Why the Meal Condition Matters
Food can change the rate or extent of drug absorption. In a fed-state bioequivalence study, both treatments are administered under a specified meal condition, and the comparison therefore answers a condition-specific question: are the test and reference products bioequivalent when administered as tested with that meal? It does not establish bioequivalence in the fasted state or explain what would happen with a different meal.
A meal is part of the study condition, not merely background detail. Meal composition, the time allowed to eat, and the interval between eating and dosing should be standardized according to the protocol. Even when the protocol is standardized, differences in meal response and execution can contribute to variation in measured exposure. A single fed-state crossover generally cannot separate this meal-related contribution from other sources of within-subject variability.
Model and Assumptions
Consider a randomized, two-period, two-sequence crossover in which each participant receives test (T) and reference (R) once, with both administrations under the same specified fed condition. Analyze AUC and Cmax separately, using their natural logarithms. For a complete, balanced study, a conventional model includes sequence, participant nested within sequence, period, and treatment effects. The treatment contrast estimates the mean difference between log(T) and log(R).
Here \(Y_{ijk}\) is the log-transformed endpoint for a participant in sequence \(i\), participant \(j\) within that sequence, and period \(k\). The treatment effect is the log-scale contrast of interest. In a complete, balanced 2×2 crossover, the same treatment contrast can be calculated as the mean of participants’ paired log differences, where each difference is log(T) minus log(R), regardless of the period in which each treatment was received.
The usual interval and test rely on randomized treatment assignment, complete paired observations, no important carryover that contaminates the second-period comparison, and an approximately normal distribution of paired log differences with a common variance. The meal and dosing procedures should be sufficiently consistent that the intended fed condition is comparable across treatments and periods. These assumptions make the paired analysis interpretable; they do not imply that all participants have the same response to food.
For \(n\) complete participants, let \(\bar{d}\) be the mean paired log difference (test minus reference) and \(s_d\) the pooled within-sequence sample standard deviation of those differences, pooled over the two sequence groups so that any period effect does not inflate it. The standard error is \(s_d/\sqrt{n}\). With \(df=n-2\) in the standard balanced 2×2 analysis, the 90% confidence interval on the log scale is then exponentiated to obtain the confidence interval for the geometric mean ratio.
When the prespecified bioequivalence limits are 80.00% and 125.00%, both endpoints of the 90% interval must lie within 0.8000 and 1.2500. Equivalently, the log interval must lie within \([\ln(0.8000),\ln(1.2500)]\). This is the confidence-interval form of two one-sided tests at significance level \(\alpha=0.05\). The limits used in an actual assessment must be the limits specified for that product and endpoint.
Why the Confidence Interval Gives the Decision
The equivalence question is not whether the data fail to show a difference. For ratio limits \(L\) and \(U\), the hypotheses for the two one-sided tests are \(H_0:\theta\leq L\) versus \(H_a:\theta>L\), and \(H_0:\theta\geq U\) versus \(H_a:\theta<U\), where \(\theta\) is the test-to-reference geometric mean ratio. Bioequivalence is concluded only if both null hypotheses are rejected. For the standard 80.00% to 125.00% limits and \(\alpha=0.05\), this is equivalent to requiring the 90% confidence interval to be entirely inside the limits.
The equivalence follows by expressing the one-sided tests on the log scale. The lower test succeeds when the lower confidence bound is greater than \(\ln(L)\); the upper test succeeds when the upper confidence bound is less than \(\ln(U)\). Both conditions together require the full interval to lie between the log limits. A point estimate inside the limits is not sufficient: uncertainty may extend beyond either boundary.
The paired standard deviation also describes precision and within-subject variability. Under the standard equal-variance crossover model, the residual mean square on the log scale is \(s_d^2/2\), where \(s_d\) is the pooled within-sequence standard deviation defined above. The corresponding within-subject coefficient of variation is estimated as \(100\sqrt{\exp(s_d^2/2)-1}\%\). This estimate summarizes within-subject variation in the fed study; it is not an estimate of meal variability alone.
Worked Example: Fed-State AUC Supports Bioequivalence
Suppose a hypothetical balanced fed-state crossover has 24 complete participants. For AUC, the mean paired log difference is \(\bar{d}=0.01980\) and the pooled within-sequence standard deviation is \(s_d=0.1600\). The mean corresponds to a test-to-reference geometric mean ratio of \(\exp(0.01980)=1.020\), or 102.0%. Assume the prespecified limits are 80.00% to 125.00%.
With \(df=24-2=22\), use \(t_{0.95,22}=1.717\). The standard error is \(0.1600/\sqrt{24}=0.03266\), and the margin of error is \(1.717(0.03266)=0.05608\). Thus the log-scale interval is \(0.01980\pm0.05608\), or \([-0.03628,\,0.07588]\).
Exponentiating the bounds gives \([\exp(-0.03628),\exp(0.07588)]=[0.9644,\,1.0788]\). The complete 90% interval is between 0.8000 and 1.2500, so this hypothetical AUC result supports bioequivalence under the tested fed condition.
The within-subject variability estimate is based on the residual mean square \(s_d^2/2=0.1600^2/2=0.01280\). Therefore \(CV_w=100\sqrt{\exp(0.01280)-1}=11.35\%\) (rounded). This value describes variability under the study’s fed procedures; it does not identify how much of that variability was caused by the meal.
Interpreting Meal-Related Variability
In a crossover, each participant contributes information under both treatments, so stable participant-level differences are largely removed from the treatment contrast. However, variation in the participant’s response to the meal, differences in meal execution between periods, and other within-participant variation can still contribute to the paired differences. If these sources increase \(s_d\), the standard error and confidence interval widen, making a conclusive result less likely for a given sample size and treatment ratio.
A fed-only study can estimate the total within-subject variability observed under its fed procedures. It cannot, by itself, assign that variability specifically to food, compare fed and fasted treatment effects, or quantify a treatment-by-meal interaction. Those questions require an appropriate design with meal condition variation. The study report should therefore describe the meal condition and interpret variability in that context, without treating a large \(CV_w\) as proof that the meal caused it.
Worked Example: Fed-State Cmax Does Not Meet the Limits
For the same hypothetical 24-participant study, suppose the Cmax paired log differences have \(\bar{d}=0.13976\) and pooled within-sequence \(s_d=0.3000\). The estimated geometric mean ratio is \(\exp(0.13976)=1.150\), or 115.0%. The point estimate is within the assumed 80.00% to 125.00% limits, but the variability is larger than in the AUC example.
The standard error is \(0.3000/\sqrt{24}=0.06124\), and the 90% margin is \(1.717(0.06124)\approx0.1051\) (0.10514 before rounding). The log-scale interval is \(0.13976\pm0.10514=[0.03462,\,0.24490]\). Exponentiation gives \([1.035,\,1.277]\), or approximately 103.5% to 127.7%.
The upper bound exceeds 1.2500, so the interval is not wholly within the bioequivalence limits. Bioequivalence is not demonstrated for Cmax under this fed condition, even though the point estimate is inside the limits. This conclusion is not evidence that the products are different; it means the required equivalence criterion was not met by this interval.
The within-subject CV estimate is \(100\sqrt{\exp(0.3000^2/2)-1}=100\sqrt{\exp(0.04500)-1}\approx21.45\%\). This higher variability contributes to the wider interval. It should be reported as variability observed in the fed crossover analysis, not attributed solely to meal response.
Worked Example: How Greater Variability Can Change the Decision
To isolate the role of precision, compare two hypothetical fed-state Cmax analyses with \(n=24\), the same mean log ratio \(\bar{d}=\ln(1.150)=0.13976\), and the same limits of 0.8000 to 1.2500. Suppose one has \(s_d=0.1600\) and the other has \(s_d=0.3000\). These are illustrative summaries, not evidence that meal variability alone explains the difference.
For \(s_d=0.1600\), \(SE=0.1600/\sqrt{24}=0.03266\); the margin is \(1.717(0.03266)=0.05608\). The log interval is \([0.08368,\,0.19584]\), which exponentiates to approximately \([1.087,\,1.216]\). It is fully within the limits.
For \(s_d=0.3000\), \(SE=0.3000/\sqrt{24}=0.06124\); the margin is \(1.717(0.06124)\approx0.1051\) (0.10514 before rounding). The log interval is \([0.03462,\,0.24490]\), or approximately \([1.035,\,1.277]\) after exponentiation. Its upper bound exceeds 1.2500, so it does not meet the criterion. The treatment estimate is unchanged; the increased variability widens the interval enough to change the decision.
Diagnostics, Alternatives, and Limits
Before interpreting the result, inspect the study’s meal and dosing records for whether the intended fed condition was implemented in both periods, and examine the paired log differences for strong skewness or influential values. A few large differences can materially affect both \(s_d\) and the interval. If meal procedures differed systematically between treatments or periods, the estimated treatment contrast may no longer represent a clean comparison under a common fed condition. Such concerns should be investigated and reported rather than hidden by relying on a single summary statistic.
The crossover is efficient when participants can receive both treatments and the relevant exposure is adequately measured in each period. A parallel-group comparison avoids within-participant carryover concerns but does not remove stable participant-to-participant variation through pairing. Neither design, by itself, separates meal-related variability from other sources; the design and collected meal information determine which questions can be answered. A fed-state result also cannot substitute for a fasted-state result when both conditions are required.
A common mistake is to say, “The ratio is 115%, so the products are bioequivalent,” or, “There was no statistically significant treatment difference, so they are equivalent.” Both statements ignore the equivalence interval. The correction is to apply the prespecified limits to the entire 90% confidence interval for each required endpoint. In addition, AUC and Cmax are separate endpoints: a conclusion for one does not establish a conclusion for the other.
Check Your Understanding
Use the stated 80.00% to 125.00% limits where applicable.
- A fed crossover estimates an AUC ratio of 1.03 with a 90% interval from 0.98 to 1.29. Does it meet the interval criterion? Explain.
- Why can a fed-only crossover not establish how much within-subject variability is specifically due to food?
- For \(n=24\), \(\bar{d}=0.0200\), and \(s_d=0.1600\), calculate the standard error of the mean paired log difference.
- A Cmax point estimate is within the limits, but the 90% interval extends above 1.2500. State the appropriate conclusion.
Solutions. 1. No. The upper bound, 1.29, exceeds 1.2500, so the full interval is not within the limits. 2. The paired differences contain total within-subject variation under the fed procedures, including meal response and other sources; a single fed condition does not separate them. 3. \(SE=s_d/\sqrt{n}=0.1600/\sqrt{24}=0.03266\) log units. 4. Bioequivalence is not demonstrated for Cmax under the tested fed condition because the full interval fails the upper limit; the point estimate alone is insufficient.