Why Design Changes the Meaning of an Exposure Contrast
A drug–drug interaction (DDI) study often compares exposure to a victim drug when it is given alone with exposure when it is given with a perpetrator. The contrast of interest may be the ratio of geometric mean AUC values, with the combination condition in the numerator. The conditions can be assigned to the same participants or to different groups; that choice affects both the precision and the assumptions needed to interpret the result.
A crossover design gives participants both conditions in different periods. A parallel-group design assigns each participant to one condition. A fixed-sequence design gives all participants the conditions in a set order, commonly victim drug alone followed by victim plus perpetrator. These designs can target the same nominal contrast, but they do not protect against the same sources of bias.
This contrast is a comparison of the specified conditions in the study population; it is not, by itself, proof that the perpetrator caused the difference. Interpretation depends on allocation, period effects, adherence, and whether the conditions are sufficiently separated to avoid relevant carryover. The introductory material on DDI exposure-ratio estimands establishes the importance of stating the numerator and denominator; here, the focus is how design supports that contrast.
Three Design Structures
| Design | Who receives each condition? | Main advantage | Main concern |
|---|---|---|---|
| Crossover | Each participant receives both conditions, usually in randomized order. | Within-subject comparison can remove stable participant differences. | Period effects, carryover, and the time needed to complete multiple periods. |
| Parallel-group | Each participant receives only one condition; participants are randomized between groups. | No within-participant carryover between conditions. | Between-subject variability can reduce precision; groups must be comparable. |
| Fixed-sequence | All participants receive conditions in the same order. | Can be operationally straightforward and uses each participant as a reference. | Treatment is confounded with order or period unless those effects are negligible or otherwise addressed. |
In a crossover study, randomized treatment sequences help separate treatment from systematic period effects across participants, provided the design and analysis account for period and sequence. The within-participant comparison also controls stable characteristics such as body size or baseline metabolic capacity. It does not automatically control changes over time, nor does it remove carryover from an earlier condition.
In a parallel study, randomization makes treatment assignment independent of participant characteristics in expectation. It does not guarantee identical groups in a particular small study. A fixed-sequence study has no randomized order: if everyone receives condition A first and condition B second, a change between periods can be attributed to treatment only if other period-related changes are absent or adequately accounted for.
Why Within-Subject Comparisons Can Be More Precise
A simple log-exposure model makes the precision distinction clear. For participant \(i\) in period \(j\), write \(Z_{ij}=\mu+\tau T_{ij}+b_i+\pi_j+\varepsilon_{ij}\), where \(T_{ij}\) indicates the combination condition, \(b_i\) is a stable participant effect, \(\pi_j\) is a period effect, and \(\varepsilon_{ij}\) is residual variation. This is a working model: it assumes the log exposure is suitable for the contrast and that residual behavior is adequately represented by the chosen variance structure.
For a participant observed in both conditions, subtracting the victim-alone log exposure from the combination log exposure cancels \(b_i\). If the residual variances in the two periods are both \(\sigma_e^2\) and their correlation is \(\rho\), the residual variance of the difference is \(2\sigma_e^2(1-\rho)\). Thus positive within-participant correlation can improve precision. This comparison is unbiased for \(\tau\) only when period effects and carryover do not create an unaccounted difference between conditions, or when the design and analysis appropriately account for them.
The cancellation is not magic: it applies to stable participant effects, not to period differences. In a balanced randomized crossover, treatment conditions occur in each period across participants, allowing period and treatment effects to be distinguished under the model. In a fixed sequence, treatment and period are aligned, so the same separation is not provided by randomization.
For a parallel design with \(n\) participants per group, suppose log exposure has independent variance \(\sigma_b^2+\sigma_e^2\) in each group, where \(\sigma_b^2\) represents between-participant variation. The variance of the difference between group means is \(2(\sigma_b^2+\sigma_e^2)/n\). A paired crossover contrast instead depends on the variance of within-participant differences. Consequently, it can be more efficient when stable between-participant variation is substantial, but it is not always preferable: feasibility and bias control matter as much as nominal precision.
Worked Example: A Crossover Exposure Ratio
In a hypothetical randomized two-sequence crossover study, 12 participants complete both conditions. Their mean log AUC is 4.200 for victim drug alone and 4.605 for the combination. Assume for this calculation that period effects and carryover are negligible or adequately controlled, and that the two conditions measure the same AUC endpoint.
The estimated log contrast is \(4.605-4.200=0.405\). Exponentiating gives \(\exp(0.405)=1.499\) to three decimal places. Thus the estimated geometric mean AUC in the combination condition is about 1.499 times that under victim drug alone, or 149.9% of it. This is an estimated contrast, not a statement about every participant; uncertainty would require the variability of the paired log differences.
The use of both conditions in each participant makes stable differences between participants less influential on this contrast. The interpretation still relies on the stated period and carryover assumptions.
Worked Example: Precision in a Parallel-Group Study
Consider a hypothetical randomized parallel study with 16 participants per group. The mean log AUC is 4.605 in the combination group and 4.200 in the victim-alone group. Suppose the observed log-scale standard deviation in each group is 0.300, and observations from different participants are independent.
The estimated contrast is \(4.605-4.200=0.405\), so the estimated ratio is \(\exp(0.405)=1.499\), or 149.9%. The standard error of the difference in group means is
This standard error reflects variation among participants as well as residual variation. Randomization supports a causal interpretation of the assigned-condition contrast in expectation, but with only 16 participants per group chance imbalances can remain. Unlike a crossover, this design does not require participants to receive both conditions and avoids within-participant carryover.
Worked Example: A Fixed-Sequence Contrast and Period Change
In a hypothetical fixed-sequence study, participants first receive victim drug alone and later receive it with a perpetrator. The mean log AUC is 4.200 in the first period and 4.550 in the second. Assume the measured endpoint and dosing conditions are otherwise comparable.
The observed within-participant mean contrast is \(4.550-4.200=0.350\). Exponentiating gives \(\exp(0.350)=1.419\), so observed geometric mean exposure in the second period is about 1.419 times the first-period value.
Now suppose, hypothetically, that an unrelated period change would have increased mean log AUC by 0.100 even without the perpetrator, and there is no other bias. Then the treatment component is \(0.350-0.100=0.250\), corresponding to \(\exp(0.250)=1.284\). The observed ratio of 1.419 would overstate the treatment contrast under these assumptions. Because everyone received the combination condition second, the study itself cannot distinguish that hypothetical period change from treatment without additional information or assumptions.
Choosing a Design in Practice
Design selection should follow the intended contrast and the practical profile of the interaction. A crossover is often attractive when exposure can be measured in both conditions within a feasible study duration, participants can safely receive both conditions, and the interaction does not persist in a way that compromises the later period. A parallel design is often more suitable when a participant should receive only one condition, the interaction may persist, or completing both periods would be impractical. A fixed sequence may be operationally useful when all participants are intended to progress through conditions in the same order, but its treatment contrast requires particular caution about time trends.
Feasibility includes more than the number of participants. Crossover studies require multiple periods and can increase participant burden, time, and dropout opportunities. Parallel studies avoid repeated conditions but may need more participants when between-subject variability is appreciable. Fixed-sequence studies may appear simple, yet the absence of randomized order can make the contrast difficult to interpret. The later material on washout periods addresses how to assess and control carryover; at the design stage, recognize that a longer washout may make a crossover impractical rather than assuming carryover can always be eliminated.
State which exposure endpoint and conditions are compared, and define the ratio direction before choosing a design.
Consider safety, duration, burden, and whether a prior condition could affect a later one.
Use within-subject information when it is valid and feasible; use randomized parallel groups when within-participant comparison is unsuitable; treat fixed sequence as vulnerable to period confounding.
For any design, address adherence and comparable endpoint measurement; also address period and carryover for repeated-condition designs and group comparability for parallel designs.
Diagnostics, Limitations, and a Common Pitfall
Before interpreting a contrast, inspect exposure data by condition, period, sequence, and participant. Look for strong period shifts, unexpectedly different patterns by sequence, missing periods, and extreme within-participant differences. In parallel groups, inspect baseline characteristics and exposure distributions by randomized group; imbalance does not automatically invalidate randomization, but may prompt a prespecified adjusted analysis. These checks can reveal problems; they cannot prove that no carryover or unmeasured period effect exists.
A crossover estimate can be especially precise yet misleading if carryover contaminates the later condition or if period effects are not handled. A parallel estimate avoids within-participant carryover but may be less precise and is sensitive to chance group imbalance. A fixed-sequence paired estimate removes stable participant differences, but does not separate the treatment effect from a systematic change between periods. No design alone guarantees a valid causal interpretation: the assumptions must fit the study conduct and the scientific question.
Common mistake: “The fixed-sequence comparison is paired, so it controls all confounding.” Pairing removes stable participant effects from a within-person difference; it does not remove period trends, time-varying factors, or carryover. The correction is to distinguish what cancels mathematically from what remains: the participant effect cancels, while the difference in period effects remains in the observed contrast.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In a crossover study, what stable source of variability is removed by subtracting each participant’s log exposure between conditions?
- A parallel study reports a combination-minus-alone mean log contrast of 0.200. What is the estimated geometric mean ratio, to three decimal places?
- Why does a fixed-sequence design not, by itself, distinguish a treatment effect from a period effect?
- Name one circumstance that can make a parallel-group design preferable to a crossover design.
Solutions. (1) The stable participant effect \(b_i\) cancels in the within-participant difference. (2) \(\exp(0.200)=1.221\), so estimated combination exposure is 1.221 times victim-alone exposure. (3) All participants receive conditions in the same order, so treatment condition is aligned with period; a period change can contribute to the observed difference. (4) For example, a long-lasting interaction or an impractical or unsafe second condition can make within-participant comparison unsuitable; parallel groups avoid giving both conditions to each participant.