Timing Interim Looks After a Delayed Effect

Why Look Timing Matters When Effects Are Delayed

An interim analysis scheduled soon after a treatment-effect lag may include many events but little information about the period in which the treatment is expected to help. Conversely, waiting longer after the lag can provide more post-lag events for efficacy monitoring, although the number accrued depends on enrollment, follow-up, censoring, and the event rate. The design question here is not simply how many total events have occurred: it is how much of the available event information arises after the prespecified lag.

The calculations below treat the lag as a fixed time since each participant’s randomization. This distinction matters with staggered enrollment: at one calendar-time look, participants can have different follow-up durations, and some may not yet have reached the lag. The introductory material on information time establishes why monitoring is indexed by information rather than calendar time; here we focus on how candidate calendar looks map to post-lag information.

Definition: Let \(\tau\) be the prespecified lag from randomization to the start of the treatment effect, and let \(T\) be a participant’s follow-up at an interim look. A post-lag event is an observed event occurring after follow-up time \(\tau\). The expected post-lag event count is the sum, over participants, of their probabilities of having an observed event after \(\tau\) by the look. It is an event-count summary, not automatically the exact information of a weighted logrank statistic.

Model and Planning Approximation

For a transparent calculation, suppose event times in the reference model are exponentially distributed with hazard \(\lambda\) per month, so survival through follow-up \(t\) is \(S(t)=e^{-\lambda t}\). Assume the treatment effect is absent through \(\tau\), and suppose the same hazard is used for both randomized groups as a planning reference. Also assume equal allocation, independent event and censoring processes, complete observation through the stated look unless otherwise specified, and a continuous event-time distribution. These assumptions make the event probabilities easy to calculate; real designs may need to account for censoring, changing hazards, unequal allocation, or treatment effects after the lag.

For a participant followed for \(t>\tau\), the probability of an event after the lag and by the look is the probability of surviving to \(\tau\) and then having an event by \(t\):

P(\tau < X \le t)=S(\tau)-S(t) =e^{-\lambda\tau}-e^{-\lambda t}.

If \(t\le\tau\), this probability is zero under a continuous event-time model. An event exactly at \(\tau\) has probability zero, so a look at the lag itself does not provide post-lag event information in this model.

For \(n\) participants with the same follow-up \(t\), the expected post-lag event count is \(n\{S(\tau)-S(t)\}\). With equal allocation, a rough null-reference approximation for the ordinary unweighted logrank information is total events divided by four. Applying the same scaling to post-lag events gives a useful planning proxy, not an exact information decomposition for every sequential statistic. For weighted logrank monitoring, information depends on the prespecified weights and the risk sets; in discrete form its variance involves terms of the form \(w_j^2V_j\), where \(w_j\) is the weight and \(V_j\) is the risk-set variance contribution at event time \(j\). Thus a post-lag event count alone cannot replace calculation of the planned weighted information.

Worked Example: A Look at the Lag Versus a Look Later

Suppose 160 hypothetical participants are randomized at the same time, equally between groups. Use \(\lambda=0.10\) events per month and a prespecified lag of \(\tau=6\) months. Compare looks at 6 and 12 months, assuming no censoring before either look.

At 6 months, no participant has follow-up beyond the lag. Therefore \(P(\text{post-lag event by 6 months})=0\), and the expected post-lag count and its event-count-based information proxy are both zero.

At 12 months, the post-lag event probability is

e^{-0.10(6)}-e^{-0.10(12)} =e^{-0.60}-e^{-1.20} \approx 0.5488-0.3012=0.2476.

Using the unrounded probability \(0.247617\), the expected post-lag event count is \(160(0.247617)=39.62\). The approximate post-lag information is \(39.62/4=9.905\). For comparison, the probability of any event by 12 months is \(1-e^{-1.20}=0.698806\), giving \(160(0.698806)=111.8\) expected total events and a total information proxy of \(111.8/4=27.95\). The post-lag share of expected events, and hence of this simple proxy, is \(39.62/111.8=35.4\%\). The later look has accumulated post-lag information, whereas the look at the lag has not.

This calculation assumes the reference event hazard applies to both groups and that all participants remain observed through 12 months. If post-lag hazards differ by treatment, or censoring removes participants, the expected post-lag count must be recalculated under those features.

Staggered Enrollment Changes the Mapping

With staggered enrollment, a calendar-time look does not give every participant the same follow-up. Suppose \(n\) participants enroll uniformly over calendar time from 0 to \(A\) months, and a look occurs at calendar time \(T\ge A\). Their follow-up durations are spread from \(T-A\) to \(T\). Under the same exponential reference model and complete observation, average the individual post-lag event probabilities over those follow-up durations:

p_{\text{post}}(T) =\frac{1}{A}\int_0^A \left(e^{-\lambda\tau}-e^{-\lambda(T-u)}\right) \mathbf{1}(T-u>\tau)\,du.

Here \(u\) is enrollment time and \(\mathbf{1}(\cdot)\) equals 1 when its condition holds and 0 otherwise. Participants with \(T-u\le\tau\) contribute no post-lag event probability at that look. Multiplying the average probability by \(n\) gives the expected post-lag count.

Worked Example: Uniform Enrollment Over the First 12 Months

Consider 100 hypothetical participants enrolled uniformly over months 0 through 12, with the interim look at month 12. Let \(\tau=6\) months and \(\lambda=0.10\) per month. This deliberately simple model assumes complete observation and a common reference hazard. Participants enrolled after month 6 have not yet reached the lag at the look; earlier enrollees have follow-up between 6 and 12 months.

Equivalently, follow-up \(x\) is uniformly distributed from 0 to 12 months at this look. Only \(x>6\) contributes post-lag events, so the average probability is

\begin{aligned} p_{\text{post}}(12) &=\frac{1}{12}\int_6^{12} \left(e^{-0.60}-e^{-0.10x}\right)\,dx\\ &=\frac{6e^{-0.60}-10\left(e^{-0.60}-e^{-1.20}\right)}{12}\\ &\approx 0.06806. \end{aligned}

The expected post-lag count is \(100(0.068058)=6.806\), and the corresponding event-count-based information proxy is \(6.806/4=1.701\). To compare with all events, the average probability of any event by the look is

1-\frac{1}{12}\int_0^{12}e^{-0.10x}\,dx =1-\frac{1-e^{-1.20}}{1.20} \approx 0.4177.

That gives 41.77 expected total events and a total information proxy of 10.44. The expected post-lag share is \(6.806/41.77\approx16.3\%\), much lower than in the fixed-cohort 12-month example. The difference arises because many participants at this calendar look have had less than 6 months of follow-up, or only a short period beyond the lag.

Comparing Candidate Looks

A useful timing comparison keeps the assumptions fixed and calculates cumulative post-lag information at each proposed look. The following example uses a fixed cohort, which isolates the effect of moving the look later. Its values are expected counts, not guaranteed event totals at a realized interim analysis.

Worked Example: How Much Post-Lag Information Accrues Between Looks?

Suppose 200 hypothetical participants are randomized together, equally between groups, with \(\lambda=0.10\) per month and \(\tau=6\) months. Compare looks at 8 and 12 months. Assume complete observation and use the common-hazard reference model.

At 8 months, the post-lag probability is

e^{-0.60}-e^{-0.80} \approx 0.5488-0.4493=0.09948.

Thus the expected post-lag count is \(200(0.099483)=19.90\), and its approximate information proxy is \(19.90/4=4.974\). The expected total event count by 8 months is \(200(1-e^{-0.80})=110.1\), giving a total proxy of 27.53; about \(19.90/110.1=18.1\%\) of expected events are post-lag.

At 12 months, the post-lag probability is \(e^{-0.60}-e^{-1.20}=0.247617\). The expected post-lag count is \(200(0.247617)=49.52\), and its proxy is \(49.52/4=12.38\). The expected total count is \(200(1-e^{-1.20})=139.8\), with proxy 34.94; the post-lag share is \(49.52/139.8=35.4\%\).

Between the two looks, the expected number of additional events occurring after month 8 and by month 12 is \(200(e^{-0.80}-e^{-1.20})=29.63\), corresponding to an incremental event-count proxy of \(29.63/4=7.407\). This increment is smaller than the increase in cumulative post-lag counts only if one mistakenly includes events from months 6 through 8 twice; here the cumulative post-lag count at 8 is already 19.90, so the difference \(49.52-19.90=29.63\) agrees with the interval calculation.

Justification and Design Checks

The central calculation follows directly from conditioning on survival to the lag. Under the exponential model, a participant must first remain event-free through \(\tau\), which has probability \(e^{-\lambda\tau}\). Conditional on reaching \(\tau\), the probability of an event before follow-up \(t\) is \(1-e^{-\lambda(t-\tau)}\). Multiplying gives

e^{-\lambda\tau}\left\{1-e^{-\lambda(t-\tau)}\right\} =e^{-\lambda\tau}-e^{-\lambda t}.

This also shows why a look just after the lag can have little post-lag information: the post-lag interval \(t-\tau\) is short, so \(1-e^{-\lambda(t-\tau)}\) is small. In a trial with staggered enrollment, that interval varies by participant, and averaging the probabilities—not plugging the average follow-up into the fixed-follow-up formula—is the appropriate calculation under the stated model.

Before relying on a projection, check whether its assumptions resemble the planned design. In particular, verify the lag scale (time since randomization versus calendar time), accrual pattern, censoring and dropout, event-rate assumptions, randomization ratio, and the analysis weights. If hazards vary over time, replace the exponential survival model with an appropriate survival curve and integrate the resulting individual post-lag event probabilities over enrollment and censoring. If post-lag treatment effects alter event rates, calculate expectations under the relevant group-specific hazards rather than treating the common-hazard reference as the actual alternative.

Specify the lag and look scale.
State whether the lag is measured from each participant’s randomization and identify each proposed calendar-time or follow-up-time look.

Project follow-up at each look.
For staggered enrollment, account for the distribution of follow-up durations rather than assigning every participant the same duration.

Estimate post-lag event contributions.
Use the specified survival and censoring assumptions to obtain expected post-lag events; calculate weighted logrank information using its own weights and risk sets.

Check the sequential design after timing changes.
Changing a look’s information fraction changes the monitoring schedule. Use the planned sequential-boundary calibration for the actual information times; do not assume an earlier calibration remains valid unchanged.

Interpretation, Alternatives, and Limitations

Calendar-spaced looks are easy to communicate and can support operational planning, but they may occur before enough participants have passed the lag. Event-driven looks instead target a number of events or a prespecified information level; they can be more directly tied to statistical precision, but the calendar date is then uncertain and the post-lag share still needs projection. A design can also specify a minimum amount of post-lag follow-up or information before an efficacy look. These are alternative scheduling principles, not substitutes for calculating the expected information under the chosen analysis.

Expected values are not guarantees. Random variation in event times, enrollment, and censoring means the realized post-lag count can be below or above its projection. The simple events-divided-by-four proxy is especially limited: it is an approximation for equal allocation and an unweighted logrank variance under a null-reference setting, not a universal formula for weighted monitoring. Weighted statistics can emphasize or downweight time periods, so two looks with the same post-lag event count need not have the same weighted information.

A common mistake is to say, “The look is after the lag, so it must contain plenty of treatment-effect information.” Being after the lag is necessary for post-lag events, but not sufficient for substantial information. The amount also depends on how long participants have been followed after the lag, how many remain at risk, event rates, censoring, and the analysis weights. The correction is to calculate post-lag event probabilities or information under explicit assumptions, then evaluate the resulting information times within the sequential design.

Key takeaway: The useful timing quantity is not simply the calendar date of a look. It is the expected post-lag information available at that look, calculated from participant follow-up and the event process, with the weighted analysis information evaluated using its actual weights and risk sets.

Check Your Understanding

Answer each question using the stated assumptions. Brief solutions follow the questions.

With a fixed cohort, exponential hazard \(\lambda=0.08\) per month, and lag \(\tau=5\) months, write the probability of a post-lag event by month 10.
Why does a look exactly at the lag have zero post-lag event probability in a continuous event-time model?
At a calendar look, why is it generally incorrect under staggered enrollment to calculate post-lag events by substituting average follow-up into the fixed-follow-up formula?
Under equal allocation, when is total events divided by four a reasonable information approximation, and why is it not exact for a weighted logrank statistic?

Solutions. (1) The probability is \(e^{-0.08(5)}-e^{-0.08(10)}=e^{-0.40}-e^{-0.80}\approx0.2210\). (2) No participant has follow-up greater than the lag, and an event occurring exactly at the lag has probability zero under a continuous distribution. (3) The post-lag probability is nonlinear in follow-up and is zero for participants who have not passed the lag; average the participant-specific probabilities over their follow-up distribution. (4) It is a rough null-reference approximation for the ordinary unweighted logrank information with equal allocation; a weighted statistic’s information also depends on the squared weights and risk-set variance contributions.