What Repetition Adds to a Crossover Study
A two-period crossover can compare test and reference formulations within the same participant: each participant receives each formulation once. That comparison is useful for estimating an average formulation difference while reducing the influence of stable differences between participants. But a single response after each formulation does not show how much that participant’s response would vary if the same formulation were given again under comparable conditions. A difference between test and reference could reflect a formulation effect, occasion-to-occasion variation, or both.
A replicate crossover design administers one or both formulations more than once to a participant in separate study periods. Each administration produces a period-level response, such as the log-transformed AUC or Cmax from that dosing occasion. Repetition gives observations of the same participant-formulation combination across occasions. This makes it possible to study repeatability as well as the average comparison, provided the design and conduct support that interpretation.
The distinction matters. Multiple concentration measurements during one dosing period describe one pharmacokinetic profile; they are not independent administrations of the formulation. Likewise, re-assaying the same stored sample can help assess analytical precision, but it does not reveal the participant’s occasion-to-occasion response variation. Replicate dosing addresses a different source of variation.
Why a Single Administration Per Formulation Is Limited
Let \(Y_{ij}\) denote a period-level response for participant \(i\) on administration occasion \(j\). For explanation, consider a simple model for repeated administrations of the same formulation:
Here \(\mu\) is the average response for that formulation, \(b_i\) represents a stable participant-specific departure from that average, and \(\varepsilon_{ij}\) represents occasion-specific variation and measurement noise. This is a conceptual model, not a complete analysis model for every crossover study. It assumes that the formulation and response scale are consistent across occasions, that the stable participant component is meaningful over the study, and that the occasion errors have mean zero and are not systematically shifted by period, carryover, or other unmodeled effects.
With only one administration of a formulation for a participant, the observed response combines \(b_i\) and \(\varepsilon_{ij}\). There is no second response under that same participant-formulation condition to help distinguish a stable participant level from occasion-specific deviation. A second administration supplies a within-participant comparison. Subtracting two responses to the same formulation cancels the stable component:
This cancellation is the basic reason replication is informative about repeatability. It does not mean that every difference is “pure noise”: period trends, residual drug effects, changes in participant condition, and treatment-by-period effects can also contribute unless controlled or modeled. The design must make those influences negligible or account for them.
Replication can therefore address questions that a simple two-period design with one administration of each formulation cannot answer directly. How similar are repeat responses to the same formulation within a participant? Is the average test-reference contrast accompanied by notably different repeatability under the two formulations? How much observed variation is between participants, and how much is occasion-to-occasion? The exact quantities that can be estimated depend on which formulations are repeated and on the design; replication alone does not guarantee that every variability component is identifiable.
Worked Example: Seeing Repeatability Within One Participant
Suppose a hypothetical participant receives the reference formulation on three separate occasions. The period-level log AUC values are 6.20, 6.35, and 6.15. Assume, for this illustration, that the response scale is consistent and there is no systematic period or carryover effect.
The participant’s mean log AUC across these administrations is:
The three responses are not identical. Their deviations from 6.233 are approximately \(-0.033\), \(0.117\), and \(-0.083\) log units. The range is \(6.35 - 6.15 = 0.20\) log units. These calculations describe the spread of this participant’s observed repeat responses; they do not establish a population variability estimate from one participant.
If only the first reference administration had been observed, the value 6.20 could not reveal whether this participant’s next response would usually be close to 6.20 or substantially different. The repeated administrations provide direct information about this participant’s repeatability on the study scale, subject to the assumptions stated above.
Average Comparison and Variability Are Different Questions
Replication can support two related but distinct aims. One is estimating the average difference between formulations. The other is characterizing the variation among repeated responses to a formulation within participants. The first concerns a location contrast; the second concerns dispersion or repeatability. A design may provide useful information about one aim without answering the other equally well.
For example, if both formulations are repeated, a participant-level average for each formulation can be compared. Repeated responses also show how much the participant’s observations vary within each formulation. If only one formulation is repeated, the study directly supplies repeat observations for that formulation, but it does not supply the same direct repeated-response information for the other formulation. The conclusions must match what was actually replicated.
Worked Example: Separating a Mean Contrast From Repeat Spread
Consider hypothetical log-AUC results from two participants, each with two reference and two test administrations. Suppose period and carryover effects are adequately controlled, so the illustrative within-person summaries are interpretable.
| Participant | Reference replicates | Test replicates |
|---|---|---|
| A | 6.00, 6.10 | 6.05, 6.15 |
| B | 6.80, 6.90 | 6.85, 6.95 |
For participant A, the reference mean is \((6.00+6.10)/2=6.05\), and the test mean is \((6.05+6.15)/2=6.10\). The test-minus-reference contrast is \(6.10-6.05=0.05\) log units. Participant B has reference mean \((6.80+6.90)/2=6.85\), test mean \((6.85+6.95)/2=6.90\), and the same contrast, \(0.05\) log units.
The mean of these two participant-level contrasts is \((0.05+0.05)/2=0.05\) log units. Separately, each participant’s two reference values differ by 0.10 log units, and the two test values also differ by 0.10 log units. Thus the table illustrates both a formulation contrast and repeated-response spread. The repeated observations do not make the average contrast itself a variability measure.
On the original multiplicative scale, a log difference of 0.05 corresponds to a ratio of \(\exp(0.05)\), approximately 1.051. In this small hypothetical illustration, the mean test response is about 5.1% higher than the reference response on that scale. This arithmetic alone is not a bioequivalence conclusion: no uncertainty interval or formal decision criterion has been assessed.
Assumptions, Diagnostics, and Failure Modes
A repeated response is informative only if administrations are sufficiently comparable. The formulation, dose, administration conditions, sampling procedures, endpoint definition, and assay approach should be consistent. Participants should be in a comparable state at each period, or relevant changes should be recorded and considered. Washout must be adequate for the scientific question, and the possibility of pharmacokinetic or pharmacodynamic carryover must be evaluated. These conditions are especially important because a systematic change across periods can masquerade as repeat variability.
Useful diagnostics include plotting each participant’s responses by period and formulation, examining period-level summaries, checking predose concentrations where appropriate, and reviewing protocol deviations and missing periods. Look for systematic increases or decreases over time, unusual observations, and response patterns that differ by formulation or sequence. A wide spread should not automatically be attributed to intrinsic formulation variability: it may reflect carryover, assay problems, changes in participant condition, or inconsistent conduct.
Replication also does not repair every design weakness. If the effect of formulation is confounded with period or sequence, adding administrations without suitable allocation and analysis may leave the central comparison biased or ambiguous. Withdrawals can leave incomplete replicate records, and unequal completion can affect precision and interpretation. The analysis must reflect the observations actually available and the study’s assignment structure.
Worked Example: A Period Trend Can Mimic Poor Repeatability
In a hypothetical study, one participant’s reference log-AUC responses are 6.10 in an early period and 6.40 in a much later period. The raw repeat difference is \(6.40-6.10=0.30\) log units. If the study assumes comparable occasions, that difference might be read as evidence of substantial within-participant variation.
Now suppose the participant’s recorded health status and a study-wide plot show that responses rose over the same period for several participants. The arithmetic difference remains 0.30 log units, but its interpretation changes: the later response may include a time-related shift as well as ordinary repeat variation. The study should investigate period effects and other time-varying explanations before describing the difference as formulation-specific repeatability.
The diagnosis is not to discard the later value simply because it is inconvenient. It is to identify whether the assumptions for interpreting the repeat difference are credible, document the evidence, and use an analysis appropriate to the design. Replication exposes patterns that one observation per formulation would not show, but it does not identify their cause automatically.
How Replicate Designs Compare With Alternatives
| Approach | What it can address | Main limitation |
|---|---|---|
| Two-period crossover with one administration of each formulation | Within-participant average test-reference comparison | No direct repeated response to the same formulation within a participant |
| Replicate crossover with one or both formulations repeated | Average comparison plus repeatability information for the formulation(s) actually repeated | More periods, participant burden, and greater exposure to period effects, carryover, and missingness |
| Parallel-group comparison | Comparison between groups without repeated crossover periods | Does not provide a within-participant formulation contrast or direct within-participant repeat observations |
| Repeat assays of one sample | Information about analytical repeatability for that sample and assay process | Does not measure variation across separate dosing occasions |
The design choice should follow the scientific question. If the objective is only an average formulation comparison and within-participant repeatability is not needed, additional administrations may add burden without corresponding value. If treatment-specific repeatability is important, a design that repeats relevant formulation administrations may be necessary. It will generally require more participant time, dosing occasions, sampling, and oversight. No design eliminates the need to define the endpoint, control period conditions, and plan for incomplete records.
Common Pitfall
The practical value of replicate crossover designs is therefore not simply a larger data set. Their distinctive contribution is observing a formulation more than once within the same participant, which can reveal repeatability and support variability questions unavailable from a single administration per formulation. The average treatment contrast remains a separate estimand, and the design must be planned around both the desired comparison and the sources of variation it is meant to characterize.
Check Your Understanding
Answer each question, then compare with the solutions below.
- Why do repeated blood samples within one dosing period not count as replicate administrations?
- In the model \(Y_{ij}=\mu+b_i+\varepsilon_{ij}\), what component cancels when two responses to the same formulation are subtracted?
- A participant has repeated log-AUC values of 5.90 and 6.10 for one formulation. Calculate their mean and their difference, and state what each describes.
- Why can a period trend make repeated responses difficult to interpret as repeatability?
- Does repeating only the reference formulation give the same direct repeat-response information for test? Explain.
Solutions. (1) Those samples belong to one dosing occasion and one pharmacokinetic profile; they do not reflect separate administrations. (2) The stable participant component \(b_i\), and the shared mean \(\mu\), cancel; the difference contains occasion-specific deviations under the conceptual model. (3) The mean is \((5.90+6.10)/2=6.00\) log units, summarizing the two responses; the difference is \(6.10-5.90=0.20\) log units, describing their observed spread. (4) A systematic time-related shift can contribute to the difference, so it cannot automatically be attributed to ordinary occasion-to-occasion variation. (5) No. Repeated reference administrations directly inform repeat responses to reference; test is not directly repeated within that participant.