Average Bioequivalence as the Default Regulatory Standard

What Does Average Bioequivalence Evaluate?

When two drug products are compared, the central question is not simply whether their measured concentrations are numerically identical. The question is whether, under specified study conditions, the test product and reference product produce sufficiently similar systemic exposure on average. This is the aim of average bioequivalence (average BE).

Average BE is a population-level comparison. It assesses the average relative exposure associated with the test product versus the reference product for a specified pharmacokinetic endpoint, such as an exposure measure. It does not assert that the products produce identical concentration profiles, nor does it show that every person has the same exposure with both products. The earlier material on equivalence claims versus therapeutic similarity distinguishes this statistical comparison from a broader claim about therapeutic outcomes.

The reference product matters: the comparison is relative to that product, not to an abstract ideal concentration. The claim also applies to the measured endpoint and study conditions. A conclusion about one endpoint does not automatically establish equivalence for every feature of the concentration–time profile.

Why Work on the Log Scale?

Exposure is positive, and regulatory comparisons are usually framed as relative differences. A change from 50 to 55 units and a change from 100 to 110 units are both increases of 10%. A log transformation turns ratios into differences, making a relative comparison convenient for statistical analysis.

For subject \(i\), let \(T_i\) and \(R_i\) be positive exposure measurements under the test and reference products. Define the subject’s log difference as \(D_i=\ln(T_i)-\ln(R_i)\). By the logarithm rule, \(D_i=\ln(T_i/R_i)\). If \(\mu_D\) is the population mean of these log differences, then \(\exp(\mu_D)\) is the population geometric mean ratio of test to reference exposure.

This relationship explains what “average” means in the usual log-scale analysis: the average is taken after the exposure values have been logged. It is not generally the arithmetic mean of the individual ratios. The geometric mean ratio describes the relative location of the two products’ exposures; it is not a statement that every subject’s ratio equals that value.

How the Comparison Is Made

A common design is a randomized crossover study in which each subject receives both products in different periods. The within-subject log difference \(D_i\) then compares the products while controlling for stable differences between subjects. A simple model for these differences is

where \(\mu_D\) is the population mean log difference and the errors \(\epsilon_i\) have mean zero. For the basic paired analysis, subjects should be independent of one another, and the distribution of the within-subject differences should be sufficiently well behaved for the chosen inference method. The study must also be conducted so that period, sequence, carryover, adherence, and measurement problems do not undermine the product comparison. In more complex designs, the analysis model must account for the design rather than treating every observation as an independent pair.

A parallel-group design compares the mean log exposure in one group receiving the test product with the mean in another group receiving the reference. It does not provide a within-subject contrast. In either design, the target remains the relative population average, but design and analysis determine how precisely that target is estimated.

A standard equivalence decision uses a pair of one-sided tests against prespecified lower and upper limits for the mean log difference. The equivalence claim is supported only if both one-sided null hypotheses are rejected. At a one-sided significance level of 0.05, this is equivalent to requiring the corresponding 90% confidence interval to lie entirely inside the prespecified limits. The limits themselves are a separate regulatory choice; the calculation does not choose them.

Worked Inference and Its Assumptions

Why This Is a Default Framework—and What It Does Not Show

For many systemic drug products, average exposure is a useful and operationally practical basis for comparing formulations. A carefully conducted pharmacokinetic study can measure exposure under controlled conditions, and a crossover design often reduces the influence of stable between-person differences. These features make an average comparison informative without requiring a large trial designed to establish identical outcomes in every possible patient.

The framework supports a regulatory conclusion about the specified test and reference products, endpoint, population, and study conditions. It is not, by itself, proof that the products have identical clinical effects, identical safety profiles, or identical concentrations at every time point. It also does not guarantee that an individual patient will have the same exposure after switching products. Those are different questions requiring evidence appropriate to those claims.

Average BE is a default framework for many systemic products, not a rule that every product must be assessed in exactly the same way. Product-specific considerations may call for additional or tailored criteria. High within-subject variability can also make the estimate less precise: the mean ratio may be near one while the confidence interval remains broad. Conversely, a precisely estimated average does not reveal the full distribution of individual treatment ratios.

Diagnostics, Alternatives, and Common Pitfalls

Before interpreting an average comparison, review whether the design and data support the model. Check the distribution of subject-level log differences, inspect influential observations, and confirm that missing measurements, protocol deviations, and assay problems have not selectively distorted one treatment. In crossover studies, examine period and sequence patterns and consider whether carryover could plausibly affect the endpoint. These checks diagnose threats to the comparison; they do not replace the prespecified analysis.

A parallel-group comparison may be appropriate when a crossover is unsuitable, but typically loses the advantage of comparing each subject with themselves. The resulting between-subject variability can reduce precision. An individual-level comparison answers a different question: how treatment differences vary across people. It should not be inferred from a population-average result alone. Product-specific approaches may also supplement or modify the default framework when average exposure is not sufficient for the regulatory question.

A second error is to treat a wide interval as proof that the products differ. A wide interval indicates imprecise evidence; it may include values both compatible and incompatible with the prespecified bounds. The appropriate conclusion depends on the interval relative to those bounds, not on whether the point estimate alone looks close to one.

A Short Justification of the Ratio-Scale Interpretation

The conversion from a mean log difference to a geometric mean ratio follows directly from the definition. For paired observations, the sample mean log difference is the mean of the log ratios. Exponentiating that mean gives the geometric mean of the ratios:

The equality holds because the exponential of a sum is the product of exponentials, and raising that product to the power \(1/n\) gives its geometric mean. Thus a log-scale estimate has a direct relative-exposure interpretation. The derivation does not imply that the ratios are constant across subjects; it identifies which average the analysis estimates.