Start With the Information the Study Must Provide
Choosing a crossover design is not simply a matter of selecting two, three, or four periods. The central question is which treatment contrasts and within-participant information must be estimable? A study concerned only with the average difference between test (T) and reference (R) may need a different layout from one that also needs a treatment-specific estimate of within-participant variability.
A two-period crossover assigns each participant to T then R, or R then T. A replicate design gives at least one treatment more than once to each participant. A reference-replicate design repeats R; a test-replicate design repeats T; a full-replicate design repeats both. The familiar examples below use two-period sequences TR and RT, three-period sequences TRR/RTR/RRT or TTR/TRT/RTT, and four-period sequences TRTR/RTRT.
The principal average comparison is the within-participant contrast between T and R, commonly analyzed on a suitable transformed outcome such as log exposure. All four designs can support that contrast when their sequence allocation and period structure make it estimable. Their key distinction is not whether they compare average T and R, but what they can learn about repeated responses to each treatment.
What Each Design Can Estimate
The design choice rests on the variability question. In a standard two-period crossover, each participant has only one response under each treatment. Under a model that assumes a common within-participant residual variance for T and R, the residual variation can be estimated from the crossover data. That estimate does not, without additional assumptions, separate within-participant variability under R from variability under T.
Replication supplies repeated responses under the same treatment. Their differences remove stable participant-specific treatment levels and can therefore inform that treatment’s within-participant variability. In a reference replicate, R-specific within-participant variability is estimable, but T-specific within-participant variability is not directly estimated from repeat T exposures. A test replicate reverses those roles. A full replicate can estimate variability for both treatments.
| Design | Average T-versus-R comparison | Treatment-specific within-participant variability | Typical reason to choose it |
|---|---|---|---|
| Two-period | Yes, under a suitable crossover model | Not separately for T and R without extra assumptions | The average comparison is the target and treatment-specific variability is not required |
| Reference replicate | Yes | R-specific variability | R variability is needed, for example for an analysis that requires it |
| Test replicate | Yes | T-specific variability | The study specifically needs within-T variability |
| Full replicate | Yes | Both T- and R-specific variability | Both variability estimates are required or their comparison is a study objective |
A design must also satisfy the assumptions that make its estimates meaningful. The assigned sequences should allocate treatments across periods so treatment and period effects can be distinguished. Period effects should be included or otherwise handled when relevant. A crossover interpretation assumes participants can receive the assigned treatments as planned, outcomes are measured comparably across periods, and carryover does not distort later-period responses. Washout, adherence, and the schedule must be adequate for the setting; replication by itself does not eliminate carryover or period trends.
Why Repeated Responses Identify Treatment-Specific Variability
A simple model makes the role of replication clear. Suppose a participant’s period-adjusted response under R is \(Y_{iRj}=\mu_R+b_{iR}+\varepsilon_{iRj}\), where \(\mu_R\) is the treatment mean, \(b_{iR}\) is a stable participant-specific deviation under R, and \(\varepsilon_{iRj}\) is occasion-level variation. Assume the occasion errors have mean zero, variance \(\sigma^2_{WR}\), and are independent across the participant’s R periods. These assumptions are a useful explanation of the design principle, not a guarantee that real data follow the model exactly.
For two R responses from participant \(i\), the difference cancels the stable participant component:
The variance equality follows because the two errors are independent and each has variance \(\sigma^2_{WR}\). Thus, after appropriate adjustment for period effects, the variance of within-participant R differences divided by two estimates \(\sigma^2_{WR}\). With no repeated R response, as in a two-period design, there is no such within-participant R difference. This is why a separate reference-specific estimate requires reference replication rather than merely a larger sample.
Worked Example: Choosing a Two-Period Design for an Average Comparison
A hypothetical study asks whether the average log-transformed exposure under T differs sufficiently little from that under R for a conventional average comparison. It does not require separate estimates of T- and R-specific within-participant variability. Participants can be randomized to TR or RT, and the plan includes a period effect.
Suppose the estimated mean log difference is \(0.025\). Exponentiating gives the ratio on the original scale:
The estimated geometric mean ratio is therefore \(1.025\), or 102.5%. This calculation describes the estimated average contrast; it does not, by itself, establish equivalence, which would require an appropriate inferential interval and prespecified criteria.
A two-period TR/RT design is a reasonable choice for the stated objective because each participant receives both treatments and the average within-participant contrast is available. The study does not need repeat exposures to either treatment. If the investigators later require a treatment-specific variability estimate, this design does not provide it without an additional variance assumption. The design decision follows the stated information need, not the numerical ratio alone.
Worked Example: Choosing a Reference-Replicate Design
A hypothetical study needs the average T-versus-R comparison and an estimate of within-participant variability under R. It does not need a direct estimate of within-participant variability under T. A three-period reference-replicate schedule such as TRR, RTR, and RRT gives every participant two R periods and one T period.
For illustration, suppose six participants’ R-response differences, after period adjustment, are \(0.08,-0.10,0.04,0.12,-0.06,-0.08\) log units. The differences sum to zero, so their sample mean is zero. Their sum of squared deviations is:
The sample variance of the differences is \(0.0424/(6-1)=0.00848\) log-squared units. Under the independent, equal-variance occasion-error assumptions in the derivation, the estimated within-R variance is half that value:
The corresponding estimated within-R standard deviation is \(\sqrt{0.00424}=0.0651\) log units. These hypothetical values illustrate the calculation, not a conclusion about any actual treatment. The schedule supplies the repeated R observations needed for this estimate while retaining exposure to T for the average comparison. It does not provide the analogous direct repeat-based estimate for T.
Worked Example: Choosing a Test-Replicate or Full-Replicate Design
A hypothetical development team first asks only for within-T variability in addition to the average T-versus-R contrast. A three-period test-replicate schedule, such as TTR, TRT, and RTT, repeats T for every participant. This is more directly matched to that question than a reference replicate, which repeats R instead.
Now suppose the objective changes: the team needs both treatment-specific variability estimates to compare response variability under T and R. Consider a four-period full replicate, such as TRTR/RTRT. For a small hypothetical illustration, let the period-adjusted repeated-response differences be \(0.08,-0.10,0.04,0.12,-0.06,-0.08\) for R and \(0.03,-0.04,0.02,0.05,-0.01,-0.05\) for T.
For R, the sum of squared deviations is \(0.0424\), giving \(s^2_{d,R}=0.0424/5=0.00848\), and the within-R variance estimate is \(0.00848/2=0.00424\) log-squared units. For T, the differences also sum to zero and their squared deviations sum to \(0.0080\). Thus:
Both estimates are available because each participant has repeat responses under both treatments. Here the estimated within-R variance exceeds the estimated within-T variance, but with only six hypothetical differences these values are illustrative and do not establish a reliable population difference. Full replication is justified by the requirement to estimate both variabilities, not simply by the fact that it produces more observations.
Diagnostics, Alternatives, and Limits
Before selecting a design, write down the primary treatment comparison and any secondary variability targets. Then check whether the sequences actually repeat each treatment the required number of times and whether period allocation permits the intended treatment and period effects to be distinguished. Sequence coverage and exposure counts should be checked directly rather than inferred from a design’s label.
Potential warning signs include incomplete or poorly balanced sequence allocation, substantial period trends that are not accounted for, inconsistent measurements across periods, and evidence that a prior treatment may influence a later response. For a replicate analysis, inspect the repeat-response differences for unusual values, treatment- or period-related patterns, and variability that differs markedly across participants. Such patterns can challenge the assumed common within-treatment variance or independent occasion errors. They should prompt investigation of data quality and model adequacy; they do not justify automatically discarding observations or interpreting a variance estimate as definitive.
The main alternative to adding periods is to retain a two-period design and make a common-variance assumption if the scientific question and analysis permit it. That may be adequate for an average comparison, but it cannot answer a question requiring separately observed repeat variability for T or R. Conversely, selecting a replicate design when treatment-specific variability is not needed adds periods without adding information essential to the stated objective. The choice is therefore an estimability decision, not a universal ranking of designs.
A replicate layout also does not guarantee a valid or efficient study. Treatment contrasts still depend on appropriate randomization, period handling, washout, measurement, and analysis. A particular regulatory procedure may impose requirements beyond the design’s ability to estimate a variance; replication alone does not make a scaled analysis appropriate or establish bioequivalence. Finally, these schedules are examples: sequence choice and allocation must be evaluated in the actual study context.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A study needs an average T-versus-R comparison but no treatment-specific variability estimate. Which design is the simplest match, and why?
- A three-period study repeats R but gives T only once. Which treatment’s within-participant variability can be estimated directly from repeated responses?
- Under independent occasion errors with variance \(\sigma^2_{WR}\), why is the variance of two R responses’ difference \(2\sigma^2_{WR}\)?
- A study needs separate within-T and within-R variability estimates. Which of the four design types meets that information requirement?
- Why does a replicate design not automatically solve period-trend or carryover problems?
Solutions. (1) A two-period crossover is the simplest match because every participant receives both treatments, supporting the average within-participant contrast, and no repeat-specific variability is required. (2) Within-R variability, because each participant has repeat R responses; T is not repeated. (3) The difference cancels the stable participant-specific component, and independent errors have \(\operatorname{Var}(\varepsilon_{iR1}-\varepsilon_{iR2})=\operatorname{Var}(\varepsilon_{iR1})+\operatorname{Var}(\varepsilon_{iR2})=2\sigma^2_{WR}\). (4) A full-replicate design repeats both treatments. (5) Replication supplies repeated treatment responses, but period trends still need suitable allocation and modeling, while carryover still requires adequate design and washout assumptions.