Conditional Forecasts From Current Trial Data

Why Condition on the Trial’s Current State?

A projection made before trial launch must represent uncertainty about who will enroll, when they will enroll, and when their events will occur. At an interim cutoff, some of that uncertainty has been replaced by observed information: the trial has a known enrollment history, a known number of events, and participants whose event-free follow-up is already known. A useful forecast should condition on that information rather than restart the projection as if the trial were still at launch.

This tutorial concerns forecasts of the remaining time until a specified event target is reached. It is not enough to calculate an event rate from the past and extend it mechanically. Participants still at risk have already remained event-free for their observed time, and future entrants have not yet accumulated any follow-up. Those are different conditional situations and generally imply different probabilities of an event by a future date.

Assumptions and Conditional Event Probabilities

Let \(S(a)\) denote the probability that a participant remains free of the event through \(a\) units of follow-up from enrollment. For a participant who is known to be event-free at the cutoff after \(a\) units of follow-up, the probability of an event in the next \(u\) units is not generally \(1-S(u)\). It is conditional on having remained event-free through \(a\).

Here \(T\) is the time from enrollment to the event. This expression follows by subtracting the conditional probability of surviving event-free for a further \(u\) units, \(P(T>a+u\mid T>a)=S(a+u)/S(a)\), from one. If the event hazard changes over follow-up, the ratio uses the survival model at the participant’s actual follow-up age. Applying a probability from enrollment to someone already known to be event-free would ignore the information in their observed history.

A forecast also needs a model for future enrollment and a clear rule for how follow-up, loss to follow-up, and administrative censoring affect event observation. The examples below assume complete event ascertainment through the forecast date, no competing event, and a specified event-time distribution. If those assumptions do not fit the endpoint or trial operations, event probabilities must account for the relevant censoring or competing-event process. Participants with observed events contribute to the current event count, not to the future risk set. A participant known to be event-free and still followed contributes conditional future event probability.

From Conditional Probabilities to a Target-Date Forecast

Suppose \(D_c\) denotes the information recorded at cutoff \(c\), and \(K\) is the target number of events. Let \(E_c\) be the number of events already observed. The number still required is \(r=K-E_c\). For each current event-free participant \(i\), let \(q_i(u)\) be the conditional probability of an event by \(u\) time units after the cutoff. If future entrants are also expected, their probabilities depend on their future enrollment times and their follow-up from enrollment.

Under the stated independence assumption, the expected additional events by \(u\) are the sum of the relevant event probabilities. Thus the expected total event count is the observed count plus that sum. A commonly reported planning quantity is the value of \(u\) at which the expected total reaches \(K\). This is a root of an expected-count equation. It is not automatically the mean, median, or a confidence limit for the random time at which the target will actually be reached.

where \(R_c\) is the set of participants who are event-free and under follow-up at cutoff \(c\). A forecast of the target date based on expected counts solves \(E\{N(c+u)\mid D_c\}=K\). If an operational decision depends on the chance of reaching the target by a particular date, the predictive distribution of the event count or target time is more informative than this expected-count crossing alone.

Conditioning on Current Participants and Future Enrollment

When enrollment continues, the forecast combines two groups. Current event-free participants are conditioned on their event-free status at the cutoff. Future participants enter at later times and accrue follow-up only after enrollment. A participant enrolling \(s\) months after the cutoff has, at a forecast horizon \(u\geq s\), only \(u-s\) months of follow-up. Counting that participant as though they had been at risk since the cutoff overstates their chance of an event by the forecast date.

Expected Crossing Versus the Distribution of Target Time

A single expected-count crossing can conceal substantial variation in when the target will actually be reached. To quantify that variation, one can calculate the distribution of future event counts at specified horizons or simulate future event times conditional on the current risk set and enrollment plan. The resulting target-time distribution can be summarized by a median or prediction interval. These summaries answer different questions from the root of the expected-count equation.

For example, under independent participant event times, the number of future events by a fixed horizon has a distribution determined by the participants’ individual conditional probabilities. With equal probabilities it is binomial; with unequal probabilities it is a sum of independent Bernoulli variables with differing probabilities. A simulated target time is the first future time at which the simulated event count, added to the observed count, reaches \(K\). A forecast interval describes uncertainty in that future time under the specified model; it is not a confidence interval for an unknown event date.

Justification and Practical Checks

The conditional survival formula and the expected-count calculation follow from basic probability. For a participant known to have survived event-free through \(a\), the conditional probability of remaining event-free for another \(u\) units is

The joint event in the numerator is simply \(T>a+u\), which already implies \(T>a\). Taking the complement yields the conditional event probability. Summing these individual probabilities gives the expected number of future events because expectation is additive; independence is not required for that expectation. Independence is needed for the simple product calculation in the previous example and for many standard count distributions, but dependence would affect predictive uncertainty.

Before interpreting a forecast, check that its inputs describe the cutoff correctly: event counts are not projected again, participants with observed events are not left in the event-free risk set, and each remaining participant’s observed event-free duration is reflected in their conditional probability unless the chosen model has a property such as exponential memorylessness. Check also that future enrollment starts after the cutoff and that follow-up time for each future entrant begins at enrollment. A negative number of events required, or an expected event count that does not increase toward the target under the assumed schedule, signals that the target definition or inputs need review.

How This Forecast Differs from a Pre-Launch Projection

A pre-launch projection averages over future enrollment and event histories that have not yet occurred. A conditional forecast uses actual enrollment and event information through the cutoff, so the observed event count is already known and participants’ current event-free status is informative. The two forecasts can differ even when they use the same underlying event-time model: conditioning updates the starting state, rather than merely changing the calendar date on an unchanged projection.

Conditional forecasting does not make future event rates known with certainty. The examples treat the event model and enrollment plan as specified to make the calculations transparent. In practice, future enrollment may depart from plan, and the assumed event-time model may be uncertain. A forecast should therefore be understood as conditional on its inputs and assumptions. When a decision requires the chance of meeting a target by a deadline, report a probability or predictive range rather than presenting the expected-count crossing as a certain date.

Common Mistakes and Limitations

• Restarting the event clock at the cutoff. This is justified by the exponential model’s memoryless property, but not by a general survival model. With a nonconstant hazard, use the conditional survival ratio at each participant’s observed follow-up age.
• Projecting the observed events again. Events already counted at cutoff are fixed historical data. Forecast only additional events and add them to the observed count once.
• Giving future entrants immediate follow-up. Their event probabilities must reflect their actual or planned enrollment times.
• Calling an expected-count date the expected target date. The solution to an expected-count equation need not equal the mean or median of the target-time distribution.
• Ignoring how events are observed. If loss to follow-up, administrative censoring, or competing events affect ascertainment, the forecast probability must represent that process as well as the event-time model.

The main alternatives answer distinct questions. A deterministic expected-count projection is direct and easy to communicate, while a predictive distribution represents the range of possible target times under the model. A pre-launch projection remains useful for planning before trial-specific data exist, but once current trial data are available it does not use the information that defines the actual risk set. None of these summaries removes the need to state the event model, enrollment assumptions, and cutoff clearly.