Design choice follows the study objective
For a narrow therapeutic index (NTI) drug, a small difference in exposure may have important clinical consequences. Tight acceptance limits and careful evaluation of relevant pharmacokinetic endpoints therefore matter. But the choice of crossover design answers a related, distinct question: which comparisons can the study support, and how much repeated information does it collect from each participant?
A two-period crossover gives each participant one exposure to test and one to reference. Replicate designs repeat exposure to at least one formulation, providing information about variation between occasions within the same participant. That extra information costs additional periods, dosing occasions, and participant time. Replication is not automatically better: it is useful when the study objective requires the information that replication supplies and when the added burden is acceptable.
The design comparisons below concern the information available from the treatment schedule. They do not prescribe a particular analysis model or regulatory decision rule. Applicable NTI requirements vary by jurisdiction, and the study should use the relevant guidance for its intended submission.
Three common crossover structures
| Design | Example sequences | Exposure per participant | Replication supported |
|---|---|---|---|
| Two-period, two-sequence | TR, RT | One T and one R | Neither formulation is repeated |
| Three-period partial replicate | TRR, RTR, RRT | One T and two R | R is repeated; T is not |
| Four-period full replicate | TRTR, RTRT | Two T and two R | Both formulations are repeated |
In a two-period design, each participant contributes a within-participant comparison of T and R. With appropriate randomization and analysis, the design can estimate the average formulation contrast while accounting for stable differences between participants. However, because no formulation is repeated within a participant, the schedule does not directly provide repeated observations of either formulation for assessing formulation-specific within-subject variability.
A three-period partial replicate repeats R. Its repeated reference exposures can inform assessment of within-subject variability for R, while the single T exposure does not provide the same repeated information for T. This can be useful when reference variability is a specific design objective and the applicable regulatory approach uses it. A partial replicate is not equivalent to a full replicate: it does not provide matched repeated exposure information for both formulations.
A four-period full replicate repeats both T and R. It supports repeated observations of each formulation within participants, allowing assessment of within-subject variability for both. The additional comparison comes with an additional period relative to the three-period design and two additional periods relative to the two-period design.
Why repeated exposure supplies different information
The design distinction can be seen with a simplified model for a log-transformed pharmacokinetic endpoint, such as log AUC. Suppose the result for participant \(i\) on an occasion receiving R is \(Y_{iRj}=\mu_R+b_i+e_{iRj}\). Here, \(\mu_R\) is the population mean under R, \(b_i\) is a stable participant-specific component, and \(e_{iRj}\) is an occasion-specific component. The same idea applies to T. This is a conceptual decomposition, not a complete analysis model: period effects, carryover, and other design features may also need to be represented.
If the same participant receives R twice, the difference between those results removes the stable participant component. Under the simplifying conditions that occasion errors have mean zero, equal variance \(\sigma_{wR}^2\), and are independent across occasions, the justification is:
This shows why replication matters: a repeated formulation creates within-participant contrasts that a nonreplicate schedule cannot create. In a real study, period effects, correlation between occasions, and the chosen model affect how such information is used. The equation is a rationale for design selection, not a substitute for the planned statistical analysis.
All three designs can be used to compare average formulation effects, subject to adequate randomization, a suitable analysis, and the assumptions needed for a crossover comparison. For a simple paired comparison, the participant-level contrast is \(D_i=Y_{iT}-Y_{iR}\); the population quantity of interest is the mean contrast, not an individual participant’s difference. Repeated observations can enrich the information about variability, but they do not automatically improve every aspect of the mean comparison.
Worked Example: Selecting a Design for a Mean Comparison
A hypothetical NTI study is intended to compare mean log AUC for T and R. Its primary design objective is the average formulation contrast; the protocol does not require repeated exposures to estimate formulation-specific within-subject variability. Which structure is a reasonable starting point?
A two-period TR/RT crossover gives each participant one exposure to each formulation. For example, suppose a hypothetical participant has log AUC values \(6.20\) under T and \(6.15\) under R. The participant’s contrast is:
This single contrast contributes to the estimate of the population mean contrast. It does not, by itself, establish the population result, and the example gives no basis for a bioequivalence conclusion: there is no sample size, uncertainty estimate, or acceptance-limit assessment. The key design point is that every participant supplies one T and one R result, but neither treatment is repeated.
The design is appropriate to consider if the objective is a mean comparison and the applicable requirements permit it. The assumption check is not merely that both treatments appear in the schedule: randomization should assign sequence, the washout should be adequate to avoid material carryover, and period effects should be accounted for in the planned analysis. If repeated variability information is a required objective, this two-period schedule does not meet it.
Choosing between partial and full replication
Selection should begin with the estimands and regulatory questions, not with the number of periods a sponsor can conveniently schedule. Ask which formulation-specific variability information is needed, whether the relevant guidance calls for or permits a particular replicate structure, and whether the additional periods are feasible without unacceptable loss of data quality.
State whether the study is intended to compare average exposure alone or also to characterize within-subject variability for one or both formulations.
A two-period design repeats neither treatment; a three-period partial replicate repeats R; a four-period full replicate repeats T and R.
Confirm the applicable jurisdiction’s requirements, then assess period count, washouts, participant burden, and the risk of missed or incomplete periods.
Specify how sequence, period, treatment, and repeated observations will be handled; a design does not remove the need for appropriate modeling.
Worked Example: Counting the Period Burden
A hypothetical protocol plans to randomize 36 participants, with each participant assigned to one sequence. Compare the scheduled treatment occasions under the three structures, assuming every participant completes every period.
For the two-period design, the total is \(36\times2=72\) treatment occasions. For the three-period partial replicate, it is \(36\times3=108\). For the four-period full replicate, it is \(36\times4=144\). Thus, the three-period schedule adds \(108-72=36\) occasions relative to two periods, and the four-period schedule adds \(144-72=72\).
These are planned occasions, not necessarily 72, 108, or 144 analyzable exposure profiles: withdrawals, missed visits, or unusable samples can reduce the data available. The arithmetic supports an operational comparison, not a statistical power calculation. To choose among the designs, the team must weigh the extra repeated information against the added participant and site burden, as well as the objective and applicable regulatory requirements.
Worked Example: When Variability for Both Formulations Matters
Suppose a hypothetical study team needs repeated exposure information for both T and R, rather than repeated information for R alone. The team considers a three-period TRR/RTR/RRT design and a four-period TRTR/RTRT design.
In the partial replicate, each participant receives T once and R twice. The repeated R occasions allow within-participant R comparisons; there is no corresponding pair of T occasions for the same participant. Therefore, this schedule does not meet the stated objective of repeated information for both formulations.
In the full replicate, each participant receives T twice and R twice. For a participant assigned TRTR, the schedule contains two T periods and two R periods; for a participant assigned RTRT, it does as well. This structure is therefore aligned with the objective of observing each formulation repeatedly, provided the study is randomized and the planned analysis accounts for period and other relevant effects.
The four-period design requires one more period per participant than the partial replicate. If repeated information for both formulations is essential, that added burden buys the needed replication. If only repeated reference information is required, the extra T period may not be justified. Neither conclusion determines the acceptance limits; those come from the applicable NTI requirements.
Assumptions, diagnostics, and failure modes
Crossover design comparisons rely on several practical and statistical conditions. Participants should be randomized to sequence, and the treatment schedule should permit the intended contrasts. Washout should be sufficient to avoid clinically relevant residual treatment effects at the next period. The planned analysis should represent treatment and period effects appropriately. In addition, the interpretation of repeated observations as within-subject variability depends on the model’s assumptions about occasion effects; unexplained trends or correlated observations can change what replication tells us.
Review the conduct and data before relying on the design’s intended information. Check sequence allocation and completion by period; inspect whether withdrawals or missed visits are concentrated in a sequence or treatment; review endpoint distributions and period patterns; and investigate protocol deviations, washout concerns, and unusual observations. These checks can reveal threats to interpretation, but a visual or operational check cannot prove that carryover is absent or that model assumptions hold.
- Carryover concern: A treatment effect persisting into a later period can contaminate the next observation. Adequate washout and protocol review are central; a sequence comparison alone does not guarantee that carryover is harmless.
- Period imbalance or missing periods: If participants disproportionately lack later observations, especially in longer designs, the realized data may not retain the planned balance or replication. Report completion by sequence and period.
- Unmodeled period effects: A systematic shift over periods may be mistaken for a formulation difference if it is ignored. The analysis plan should address period effects rather than assuming they cannot occur.
- Excessive participant burden: More periods create more opportunities for withdrawal and missed visits. A nominally richer design may yield less useful information if completion is poor.
Comparing the alternatives and avoiding a common error
| Question | Two-period | Three-period partial replicate | Four-period full replicate |
|---|---|---|---|
| Can each participant receive both T and R? | Yes | Yes | Yes |
| Does each participant repeat R? | No | Yes | Yes |
| Does each participant repeat T? | No | No | Yes |
| Main design trade-off | Lowest period burden; no treatment replication | Reference replication with an intermediate burden | Replication of both treatments with the highest period burden |
A frequent mistake is to assume that an NTI drug automatically requires the most replicated design, or that adding periods makes the acceptance limits less strict. The correction is to separate design information from the bioequivalence criterion. Replication determines which repeated within-participant comparisons are available. The applicable limits and decision rules are specified by the relevant regulatory framework; they are not loosened merely because a replicate design is used.
No design is universally best. A two-period crossover may be sufficient for a mean-comparison objective if permitted by the applicable requirements. A partial replicate is suited to a design objective that specifically needs repeated reference observations. A full replicate is the option among these three when repeated observations of both formulations are needed. All require careful attention to washout, period effects, randomization, and missing data, and all have limits: a richer schedule cannot compensate for poor conduct, an unsuitable analysis, or an objective the design was not built to answer.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a TR/RT two-period crossover, how many times does each participant receive T and R?
- Which formulation is repeated for each participant in a TRR/RTR/RRT partial replicate?
- A study requires repeated within-participant observations of both formulations. Which of the three designs directly matches that objective?
- For 24 participants completing all periods, how many treatment occasions are scheduled in a four-period design?
- Does a four-period full-replicate design, by itself, change the applicable bioequivalence acceptance limits?
Solutions. (1) Each participant receives T once and R once. (2) R is repeated twice; T is given once. (3) The four-period full replicate, because each participant receives T twice and R twice. (4) \(24\times4=96\) treatment occasions. (5) No. Replication changes the repeated information available from the design, not the acceptance limits; the applicable regulatory framework determines those limits.