Why Randomize Within a Crossover Design?
In a two-period, two-sequence crossover study, participants are assigned to one of two planned treatment orders: AB or BA. Randomization determines which eligible participant receives which sequence. It does not mean choosing each participant’s sequence by judgment, nor does it mean that a participant’s sequence can be changed later to make the numbers look balanced. The assignment process should be set in advance and applied consistently.
Random assignment reduces the risk that sequence membership is systematically related to participant characteristics or enrollment decisions. It does not guarantee that the two sequence groups will be identical at baseline, especially in a small study. A separate practical goal is to keep the number assigned to each sequence reasonably close to the planned allocation as recruitment proceeds. Simple randomization and blocked randomization address that goal differently.
The sequence is assigned to a participant after eligibility and consent have been established, using the study’s prespecified procedure. Randomization is meaningful only if assignments are generated by chance and the procedure is followed. Alternating AB, BA, AB, BA is predictable, so it is not randomization: a recruiter who knows the pattern can anticipate the next assignment.
Simple Randomization and Its Imbalance
With a planned 1:1 allocation, simple randomization independently assigns each participant to AB or BA with probability \(0.5\). Let \(K\) be the number assigned to AB among \(N\) participants. Then \(K\sim\operatorname{Binomial}(N,0.5)\), and the number assigned to BA is \(N-K\). The expected number in either sequence is \(N/2\), but the realized counts need not be equal.
The size of this random fluctuation can be derived directly. Write \(I_i=1\) if participant \(i\) is assigned to AB and \(I_i=0\) otherwise. Under independent assignments, \(K=\sum_{i=1}^{N} I_i\), with \(E(I_i)=0.5\) and \(\operatorname{Var}(I_i)=0.25\). Independence gives:
Thus, an imbalance is not by itself evidence that randomization failed. It is a possible result of a valid random process. However, large imbalances can arise by chance, and the probability of imbalance increases with enrollment size in absolute terms even though its typical size is smaller relative to the total sample.
Worked Example: A Chance Imbalance Under Simple Randomization
Suppose a hypothetical 12-participant crossover study uses independent 1:1 assignments. One valid random draw assigns 8 participants to AB and 4 to BA. The signed difference is \(8-4=4\) participants, and the allocation ratio observed is \(8:4=2:1\), rather than the planned 1:1.
This outcome is possible under the stated procedure. In fact, if \(K\) counts AB assignments, the probability of exactly 8 AB assignments is:
The calculation does not imply that this draw was desirable or that every such imbalance should be accepted operationally. It shows why simple randomization does not promise equal counts. If keeping counts close during recruitment is important, a blocked procedure can be specified before the first assignment.
Permuted Blocks for a Planned Allocation
In permuted-block randomization, assignments are generated in blocks with a prespecified composition, and the order within each block is randomized. For a 1:1 allocation and block size 4, each block contains two AB and two BA assignments in a randomly selected order. A possible block might be AB, BA, BA, AB; the order should come from a chance-based procedure, not from a recruiter.
If every block is completed, each contributes equal counts to the two sequences. After \(m\) complete blocks of size 4, there are \(2m\) assignments to AB and \(2m\) to BA. The allocation is therefore exactly balanced at those points. This is the main benefit over simple randomization when enrollment occurs over time.
Worked Example: Generating and Checking Two Blocks
For a hypothetical 8-participant study, a statistician generates two permuted blocks of size 4. Suppose the randomly selected orders are AB, BA, BA, AB and BA, AB, AB, BA. Participant IDs are assigned sequentially to these slots after eligibility is confirmed.
The first block contributes 2 AB and 2 BA assignments. The second also contributes 2 AB and 2 BA. Across both blocks, the totals are \(2+2=4\) for AB and \(2+2=4\) for BA. The planned 1:1 allocation is preserved at the end of each complete block and overall.
A valid implementation also records the generated list, its version, and the participant ID assigned to each slot. These records allow an independent check that the order was followed. They do not justify revealing the unused part of the list to enrollment staff.
For a planned allocation other than 1:1, block composition should match that ratio. For example, with a planned 2:1 allocation favoring AB, a block of size 6 can contain four AB and two BA assignments. Exact balance at completed block boundaries is possible when the block size accommodates the ratio. The allocation ratio and block sizes must be chosen prospectively, not adjusted in response to observed participant characteristics or outcomes.
Uneven Enrollment and Recruitment Batches
A study may stop before a block is complete, reach its target with an odd number of participants, or recruit in batches whose sizes do not match the block size. In these situations, exact equality at every interim point is not generally possible. The important distinction is between a planned method that has a known balance property and an improvised correction to counts after enrollment.
For an even block of size \(b\) with equal numbers assigned to each sequence, any prefix of the block can differ by at most \(b/2\) assignments between sequences. The maximum occurs if the first \(b/2\) positions all contain the same sequence. Once the entire block is assigned, the difference returns to zero. This bound is a property of the block composition; it is not a promise that each recruitment batch will be balanced.
Worked Example: Tracking Enrollment Across Batches
Consider a hypothetical study using blocks of size 4. Three generated block orders are AB, BA, BA, AB; BA, AB, AB, BA; and AB, BA, AB, BA. Enrollment occurs in batches of 5, 3, and 3 participants, for 11 participants total. The allocation list continues across batches; it is not restarted at the beginning of each batch.
The first five assignments are the first complete block plus the first position of block 2. The first block contributes 2 AB and 2 BA, and the fifth assignment is BA, so the cumulative totals are 2 AB and 3 BA.
The next three assignments complete block 2: BA, AB, AB. They contribute 2 AB and 1 BA. The cumulative totals become 4 AB and 4 BA. The final batch uses the first three positions of block 3: AB, BA, AB. These contribute 2 AB and 1 BA, giving final totals of 6 AB and 5 BA.
The final one-participant difference is expected to be possible because enrollment stopped partway through a block. It is not a reason to reassign the eleventh participant or alter the completed assignments. If the target enrollment had been even and the final block completed, the planned 1:1 count would be restored at the end of that block.
When enrollment is expected to occur in batches, the randomization plan should specify whether a single list continues across all batches and how incomplete blocks are handled. Resetting the list for each batch can produce avoidable imbalances, particularly if batch sizes are uneven. If batches are deliberately treated as separate allocation units, their sizes and block compositions must be designed to support the intended overall ratio.
Choosing and Safeguarding the Procedure
| Approach | Strength | Limitation |
|---|---|---|
| Simple randomization | Easy to generate and difficult to predict when assignments are independent. | Can produce substantial sequence-count imbalance, especially in a small study or during early recruitment. |
| Fixed-size permuted blocks | Maintains the planned ratio at completed block boundaries. | If block size and prior assignments are known, the final assignment in a block can become predictable. |
| Randomly varying block sizes | Can retain approximate balance while making the next assignment harder to infer. | Requires a carefully generated and safeguarded list; it does not guarantee balance at every interim point. |
A practical choice depends on the planned sample size, recruitment pattern, allocation ratio, and the need to limit interim imbalance. Fixed-size blocks are simple to audit, but their predictability is a concern if the block size and assignments are visible to staff who enroll participants. Randomly varying block sizes can reduce that predictability. In either approach, concealment is essential: the person enrolling a participant should receive only the assignment for that participant, after enrollment is irrevocably established.
The generation procedure should use a suitable random-number method, and the assignment list should be stored with access restricted to authorized personnel. A secure centralized system or sequentially numbered, opaque, sealed envelopes prepared and handled appropriately can support concealment. A list that is visible in advance, envelopes that can be opened or read through, or assignments issued before eligibility is confirmed can undermine the procedure even if the original sequence was generated randomly.
Allocation reports should show the planned ratio, the number assigned to each sequence, the recruitment status, and any deviations from the list. Counts should be checked against the allocation records, not “corrected” by moving participants between sequences. Participants who discontinue after assignment remain part of the allocation record; their assignment is not erased or reused. Handling incomplete treatment-period records is a separate analysis issue.
Diagnostics, Limitations, and Common Pitfalls
Useful checks include confirming that the specified ratio and block composition were used, verifying that assignments were made in sequence, comparing participant IDs with the allocation log, and documenting any deviations. Unexpected duplicate assignments, skipped list positions, assignments made before eligibility confirmation, or an enrollment pattern that repeatedly seems to anticipate upcoming assignments warrant investigation. A count imbalance alone is not a diagnostic of failed randomization: it must be interpreted against the randomization method and whether blocks were complete.
Randomization balances measured and unmeasured participant characteristics only in expectation across repetitions of the assignment process. It does not guarantee balance in one realized study. Blocking controls the sequence counts at specified points; it does not ensure that age, body weight, or other baseline characteristics are equal across sequence groups. Nor does a balanced allocation fix problems such as inadequate concealment or post-assignment changes to the assigned sequence.
Common mistake: “The first five participants produced three AB and two BA assignments, so the next participant must be assigned to BA to restore balance.” This reasoning turns a planned random procedure into an unplanned deterministic correction. The next assignment should instead be the next protected entry in the prespecified randomization list. If interim balance is important, use a blocking plan designed in advance; do not improvise assignments based on accumulated counts.
Check Your Understanding
Answer each question using the stated randomization plan.
- Under simple 1:1 randomization, what are the expected number and standard deviation of AB assignments for \(N=16\)?
- A 1:1 study uses blocks of size 4 and stops after 6 assignments. What is the largest possible absolute difference between the sequence counts at that point?
- A study plans a 2:1 allocation favoring AB. Give one block size and composition that can achieve the planned ratio at a completed block boundary.
- Why is assigning AB to every other participant not a valid substitute for randomization?
- A study reaches its target after an incomplete equal-allocation block and has one more AB than BA. Should a participant be reassigned to make the final counts equal? Explain.
Solutions. (1) \(E(K)=N/2=16/2=8\), and \(SD(K)=\sqrt{N}/2=\sqrt{16}/2=2\) participants. (2) The first six positions consist of one complete block, balanced 2:2, plus two positions in the next block. Those two may both be the same sequence, so the largest possible difference is 2 participants. (3) A block of size 6 containing 4 AB and 2 BA assignments gives the 2:1 ratio when complete. (4) Alternation is predictable; the next sequence is known from the prior assignment, so enrollment decisions can potentially be influenced and assignments are not independently determined by chance. (5) No. The one-participant difference can occur when enrollment stops mid-block; the assigned sequence must not be changed to force equality. The study should document the final counts and any deviations from the prespecified procedure.