Random Assignment Does Not Mean Balanced Enrollment Times
Random assignment is essential to a response-adaptive trial, but it does not guarantee that the realized treatment groups have the same distribution of enrollment times. This distinction matters when prognosis changes over calendar time. If participants who enroll earlier tend to have different outcomes from those who enroll later, a time imbalance can make the observed treatment groups look different even when treatment has no effect.
The key issue is not that randomization has failed. Rather, random assignment governs how each participant is assigned given the information available at that decision. It does not force the assignments that actually occur to balance every characteristic, including time. Moreover, when allocation probabilities change substantially, the assignment mechanism can make a time imbalance more likely in one direction than another.
The history \(H_i\) can include prior assignments and observed responses. Thus, in a response-adaptive trial, \(p_i\) may change as feedback accumulates. Even if the protocol is followed exactly and each assignment is randomized correctly, the resulting \(A_i\) values may cluster unevenly across calendar periods.
It is useful to separate two sources of imbalance. Chance imbalance occurs because random draws can produce unequal groups within a period, even when the allocation probabilities are constant. Probability-driven imbalance occurs when the allocation probabilities themselves favor one arm at some times and the other arm at other times. A trial can have both.
What the Allocation Probabilities Imply
Consider a set of participants whose enrollment times and treatment probabilities are fixed in advance. For any calendar-time window \(W\), the expected number assigned to treatment is the sum of their treatment probabilities. The expected number assigned to control is the sum of their control probabilities. These expectations need not be equal, even if the expected total number assigned to each arm over the whole trial is equal.
A useful signed measure of the allocation pattern is \(\sum_i(2A_i-1)(t_i-\bar t)\), where \(\bar t\) is the mean enrollment time for all participants. Positive values indicate that treatment assignments tend to occur later than control assignments; negative values indicate the reverse. With fixed probabilities, the expected value of this measure is \(\sum_i(2p_i-1)(t_i-\bar t)\). If probabilities are higher earlier and lower later, that expectation tends to be negative. This is a statement about an allocation-weighted contrast, not an exact formula for the expected difference between the two arms’ mean enrollment times, whose denominators are random.
Why Correct Randomization Can Still Leave Imbalance
A short derivation clarifies what randomization does guarantee. At the \(i\)th assignment, conditional on the history \(H_i\), the assignment has mean \(p_i\). Therefore, the assignment’s deviation from its probability, \(A_i-p_i\), has conditional mean zero. If enrollment time \(t_i\) is known at that assignment, multiplying this deviation by the centered time does not change that conditional mean.
So the random draw does not systematically depart from its stated probability, conditional on the history. But \(A_i-p_i\) is not always zero: the realized assignment can differ from its probability. Also, the probabilities \(p_i\) can vary with time. The realized time pattern therefore combines the allocation probabilities with random deviations around them. Neither feature forces the treatment and control groups to have identical enrollment-time distributions.
When \(p_i\) depends on prior outcomes and assignments, the probability sequence can itself be random. The conditional mean-zero statement still applies at each assignment, but one cannot treat the entire realized sequence of probabilities as if it had necessarily been fixed independently of the earlier random draws. The safe interpretation is sequential: each assignment follows its conditional probability; balance across time is not guaranteed by that fact.
Worked Example: Equal Expected Totals, Different Expected Time Patterns
Suppose six hypothetical participants enroll in an early period and six in a late period. After interim feedback, the protocol uses \(p_i=0.8\) in the early period and \(p_i=0.2\) in the late period. Assume these probabilities are fixed for the periods in this example and assignments are generated by independent random draws.
Expected counts. The expected early treatment count is \(6(0.8)=4.8\), and the expected late treatment count is \(6(0.2)=1.2\). Thus the expected total treatment count is \(4.8+1.2=6\). The expected control counts are \(6(0.2)=1.2\) early and \(6(0.8)=4.8\) late, also totaling 6.
Equal expected total counts do not mean equal time patterns. The expected counts place most treatment assignments early and most control assignments late. If early and late are coded as times 0 and 1, the allocation-weighted mean time for treatment is \(1.2/6=0.2\), while the corresponding mean from the expected control counts is \(4.8/6=0.8\). These are summaries of the expected allocation table, not claims that the expectation of a random arm mean is exactly a ratio of expectations.
A particularly separated realized pattern is all six early participants assigned to treatment and all six late participants assigned to control. Its probability under these fixed probabilities is \((0.8)^6(0.8)^6=(0.8)^{12}=0.06872\). This is not a rare impossibility: it is one possible result of correctly randomized assignments, and it produces complete separation of the arms by period.
How Time Imbalance Can Mimic an Outcome Difference
A time imbalance matters for a treatment comparison when outcomes also vary with time. Suppose, for illustration, that treatment has no effect and that the outcome is 10 units for each early participant and 20 units for each late participant. If the treatment group contains more early participants, its average outcome can be lower simply because of the enrollment-time pattern. The outcome difference is not, by itself, evidence of a treatment effect.
Worked Example: A Null Treatment Effect With an Apparent Difference
Consider eight hypothetical participants: four enroll early and four late. Assume every participant’s outcome would be 10 units if enrolled early and 20 units if enrolled late, regardless of assignment. Thus, there is no treatment effect. Suppose the allocation probabilities are \(0.75\) for treatment early and \(0.25\) for treatment late. One possible realized allocation has three early participants and one late participant assigned to treatment, with the remaining participants assigned to control.
Check the probability of this allocation pattern. The probability of exactly three of the four early participants receiving treatment is \(\binom{4}{3}(0.75)^3(0.25)=0.4219\). The probability of exactly one of the four late participants receiving treatment is \(\binom{4}{1}(0.25)(0.75)^3=0.4219\). Under the independent draws assumed here, the probability of both count patterns is \(0.4219^2=0.1780\).
Calculate the observed outcome means. The treatment group has three outcomes of 10 units and one of 20 units, so its mean is \((3(10)+1(20))/4=12.5\) units. The control group has one outcome of 10 units and three of 20 units, so its mean is \((1(10)+3(20))/4=17.5\) units. The treatment-minus-control difference is \(12.5-17.5=-5.0\) units.
The observed difference is entirely explained by the groups’ different time composition in this example: treatment has one late participant out of four, while control has three. Since the hypothetical outcome is identical under treatment and control for every participant, the difference does not represent a treatment effect.
This example shows why an observed imbalance is not a paradox and does not prove that randomization was violated. It does show why a treatment comparison can be misleading if temporal drift is relevant and the analysis ignores the assignment-time pattern. The earlier material on calendar-time drift and allocation-probability timelines provides useful context for recognizing these patterns.
Diagnostics and Comparisons
A practical diagnostic begins with the allocation probabilities and the enrollment record, not only with the final outcome comparison. An arm-by-time enrollment table reveals whether the realized groups differ across periods. A probability timeline shows whether the design itself favored different arms at different stages. Read together, they help separate a chance departure from the probabilities from imbalance that was encouraged by the probability schedule.
- Compare counts by period. Look at the number assigned to each arm in prespecified or substantively meaningful enrollment windows. Overall arm totals can match while period-specific counts differ sharply.
- Review the probability history. Identify whether \(p_i\) changed substantially and whether higher probabilities for one arm occurred mainly earlier or later.
- Compare observed counts with expected counts. For a fixed probability schedule, sum \(p_i\) and \(1-p_i\) within each window. A discrepancy from those expectations is random variation; the expectations themselves can still imply a time imbalance.
- Relate the pattern to outcome time trends. If outcomes differ by calendar period, an arm-by-time imbalance has a plausible route to affecting the unadjusted arm comparison.
Simple randomization with a constant treatment probability can still produce chance imbalance in enrollment time. In contrast, changing allocation probabilities can make the expected distribution of assignments vary with time, as well as leaving chance imbalance around that pattern. Methods that explicitly target balance across time, such as temporal blocking, address a different design objective; ordinary sequential randomization alone does not do so.
| Feature | What it supports | What it does not guarantee |
|---|---|---|
| Random assignment conditional on \(H_i\) | Each assignment follows its stated conditional probability | Equal treatment and control counts in every time period |
| Equal expected total arm counts | The allocation plan is balanced in expectation overall | Equal arm-specific distributions of enrollment time |
| Constant allocation probability | No planned time shift in allocation probability, if the probability is constant over time | No realized time imbalance |
| Changing allocation probabilities | Adaptation to the information used by the protocol | Temporal comparability of the realized arms |
Common Mistake and Limitations
Wrong reasoning: “Treatment was randomly assigned, so the treatment and control groups must be balanced in enrollment time. Any observed time difference must therefore be negligible or irrelevant.”
Correction: Random assignment makes assignments random according to the specified probabilities; it does not guarantee balance in a finite realized sample. If those probabilities vary over time, the expected allocation pattern can also vary over time. When outcomes drift with calendar time, the resulting imbalance can matter even if the overall treatment and control counts are equal.
An arm-by-time table is a diagnostic, not a proof of bias or evidence that the randomization procedure was improperly implemented. A chance imbalance can occur under a valid design. Conversely, equal total counts do not establish temporal balance. This tutorial does not prescribe a particular adjustment method: the appropriate analysis depends on the estimand, the allocation mechanism, the outcome process, and the trial protocol. The essential conclusion is narrower: randomization does not erase time imbalance after it occurs.
Worked Example: Checking Whether Balance Is Guaranteed
A hypothetical trial has four early and four late enrollees. Following interim feedback, the protocol assigns with probability \(0.8\) to treatment early and \(0.2\) late. Consider the question: does this rule guarantee exactly two treatment assignments in each period?
Calculate the probability of exactly two treatment assignments early. It is \(\binom{4}{2}(0.8)^2(0.2)^2=6(0.64)(0.04)=0.1536\).
Calculate the probability of exactly two treatment assignments late. It is \(\binom{4}{2}(0.2)^2(0.8)^2=0.1536\).
Because the assignment draws are independent under the stated fixed schedule, the probability of exactly two treatment assignments in each period is \(0.1536(0.1536)=0.02359\), or about 2.36%. Thus, the rule does not guarantee the desired balance. For comparison, with a constant probability of \(0.5\) in both periods, the probability of exactly two treatment assignments in each period would be \([\,\binom{4}{2}(0.5)^4\,]^2=0.1406\). Neither rule guarantees balance, but the changing probabilities make this particular balanced pattern substantially less likely.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a time window, five participants each have \(p_i=0.6\). What are the expected treatment and control counts in that window?
- Why does a constant treatment probability fail to guarantee equal enrollment times between arms?
- Suppose outcomes improve over calendar time and treatment assignments are more common early. Give one reason an unadjusted treatment-minus-control outcome difference could be negative even if treatment has no effect.
- What two records would you examine to distinguish chance imbalance from imbalance encouraged by changing probabilities?
Solutions. (1) The expected treatment count is \(5(0.6)=3\), and the expected control count is \(5(0.4)=2\). Expected counts are not guarantees about the realized counts. (2) Random draws with a constant probability can still cluster unevenly by time; randomization controls probabilities, not every realized arrangement. (3) If outcomes are higher later and treatment assignments occur more often early, treatment participants may have lower outcomes because of their earlier enrollment, producing a negative contrast without a treatment effect. (4) Examine the arm-by-time enrollment table and the allocation-probability timeline; together they show the realized pattern and whether the design’s probabilities favored that pattern across time.