Why Compare Treatments Within the Same Participant?
Participants can differ substantially in characteristics that affect pharmacokinetic outcomes, including body size, metabolism, and renal function. In a crossover study, each participant receives more than one treatment, so the treatment outcomes can be compared within that participant. The participant’s own outcome under one treatment serves as a control for their outcome under another.
This pairing can make a treatment contrast more precise than a comparison between separate groups. It does not, however, automatically make the contrast unbiased or valid. The comparison is informative only when the outcomes under the different treatments are comparable apart from the treatment difference of interest and the relevant design assumptions are reasonable.
The contrast is built from two measurements on one person, usually obtained in different study periods. It is not the difference between two independent observations. The earlier material on the bioequivalence crossover estimand identifies the population treatment contrast of interest; here, the focus is on why observed within-participant contrasts can estimate it and what can undermine that interpretation.
How Stable Participant Differences Cancel
A simple model makes the logic explicit. Suppose the outcome under treatment \(k\) for participant \(i\) can be written as
where \(\mu_k\) is the population mean under treatment \(k\), \(b_i\) is a stable participant-specific contribution shared across treatments, and \(\varepsilon_{ik}\) is the remaining deviation for that participant and treatment. Subtracting the reference outcome from the test outcome gives
The stable contribution \(b_i\) disappears. Thus, a participant who tends to have higher exposure under both treatments does not, by that stable tendency alone, create a difference between the treatments. This cancellation is the central benefit of pairing.
The cancellation does not remove every source of variation. Short-term biological fluctuations, measurement error, changes between periods, and treatment-specific responses remain in the paired contrast. Nor does pairing erase genuine variation in treatment effects between participants. Those differences contribute to the distribution of \(d_i\) and must be represented in uncertainty about the population contrast.
For a pharmacokinetic outcome, the scale matters. If \(Y\) is log AUC, then \(d_i\) is the log of the participant’s test-to-reference AUC ratio. The average contrast on the log scale can be exponentiated to express a geometric mean ratio. This does not make the measurements unit-free in every practical sense: the original outcome and its units should still be stated when describing what was measured.
Why Pairing Can Improve Precision
The precision advantage depends on how strongly the two outcomes from the same participant are correlated. The variance of the paired contrast is
If the two treatment outcomes have the same variance \(\sigma^2\) and correlation \(\rho\), this becomes
The result follows by substituting \(\operatorname{Var}(Y_{iT})=\operatorname{Var}(Y_{iR})=\sigma^2\) and \(\operatorname{Cov}(Y_{iT},Y_{iR})=\rho\sigma^2\) into the variance formula. When \(\rho\) is positive, the covariance term reduces the variance of the difference. When \(\rho=0\), there is no such reduction; if \(\rho\) is negative, the paired difference can be more variable than it would be with uncorrelated outcomes.
For \(n\) independent participants, the estimated mean paired contrast is \(\bar{d}\), and its standard error is \(s_d/\sqrt{n}\), where \(s_d\) is the sample standard deviation of the paired contrasts. A parallel-group comparison with \(n\) participants per treatment, under equal outcome variances and independence between groups, has variance \(2\sigma^2/n\) for the difference in sample means. Under the equal-variance paired model, the variance of the mean paired contrast is \(2\sigma^2(1-\rho)/n\). The comparison shows why positive within-participant correlation can improve precision; it does not establish that every crossover is preferable, because crossover assumptions and practical costs also matter.
Worked Example: Computing a Paired Contrast
In a hypothetical crossover study, eight participants have log-transformed AUC measured under both treatments. Suppose the participant-level test-minus-reference contrasts, in log units, are \(0.08,-0.02,0.05,0.01,0.10,-0.04,0.03,0.03\). Assume for this illustration that the paired contrasts are independent across participants, have no influential outliers, and are reasonably consistent with a normal distribution. These assumptions would need assessment in an actual analysis.
The mean contrast is
The squared deviations from \(0.0300\) sum to \(0.0156\), so the sample standard deviation and standard error are
Using a two-sided 95% \(t\) interval with \(df=7\) and critical value \(2.365\), the interval is
Exponentiating the mean contrast gives \(\exp(0.0300)=1.030\), a point estimate of a 1.030 geometric mean test-to-reference ratio. Exponentiating the interval endpoints gives approximately \(0.991\) to \(1.072\). Under the stated assumptions, the interval is a plausible range for the population log-scale contrast, or equivalently for its exponentiated ratio after transformation. This example illustrates the paired calculation; any regulatory conclusion would require the applicable prespecified criteria and analysis.
Worked Example: The Precision Gain From Positive Correlation
Suppose a hypothetical study has log-AUC outcomes with standard deviation \(0.20\) under each treatment and within-participant correlation \(0.75\). Assume the two treatment outcomes have equal variances and that participants are independent. For \(n=16\) participants, the standard deviation of a paired contrast is
The corresponding standard error of the mean paired contrast is \(0.1414/\sqrt{16}=0.0354\) log units. For comparison, if two independent groups each contained 16 participants and had the same outcome standard deviation, the standard error for their difference in means would be
In this hypothetical calculation, the paired standard error is half as large because the repeated outcomes are strongly positively correlated. This is a variance comparison under the stated assumptions, not a universal promise of a twofold precision gain. If the correlation were near zero, the covariance would provide little or no advantage.
When the Within-Participant Comparison Can Mislead
The cancellation argument assumes that the participant-specific contribution is stable across the measurements. A difference in conditions between measurements can instead enter the contrast alongside the treatment effect. For example, a systematic change over time may be mistaken for a treatment difference if treatment and measurement occasion are not adequately separated by the design or analysis.
Worked Example: A Time Trend Masquerading as a Treatment Difference
Consider a hypothetical illustration in which every participant receives the reference measurement first and the test measurement later. Suppose the true treatment contrast on the log-AUC scale is \(0.00\), but a time-related change increases the later measurement by \(0.12\) log units. For participant \(i\), write the outcome as treatment mean plus a stable participant contribution, a time contribution, and random error. The observed contrast is then
If the true treatment contrast is zero and the average error difference is zero, the average observed contrast is \(0.00+0.12=0.12\) log units. Its exponentiated value is \(\exp(0.12)=1.128\), even though the hypothetical treatment effect is null. Pairing cancels the stable participant contribution, but it does not cancel a systematic change that occurs between measurements and is aligned with treatment.
The practical lesson is to examine whether study conduct and the design make treatment comparisons separable from time-related changes. Relevant checks include reviewing treatment and measurement schedules, assessing whether washout and sampling conditions are appropriate, and looking for evidence that residual drug or changing conditions could affect later outcomes. A diagnostic cannot rescue a design that does not permit the treatment contrast to be distinguished from these influences.
Carryover is a related concern: an earlier treatment may affect an outcome recorded under a later treatment. If that happens, \(Y_{ik}\) cannot be understood simply as the outcome under treatment \(k\) alone, and subtracting two measurements does not guarantee an uncontaminated treatment contrast. The relevant assumption is not merely that a washout interval was scheduled, but that the prior treatment has no material residual influence on the outcome used for the comparison.
Other warning signs include missing one of a participant’s paired outcomes, inconsistent measurement procedures across treatments, substantial departures from the assumptions used for inference, or changes in participant health that coincide with treatment. Missing a member of a pair can prevent that participant from contributing to a simple paired contrast; analyzing only complete pairs may also be problematic if missingness is related to treatment or outcome. These issues call for design review and an analysis appropriate to the missing-data pattern, not an assumption that pairing has removed the problem.
Choosing Between Within-Participant and Between-Participant Comparisons
| Feature | Within-participant comparison | Parallel-group comparison |
|---|---|---|
| Source of comparison | Different treatment outcomes from the same participant | Outcomes from participants assigned to different treatments |
| Stable participant differences | Cancel from the paired contrast if their effects remain stable across treatments | Contribute to between-participant variation; randomization supports group comparability on average |
| Key vulnerability | Carryover, period-related changes, or other differences between measurements can contaminate the contrast | Imbalance between groups can occur by chance, particularly in small samples |
| Practical requirement | Each participant must be able to receive the treatments and provide usable outcomes under each | Each participant generally receives one assigned treatment |
A crossover is attractive when within-participant correlation is positive and participants can safely and feasibly receive each treatment with outcomes that remain comparable across measurement occasions. A parallel-group design avoids the need to interpret repeated outcomes in light of prior treatment, but it does not benefit from cancelling stable participant differences within the same person. The choice depends on the scientific question, treatment characteristics, participant burden, feasibility, and the plausibility of the assumptions—not on a general rule that one design is always more efficient.
Common Mistake
- Wrong reasoning: “Each participant is their own control, so participant differences and all other sources of bias are removed.”
- Correction: Pairing cancels stable participant contributions shared across treatments. It does not remove carryover, time-related changes, treatment-specific measurement problems, or selection and missing-data concerns. The comparison is informative only if the design and assumptions address those sources.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In the model \(Y_{ik}=\mu_k+b_i+\varepsilon_{ik}\), which component cancels from \(Y_{iT}-Y_{iR}\), and which components remain?
- If the two treatment outcomes have equal variance \(\sigma^2\) and correlation \(\rho=0.40\), express the variance of the paired contrast in terms of \(\sigma^2\).
- Why can a time-related change create an apparent treatment contrast even when each participant serves as their own control?
- Name two conditions that can make a within-participant comparison uninformative or misleading.
- Does a paired design always have a smaller standard error than a parallel-group design? Explain.
Solutions. (1) The stable participant contribution \(b_i\) cancels; the contrast contains \(\mu_T-\mu_R\) and \(\varepsilon_{iT}-\varepsilon_{iR}\). (2) \(\operatorname{Var}(d_i)=2\sigma^2(1-0.40)=1.20\sigma^2\). (3) If measurement occasion is associated with treatment, the contrast can contain both the treatment difference and the time-related change; subtracting outcomes removes only contributions shared stably across the measurements. (4) Examples include material carryover, systematically different conditions between measurements, noncomparable measurement procedures, and informative missingness of one outcome in a pair. (5) No. Positive within-participant correlation can reduce the variance of paired contrasts, but if correlation is near zero there may be little advantage, and negative correlation can increase the variance. Design assumptions and practical feasibility also matter.