Why Compare Alternative Four-Period Schedules?
A four-period full-replicate crossover gives each participant two administrations of the test formulation and two of the reference formulation. That replication is useful only if the order of those administrations is also considered carefully. A schedule can meet the exposure-count requirement while placing one formulation disproportionately in particular periods, especially if its sequence groups are not balanced.
Period matters because responses may change over time for reasons not caused by formulation: for example, a common time trend or changing study conditions. If formulation assignment and period are systematically associated, a treatment contrast can partly reflect those period differences. The design question here is how to compare alternative four-period sequences so that treatment is not systematically aligned with period across the enrolled sequence groups.
The six possible sequences are \(TTRR\), \(TRTR\), \(TRRT\), \(RTTR\), \(RTRT\), and \(RRTT\). Each is full-replicate, but their order patterns differ. The central comparison is not whether each sequence contains two of each formulation—it does by definition—but whether the sequence groups together protect comparisons from period imbalance.
Complementary Sequence Pairs
A convenient way to construct a period-balanced set is to pair each sequence with its complement: replace every \(T\) with \(R\) and every \(R\) with \(T\). The complementary pairs are \(TTRR/RRTT\), \(TRTR/RTRT\), and \(TRRT/RTTR\). If participants are allocated equally to the two sequences in a pair, each period has equal numbers assigned to \(T\) and \(R\).
For instance, in the pair \(TRRT/RTTR\), period 1 is \(T\) for the first sequence and \(R\) for the second; period 2 is \(R\) and \(T\); period 3 is \(R\) and \(T\); period 4 is \(T\) and \(R\). The same complementarity holds at every period. Thus, equal group sizes give period balance regardless of which of the three complementary pairs is chosen.
Worked Example: Checking a Complementary Pair
Suppose a hypothetical study plans to enroll 24 participants, allocating 12 to \(TRRT\) and 12 to \(RTTR\). Check the planned treatment count in each period.
Period 1: the \(TRRT\) group contributes 12 participants on \(T\), and the \(RTTR\) group contributes 12 on \(R\). The counts are 12 \(T\) and 12 \(R\).
Period 2: the assignments are reversed, giving 12 \(R\) and 12 \(T\). The counts are again 12 and 12.
Period 3: the assignments are also \(R\) and \(T\), giving 12 on each formulation.
Period 4: the assignments are \(T\) and \(R\), again giving 12 on each formulation. The planned design is period-balanced in all four periods. This conclusion follows from the schedule and equal allocation; it does not establish that the observed data will remain balanced after withdrawals.
Why Equal Allocation Cancels Period Imbalance
The balancing property can be stated as a simple design-matrix argument. Code treatment within a period as \(+1\) for \(T\) and \(-1\) for \(R\). For any sequence, its four treatment codes form a vector \(z\). Its complement has codes \(-z\). If the two sequences have equal group sizes, their contributions to the treatment code at each period sum to zero.
Here \(n\) is the number assigned to each sequence and \(j\) indexes period. A common period effect is represented by a column that identifies each period. The treatment code sums to zero among participants in every period, so the planned treatment contrast is orthogonal to those period indicators. In practical terms, a common period effect cannot create a systematic treatment-versus-period imbalance in the planned, equally allocated groups.
This is a design property, not a claim that period effects do not exist. It also does not by itself guarantee an unbiased treatment interpretation under every possible response pattern. For a common treatment contrast to be meaningful, the analysis must use an appropriate model, and important treatment-by-period differences or residual effects of earlier administrations may require attention. Period balance addresses one specific risk: confounding caused by unequal treatment assignments across periods.
What the Three Complementary Pairs Emphasize
Although all three complementary pairs give equal treatment counts in every period when equally allocated, their order patterns differ. Those differences can matter for operational planning and for the way exposures are distributed across adjacent periods. A schedule should therefore be described by its actual sequences, not just by saying that it is “a four-period replicate.”
| Complementary pair | Order pattern | Design feature |
|---|---|---|
| \(TRTR/RTRT\) | Alternating | No same-formulation administrations in adjacent periods. |
| \(TRRT/RTTR\) | Mixed order | Each sequence contains one adjacent repeat; the two sequences reverse the assignments. |
| \(TTRR/RRTT\) | Grouped order | Each sequence gives two consecutive periods of one formulation followed by two of the other. |
This comparison does not establish that one pair is universally best. All three have the same planned period-specific treatment counts under equal allocation, but they distribute orders differently. Alternation avoids adjacent repetitions; grouped schedules include long runs; mixed-order schedules have a different placement of repeated administrations. Selection should reflect the study’s practical constraints and the treatment-order risks that need to be managed.
Worked Example: Comparing Order Counts
Consider two hypothetical designs with equal group sizes: \(TRTR/RTRT\) and \(TRRT/RTTR\). Compare the adjacent-period transitions within their sequence pairs. This is a descriptive check of schedule order, not a test of an outcome.
For \(TRTR\), the three adjacent transitions are \(T\to R\), \(R\to T\), and \(T\to R\). For its complement \(RTRT\), they are \(R\to T\), \(T\to R\), and \(R\to T\). Pooling one participant from each sequence gives three \(T\to R\) and three \(R\to T\) transitions, with no \(T\to T\) or \(R\to R\) transition.
For \(TRRT\), the transitions are \(T\to R\), \(R\to R\), and \(R\to T\). For \(RTTR\), they are \(R\to T\), \(T\to T\), and \(T\to R\). Pooling the pair gives two \(T\to R\), two \(R\to T\), one \(T\to T\), and one \(R\to R\) transition.
Both pairs remain period-balanced under equal allocation. The transition counts show that period balance alone does not make the sequence patterns identical. These counts can inform design review, but they do not establish that a particular transition causes a response or that carryover has been eliminated.
Unequal Allocation and Incomplete Records
Complementary sequences balance periods only to the extent that their group sizes are equal. The algebra makes the consequence of imbalance clear. If \(n_1\) participants are assigned to a sequence and \(n_2\) to its complement, their combined treatment code at period \(j\) is \((n_1-n_2)z_j\), not zero unless the group sizes match. This imbalance may align treatment with a pattern of periods.
Worked Example: Quantifying Unequal Sequence Groups
In a hypothetical study using \(TRTR/RTRT\), suppose 15 participants are assigned to \(TRTR\) and 9 to \(RTRT\). Count the treatment assignments by period.
In period 1, \(TRTR\) assigns 15 to \(T\) and \(RTRT\) assigns 9 to \(R\). The counts are 15 \(T\) and 9 \(R\). In period 2, they are 15 \(R\) and 9 \(T\). Period 3 has 15 \(T\) and 9 \(R\), and period 4 has 15 \(R\) and 9 \(T\).
The treatment imbalance alternates by period: the excess of \(T\) is 6 in periods 1 and 3, while the excess of \(R\) is 6 in periods 2 and 4. If responses have an alternating period pattern, the observed treatment contrast can be entangled with it. The planned sequences are each valid full-replicate sequences, but their unequal allocation removes the complementary-pair balance.
Withdrawals can create a related issue in the observed records even when assignment began equally. A participant may contribute some periods but not others, so the number with observed \(T\) and \(R\) responses can differ by period. This does not automatically invalidate the design or analysis, but it means planned balance should not be confused with observed balance. Report sequence allocation and record availability by period, and assess whether missingness is associated with sequence, period, treatment, or response history.
Design Checks, Alternatives, and Limitations
Before enrolling participants, verify the sequence list and allocation ratio rather than relying on the design label. At minimum, make a table with periods as columns and planned \(T\) and \(R\) counts as rows. Confirm that every sequence contains two administrations of each formulation, and that the proposed sequence groups give the intended count in every period. After data collection, repeat the count using observed records, noting where participants are missing.
A further check is whether the planned treatment and period columns are distinguishable in the analysis. Equal allocation to complementary sequences ensures no marginal treatment imbalance by period, but a more complex model may include additional terms. Their estimability depends on the actual sequence allocation and records. A numerical design check or model-matrix rank check can reveal exact aliasing; it cannot determine whether the scientific assumptions behind the model are reasonable.
The main alternative is not a different label for the same design, but a different sequence pair or allocation plan. The alternating pair may be chosen when alternating administration is operationally acceptable; the mixed pair offers a different arrangement; the grouped pair concentrates each formulation into two-period runs. Any comparison should keep the objective clear: period balance is achieved by equal complementary allocation, while order-pattern properties vary among pairs. A two-period crossover cannot provide the same two administrations of each formulation per participant, and a parallel design does not distribute each participant across formulations; those are different design choices rather than substitutes that preserve all features of the four-period replicate.
There are important limitations. Period balance does not remove individual period variation, prove the absence of carryover, or protect against treatment-by-period interaction. It also does not ensure balance after differential withdrawal. Finally, balancing the marginal counts is not a substitute for random allocation to sequence groups: sequence assignment should be randomized or otherwise justified so that systematic participant differences do not line up with sequence.
Common Pitfall
A frequent error is to reason, “Every participant receives \(T\) twice and \(R\) twice, so treatment cannot be confounded with period.” That conclusion does not follow. Replication counts exposures within participants; period balance is a property of the sequence groups together. If most participants receive \(T\) in periods 1 and 3 and \(R\) in periods 2 and 4, the treatment allocation is still associated with period despite every participant receiving both formulations twice.
Check Your Understanding
Answer each question, then compare with the solutions below.
- List the complementary sequence of \(TTRR\), and state the condition under which the pair is period-balanced.
- A study allocates 10 participants to \(TRTR\) and 10 to \(RTRT\). How many are planned on \(T\) in each period?
- Why does giving each participant two \(T\) and two \(R\) administrations not, by itself, establish period balance?
- For \(TRRT/RTTR\), name one order-pattern difference from \(TRTR/RTRT\).
- What should be checked if sequence groups were equally allocated at enrollment but some participants later miss periods?
Solutions. (1) The complement is \(RRTT\); equal numbers allocated to the two sequences give equal \(T\) and \(R\) counts in every period. (2) Each period has 10 participants on \(T\), because the assignments alternate between the two equally sized groups. (3) The statement concerns each participant’s exposure totals, not how treatment assignments are distributed across participants at each period. (4) \(TRRT/RTTR\) includes same-formulation adjacent transitions, whereas the alternating pair has none. (5) Recount observed treatment assignments by period and sequence, and assess the pattern of missing records rather than assuming planned balance remains in the observed data.