Delay Mechanisms and Their Consequences

Why the Delay Mechanism Matters

In response-adaptive randomization, an allocation decision can use only information available at the time it is made. A participant’s endpoint may be defined, and eventually observed, but still be pending when the next participant is assigned. The delay mechanism determines which earlier responses have arrived and whether the responses currently available are a neutral subset of all responses.

The key distinction is not simply whether outcomes are delayed. It is how the time to outcome availability is determined. A delay can be fixed by the study schedule, vary randomly for reasons unrelated to the response, or vary with the response itself. Those mechanisms create different information sets at allocation decisions. This tutorial classifies them and works out their consequences without prescribing a particular update rule.

Definition: Let $T_i$ be the time participant $i$ enters the trial, $D_i$ the time from entry until that participant’s outcome is finalized and available, and $Y_i$ the eventual outcome. At a decision time $t$, define the availability indicator $$ R_i(t)=\mathbf{1}\{T_i+D_i<t\}. $$ Thus $R_i(t)=1$ means that participant $i$'s response is available immediately before the allocation decision at time $t$; $R_i(t)=0$ means it is still pending. This tutorial uses the convention that an outcome finalized exactly at $t$ is not available for that decision. A study may instead specify another rule for ties, but it must apply that rule consistently.

The indicator describes availability, not the outcome value: if $R_i(t)=0$, $Y_i$ still exists conceptually but has not yet been observed by the allocation process. In practice, an outcome may be finalized only after required follow-up, review, or data processing is complete. The relevant $D_i$ is the time until the response can actually be used, not necessarily the time until the underlying event occurred.

Three Delay Mechanisms

A delay is fixed when its duration is set by design rather than varying randomly across participants. For example, if every participant’s endpoint is finalized 14 days after enrollment, then $D_i=14$ days for all $i$. Fixed duration does not imply that all participants’ responses become available together: entry times $T_i$ may differ.

A delay is random when its duration varies stochastically. For the distinction in this tutorial, call the delay random but not outcome-dependent when, within the relevant treatment and design or enrollment-time stratum, its distribution does not depend on the eventual response. In symbols, this means the conditional distribution of $D_i$ is the same for different values of $Y_i$ in that stratum. The delay may still vary from person to person, and the probability of availability can depend on how long each person has been followed.

A delay is outcome-dependent when its distribution varies with the eventual response, after making the relevant design and follow-up conditions comparable. For instance, a symptom endpoint might be recorded promptly after a severe event but only after a scheduled visit if no event occurs. This is still a random delay in the broad sense, but it is useful to distinguish it because the available responses can become enriched for some outcome values.

Mechanism	What determines the delay?	Consequence for availability
Fixed	A specified duration or schedule	Availability follows from entry time and the specified schedule.
Random, not outcome-dependent	Stochastic variation unrelated to the response within comparable strata	The number available varies, but availability does not preferentially select response values within those strata.
Outcome-dependent	The delay distribution differs according to the eventual response	The responses available at a decision can have a different outcome mix from the responses still pending.

These descriptions concern the mechanism generating the delay, not whether every realized delay looks typical. A single unusually long delay does not establish outcome dependence. Conversely, a delay schedule can appear similar across the trial overall while still depending on the response within a relevant subgroup. Comparisons should therefore account for the design features that determine follow-up opportunity, such as time since entry.

Fixed Delays: Availability Follows the Calendar

Worked Example: A Fixed Three-Day Delay

Suppose three hypothetical participants enter at days $T_1=0$, $T_2=2$, and $T_3=4$. Every outcome is finalized exactly three days after entry, so $D_i=3$. Which responses are available immediately before an allocation decision on day $7$?

For participant 1, $T_1+D_1=0+3=3<7$, so $R_1(7)=1$. For participant 2, $T_2+D_2=2+3=5<7$, so $R_2(7)=1$. For participant 3, $T_3+D_3=4+3=7$, which is not strictly less than 7, so $R_3(7)=0$ under the stated tie convention.

Two of the three responses are available; the third is pending. The delay is fixed, but the participants do not all become available at the same calendar time because they entered on different days. If the third response were finalized just before the decision, it would be available under the definition; a timestamp and tie rule are therefore operationally important.

With fixed delays, the available set can often be anticipated from enrollment times and the study schedule. However, predictable timing does not mean there is no information limitation: participants who entered recently will still have pending outcomes. If outcome distributions vary across calendar time or enrollment cohorts, availability can also be associated with those features even though the duration itself is fixed. The mechanism should be described conditional on such relevant features rather than assumed independent of everything.

Random Delays: Availability Varies Across Participants

For random delays, availability at time $t$ depends on the probability that a participant’s delay has ended by then. If participant $i$ has been followed for $a_i=t-T_i$ time units, and the delay distribution in that participant’s relevant stratum is $F_D$, then the probability of availability is $P(D_i<a_i)$, with the strict inequality reflecting the tie convention. Participants with different follow-up ages can have different availability probabilities even under the same delay distribution.

Worked Example: Expected Availability With Random Delays

Consider eight hypothetical participants whose outcomes are all pending at entry. At a particular allocation decision, four have been followed for five days and four for two days. Suppose each delay is independent of the eventual response within this example’s single stratum, and the delay distribution gives $P(D_i<5)=0.75$ and $P(D_i<2)=0.25$.

For the four participants followed five days, the expected number available is $4(0.75)=3$. For the four followed two days, it is $4(0.25)=1$. Therefore the expected total number available is $$ 4(0.75)+4(0.25)=3+1=4. $$

Four is an expectation, not a guarantee that exactly four outcomes will be available in a realized trial. The actual number can be smaller or larger. Under the stated independence from response, random variation in reporting changes which participants are available but does not, within this stratum, favor one response value over another. The assumptions matter: the same conclusion need not hold if delay probabilities differ by response or if the relevant strata are mixed in a way that also relates to response.

A useful basic check is to compare availability rates at similar follow-up ages and under comparable design conditions. A raw comparison that combines participants observed for very different lengths of time can be misleading: those with more time to complete follow-up naturally have more opportunity to become available, even when the delay mechanism does not depend on outcome.

Outcome-Dependent Delays: Available Responses Can Be Selected

To see the selection effect precisely, consider participants in a fixed treatment, design, and follow-up stratum at a particular decision. Let $Y_i=1$ denote one response category, let $p=P(Y_i=1)$, and define $q_1=P(R_i(t)=1\mid Y_i=1)$ and $q_0=P(R_i(t)=1\mid Y_i=0)$. These availability probabilities can differ when delay depends on outcome.

Result: If $P(R_i(t)=1)=p q_1+(1-p)q_0>0$, then the proportion of response category $Y_i=1$ among available outcomes is $$ P(Y_i=1\mid R_i(t)=1) = \frac{p q_1}{p q_1+(1-p)q_0}. $$ If $p q_1+(1-p)q_0=0$, no outcome is available with positive probability in this stratum, so the conditional response proportion among available outcomes is undefined.

The denominator condition is essential: conditional probability given availability is defined only if availability has positive probability. To justify the result when that condition holds, the joint probability of response category 1 and availability is $P(Y_i=1,R_i(t)=1)=P(Y_i=1)P(R_i(t)=1\mid Y_i=1)=p q_1$. The total probability of availability is the sum over the two response categories, $p q_1+(1-p)q_0$. Dividing the joint probability by this positive total gives the displayed conditional probability.

If $q_1=q_0=q>0$, the expression reduces to $p q/q=p$, so the available outcomes preserve the response proportion in this stratum in expectation. If $q_1\ne q_0$, that equality generally fails. If $q=0$, the expression would be $0/0$, not $p$: there are no available outcomes from which to calculate a proportion.

Worked Example: Faster Availability for One Outcome

Suppose, hypothetically, that among participants in a comparable stratum, $p=0.40$ have response category 1. At the decision time, the availability probabilities are $q_1=0.80$ for category 1 and $q_0=0.50$ for category 0. The probability of being available is $$ p q_1+(1-p)q_0 =(0.40)(0.80)+(0.60)(0.50) =0.32+0.30=0.62. $$

Because $0.62>0$, the conditional proportion among available outcomes is defined. It is $$ P(Y_i=1\mid R_i(t)=1) =\frac{(0.40)(0.80)}{0.62} =\frac{0.32}{0.62} \approx 0.5161. $$

Thus the eventual response proportion is 0.40 in the hypothetical stratum, while the response proportion among outcomes available at this decision is about 0.516. The available outcomes overrepresent category 1 because category 1 is more likely to be reported by this time. This calculation describes the information available to an allocation rule; it does not say that the allocation rule must use this proportion or specify how pending outcomes should be handled.

Consequences, Diagnostics, and Limits

The immediate consequence of a delay mechanism is the composition and size of the information set at an allocation decision. Fixed delays create a schedule determined by entry times. Random delays make the size and membership of the available set variable. Outcome-dependent delays can change not only which responses are available but also their response distribution. Because an adaptive allocation rule responds to the information it receives, these mechanisms can alter the sequence of updates and allocations even if the eventual outcomes are unchanged.

Several practical checks help identify whether the assumed mechanism is plausible:

Verify that entry, endpoint, finalization, and allocation timestamps use a consistent time origin and that the tie rule is explicit.
Summarize the number pending and available at decision times, separated where appropriate by time since entry and treatment group.
Compare time to finalization across eventual response categories when outcomes can later be determined, while accounting for differing follow-up opportunity.
Investigate whether review procedures, visit schedules, or operational rules could make reporting speed depend on the response.
Check whether participants with longer follow-up are being compared with those who have had less time to report; they do not have the same opportunity to become available.

These checks have limits. A small trial may provide little evidence about differences in reporting times, and an absence of an obvious difference does not establish independence. In particular, if many outcomes are still pending, their eventual response categories are not yet known, so the mechanism cannot be fully assessed from the current allocation-time data alone. The delay classification is an assumption about the data-generating and reporting process, not a conclusion established merely by counting pending records.

The three mechanisms are useful idealizations, not necessarily mutually exclusive descriptions of every operational detail. A trial can have a fixed minimum follow-up period plus random processing time; its total delay then varies, even if part of the delay is fixed. A delay can also vary randomly and depend on response. When describing such a setting, state which part is scheduled, which part varies, and whether the conditional delay distribution differs by response. A single label should not conceal those distinctions.

Comparing the Mechanisms

Question	Fixed delay	Random, not outcome-dependent delay	Outcome-dependent delay
Can the exact available set be known in advance?	Often, if entry times and the schedule are known.	Usually not; only availability probabilities may be known.	Usually not; availability can differ by eventual response.
Does availability vary by follow-up age?	Yes, through entry time and the fixed duration.	Yes, through the chance the random delay has ended.	Yes, and also potentially by response.
Can the available response mix differ from the eventual response mix?	Not from response-dependent delay within a comparable stratum, but other time-related differences can matter.	Not from delay-response dependence within a comparable stratum, by the stated assumption.	Yes; the availability probabilities may differ by response.

No mechanism removes the basic limitation that pending outcomes cannot be directly read at an allocation decision. A predictable fixed schedule may help with planning but does not make pending responses observed. Random, response-independent delays still create uncertainty about when information will arrive. Outcome-dependent delays add the further concern that available responses can be a selected subset. Identifying the mechanism is therefore a prerequisite for interpreting what an adaptive allocation process has actually been able to use.

Common pitfall: “The delay is random, so the available responses are representative.” Randomness alone does not justify that conclusion. The relevant condition is that, within comparable treatment, design, and follow-up strata, availability does not depend on the eventual response. If $q_1\ne q_0$, the available response proportion can differ from the eventual proportion; if availability probability is zero, that conditional proportion is not defined at all.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

All participants have outcomes finalized exactly 10 days after enrollment. Is the delay fixed, and why might participants still become available on different calendar dates?
At a decision time, $P(Y_i=1)=0.30$, $q_1=0.60$, and $q_0=0.60$ in a comparable stratum. Calculate the availability probability and the response proportion among available outcomes.
Suppose $p=0.25$, $q_1=0.40$, and $q_0=0.80$. Calculate the availability probability and $P(Y_i=1\mid R_i(t)=1)$.
If $p=0$ and $q_0=0$, is the response proportion among available outcomes defined? Explain.
Name one useful diagnostic for a possible outcome-dependent delay and one reason it may not settle the question.

Solutions. (1) Yes. The duration is fixed at 10 days, but different enrollment times give different finalization times $T_i+10$. (2) Availability is $0.30(0.60)+0.70(0.60)=0.60$. Since the availability probabilities are equal and positive, the available response proportion is $0.30(0.60)/0.60=0.30$. (3) Availability is $0.25(0.40)+0.75(0.80)=0.10+0.60=0.70$. The response proportion among available outcomes is $0.10/0.70\approx0.1429$. (4) No. The availability probability is $0( q_1)+1(0)=0$; conditioning on availability is undefined because no outcomes are available with positive probability. (5) Compare time to finalization across eventual response categories among participants with comparable follow-up opportunity. Such a comparison may have limited evidence in a small trial, and unresolved outcomes may prevent assessment of their eventual categories.