Why Allocate Participants to Both Sequences?
In a two-period crossover study, each participant follows a treatment sequence. The labels TR and RT indicate the order in which the test treatment \(T\) and reference treatment \(R\) are assigned: TR means T in period 1 and R in period 2, while RT means R in period 1 and T in period 2. Sequence allocation is the assignment of participants to one of these two orders.
The allocation matters because treatment and period can otherwise be entangled. A participant's response may differ between periods for reasons unrelated to formulation, such as a general period-to-period change. If everyone receives T in period 1 and R in period 2, any observed difference between treatments is also a difference between periods. Assigning participants to both sequences places each treatment in each period, making those effects distinguishable under the study model.
Allocation to both sequences is not the same as independently choosing a treatment for each period. A participant receives the two treatments in the order set by the assigned sequence. Nor does allocation guarantee that the numbers assigned to each sequence will be exactly equal: that depends on the planned allocation and the random assignments actually made.
How Both Sequences Separate Treatment from Period
A simple additive model helps explain the design logic. Suppose a measured response can be represented by an overall level, a participant-specific contribution, a treatment contribution, a period contribution, and random error. The treatment contribution is assumed not to change from one period to the other. This explanation also assumes no material carryover into the next period and no treatment-by-period interaction; those assumptions are necessary for the simple separation described here.
With only TR participants, T is always given in period 1 and R is always given in period 2. A difference between their period averages therefore combines a treatment difference and a period difference. In symbols, if \(\tau_T\) and \(\tau_R\) are treatment contributions and \(\pi_1\) and \(\pi_2\) are period contributions, the difference between the R-in-period-2 average and the T-in-period-1 average contains both \((\tau_R-\tau_T)\) and \((\pi_2-\pi_1)\). One observed difference cannot by itself identify these two components.
Adding RT participants changes the pattern: T occurs in period 1 for RT and in period 2 for TR; R occurs in period 1 for TR and in period 2 for RT. Treatment and period are no longer tied together in a single way. The study can therefore compare treatments while accounting for a common period difference, subject to its model assumptions. This is the design rationale for using both sequences; the precise treatment contrast depends on the analysis model.
The key is not simply that the treatment orders differ. It is that the treatment-period combinations are represented by participants in both sequences. Equal sequence counts make the counts in those combinations equal in the planned allocation, but exact equality is not required for treatment and period to be distinguishable.
Worked Example: Planning an Equal Allocation
A hypothetical study plans to assign 40 participants equally to TR and RT. Determine the planned sequence counts and how many participants are planned to receive each treatment in each period.
An equal allocation divides the total by two:
The TR participants contribute 20 assignments to T in period 1 and R in period 2. The RT participants contribute 20 assignments to R in period 1 and T in period 2. Thus, the planned counts are 20 for T in period 1, 20 for R in period 1, 20 for T in period 2, and 20 for R in period 2.
This count pattern supports separation of treatment and period: both treatments appear in both periods, and each treatment-period combination has the same planned number of participants. It does not establish that observed responses will be equal or that the study will meet any particular equivalence criterion. Those questions depend on the outcome data and the analysis.
Planned Allocation and Realized Counts
An allocation plan describes the intended numbers or probabilities of assignment, while the realized counts are the numbers actually assigned. For example, assigning each participant to TR or RT with probability one-half gives equal expected counts, but a finite sample can still end up unequal. Random assignment does not mean that every possible study will have exactly half its participants in each sequence.
A small difference in realized counts does not automatically invalidate a study. It does mean that the design is not perfectly balanced in sequence counts, and the analysis must use the assignments and observations actually available rather than treating the planned counts as if they occurred. Larger imbalances can reduce precision, particularly if the smaller sequence has fewer usable participants. The allocation should be reviewed alongside completion and missing-data patterns.
Worked Example: Interpreting Unequal Realized Counts
In a hypothetical study, 31 participants are assigned with equal probability to TR or RT. The realized assignment is 17 to TR and 14 to RT. Calculate the allocation proportions and interpret the difference.
The proportions are:
The sequence-count difference is \(17-14=3\) participants. The allocation is therefore somewhat unequal, even though the plan used equal probabilities. This is a possible random outcome, not by itself evidence that the assignment procedure failed.
Because participants in both sequences were assigned, each treatment still occurs in both periods. The study should report the realized counts and retain the actual sequence assignments in its analysis. It should not discard three participants or reassign them after seeing outcomes merely to make the totals equal; such outcome-dependent changes can compromise the planned comparison.
What Goes Wrong with One Sequence?
If every participant is assigned to TR, then T is observed only in period 1 and R only in period 2. If every participant is assigned to RT, the reverse is true. Either way, treatment and period are perfectly linked: a treatment difference cannot be separated from a period difference using the crossover allocation alone. Increasing the number of participants in that one sequence would provide more observations of the same confounded pattern, not the missing treatment-period combinations.
A numerical illustration makes the ambiguity concrete. Suppose a hypothetical TR-only study has an average response of 100 units in period 1 under T and 110 units in period 2 under R. The observed increase is 10 units. It could reflect a treatment difference, a period difference, or a combination of both. Without participants receiving the opposite order, the design supplies no independent comparison that distinguishes these explanations.
Worked Example: Checking the Treatment-Period Pattern
A hypothetical crossover study has 12 participants in each sequence. Its observed averages, in arbitrary response units, are as follows. These values are illustrative only.
| Period | Average under T | Average under R |
|---|---|---|
| 1 | 100 | 104 |
| 2 | 103 | 107 |
The table shows that both treatments are represented in both periods: period 1 includes T observations from RT and R observations from TR; period 2 includes T observations from TR and R observations from RT. The within-period differences between the displayed averages are 104 − 100 = 4 units in period 1 and 107 − 103 = 4 units in period 2. The period-to-period change in the displayed T averages is 103 − 100 = 3 units, and the corresponding change under R is 107 − 104 = 3 units.
This pattern illustrates why both sequences are useful: the treatment comparison is not confined to one period, and the period change is visible under both treatments. The table alone is not a complete analysis of crossover data, and equal sequence counts do not prove that model assumptions hold. It does, however, confirm that the allocation provides the treatment-period combinations needed for the intended comparison.
Choosing and Monitoring an Allocation
Equal allocation is a common choice when the study's main aim is to compare the two treatments and there is no compelling reason to prefer one sequence. It gives the same planned number of participants to each order and, for a fixed total, makes the sequence counts as close as possible. An unequal allocation can still include both sequences and preserve the basic treatment-period pattern, but it gives fewer participants to one order and may make some aspects of the analysis less precise.
| Allocation choice | What it provides | Main consideration |
|---|---|---|
| Equal sequence allocation | The same planned count in TR and RT | A straightforward default when no other design objective favors an imbalance |
| Unequal allocation to both sequences | Both treatment orders remain represented | The smaller sequence may contribute less information, especially after losses |
| Allocation to one sequence only | Only one treatment order is observed | Treatment and period are confounded in the two-period pattern |
Sequence counts are only one part of design monitoring. Investigators should distinguish participants assigned to a sequence from those who complete both periods and contribute usable measurements. If withdrawals, missed periods, or protocol deviations differ between sequences, the completed-data counts may be more unequal than the assigned counts. Reporting both helps readers understand the data available for the comparison.
A common mistake is to reason, “The study planned equal numbers, so it is balanced,” even when the realized counts or completed observations differ. The correction is to inspect the actual assignments and available data. Conversely, a modest realized imbalance does not mean treatment and period are automatically confounded: the essential design feature is representation of both sequences, while the size and consequences of imbalance affect precision and should be assessed in context.
Allocation is also not a substitute for the assumptions behind the analysis. The simple separation argument assumes a common treatment effect across periods and no material carryover that distorts the response after a participant switches treatments. If those assumptions are not plausible, having both sequences does not by itself repair the model. Sequence allocation creates a useful comparison structure; it does not guarantee that every source of bias has been eliminated.
Check Your Understanding
Answer each question, then check the brief solutions below.
- A study assigns 28 participants equally to TR and RT. How many are planned for each sequence, and how many are planned for each treatment in each period?
- A study assigns 22 participants to TR and 18 to RT. Does unequal allocation mean treatment and period are necessarily confounded? Explain.
- In a TR-only study, T is given in period 1 and R in period 2. Why can an observed difference not be attributed uniquely to treatment?
- A study plans equal random assignment but ends with 15 participants in TR and 13 in RT. Is this alone proof that randomization failed? What should be reported?
- Why is it not appropriate to reassign participants after looking at their measured outcomes solely to equalize sequence counts?
Solutions. (1) \(28/2=14\) participants are planned for each sequence. The planned treatment count is 14 per treatment in each period. (2) No. Both sequences are represented, so both treatments occur in both periods; the unequal counts may affect precision but do not by themselves create perfect confounding. (3) The observed difference combines a treatment difference with a period difference, and the TR-only pattern does not provide the opposite order to separate them. (4) No. Random assignment can produce unequal realized counts; report the planned assignment method and actual counts, along with completion information. (5) Outcome-dependent reassignment changes the allocation after observing responses and can undermine the intended comparison; retain and analyze participants according to their actual assignments under the prespecified analysis plan.