Why Carryover Matters
A crossover study relies on each participant receiving both treatments in separate periods. The within-participant comparison is useful because it controls for stable differences between people. But that comparison becomes difficult to interpret if drug from the first period remains when the next period begins. The second period may then reflect both the treatment assigned for that period and residual exposure from the earlier treatment.
A carryover signal is evidence that this assumption may not hold; it is not, by itself, proof that the treatment comparison is biased. Predose concentrations, washout calculations, and sequence-by-period patterns help assess plausibility. Their interpretation requires context: assay performance, the drug’s elimination, the timing of samples, and whether residual concentrations could affect the pharmacokinetic endpoint.
What the Crossover Comparison Assumes
For a standard two-treatment, two-period crossover, participants are assigned to one of two sequences: AB or BA. A common analysis compares log-transformed pharmacokinetic endpoints, such as \(\ln(\mathrm{AUC})\), using a model that accounts for treatment, period, sequence, and participant within sequence. The treatment estimate is then expressed as a geometric mean ratio after exponentiation.
The interpretation of this estimate depends on the study design and model being appropriate. In particular, the usual within-participant treatment comparison assumes that the first-period treatment does not materially affect the response in the next period. It also relies on correct treatment and period records, appropriate endpoint calculation, and a suitable model for the data. Carryover is a threat to the first of these assumptions: a later-period observation can partly reflect what happened before its assigned treatment.
Washout duration is chosen to make residual drug unlikely, but a long interval is not a guarantee. Conversely, a detectable predose concentration does not automatically mean that the bioequivalence comparison is unusable: the concentration may be negligible relative to postdose exposure, or it may be an assay or recording issue that needs investigation. The central question is whether the evidence makes a material influence on the endpoint plausible.
Use Several Signals, Not One Test
Begin by checking the data and the study timeline. Confirm sequence, period, dose, sampling times, washout duration, and predose sample labels. A misplaced period label or an incorrect sample time can create an apparent carryover pattern. Then inspect predose concentrations by sequence and period, including values below the quantification limit and assay precision near that limit. Examine the concentration-time profiles and relevant pharmacokinetic endpoints by sequence and period as well.
A predose concentration after washout is more concerning when it is reproducible, clearly quantifiable, and consistent with the drug’s expected persistence. A concentration detected in only one sample near the assay’s lower limit is weaker evidence. Predose values should be compared with relevant postdose concentrations and interpreted in their units; a small absolute value may still matter for a low-exposure product or an endpoint sensitive to baseline contribution.
Period patterns are also informative. If later-period measurements are systematically elevated in participants who received a particular treatment first, that pattern may be consistent with residual drug. But period patterns can also arise from chance imbalance, time trends, operational differences, or other study features. Sequence and period comparisons are diagnostics, not a substitute for checking the source data and understanding the pharmacology.
Worked Example: Estimating Whether a Washout Is Plausible
Suppose a hypothetical drug has an elimination half-life of 8 hours, and the washout between periods is 48 hours. Under a one-compartment, first-order elimination approximation, calculate the fraction of drug remaining after washout.
The elimination rate constant is \(k=\ln(2)/t_{1/2}\), where \(t_{1/2}\) is the half-life. The fraction remaining after time \(\tau\) is \(e^{-k\tau}\). Equivalently, each half-life halves the amount remaining.
The calculation predicts about 1.563% of the reference amount remains after 48 hours. If the hypothetical concentration at the start of washout were 10.0 mg/L and elimination followed this approximation throughout, the predicted residual concentration would be \(10.0 \times 0.015625 = 0.1563\) mg/L.
This is a plausibility calculation, not a direct measurement or a universal acceptance rule. It assumes the stated half-life applies over the relevant concentration range and that elimination is adequately described by first-order kinetics. If the observed predose concentration is much higher than expected, investigate the assay, timing, adherence, and pharmacokinetic assumptions. If it is lower, that supports—but does not prove—the absence of a material effect.
Worked Example: Reading Predose Concentrations by Sequence
In a hypothetical study, predose concentrations for a drug normally expected to be absent before dosing are summarized as follows. The assay quantification limit is 0.05 mg/L.
| Sequence | Period 1 predose mean (mg/L) | Period 2 predose mean (mg/L) | Period 2 observations above 0.05 mg/L |
|---|---|---|---|
| AB | 0.02 | 0.31 | 6 of 8 |
| BA | 0.02 | 0.04 | 1 of 8 |
Period 1 means are similar and below the assay limit. In period 2, the AB mean is \(0.31\) mg/L, compared with \(0.04\) mg/L in BA, a difference of \(0.31-0.04=0.27\) mg/L. Six of eight AB participants, but one of eight BA participants, have values above the limit. In AB, participants received A in period 1; this pattern is therefore compatible with residual A before the period 2 dose.
The pattern warrants investigation, but it does not establish that A caused the measured values or that the bioequivalence estimate is biased. Check individual values and profiles, sample timing, dose records, assay performance, and whether the observed residual level could materially affect AUC or Cmax. Do not treat a group mean alone as proof, and do not silently subtract the predose value from the concentration-time profile: a residual drug contribution may be declining over time rather than forming a constant baseline.
How Differential Carryover Can Distort the Estimate
A short derivation shows why the direction of the distortion depends on which treatment was given first. Let \(D\) be the log-scale direct treatment difference, defined as the response under A minus the response under B. For illustration, suppose the period effect is zero, participants are balanced across sequences, and a residual effect contributes \(c_A\) to the period 2 log response after A and \(c_B\) after B. These are simplifying assumptions used to show the mechanism, not a complete analysis model.
For sequence AB, the within-participant contrast (period 1 A minus period 2 B) is \(D-c_A\): the later B response includes residual A. For sequence BA, the contrast (period 2 A minus period 1 B) is \(D+c_B\): the later A response includes residual B. Averaging the two sequence contrasts gives:
Thus, equal residual effects (\(c_A=c_B\)) cancel in this simplified balanced contrast, whereas unequal residual effects shift it. In practice, effects need not be equal, the design may be unbalanced, and other model terms matter. The derivation explains why “some residual drug” does not specify the direction or size of bias. The relevant concern is whether residual effects differ in a way that changes the comparison.
Worked Example: A Differential Carryover Effect on the Log Scale
Consider hypothetical log-scale values with a true direct treatment difference \(D=\ln(1.05)=0.04879\), corresponding to a 5% higher geometric mean under A. Suppose residual effects are \(c_A=0.12\) after A and \(c_B=0.02\) after B. Assume equal numbers in AB and BA, no period effect, and the simplified model above.
The average crossover contrast is:
Exponentiating gives an apparent ratio of \(e^{-0.00121}=0.9988\), or about 99.88%, rather than the direct ratio of 105%. In this constructed example, differential residual effects nearly erase the direct difference and slightly reverse its direction. The calculation is not an estimate from a real study; it demonstrates why an apparently reassuring crossover ratio cannot by itself rule out distortion when material carryover is plausible.
Consequences for the Bioequivalence Conclusion
A confidence interval that falls within the prespecified bioequivalence limits is ordinarily interpreted under the assumptions of the analysis. If there is credible evidence that residual drug materially affects later-period endpoints, the interval may be precise for the fitted model but not a reliable estimate of the intended direct treatment comparison. Precision does not repair a violated design assumption.
The consequence is not automatically “the products are not bioequivalent.” Rather, the study may not provide a sufficiently interpretable comparison to support a confident conclusion. Report the signal, its strength, its likely direction or uncertainty, and how it relates to the endpoint. Distinguish a weak, isolated detection from a consistent, pharmacologically consequential pattern.
A first-period-only comparison can serve as a diagnostic alternative because period 1 observations precede any study-treatment carryover. In a randomized AB/BA study, the period 1 comparison is between participants assigned to A and B, so it avoids carryover from an earlier study period. However, it is a parallel-group comparison with less within-participant information and often much lower precision. It should not be selected after looking at the results simply because it gives a preferred answer. A materially different first-period estimate may raise concern, but its wider uncertainty must also be considered.
A formal statistical test for carryover is not a dependable gate for deciding whether to use the crossover analysis: such tests may have low power, and a nonsignificant result does not demonstrate that carryover is absent. Likewise, switching to a first-period analysis only after a significant carryover test can produce a data-dependent analysis strategy. The primary conclusion should instead reflect the planned design, the pharmacokinetic evidence, and the plausibility of material impact. The detailed treatment of predose concentrations as an analysis topic is separate; here they are one part of a broader interpretability assessment.
Diagnostics, Alternatives, and Limits
Confirm treatment sequence, period labels, dose times, washout duration, and predose sample timing before interpreting a pattern.
Compare expected elimination with the washout and inspect the observed concentration-time and predose data, accounting for assay limits and units.
Ask whether the residual concentration or effect could materially influence the endpoint, and whether the pattern differs by prior treatment or sequence.
Present the planned crossover estimate alongside the carryover evidence and relevant diagnostics; do not interpret a nominal interval in isolation.
The main alternative diagnostic is the period 1 comparison, which avoids within-study carryover but gives up the efficiency of paired observations. A more complex model that includes carryover terms may be considered only when justified by the design and data; in a small two-period study, treatment and carryover components can be difficult to separate reliably. Neither alternative converts weak data into definitive evidence. If residual effects are substantial and cannot be distinguished from the direct treatment effect, the central limitation is interpretability, not merely a choice of statistical model.
Common warning signs include repeated quantifiable predose values after one treatment, profiles that begin elevated and decline, and later-period outcomes that differ by sequence in a manner consistent with the earlier treatment. Warning signs also include unexplained discrepancies between measured predose concentrations and expected elimination. These clues can be misleading if sample timing, assay behavior, or data records are wrong, so diagnostics should be checked rather than read as automatic proof.
Common Mistakes
- “The carryover test was nonsignificant, so carryover is absent.” A nonsignificant test may simply lack power. Consider pharmacokinetic plausibility and the observed data.
- “Any detectable predose concentration invalidates the study.” Detection is evidence to assess, not a universal invalidation rule. Its magnitude, reliability, and possible effect on the endpoint matter.
- “A passing confidence interval settles the question.” The interval relies on the analysis assumptions. If credible carryover undermines the estimand, a nominally passing interval may not support an interpretable direct comparison.
- “Subtract the predose concentration from every later sample.” Residual drug generally changes with time; subtracting a fixed value can create a false concentration profile unless a justified method supports that correction.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A drug has a 6-hour half-life and a 30-hour washout. Under first-order elimination, what fraction is expected to remain?
- Why does a nonsignificant formal carryover test not establish that carryover is absent?
- In a balanced AB/BA design, why can different residual effects after A and B distort the average treatment contrast?
- A predose result is just above the assay quantification limit in one participant. Name two checks before interpreting it as carryover.
- Why can a first-period-only analysis be useful, and what does it lose compared with a crossover analysis?
Solutions. (1) There are \(30/6=5\) half-lives, so the expected fraction is \((1/2)^5=0.03125\), or 3.125%. (2) The test may have low power; failure to reject its null hypothesis is not proof that the effect is zero or immaterial. (3) In the simplified balanced derivation, the contrast is \(D+(c_B-c_A)/2\), so unequal residual effects shift the estimate. (4) Check the sample timing and sequence/period records, and review assay performance and nearby individual concentration-time values; other relevant checks include dose records and expected elimination. (5) Period 1 precedes any study-treatment carryover, so the comparison avoids that source of contamination. It loses the within-participant comparison and is usually less precise because it uses a parallel-group comparison.