Similar Mean Exposure Does Not Guarantee a Precise Result
In a conventional average bioequivalence assessment, the estimated test-to-reference exposure ratio can be close to 1 while its confidence interval is wide. That distinction matters: the point estimate describes the estimated average difference, whereas the interval also reflects how precisely that difference has been estimated. High within-subject variability can reduce precision and make it harder to demonstrate that the ratio lies within the conventional limits, even when the formulations have similar mean exposure.
For a positive pharmacokinetic measure such as AUC or \(C_{\max}\), analyses commonly use logarithms. On this scale, a difference between test and reference corresponds to a ratio on the original scale. The conventional limits are fixed at 80% and 125%: a 90% confidence interval for the test-to-reference ratio must lie wholly within \([0.80,1.25]\) to meet the usual confidence-interval criterion. The limits concern the estimated ratio and its uncertainty; they do not change merely because a drug is variable.
A log-normal model is a useful way to understand the issue. If positive exposure measurements are approximately log-normal, their logarithms are approximately normal. In a simple paired crossover analysis, suppose the subject-level contrasts \(D_i\) are independent observations from a normal distribution with mean \(\delta\) and variance \(\sigma_D^2\). The estimate is \(\bar D\), its standard error is \(s_D/\sqrt n\), and the estimated geometric mean ratio is \(\exp(\bar D)\). As the variability of the paired contrasts increases, the standard error and confidence-interval width increase unless the sample size also increases.
Why the Confidence Interval Widens
For this simple paired model, the 90% confidence interval on the log scale is \(\bar D \pm t_{0.95,df}\,s_D/\sqrt n\), with \(df=n-1\). Exponentiating both endpoints gives the confidence interval for the geometric mean ratio. The interval is wider when \(s_D\) is larger and narrower when \(n\) is larger. The fixed equivalence limits do not widen with the data’s variability.
This result follows directly from the paired normal model. The sample mean \(\bar D\) estimates \(\delta\). When the contrasts are independent and normally distributed, the statistic \((\bar D-\delta)/(s_D/\sqrt n)\) follows a \(t\) distribution with \(n-1\) degrees of freedom. The central 90% of that distribution lies between \(-t_{0.95,n-1}\) and \(t_{0.95,n-1}\). Rearranging those inequalities for \(\delta\) gives the stated log-scale interval. Since the exponential function is increasing, exponentiating preserves the endpoint order and produces the ratio interval.
In a simple crossover setting, if the log measurement under each formulation has within-subject residual variance \(\sigma_w^2\), and the two residual errors are independent with the same variance, then the variance of their difference is \(\sigma_D^2=2\sigma_w^2\). This illustrates why greater within-subject variation can produce more variable paired contrasts. Actual crossover analyses may use a model that accounts for the design and its degrees of freedom; the paired calculation below is intended to isolate the precision mechanism.
Worked Example: Lower Variability and a Narrow Interval
Suppose a hypothetical, balanced crossover study measures AUC on the log scale in \(n=24\) subjects. The paired log contrasts have mean \(\bar D=\ln(1.02)=0.01980\) and sample standard deviation \(s_D=0.16\). Assume the subject-level contrasts are independent, approximately normal, and have no influential outliers, and that the crossover analysis is otherwise appropriate. With \(df=23\), use \(t_{0.95,23}=1.714\).
The standard error is \(SE=s_D/\sqrt n=0.16/\sqrt{24}=0.03266\). The 90% interval on the log scale is \[ 0.01980\pm 1.714(0.03266) = 0.01980\pm 0.05598 = [-0.03618,\ 0.07578]. \]
Exponentiating the endpoints gives \([\exp(-0.03618),\exp(0.07578)]=[0.9645,1.0787]\). This entire interval is within \([0.80,1.25]\), so the result meets the conventional confidence-interval criterion for average bioequivalence. The estimated geometric mean ratio is 1.02, or 102%; the interval describes uncertainty around that estimate, in ratio units.
High Variability With a Similar Estimated Ratio
The next hypothetical example keeps the sample size and point estimate unchanged but increases the standard deviation of the paired log contrasts. The example shows that the same estimated ratio can lead to a different conclusion because its uncertainty is greater. It does not show that the true mean ratio has changed.
Worked Example: A Wide Interval Crosses the Fixed Limits
Suppose another hypothetical study has \(n=24\), \(\bar D=\ln(1.02)=0.01980\), and \(s_D=0.72\). Assume the same model conditions as before: independent, approximately normal paired contrasts without influential outliers, and an appropriate crossover analysis. The degrees of freedom and critical value are again \(23\) and \(1.714\).
Here \(SE=0.72/\sqrt{24}=0.14697\), so the 90% log-scale interval is \[ 0.01980\pm 1.714(0.14697) = 0.01980\pm 0.2519 = [-0.2321,\ 0.2717]. \]
On the ratio scale, the interval is \([\exp(-0.2321),\exp(0.2717)]=[0.7929,1.312]\). It extends below 0.80 and above 1.25. Therefore, it does not meet the conventional confidence-interval criterion for average bioequivalence. The estimated ratio remains 1.02, but the data are not precise enough to establish that the ratio lies within the fixed limits. This is an inconclusive bioequivalence result, not proof that the formulations’ mean exposures differ.
The contrast between these examples can be summarized by the standard error. Both use the same \(n\), the same estimated log ratio, and the same confidence level. The second has a much larger \(s_D\), so its standard error is larger and its interval spans a wider range of plausible ratios. The fixed limits remain 0.80 and 1.25. Thus, high variability undermines the conventional assessment through uncertainty relative to fixed bounds, rather than by automatically moving the estimated mean ratio away from 1.
Worked Example: More Subjects Can Improve Precision
Consider a hypothetical study with the high paired-contrast standard deviation \(s_D=0.72\), but with \(n=48\) subjects and the same estimated ratio, \(\exp(\bar D)=1.02\). Assume the paired-model conditions remain reasonable. Using \(df=47\) and \(t_{0.95,47}=1.678\), the standard error is \(0.72/\sqrt{48}=0.1039\).
The 90% log-scale interval is \[ 0.01980\pm 1.678(0.1039) = 0.01980\pm 0.1743 = [-0.1545,\ 0.1941]. \]
Exponentiating gives \([0.8568,1.214]\), which lies within \([0.80,1.25]\). Under these assumptions, this hypothetical result meets the conventional confidence-interval criterion. The interval is narrower than in the 24-subject example because the standard error decreased as \(n\) increased. Adding subjects can improve precision, but it does not reduce the underlying within-subject variability or guarantee a favorable result.
Assumptions, Diagnostics, and Failure Modes
The paired \(t\)-interval is justified when the subject-level log contrasts are independent across subjects and approximately normally distributed, or when the sample is large enough for the mean’s sampling distribution to be adequately approximated. The log transformation requires positive measurements. The design and analysis must also support a meaningful paired test-to-reference comparison; period, sequence, missing-data, or other design features cannot be ignored if they affect that comparison. In a more complex crossover analysis, use the model appropriate to the design rather than automatically applying \(n-1\) degrees of freedom.
Useful checks include examining the paired log contrasts for outliers, skewness, and patterns associated with sequence or period, and confirming that the standard deviation used is the one appropriate to the analysis. A few extreme contrasts can inflate \(s_D\) and the interval width; they should be investigated rather than removed automatically. A mismatch between the reported standard error and \(s_D/\sqrt n\) can indicate that a different model or variance estimate was used, in which case the interval must be interpreted using that model’s assumptions.
Another diagnostic distinction is between within-subject and between-subject variability. Subjects may have very different typical exposure levels, while each subject’s repeated measurements remain relatively consistent. In a crossover comparison, subject pairing can account for stable differences between subjects; it is the variability of the relevant within-subject contrast that directly drives the simple paired standard error. Treating a large spread across all subjects as if it were the paired-contrast standard deviation can therefore misstate precision.
How Conventional Limits Compare With Other Approaches
Conventional average bioequivalence uses fixed ratio limits, commonly 0.80 and 1.25, regardless of the observed variability. Its central question is whether the confidence interval for the average ratio fits inside those limits. A high-variability product can consequently require a larger study to obtain a sufficiently narrow interval. A nonsignificant test of equality of means is not a substitute: failure to detect a difference does not show that the entire confidence interval is within prespecified equivalence limits.
Reference-scaled approaches are an alternative framework used for certain highly variable drugs. Their general rationale is to relate the allowed assessment to variability under specified regulatory rules, rather than to rely only on the same fixed limits. They require different criteria and supporting data; they should not be treated as a way to relabel a wide conventional interval as a passing result. The choice of framework and whether a drug qualifies for it are separate questions from the calculation shown here.
High variability also has limits as an explanation. It can account for an imprecise interval, but it does not establish that the formulations are equivalent, nor does it establish that their mean exposures differ. The interval and its relation to the prespecified limits determine the conventional decision. The result is a statement about the study’s evidence under its model and design, not a probability that the unknown ratio lies inside the calculated interval.
Common Mistake
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If the estimated test-to-reference ratio remains 1.00 while the standard deviation of paired log contrasts increases, what happens to the confidence interval, all else equal?
- A 90% ratio confidence interval is \([0.82,1.24]\). Does it meet the conventional fixed-limit criterion? Explain.
- Why is a large between-subject spread not automatically the same as high within-subject variability in a crossover comparison?
- A study reports a ratio of 0.99 and a 90% interval of \([0.78,1.18]\). What conclusion follows under conventional limits, and what does not follow?
Solutions. 1. The standard error and interval width increase because \(SE=s_D/\sqrt n\); the interval remains centered at the same estimated log ratio in this simplified comparison. 2. Yes. Both endpoints lie within \([0.80,1.25]\). 3. Between-subject spread reflects differences in typical exposure across people, while within-subject variability concerns changes within the same person; pairing can account for stable subject differences. 4. The interval crosses the lower limit, so the study does not meet the conventional criterion. This does not prove that the mean ratios differ or that the formulations are non-equivalent; it means equivalence has not been established by this interval.