Period Balance in Replicate Designs

Why Period Balance Matters

In a replicate crossover design, participants receive multiple administrations of the test formulation (T), the reference formulation (R), or both, over ordered study periods. A schedule determines which formulation is assigned in each period. Replication supplies repeated exposures, but it does not by itself ensure that T and R occur at comparable points in the study. If one formulation is disproportionately assigned early and the other late, a change in response over time can contribute to their observed difference.

A period-count audit is a direct design check: for every period, count how many participants are assigned to each formulation. The important comparison is not always whether the two counts are equal. Some replicate designs deliberately give one formulation more exposures overall. The general question is whether the formulations have comparable distributions across periods.

Definition: Let \(n_{jp}\) be the number of participants assigned formulation \(j\) in period \(p\), where \(j\) is T or R. Let \(N_j=\sum_p n_{jp}\) be the total number of assigned exposures to formulation \(j\), and define its period weight as \(w_{jp}=n_{jp}/N_j\). Treatment-period balance means the period weights are the same for T and R in every period: \(w_{Tp}=w_{Rp}\). When the total exposure counts are equal, this is equivalent to equal T and R counts within every period. When total exposure counts differ, proportional balance—not equal raw counts—is the relevant criterion.

How Period Imbalance Can Affect a Contrast

The following additive model is a simple way to isolate the issue:

Y_{ip}=\mu+b_i+\tau_{j(ip)}+\pi_p+\varepsilon_{ip}.

Here \(Y_{ip}\) is the response for participant \(i\) in period \(p\), \(b_i\) is a participant effect, \(\tau_{j(ip)}\) is the effect of the formulation assigned in that period, and \(\pi_p\) is a common period effect. For the explanation below, assume the treatment effect is additive across periods, the period effect is common to T and R, and the errors have mean zero conditional on the assigned schedule. This is an illustrative mean model, not a complete inferential analysis: repeated observations from the same participant are generally correlated, and a schedule audit does not estimate standard errors or test bioequivalence.

If we average observations separately by formulation, their expected period contributions are weighted by the periods in which each formulation occurs. Under the additive model, the period component of the T-minus-R contrast is therefore:

\sum_p w_{Tp}\pi_p-\sum_p w_{Rp}\pi_p = \sum_p (w_{Tp}-w_{Rp})\pi_p.

This term is zero for every possible set of period effects when the period weights match. If the weights differ, a nonzero period trend can contribute to the crude formulation contrast even when the underlying formulation effects are identical. This calculation isolates one source of distortion; it does not say that every imbalance will produce a large bias. The size and direction depend on both the pattern of imbalance and the period effects.

Justification: Under the model, the expected mean response for formulation \(j\) is \(\mu+\tau_j+\sum_p w_{jp}\pi_p\), apart from participant-effect contributions. Subtracting the R mean from the T mean cancels \(\mu\) and leaves \(\tau_T-\tau_R+\sum_p(w_{Tp}-w_{Rp})\pi_p\). Thus matching the period weights removes the additive period contribution to this contrast. This argument concerns the period component; it does not guarantee that other design or analysis problems are absent.

Auditing a Replicate Schedule

Make a table with one row per period and columns for T count, R count, and, when totals differ, the corresponding period weights. Counts should refer to the participants actually assigned to the schedule being audited. If the audit concerns the planned design, use planned sequence allocations; if it concerns the realized study, use the actual assignments and report withdrawals or missing periods separately rather than silently treating them as completed exposures.

List the sequences and allocations.
Record each sequence and the number of participants assigned to it. Check that each sequence is written in the intended period order.

Count each formulation by period.
For each period, sum the participants assigned T and the participants assigned R across all sequences.

Compare period weights.
Calculate \(w_{Tp}\) and \(w_{Rp}\). Equal weights indicate proportional balance, even when the raw counts differ.

Interpret any differences.
Describe which periods are overrepresented for each formulation and consider whether plausible time changes could matter. Do not treat balance as proof that the full analysis is valid.

Worked Example: Four-Period Full-Replicate Balance

Hypothetical schedule. Eight participants are assigned to a four-period full-replicate design: four receive TRTR and four receive RTRT. Audit the treatment counts by period.

In period 1, the four TRTR participants receive T and the four RTRT participants receive R, giving counts T = 4 and R = 4. Period 2 reverses those assignments, also giving T = 4 and R = 4. Periods 3 and 4 repeat the period 1 and period 2 patterns, respectively.

Period	T count	R count	T weight	R weight
1	4	4	4/16 = 0.25	4/16 = 0.25
2	4	4	4/16 = 0.25	4/16 = 0.25
3	4	4	4/16 = 0.25	4/16 = 0.25
4	4	4	4/16 = 0.25	4/16 = 0.25

Each participant contributes two T and two R exposures, so \(N_T=N_R=16\). The period weights match in every period. Consequently, under the illustrative additive model, any common period effects cancel from the pooled T-minus-R mean contrast. This is a statement about period representation; it is not a claim that the observations are independent or that the schedule alone settles the choice of statistical analysis.

Equal Counts Are Not Always Required

Some replicate schedules assign one formulation more often than the other. In those cases, an audit based only on whether the raw counts match can incorrectly label a well-distributed schedule as imbalanced. Compare the fraction of each formulation’s exposures occurring in each period.

Worked Example: Proportional Balance in a Test-Replicate Schedule

Hypothetical schedule. Three participants are assigned to each of the sequences TTR, TRT, and RTT in a three-period test-replicate design. Each participant therefore has two T exposures and one R exposure.

In period 1, T appears in TTR and TRT, so six participants receive T; three receive R. In period 2, T appears in TTR and RTT, again giving six T and three R. In period 3, T appears in TRT and RTT, also giving six T and three R.

Period	T count	R count	T weight	R weight
1	6	3	6/18 = 0.3333	3/9 = 0.3333
2	6	3	6/18 = 0.3333	3/9 = 0.3333
3	6	3	6/18 = 0.3333	3/9 = 0.3333

The raw counts are not equal, but each formulation has one-third of its exposures in each period. Thus \(w_{Tp}=w_{Rp}\) for every period, and an additive period trend contributes equally to the two formulation means. Calling this schedule period-imbalanced merely because there are twice as many T exposures as R exposures would confuse overall replication frequency with period distribution.

When Sequence Allocation Creates Imbalance

A balanced set of possible sequences does not guarantee a balanced realized schedule if very different numbers of participants are assigned to those sequences. For this reason, an audit should use sequence allocations, not just the sequence names printed in a design description.

Worked Example: Unequal Sequence Allocation and a Time Trend

Hypothetical schedule. In a four-period full-replicate design, six participants receive TRTR and two receive RTRT. Suppose the true formulation effects are equal, and, solely to illustrate the arithmetic, the common period effects are \(\pi_1=0\), \(\pi_2=1\), \(\pi_3=2\), and \(\pi_4=3\) response units.

Periods 1 and 3 have T counts of 6 and R counts of 2. Periods 2 and 4 have T counts of 2 and R counts of 6. There are 16 exposures to each formulation. The period weights are therefore T: \((0.375, 0.125, 0.375, 0.125)\) and R: \((0.125, 0.375, 0.125, 0.375)\).

The period contribution to the T-minus-R contrast is:

\begin{aligned} \sum_{p=1}^{4}(w_{Tp}-w_{Rp})\pi_p &=(0.375-0.125)(0)+(0.125-0.375)(1)\\ &\quad +(0.375-0.125)(2)+(0.125-0.375)(3)\\ &=0-0.25+0.50-0.75\\ &=-0.50\ \text{response units}. \end{aligned}

Although the true formulation effects were set equal, the pooled T mean would be lower by 0.50 response units from the period contribution alone. The direction follows the schedule: T is more represented in periods 1 and 3, while R is more represented in periods 2 and 4. The example does not establish that these particular period effects are realistic; it demonstrates how an imbalance can let a time pattern contribute to a formulation contrast.

Diagnostics, Alternatives, and Limits

A clear diagnostic is a treatment-by-period table accompanied by the weights \(w_{jp}\). If the weights differ, report the differences \(w_{Tp}-w_{Rp}\), not just a single label such as “balanced” or “unbalanced.” A difference of zero in every period is exact proportional balance. Small differences indicate near balance, but whether they are practically important depends on the plausible size and shape of period effects. Also inspect sequence allocations: unequal allocation can produce imbalance even when the intended sequence set appears symmetric.

Design-stage balancing and analysis-stage adjustment address the issue in different ways. Allocating participants across sequences to obtain comparable period weights reduces the opportunity for common time trends to contribute to a crude treatment contrast. Including period in an analysis can account for modeled period differences, but it does not make a poorly represented schedule equivalent to a balanced one. In particular, model adjustment relies on the model being appropriate and cannot create information where treatment-period combinations are absent or sparsely represented. The two approaches can complement one another; neither should be presented as a substitute for checking the schedule.

Period balance also has limits. It does not establish that there is no treatment-by-period interaction, does not resolve every source of bias, and does not address the correlation among repeated observations from a participant. Nor does a balanced assignment table necessarily describe the data available for analysis if participants miss or discontinue periods. The audit is a design and data-description tool, not a bioequivalence conclusion.

Common Mistakes

Requiring equal raw counts in every period. That is appropriate only when total T and R exposure counts are equal. With unequal replication totals, compare the within-formulation period weights.
Checking only the sequence list. Sequence coverage does not show how many participants were assigned to each sequence. Count participants in each sequence before aggregating by period.
Assuming imbalance proves bias. Imbalance permits a period trend to contribute to a crude contrast; the size and direction depend on the actual period effects and the pattern of weights.
Assuming balance solves the whole analysis. Matching period weights removes the common additive period contribution in the stated model, but it does not validate other assumptions or address repeated-measure dependence.

Key takeaway: Audit treatment representation by period using both counts and normalized weights. Proportional balance makes T and R share the same period distribution, so a common additive time trend does not contribute to their pooled mean contrast under the stated model.

Check Your Understanding

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

A design has 12 T exposures and 12 R exposures. What condition on the counts gives period balance?
A schedule has 18 T exposures and 9 R exposures. In each of three periods, there are 6 T and 3 R exposures. Is it proportionally balanced? Show the weights.
In a two-period audit, T has period weights \((0.7,0.3)\), while R has weights \((0.3,0.7)\). The common period effects are 0 and 2 response units. Find the period contribution to the T-minus-R contrast.
Why is it insufficient to check that every planned sequence appears at least once?

Solutions. (1) Because the exposure totals are equal, period balance requires equal T and R counts in each period. (2) Yes. The weights are \(6/18=1/3\) for T and \(3/9=1/3\) for R in each period. (3) The contribution is \((0.7-0.3)(0)+(0.3-0.7)(2)=-0.8\) response units. (4) Sequence presence does not reveal allocation sizes: unequal numbers assigned to the sequences can give different T and R period weights even when every sequence is represented.