Bioequivalence for highly variable drugs

Between 2003 and 2005, FDA's Office of Generic Drugs found 1,010 bioequivalence studies acceptable, covering 180 different drugs. Agency scientists later went back through the data and found that 57 of those drugs, or 31 percent, were highly variable, meaning a person's blood levels swung widely from one dose to the next even on the same product. The results were published in the AAPS Journal in 2008¹.

That variability has a price under the standard test. For drugs highly variable in total exposure, the studies FDA reviewed averaged 55 volunteers each, against 32 for less variable drugs. Some enrolled as many as 134.

FDA's answer is reference-scaled average bioequivalence, usually shortened to RSABE, which widens the acceptance boundary in proportion to how variable the brand itself is. The method is now written into FDA's May 2026 guidance on statistical approaches to bioequivalence⁵. This piece explains the problem, how the scaling works, the study designs it requires and how FDA's rule compares with Europe's.

Why variability inflates a bioequivalence study

A standard test, described in GxRadar's explainer on the 80-125 rule, requires the 90 percent confidence interval around the generic-to-brand ratio to sit inside 80 to 125 percent, for both peak concentration (Cmax) and total exposure (AUC). The width of that interval "depends upon the number of subjects in the study and the variability of the BE measure," FDA scientists wrote in a 2012 paper on how the agency implemented RSABE². When variability is high, the only way to narrow the interval is to enroll more people.

Their table shows how fast the numbers climb. Figures are volunteers needed for 80 percent power under the unscaled 80-125 test.

Source: Davit and colleagues, AAPS Journal 2012, Table I.

A 2008 simulation by FDA scientists, also in the AAPS Journal³, put the problem starkly. At 60 percent variability under the standard test, study power was about 22 percent, so "about one study in five would pass BE. This is true for two products with identical bioavailability." The 2012 paper adds a clinical argument for relaxing the limits: the brand itself was shown safe and effective in its original trials "despite high PK variability." Clinical data, the authors wrote, "strongly support a conclusion that HV drugs have wide therapeutic indices."

How reference scaling works

FDA's 2026 guidance⁵ defines highly variable drugs as those with within-subject variability "30 percent or greater" that are not narrow therapeutic index drugs. Each drug's product-specific guidance says which approach FDA expects. To use scaling, the study has to give each volunteer the brand at least twice, so the brand's own variability can be measured. The analysis then follows three steps, set out in the guidance's Appendix G.

1. Estimate the brand's within-subject standard deviation, called sWR, for each measure. If it is below 0.294 (about 30 percent CV), the ordinary 80-125 test applies.
2. If sWR is 0.294 or higher, test whether the squared difference between generic and brand means (on a log scale), minus a scaling term built from the brand's variability and a regulatory constant of 0.25, has a 95 percent upper confidence bound at or below zero.
3. Separately, the generic-to-brand point estimate must still fall between 80.00 and 125.00 percent.

That third step is a backstop. However variable the brand, a generic cannot be approved on scaling alone if its average sits outside the usual range. FDA's 2026 guidance also recommends a minimum of 24 volunteers for highly variable drug products.

A scaled test does not produce a fixed interval, but its formula implies one. FDA's 2012 paper calls these the "implied limits." GxRadar calculated them from the guidance's constants and set them beside the European Medicines Agency's limits, which follow a similar formula with a cap.

GxRadar's FDA column is exp(±ln(1.25)/0.25 × sWR), with sWR converted from CV. A real study must clear these limits with statistical room to spare, because FDA tests an upper confidence bound. EMA figures are from its 2010 bioequivalence guideline⁸.

Europe's system differs from FDA's in three ways. EMA widens limits only for Cmax, keeping AUC at 80 to 125 "regardless of variability," while FDA's guidance applies scaling to both measures. EMA caps the widening at 50 percent variability; FDA's formula has no cap beyond the point-estimate rule. And EMA requires a "sound clinical justification" that a wider Cmax difference is clinically irrelevant.

Replicate study designs

An ordinary bioequivalence study gives each volunteer each product once. Scaling needs repeats. FDA's guidance recommends a four-period, two-sequence full replicate design (test, reference, test, reference, or the reverse) and accepts a three-period full replicate or a three-period partial replicate in which only the brand is repeated (sequences TRR, RTR and RRT). The three-period designs need more volunteers for the same statistical power.

FDA's guidance notes the trade. A replicate design "can have the advantage of using fewer subjects compared to a nonreplicate design, although each subject in a replicate design would receive more treatments." The companion May 2026 guidance on pharmacokinetic bioequivalence studies⁷ covers how those studies are run.

How FDA got here

Getting there took years. FDA's Office of Generic Drugs took the question to its pharmaceutical science advisory committee in April 2004, and an internal working group presented a reference-scaling proposal to the committee in April 2006, according to the 2012 paper. FDA scientists described the approach, including the point-estimate constraint, in Pharmaceutical Research in 2008⁴, and their companion simulation study found that a regulatory constant of 0.25 "appears to work best."

The first application using RSABE arrived in 2007. The first tentative approval came in 2009 and the first full approval in 2011. By the time of the 2012 paper, FDA had reviewed 46 generic applications containing 64 RSABE studies, and 62 met the scaled criteria. The 2026 statistical guidance⁶ replaced FDA's February 2001 version, which predated the method, and finalized a December 2022 draft. FDA listed both 2026 bioequivalence guidances under its Drug Competition Action Plan¹⁰.

Scaling also runs the other way. For narrow therapeutic index drugs, FDA uses reference scaling to tighten the limits, and a 2025 FDA analysis in the AAPS Journal⁹ of 93 such applications found that of 20 failed studies, 90 percent failed the reference-scaled criterion alone.

International alignment is the open question, and it is on the research agenda described in GxRadar's piece on FDA's generic drug science program. FDA adopted ICH M13A¹¹ for immediate-release tablets and capsules in October 2024, and its fiscal 2027 research priorities¹² list "global harmonization" of efficient bioequivalence approaches as a goal, alongside "innovative study designs for PK BE studies in patients," including adaptive designs. For individual patients, questions about whether a particular generic suits them remain a matter for a pharmacist or prescriber. More on how generics are reviewed is on GxRadar's approval desk and in its glossary.