Period Effects in Crossover Studies

Why Period Matters

A crossover study observes each participant under more than one treatment, but those observations occur at different times. A participant’s response can change between occasions for reasons other than the treatment: the participant’s health or behavior may change, study procedures may differ, or calendar conditions such as temperature may shift. Such influences can make the response in one period systematically higher or lower than in another.

A period effect is not automatically a treatment effect, nor does it imply that treatment caused a change over time. It describes a systematic difference associated with the scheduled occasion of measurement. In bioavailability studies, the response might be a pharmacokinetic endpoint such as log-transformed AUC or \(C_{\max}\). The period label refers to when the treatment was administered and measured, not to a new treatment.

An Additive Model and Its Assumptions

For a two-period study, an additive fixed-effects model can be written as

Here \(Y_{ij}\) is the response for participant \(i\) in period \(j\); \(\mu\) is the overall level; \(s_i\) is the participant contribution; \(T_{ij}\) identifies the treatment received; \(\tau_{T_{ij}}\) is its treatment effect; \(\pi_j\) is the period effect; and \(\varepsilon_{ij}\) is the remaining error. The model is a useful description when treatment and period contributions are additive: the period difference is common across treatments, rather than depending on which treatment is given.

For two treatments, a conventional constraint is \(\tau_A+\tau_B=0\); for two periods, use \(\pi_1+\pi_2=0\). If participant effects are also treated as fixed effects, those constraints alone do not identify the overall level: adding a constant to \(\mu\) and subtracting it from every \(s_i\) leaves fitted values unchanged. One can additionally impose \(\sum_i s_i=0\), or simply note that \(\mu\) and the fixed participant effects are not separately identified. Alternatively, participant effects can be modeled as random with mean zero. These choices do not change the central distinction between treatment and period contrasts.

The additive model assumes that the treatment and period effects are meaningfully separable for the question being asked. A balanced schedule supports that separation, but cannot guarantee the model is true. For ordinary model-based inference, the error structure must also be appropriate for repeated observations on the same participant; for example, the analysis must account for within-participant dependence. A set of arithmetic means can illustrate balance without making a claim about statistical significance.

Why Balance Helps

Treatment-period balance means that each treatment is observed equally often in each period. In the standard two-treatment, two-period layout, participants receive A then B or B then A. With equal numbers assigned to the two sequences and complete records, each treatment appears equally often in period 1 and period 2. Thus, a higher average in period 2 is not simply the result of period 2 containing more observations under one particular treatment.

The logic can be seen from the model. Within an AB participant, the period-2 minus period-1 difference includes the treatment contrast and the period contrast. Within a BA participant, the same subtraction includes the opposite treatment contrast but the same period contrast. If the two sequences are equally represented, averaging those two types of difference cancels the treatment contributions and leaves the period contrast, apart from error. Conversely, balancing the treatments across periods prevents a marginal treatment comparison from being just a comparison of different occasions.

This is a design property, not a post hoc repair. If a treatment is observed only in period 1 and another only in period 2, their effects are perfectly confounded in this model: the data cannot determine whether the response difference arose from treatment or period. Statistical adjustment cannot recover information the schedule did not provide.

A Short Justification Using Within-Participant Differences

The cancellation argument can be made explicit under the additive model. Let \(\Delta\tau=\tau_B-\tau_A\) and \(\Delta\pi=\pi_2-\pi_1\). For an AB participant, subtracting period 1 from period 2 removes the stable participant contribution:

For a BA participant, the treatment order is reversed, so the corresponding difference is

If the sequence groups have equal weight, the treatment terms cancel when the mean period-2-minus-period-1 difference is taken across the two groups:

This is a justification of the design’s ability to distinguish the additive period contrast from the treatment contrast; it is not a claim that observed data will equal their model expectations exactly. Random error, unequal group sizes, missing records, or departures from the additive model can alter the observed estimates.

When Conditions Change and What to Check

A period difference may be scientifically plausible even when the treatment schedule is balanced. Participants might change their diet, activity, clinical state, or adherence between visits. Study operations may also change, such as assay procedures, staff, or environmental conditions. These are reasons to standardize procedures across periods and record relevant information. Balance protects the treatment comparison against a common additive period shift; it does not make changing conditions disappear.

Useful checks begin with the design and records rather than a significance test. Confirm that each treatment was scheduled in each period and inspect sequence counts and completion by period. Summarize responses by treatment-period cell, and plot paired observations or within-participant changes by sequence. A period pattern that appears similarly under both treatments is compatible with a common additive period effect. Markedly different patterns across treatments may indicate that the common additive description is inadequate, though a small study can show irregular patterns by chance.

Also examine whether a period difference could be explained by operational changes or selective missingness. If participants with certain responses are more likely to miss a later visit, the observed period means may not represent comparable sets of participants. An imbalance caused by missing records can weaken the protection provided by the originally balanced schedule. Model-based adjustment may be useful under stated assumptions, but it does not replace examining who contributed data and why.

Alternatives, Limits, and a Common Pitfall

A balanced crossover schedule is the primary design-based safeguard when the aim is to distinguish a common period shift from a treatment difference. An analysis that includes both treatment and period terms is a complementary model-based approach: it estimates their contributions under the specified additive structure. A crude comparison of treatment-specific means that ignores period is less defensible when there is a period effect and the treatment counts across periods are unequal. None of these approaches can identify separate treatment and period effects under perfect confounding.

A period term also has limits. It represents an average systematic difference across occasions, not a diagnosis of its cause. It cannot by itself establish that participants changed biologically, that the calendar caused the change, or that study operations were responsible. Nor does balance protect against every design complication: treatment effects may vary over periods, data may be missing selectively, or the model may omit relevant structure. Those possibilities require separate evaluation and should not be hidden by describing a design as “balanced.”

• Pitfall: “The period-2 mean is higher, so treatment B must raise the response.” This is wrong when period and treatment are not separated or when the schedule is imbalanced. The correction is to compare treatments across periods using the design and an appropriate model, and to state when the design cannot identify separate effects.
• Pitfall: “Balanced assignment guarantees no period effect.” Balance does not remove a period effect. It helps distinguish a common period effect from treatment under the model; the period effect may still be present and should be described.
• Pitfall: “A nonsignificant period test proves there is no period effect.” A nonsignificant result is not proof of equality. It may reflect limited precision, and period testing should not substitute for design-based reasoning.

In bioequivalence work, the central practical lesson is to preserve treatment-period balance, standardize procedures across visits, and interpret period differences as occasion-related evidence rather than as a treatment conclusion. The design supports a cleaner treatment comparison, while assumptions and observed data determine how credible that comparison is.