t-\frac{1-e^{-\lambda t}}{\lambda}\right].

The event rate is therefore not constant from the start: differentiating the expected count gives \(r(1-e^{-\lambda t})\), which increases with \(t\). The derivation follows by summing the event probabilities of participants enrolled at each time \(u\), whose follow-up by time \(t\) is \(t-u\). It requires constant enrollment and hazard, and no censoring; changing either assumption changes the forecast.

Worked Example: A Linear Extrapolation Reaches the Target Too Soon

Assume enrollment continues at 100 participants per month, the event hazard is 0.10 per month, and there is no censoring. The target is 100 events. Using the formula above, the expected count at month 6 is

E\{D(6)\} =100\left[6-\frac{1-e^{-0.6}}{0.10}\right] =100[6-4.512] \approx148.8.

A linear extrapolation from zero to 148.8 events over six months would project the 100-event target at \(6(100/148.8)=4.032\) months. That calculation assumes events accumulate uniformly from the start. The cohort-based model instead solves

100\left[t-\frac{1-e^{-0.10t}}{0.10}\right]=100.

Dividing by 100 and rearranging gives \(t+10e^{-0.10t}=11\). At \(t=4.83\), the left side is approximately \(4.83+10(0.617)=11.00\), so the expected target date is about month 4.83. The linear shortcut is about 0.80 months early in this hypothetical example. Both calculations use the stated enrollment rate, hazard, and target; the difference arises because the shortcut ignores the gradual accumulation of follow-up.

Censoring Changes the Number of Observable Events

A forecast that counts everyone as remaining under observation may overstate the number of events available by a target date. Censoring includes loss to follow-up or withdrawal that prevents further event ascertainment; administrative censoring occurs when observation ends at a planned cutoff. If a participant is censored before experiencing the endpoint, the participant cannot contribute a subsequently observed event in that data window. A forecast should therefore represent both event occurrence and the opportunity to observe it.

For example, suppose event times have constant hazard \(\lambda\), and independent dropout times have constant hazard \(\delta\). Treating event and dropout as competing ways to leave the event-free, observed state, the probability of an observed event by follow-up time \(\tau\) is

P(\text{event observed by }\tau) =\int_0^\tau \lambda e^{-(\lambda+\delta)s}ds =\frac{\lambda}{\lambda+\delta} \left[1-e^{-(\lambda+\delta)\tau}\right].

This expression assumes constant hazards and independence of the event and dropout times. It describes events observed before dropout, not the probability of an event in a hypothetical world where dropout did not occur. In settings with time-varying hazards, informative dropout, or other competing events, this simple expression is not adequate.

Worked Example: Quantifying the Effect of Dropout

Consider 100 participants all enrolled at the same time. Assume an event hazard of \(\lambda=0.10\) per month, an independent dropout hazard of \(\delta=0.05\) per month, and a six-month follow-up window. First calculate the probability of an event by six months if dropout is ignored: \(1-e^{-0.10(6)}=1-e^{-0.6}\approx0.4512\), or 45.12 events expected among 100 participants.

Accounting for dropout, the probability of an event observed before dropout and by month 6 is

\frac{0.10}{0.10+0.05} \left[1-e^{-(0.10+0.05)6}\right] =\frac{2}{3}(1-e^{-0.9}) \approx\frac{2}{3}(0.5934) =0.3956.

The corresponding expected observed count is \(100(0.3956)=39.56\) events. Under the stated assumptions, omitting dropout overstates the expected observed count by about 5.56 events. Before using this calculation, check that the hazards are plausible and that treating dropout as independent is defensible; if dropout is related to event risk, this model may be biased.

Keep Event Dates Separate From Report Dates

A forecast may correctly estimate when the target number of events will occur and still give the wrong operational date if it labels that date as the analysis date. Event occurrence, database availability, adjudication, and analysis readiness are distinct stages. The delay between them may vary by site, event type, or review workload. A fixed lag is only an approximation when the underlying process is reasonably stable.

Worked Example: Translating an Event Target Into a Report Date

Suppose a projection estimates that the 100th event will occur at the end of month 10. Assume, for planning purposes, that each event takes two months to be cleaned and adjudicated and that the lag is fixed. The event-target date is month 10, but the corresponding date when the 100th event is expected to be available for analysis is month \(10+2=12\). Reporting month 10 as the analysis-ready date would be two months early under these assumptions.

The arithmetic does not account for uncertainty in event timing or variation in processing delays. If the event records are reported in batches, or if adjudication delays differ materially across events, a single fixed lag may not describe when the full target becomes available. Track event occurrence and data processing separately rather than silently adding the lag to the event forecast.

Diagnostics, Alternatives, and Limits

When a projection appears implausible, first identify which quantity is being forecast and check whether the assumptions match the data. Useful diagnostics include comparing observed accrual with the planned enrollment pattern, reviewing follow-up time by enrollment cohort, checking dropout and administrative censoring, and measuring the delay from event occurrence to data availability. A forecast can be internally calculated correctly and still be wrong because one of these inputs is outdated or inappropriate.

Approach	Useful when	Main risk
Linear calendar-time extrapolation	Accrual has been stable and the forecast interval is short	Assumes a constant event rate despite changing follow-up or enrollment
Cohort-based calculation	Entry dates and an event-time model are available	Depends on the model and on assumptions about censoring and future enrollment
Simulation across scenarios	Enrollment, event rates, or delays may vary and uncertainty matters	Results depend on the scenarios and distributions supplied to the simulation

A cohort-based calculation makes unequal follow-up explicit. Simulation can extend this approach by varying enrollment dates, event times, censoring, and reporting delays, but it does not remove the need to justify those inputs. Historical event rates may also be a poor guide when the current population, treatment, endpoint definition, or follow-up procedures differ. No projection method guarantees the future target date; forecasts should be updated when observed accrual or operational conditions change.

Common Mistakes

Assuming events accrue uniformly because recruitment is uniform. Recruitment counts participants; event counts also depend on time at risk. Use follow-up from each participant's entry date.
Giving every participant the same follow-up duration. This overstates expected events when enrollment is staggered. Calculate each participant's event probability over available follow-up, or integrate over the enrollment pattern.
Projecting events as though nobody is censored. State how dropout and administrative censoring affect observable events, and check whether the censoring assumptions are credible.
Calling the event-target date the report date. Add the expected processing delay only when translating to data availability, and state what that delay represents.
Reporting a precise date without checking the inputs. A precise numerical solution to a misspecified model is not a reliable forecast. Check enrollment, event-rate, censoring, and delay assumptions.

Key takeaway: A defensible event-date forecast follows participants from their actual entry dates, models the event opportunities they have by the forecast date, accounts for censoring, and keeps event occurrence separate from reporting and analysis readiness.

Check Your Understanding

Answer each question, then compare with the brief solutions.

Why can a forecast based on a constant monthly event count be misleading while recruitment is still increasing?
In the staggered-entry example, what assumption caused the shortcut to overstate expected events?
A target is forecast to occur in month 8, and the assumed data-processing lag is 1.5 months. What date does the forecast imply for event occurrence, and what date for analysis availability?
Under the independent constant-hazard model, explain qualitatively what happens to the probability of an observed event by a fixed follow-up time as the dropout hazard increases.

Solutions. (1) Participants have different and often increasing amounts of follow-up, so calendar-time event counts need not accrue at a constant rate. (2) It treated all participants as if they had been enrolled at the beginning of the 10-month period, giving later entrants too much follow-up. (3) The event target is forecast for month 8; with a 1.5-month lag, analysis availability is forecast for month 9.5. (4) The probability decreases because more participants leave observation before an event can be recorded; this conclusion uses the stated competing-hazard model.