‘Tests for Individual and Population Bioequivalence Based on Generalized p-Values Richard J. McNally and Hari Iyer ‘Thomas Mathew Department of Statistics Department of Mathematics and Statistics Colorado State University University of Maryland-Baltimore County Fort Collins, CO 80588 U.S.A. Baltimore, MD 21250, U.S.A. SUMMARY ‘The US Food and Drug Administration (FDA) has proposed new regulations that address the ‘prescribability’ and ‘switchability’ of new formulations of already-approved drugs. ‘These new criteria are known, respectively, as population and individual bioequivalence. Two methods have been proposed for assessing population and individual bioequivalence that calculate an upper 95% confidence bound for the bioequivalence criterion in question, and then test bioequivalence by -comparing this bound to the limit established by the FDA. In this paper we propose applying the generalized test function (GIF) methodology of Tsui and Weerahandi (JASA, 1989, pp 602-607) to this problem to produce tests based on a generalized p-value (GPV). ‘This methodology allows us to construct hypothesis tests in the presence of nuisance parameters. Using simulation we show ‘that these tests perform well in comparison to the confidence interval methods, and have superior power for assessing population bioequivalence. email: menally@stat.colostate.edu1 Introduction In the past decade, the focus of bioequivalence testing has shifted from average bioequivalence (ABE), which emphasizes comparing the means of two drug formulations, to population bioequiv- alence (PBE) and individual bioequivalence (IBE). These new criteria are intended to address the issues of prescribability and switchability, respectively, of different formulations of the same drug. Chow and Liu® discuss the criteria for PBE and IBE that have been proposed in the literature by various authors.'4"! These methods can be divided into two categories ~ probability-based crite- ria and moment-based criteria, The methods currently recommended by the US Food and Drug Administration (FDA) are moment-based criteria, which Chow and Liu® refer to as Mpa for PBE, and Mjs for IBE. At first, the FDA recommended using the bootstrap to evaluate IBE and PBE.* But more recently, they have advocated an approach developed by Hyslop, Hsuan, and Holder (HHH).!? The HHH method is based on the work of Howe,!? who developed an extension of the Cornish-Fisher expansion to compute percentiles for the sum of two random variables with tabulated distribution functions. Howe's method involved, for confidence intervals, calculating an upper bound for each component, and then using the individual bounds to derive a bound for the sum, Howe examined some of the properties of the confidence intervals produced by his method, which he applied to computing percentiles of the non-central tdistribution and confidence intervals for the between- group variance component in a one-way random effects model. He speculated that this method could be applied to general linear combinations, which is what HHH have done, We will refer to this method as the Howe Small-sample Confidence Interval, or HSCI. Quiroz et al!8 (QTWB) have proposed alternative approaches for assessing PBE and IBE using a generalization of what is now known as the modified large sample (MLS) method originally introduced by Graybill and Wang," and discussed in detail by Burdick and Graybill.'* QTWB showed this method to be comparable with the HSCT method, with MLS being somewhat more powerful with respect to PBE, while HSCI is somewhat better for IBE. Like the HSCI, MLS gives exact intervals when all but one component of the problem is known, Unlike the HSCI method, which is based on a linearized version of the bioequivalence criterion, the MLS method yields upper confidence bounds for the actual criterion parameters Mp3 and Mrs. The paper by Quiroz et al provides a detailed comparison of the HSCI method along with their proposed methods, hereafter referred to as MLS method. ‘Tsui and Weerahandi!® and Weerahandi!’ have proposed a general approach for constructing tests for problems in which traditional methods produce test statistics whose distributions depend on one or more nuisance parameters. Weerahandi’® has also studied closely related methods for constructing confidence intervals. These methods are referred to as Generalized p-values (GPV) and Generalized confidence intervals (GCI), respectively. The approach has been successfully applied to a wide variety of problems by Zhou and Mathew.!° The book by Weerahandi” gives further examples of these methods. In this investigation we examine the generalized p-value approach for making inferences concern- ing the FDA-recommended PBE and IBE criteria. The underlying statistical model is described in Section 2. The Generalized p-value approach is described, along with the notation we use, in Section 3. Sections 4 and 5 present the generalized test functions for IBE and PBE, respectively. Examples are also given for illustrating the application of the proposed methods, with direct com- parisons to the HSCI and MLS methods. Section 6 discusses the results of a simulation study weconducted for evaluating the performances of the GPV methods. Comparisons to the HSCI and the MLS methods are included. Section 7 discusses some of the advantages of the GPV approach over the existing methods. Hyslop et al”? only applied the Howe methodology to TBE. The extension of this method to PBE is given in the FDA guidance document,” using, for the most part, the notation of HHH. QTWB" have addressed both IBE and PBE, using their own notation. To address both IBE and PBE in a unified fashion, we have developed notation which differs from that of HHH/FDA and QTWB. To attempt to reconcile the different notations and minimize any resulting confusion, we have summarized, in Tables A and B in the appendix, how the differing notations correspond. It is assumed that the reader is familiar with or can refer to the methods in the FDA guidance document (which are largely the methods of HHH), and that the reader is also familiar with MLS methods for constructing confidence intervals for variance components. For information on the latter, one can refer to the book by Burdick and Graybill."® Tt is further assumed that the reader is familar with the concepts relating to the evaluation of pharmacokinetics and bioequivalence. 2 THE STATISTICAL MODEL ‘The decision criteria mentioned in the current FDA guidance document” on population and indi- vidual BE are based on four-period crossover designs. In this paper we only consider the 2-sequence 4-period design and the 4-sequence 4-period design described in Tables 1a and 1b, respectively. Table 1a FDA-recommended 2-sequence 4-period design Sequencel TR T R Sequene2 RT RT Table 1b FDA-recommended 4-sequence 4-period design eee eee Sequence? RT TR Sequencee3 T T RR ple ng a Sas a ‘The observational model is given below. Yaga = se + ta + nage + pnt aid sm RETR, 11,2 where Yiji is the response from subject j in sequence i to the |** application of treatment k, and s is the number of sequences. The parameters jy and 1, are the population mean responses to treatments T and R and yu: is the (unknown) fixed effect corresponding to the J* application of treatment k in sequence i. The following estimability conditions are imposed on the iki (see Chincbilli and Esinhart”!), : EDs a filial 0 fork=T,R