Fda On Statistical Approaches

Statistical Approaches
to Establishing
Bioequivalence
Guidance for Industry
DRAFT GUIDANCE
This guidance document is being distributed for comment purposes only.
Comments and suggestions regarding this draft document should be submitted within 60 days of
publication in the Federal Register of the notice announcing the availability of the draft
guidance. Submit electronic comments to https://www.regulations.gov. Submit written
comments to the Dockets Management Staff (HFA-305), Food and Drug Administration, 5630
Fishers Lane, Rm. 1061, Rockville, MD 20852. All comments should be identified with the
docket number listed in the notice of availability that publishes in the Federal Register.
For questions regarding this draft document, contact (CDER) David Coppersmith at 301-796-
9193.
U.S. Department of Health and Human Services

Food and Drug Administration
Center for Drug Evaluation and Research (CDER)
December 2022
Biopharmaceutics
Revision 1
Statistical Approaches
to Establishing
Bioequivalence
Guidance for Industry
Additional copies are available from:

Office of Communications, Division of Drug Information
Center for Drug Evaluation and Research
10001 New Hampshire Ave., Hillandale Bldg., 4th Floor
Silver Spring, MD 20993-0002
Phone: 855-543-3784 or 301-796-3400; Fax: 301-431-6353
Email: [email protected]
https://www.fda.gov/drugs/guidance-compliance-regulatory-information/guidances-drugs
U.S. Department of Health and Human Services

Center for Drug Evaluation and Research (CDER)
December 2022
Biopharmaceutics
Revision 1
TABLE OF CONTENTS
I. INTRODUCTION............................................................................................................. 1
A. Overview.................................................................................................................................. 1
B. Statistical Guidance Background .......................................................................................... 2
II. GENERAL CONSIDERATIONS ................................................................................... 3
A. Study Design ........................................................................................................................... 3
B Data Preparation .................................................................................................................... 8
C. Statistical Models .................................................................................................................. 11
III. SPECIFIC SITUATIONS .............................................................................................. 14
A. In Vitro Bioequivalence and Population Bioequivalence.................................................. 14
B. Statistical Methods for Narrow Therapeutic Index and
Highly Variable Drug Products .......................................................................................... 19
C. Comparative Clinical Endpoint Bioequivalence Studies .................................................. 21
D. Studies in Multiple Groups .................................................................................................. 22
E. Bioequivalence Statistics for Adhesion and Irritation Studies ......................................... 23
F. Dose Scale for Bioequivalence Assessment ......................................................................... 24
G. Bioequivalence Studies Using Multiple References............................................................ 24
V. APPENDICES ................................................................................................................. 25
A. Choice of Specific Replicated Crossover Designs .............................................................. 25
B. Rationale for Logarithmic Transformation of Pharmacokinetic Data ........................... 27
C. SAS Program Statements for Average Bioequivalence Analysis of
Replicated Crossover Studies .............................................................................................. 28
Contains Nonbinding Recommendations
Draft – Not for Implementation
1 Statistical Approaches to Establishing Bioequivalence

2 Guidance for Industry 1
3
4
5 This draft guidance, when finalized, will represent the current thinking of the Food and Drug
6 Administration (FDA or Agency) on this topic. It does not establish any rights for any person and is not
7 binding on FDA or the public. You can use an alternative approach if it satisfies the requirements of the
8 applicable statutes and regulations. To discuss an alternative approach, contact the FDA staff responsible
9 for this guidance as listed on the title page.
10
11
12
13 I. INTRODUCTION
14
15 Requirements for submitting bioavailability (BA) and bioequivalence (BE) data in
16 investigational new drugs (INDs), new drug applications (NDAs), abbreviated new drug
17 applications (ANDAs), and supplements; the definitions of BA and BE; and the types of in vitro
18 and in vivo studies that are appropriate to measure BA and establish BE are set forth in part 320
19 (21 CFR part 320). This guidance provides recommendations on how to meet provisions of part
20 320 for all drug products.
21
22 In general, FDA’s guidance documents do not establish legally enforceable responsibilities.
23 Instead, guidances describe the Agency’s current thinking on a topic and should be viewed only
24 as recommendations, unless specific regulatory or statutory requirements are cited. The use of
25 the word should in Agency guidances means that something is suggested or recommended,
26 but not required.
27
28 A. Overview
29
30 This guidance provides recommendations to sponsors and applicants who intend to use
31 equivalence criteria in analyzing in vivo or in vitro BE studies for INDs, NDAs, ANDAs, and
32 supplements to these applications. This guidance discusses statistical approaches for BE
33 comparisons and focuses on how to use these approaches both generally and in specific
34 situations. When finalized, this guidance will replace the guidance for industry Statistical
35 Approaches to Establishing Bioequivalence, which was issued in February 2001 (2001
36 guidance). This guidance provides recommendations on the topics covered in the 2001 guidance
37 as well as recommendations on additional topics, including missing data and intercurrent events,
38 adaptive design, and specific situations, such as narrow therapeutic index drugs and highly
39 variable drugs.
40
1
This guidance has been prepared by the Office of Generic Drugs in the Center for Drug Evaluation and Research
(CDER) in cooperation with CDER’s Office of Translational Sciences and Office of Pharmaceutical Quality at the
Food and Drug Administration.
1
41 Defined as relative BA, the assessment of BE involves comparison between a test (T) and
42 reference (R) drug product, where T and R can vary depending on the comparison to be
43 performed (e.g., to-be-marketed formulation versus clinical trial formulation, generic drug versus
44 reference listed drug (RLD), originally approved formulation versus postapproval formulation
45 changes). Although BA and BE are closely related, BE comparisons normally rely on (1) a
46 criterion, (2) a confidence interval for the criterion, and (3) a predetermined BE limit. BE
47 comparisons could also be used in certain pharmaceutical product line extensions, such as
48 additional strengths, new dosage forms (e.g., changes from immediate release to extended
49 release), and new routes of administration. 2 In these contexts, the approaches described in this
50 guidance can be used to determine BE. The general approaches discussed in this guidance may
51 also be useful when assessing pharmaceutical equivalence (i.e., the identical dosage form and
52 route(s) of administration that contain identical amounts of the identical active drug ingredient)
53 or performing equivalence comparisons in clinical pharmacology studies and other areas.
54
55 This guidance is intended to encourage the use of science-based approaches to making statistical
56 BE assessments. Given the evolving nature of statistical approaches and technologies, FDA
57 encourages generic and new drug applicants to propose and discuss novel methodologies (e.g.,
58 model-based BE and novel adaptive designs for comparative clinical endpoint BE studies) with
59 the Agency through appropriate regulatory meetings, as described below.
60
61 B. Statistical Guidance Background
62
63 In the July 1992 guidance on Statistical Procedures for Bioequivalence Studies Using a Standard
64 Two-Treatment Crossover Design (the 1992 guidance), the Center for Drug Evaluation and
65 Research (CDER) recommended that a standard in vivo BE study design be based on the
66 administration of either single or multiple doses of the T and R products to healthy subjects on
67 separate occasions, with random assignment to the two possible sequences of drug product
68 administration. The 1992 guidance further recommended that statistical analysis for
69 pharmacokinetic (PK) measures, such as area under the curve (AUC) and peak concentration
70 (Cmax), be based on the two one-sided tests procedure to determine whether the average values
71 for the PK measures determined after administration of the T and R products were comparable.
72 This approach is termed average BE (ABE) and involves the calculation of a 90% confidence
73 interval for the ratio of the averages (population geometric means) of the measures for the T and
74 R products. To establish BE, the calculated confidence interval should fall within a BE limit,
75 usually 80 to 125% for the ratio of the product averages. 3 In addition to this general approach,
76 the 1992 guidance provided specific recommendations for (1) logarithmic transformation of PK
77 data, (2) methods to evaluate sequence effects, and (3) methods to evaluate outlier data.
2
For example, to submit an ANDA that is not the same as its RLD because it has a different strength, dosage form,
or route of administration than that of the RLD, an applicant first must obtain permission from FDA through the
citizen petition process. See section 505(j)(2)(C) of the Federal Food, Drug and Cosmetic Act (21 U.S.C.
355(j)(2)(C)); 21 CFR 314.93(b). Such petitions are referred to as suitability petitions.
3
For a broad range of drugs, a BE limit of 80 to 125% for the ratio of the product averages has been adopted for use
of an average BE criterion. Generally, the BE limit of 80 to 125% is based on a clinical judgment that a test product
with BA measures outside this range should be denied market access.
2
78
79 In addition to reiterating the key points from the 1992 guidance and replacing that guidance, the
80 2001 guidance introduced two additional approaches to assessing BE: population BE and
81 individual BE. Both of these approaches, unlike the average BE approach, include a comparison
82 of the variabilities of the PK metrics of the two products being compared, as well as the average
83 responses. However, the individual BE approach is not currently used in the regulatory setting
84 while the population BE approach is mainly used for certain in vitro BE studies. The 2001
85 guidance also includes discussion of replicated crossover designs — crossover designs in which
86 at least some of the subjects receive at least one of the products more than once. The discussion
87 of these designs in that guidance included their implications for possible carryover effects and
88 their use in screening for outliers.
89
90 This guidance provides recommendations on the topics covered by the 1992 guidance and the
91 2001 guidance, as well as recommendations on some additional topics. As noted in the
92 Overview section above, when finalized, this guidance will replace the 2001 guidance.
93
94
95 II. GENERAL CONSIDERATIONS
96
97 A. Study Design
98
99 1. Experimental Design
100
101 a. Nonreplicated designs
102
103 A conventional nonreplicated design, such as the standard two-formulation, two-period, two-
104 sequence crossover design, can be used to generate data when an average or population approach
105 is chosen for BE comparisons. Under certain circumstances, such as products with apparent,
106 long half-lives where crossover studies are impractical, parallel designs can be used.
107
108 b. Replicated crossover designs
109
110 Replicated crossover designs can be used irrespective of which BE approach is selected to
111 establish BE, although they are not necessary when an average or population BE approach is
112 used. When a reference-scaled BE approach is used, replicated crossover designs are critical to
113 allow estimation of within-subject variances for the R (and T if a fully replicated study is used)
114 measures. In particular, the following four-period, two-sequence, two-formulation design is
115 recommended for fully replicated BE studies (see Appendix A for further discussion of
116 replicated crossover designs).
117
Period
1 2 3 4
Sequence 1 T R T R
2 R T R T
3
118
119
120 For this design, the same lots of the T and R formulations should be used for the replicated
121 administration. Each period should be separated by an adequate washout period.
122
123 Other fully replicated crossover designs are also possible. For example, a three-period design, as
124 shown below, could be used. A fully replicated design can estimate the subject-by-formulation
125 interaction variance components.
126
Period
1 2 3
Sequence 1 T R T
2 R T R
127
128 The following three-period, three-sequence, two-formulation, partially replicated design can also
129 be used for assessing reference-scaled BE, though it cannot fully estimate the subject-by-
130 formulation interaction variance component (as a fully replicated design can).
131
Period
1 2 3
Sequence 1 T R R
2 R T R
3 R R T
132 A greater number of subjects would be needed for the three-period designs compared to the
133 recommended four-period design to achieve the same statistical power to conclude BE.
134
135 c. Adaptive design
136
137 An adaptive design is a clinical trial design that allows for prospectively planned modifications
138 to one or more aspects of the design based on accumulating data from subjects in the trial. An
139 adaptive design can be a group sequential design, or other design with one or more adaptive
140 features. 4 For example, Potvin’s methods (Potvin et al. 2008, Xu et al. 2016) 5 are a combination
141 of a group sequential design and an adaptive design with sample size re-estimation.
142
4
See the guidance for industry Adaptive Designs for Clinical Trials of Drugs and Biologics (November 2019). We
update guidances periodically. For the most recent version of a guidance, check the FDA guidance web page at
https://www.fda.gov/regulatory-information/search-fda-guidance-documents.
5
Potvin, D., C.E. DiLiberti, W.W. Hauck, A.F. Parr, D.J. Schuirmann, and R.A. Smith, 2008, Sequential Design
Approaches for Bioequivalence Studies With Crossover Designs, Pharmaceutical Statistics: The Journal of Applied
Statistics in the Pharmaceutical Industry 7, no. 4: 245-262; Xu, J., C. Audet, C.E. DiLiberti, W.W. Hauck, T.H.
Montague, A.F. Parr, D. Potvin, and D.J. Schuirmann, 2016, Optimal Adaptive Sequential Designs for Crossover
Bioequivalence Studies, Pharmaceutical Statistics (15) 1:15-27.
4
143 Adaptive design can provide ethical advantages 6 and statistical efficiency. When appropriately
144 implemented, adaptive designs can reduce resources used, decrease time to study completion,
145 and increase the chance of study success, especially when the prior information needed for the
146 study design is limited. However, use of adaptive designs can also have limitations. For
147 example, adaptive designs may call for certain statistical methods to avoid increasing the chance
148 of erroneous conclusions and introducing bias in estimates and for complex adaptive designs,
149 such methods may not be readily available. 7 The decision to use or not use an adaptive design is
150 at the applicant’s discretion.
151
152 In general, the design, conduct, and analysis of a proposed adaptive study design should satisfy
153 the following recommendations:
154
155 • The details of the adaptive design should be completely specified prior to initiation of the
156 study and documented accordingly. For example, prospective planning should include
157 prespecification of the anticipated number and timing of interim analyses, the type of
158 adaptation, the statistical inference methods to be used and the specific algorithm
159 governing the adaptive decision. If a study should be stopped early (e.g., for futility or
160 for success in demonstrating BE), detailed stopping criteria should be pre-specified and
161 scientifically justified.
162
163 • The applicant should establish that estimation of treatment effect will be sufficiently
164 reliable, and the chance of erroneous conclusions will be adequately controlled. The
165 Agency will accept appropriately designed BE studies that are scientifically justified.
166 Support might include published literature in peer-reviewed journals in which the
167 applicant’s proposed approach is validated or simulation results meeting desired criteria
168 (e.g., the Type I error probability of the proposed approach is controlled at a nominal
169 level of 0.05 for a BE test). Appropriate details (e.g., literature references, proofs,
170 simulation codes/results) for the methodology should be submitted.
171
172 • The applicant should ensure that study integrity will be appropriately maintained. A
173 comprehensive written data access plan defining how study integrity will be maintained
174 in the presence of the planned adaption should be included in the protocol or statistical
175 analysis plan (SAP). This applies to both adaptive comparative clinical endpoint BE
176 studies and PK BE studies, whether blinded or unblinded by design.
177
178 For details, refer to the guidance for industry Adaptive Design for Clinical Trials of Drugs and
179 Biologics (November 2019).
6
See footnote 4. For example, the ability to stop a trial early if it becomes clear that the trial is unlikely to
demonstrate equivalence can reduce the number of patients exposed to the unnecessary risk of an ineffective
investigational treatment and allow subjects the opportunity to explore more promising therapeutic alternatives.
7
See footnotes 4 and 5.
5
180 Due to the increased complexity of adaptive studies and uncertainties regarding their operating
181 characteristics, applicants are encouraged to contact the Agency early to discuss their proposed
182 adaptive study designs and statistical methods via the controlled correspondence, 8 pre-ANDA
183 meeting, 9 pre-IND meeting, or pre-NDA meeting pathway. 10
184
185 d. Design with sparse sampling
186
187 For certain generic products, a sparse BE design is used, where the sampling for each subject is
188 done at a single or very limited number of time points rather than the number needed to get a full
189 concentration profile. For example, some ophthalmic products are studied using a sparse BE
190 design, where only a single sample is collected from a single eye of each subject, at one assigned
191 sampling time point for that subject. More generally, a sparse BE study design can be a parallel
192 design where each subject should receive only one treatment, T or R, but not both. Alternatively,
193 a crossover sparse study design can be used where each subject receives both test and reference
194 treatments (e.g., in subjects undergoing indicated cataract surgery for both eyes).
195
196 For a sparse BE study design, the mean concentration for each product at each time point of
197 measurement is calculated by using the mean concentration of the subjects measured at each time
198 point to derive the mean profile for each product. Based on the trapezoid rule, the AUC0−t for
199 each product is computed as a weighted linear combination of these mean concentrations at each
200 time point through time t. The AUC0−t is the area under the concentration – time curve from
201 zero to the time t. Cmax and Tmax (time to maximum observed concentration) can be determined
202 accordingly. The ratios of AUC0−t and Cmax between the test and the reference product are used
203 to assess BE. Estimation of the standard deviation and confidence interval for the ratio of
204 AUC0−t may be done by bootstrap or parametric methods (e.g., Bailer’s methods (Bailer 1988) 11
205 for a parallel study design), and that for the ratio of Cmax may be done by bootstrap methods. BE
206 is supported if the 90% confidence interval for the ratio of AUCt between the test and the
207 reference product lies within the BE margin (80.00%, 125.00%). Model-based approaches can
208 be considered when they can reliably control the error rate of concluding BE for bio inequivalent
209 products (Type I error). 12
210
211 For complicated issues such as other forms of sparse design or alternative statistical methods,
212 applicants are encouraged to contact the Agency early to discuss their proposed study design and
213 statistical methods via the controlled correspondence, pre-ANDA meeting, pre-IND meeting, or
214 pre-NDA meeting pathway. 13
8
See the guidance for industry Controlled Correspondence Related to Generic Drug Development (December
2020).
9
See the guidance for industry Formal Meetings Between FDA and ANDA Applicants of Complex Products Under
GDUFA (October 2022).
10
See the draft guidance for industry Formal Meetings Between the FDA and Sponsors or Applicants of PDUFA
Products (December 2017). When final, this guidance will represent FDA’s current thinking on this topic.
11
Bailer, A.J., 1988, Testing for the Equality of Area Under the Curves When Using Destructive Measurement
Techniques, Journal of Pharmacokinetics and Biopharmaceutics, 16(3): 303-309.
12
Zhao, L., M.-J. Kim, L. Zhang, and R. Lionberger, 2019, Generating Model Integrated Evidence for Generic Drug
Development and Assessment, Clinical Pharmacology and Therapeutics, 105(2): 338-349.
13
See footnotes 8, 9, and 10.
6
215
216 2. Sample Size Determination
217
218 It is an applicant’s responsibility to design an adequately powered BE study for the proposed
219 study. We recommend that applicants enroll enough subjects to power the study at a level of 0.8
220 or higher, for a BE test to be carried out with a type 1 error rate of 0.05 (see section III.C.1.a for
221 more details). When determining the sample size, rates of attrition and noncompliance (e.g.,
222 protocol violation) should be taken into consideration. Enough subjects should be recruited,
223 randomized, and dosed at the beginning of the study to ensure that the desired number of
224 evaluable subjects will be available for analysis. All eligible subjects who were dosed should be
225 included in the analysis. For BE studies, add-on subjects after the pre-specified number of
226 subjects have been reached are generally not encouraged except in an adaptive study design with
227 a pre-specified adaptation to add subjects and statistical methods to control the Type I error rate
228 under the nominal level.
229
230 The number of subjects to be included in a study should be based on an appropriate sample size
231 calculation for the proposed study design. 14,15,16 For example, the standard 2×2 cross-over study
232 will use a particular calculation while studies with a different design or set of endpoints will use
233 different calculations. For sample size re-estimation in an adaptive study design, refer to Section
234 II.A.1.c. Adaptive Design.
235
236 Sample size and power calculation should be supported by established scientific practice. For
237 complex study designs with no analytical solutions for sample size calculation, simulation can be
238 used to estimate the needed sample size in order to reach a desired power. The method by which
239 the sample size is determined should be given in the protocol, together with the estimates of any
240 quantities used in the calculations (such as variances, mean values, response rates, the assumed
241 effect size). The basis for these estimates should also be given. For example, variance estimates
242 can be obtained from the biomedical literature and/or pilot studies. It is important to investigate
243 the sensitivity of the sample size calculated to a variety of deviations from the assumed
244 estimates. This may be facilitated by providing a range of sample sizes appropriate for a
245 reasonable range of deviations from the assumptions or alternative approaches supported by
246 published peer-reviewed literature.
247
248 Applicants should enter a sufficient number of subjects in the study to allow for dropouts.
249 Dropouts generally should not be replaced because replacement of subjects during the study
250 could complicate the statistical model and analysis. Applicants who wish to replace dropouts
251 during the study should indicate this intention in the protocol. The protocol should also state
252 whether samples from replacement subjects, if not used, will be assayed. If the dropout rate is
253 high and applicants wish to add more subjects, a modification of the statistical analysis may be
14
Chow, S.-C. and J.-P. Liu, 2008, Design and Analysis of Bioavailability and Bioequivalence Studies, 3rd Edition,
New York: Chapman and Hall/CRC.
15
Draft guidance for industry Bioequivalence Studies with Pharmacokinetic Endpoints for Drugs Submitted Under
an ANDA (August 2021). When final, this guidance will represent FDA’s current thinking on this topic.
16
Patterson, S.D. and B. Jones, 2017, Bioequivalence and Statistics in Clinical Pharmacology, 2nd Edition, New
York: Chapman and Hall/CRC.
7
254 recommended. Additional subjects should not be included after data analysis unless the study
255 was designed from the beginning as an adaptive design.
256
257 In general, for PK BE or in vitro BE studies, sample size calculation should be based on BE
258 metrics (e.g., AUC, Cmax) after log-transformation; for comparative clinical endpoint BE studies,
259 sample size calculation should be based on the un-transformed comparative clinical endpoints
260 unless otherwise noted in the relevant FDA product-specific guidance (PSG). 17 The number of
261 evaluable subjects in a PK BE study should not be less than 12. For highly variable drug
262 products, a minimum of 24 subjects are recommended for BE assessment. 18
263
264 B. Data Preparation
265
266 The drug concentration in biological fluid determined at each sampling time point should be
267 furnished on the original scale for each subject participating in the study. The PK measures of
268 systemic exposure should also be furnished on the original scale. The variables for a
269 comparative clinical endpoint BE study should also be furnished on the original scale. The
270 mean, standard deviation, and coefficient of variation for each variable should be computed and
271 tabulated in the final report.
272
273 1. Log-Transformation
274
275 A general approach to assessing BE is to compare the log-transformed BA measures after
276 administration of the T and R products.
277
278 a. Logarithmic transformation for PK measures
279
280 This guidance recommends that PK BE measures (e.g., AUC and Cmax) be log-transformed (see
281 Appendix B). The choice of common or natural logs should be consistent and should be stated in
282 the study report. The limited sample size in a typical BE study precludes a reliable
283 determination of the distribution of the data set. Sponsors and/or applicants are not encouraged
284 to test for normality of error distribution after log-transformation, nor should they use normality
285 of error distribution as a reason for carrying out the statistical analysis on the original scale.
286 Justification should be provided if sponsors or applicants believe that their BE study data should
287 be statistically analyzed on the original rather than on the log scale.
288
289
290
291
17
For the most recent version of a product-specific guidance, check the product-specific web page at
https://www.accessdata.fda.gov/scripts/cder/psg/index.cfm.
18
Davit, B. and D. Conner, 2010, Reference-Scaled Average Bioequivalence Approach. In: I. Kanfer and L.
Shargel, editors. Generic Drug Product Development — International Regulatory Requirements for Bioequivalence,
New York, NY: Informa Healthcare, 271-272; Food and Drug Administration, Advisory Committee for
Pharmaceutical Science, October 5-6, 2006.
8
292 b. Data transformation for comparative pharmacodynamic and clinical

293 endpoint BE study
294
295 The decision on whether and how to transform a variable for a comparative pharmacodynamic
296 (PD) or comparative clinical endpoint BE study should be specified in the protocol, especially
297 for the primary variable(s). The basis for the variables should also be given in the protocol. For
298 example, these variables can be obtained from the biomedical literature and/or pilot studies.
299 Similar considerations apply to other derived variables, such as the use of change from baseline,
300 percentage change from baseline, the area under the curve of repeated measures, or the ratio of
301 two different variables. Subsequent clinical interpretation should be carefully considered.
302 Regarding comparative clinical endpoint studies, in general the log-transformation is not
303 used. For example, in the case of the Fieller’s confidence interval for the ratio of two means, the
304 raw (untransformed) data are used for the confidence interval derivation. 19
305
306 c. Negative values for baseline corrected PK or PD endpoints
307
308 Because data transformation and scales might affect BE conclusions, they should be chosen
309 carefully and appropriately justified in the protocol. 20 If a baseline correction results in a
310 negative plasma concentration value, the value should be set equal to 0 before calculating the
311 baseline-corrected AUC.
312
313 2. Missing Data and Intercurrent Events
314
315 Subjects may have missing data in the study for various reasons (e.g., subject’s refusal to
316 continue in the study, worsening of conditions or emergence of adverse events, subject’s failure
317 to meet scheduled appointments for evaluation). Subjects may also have intercurrent (post-
318 randomization) events that affect either the interpretation or the existence of the measurements
319 associated with the question of interest (e.g., noncompliance with the protocol for various
320 reasons, use of rescue medication due to lack of efficacy, death). Missing data and intercurrent
321 events can introduce problems such as bias, misleading inference, loss of precision and loss of
322 power, which make it hard to interpret the trial outcome.
323
324 The ICH (Internal Council for Harmonization) E9(R1) Addendum introduces the concept of an
325 estimand, which is a precise description of the treatment effect reflecting the clinical question
326 posed by a particular study objective. 21 The trial protocol of a BE study should include the
327 following components of an estimand: (1) the treatment of interest and alternative treatment(s) to
328 which comparison will be made: e.g., test drug compared with reference drug; (2) the analysis
329 population for BE assessment; (3) the variable (or endpoint) to be measured for each subject
330 (e.g., AUC or Cmax); (4) the specification of how to account for intercurrent events in assessing
331 the scientific question of interest (for example, in a comparative clinical endpoint BE study with
19
Fieller, E., Some Problems in Interval Estimation, 1954, Journal of the Royal Statistical Society, 16(2): 175-185.
20
For example, see Sun, W., S. Grosser, and Y. Tsong, 2017, Ratio of Means vs. Difference of Means as Measures
of Superiority, Noninferiority, and Average Bioequivalence, Journal Biopharmaceutical Statistics, 27(2): 338-355.
21
Guidance for industry E9(R1) Statistical Principles for Clinical Trials: Addendum: Estimands and Sensitivity
Analysis in Clinical Trials, Revision 1 (May 2021).
9
332 a binary endpoint, subjects who discontinue study treatment early due to lack of treatment effect
333 should be included as treatment failures); and (5) the population-level summary for the variable
334 to compare between treatment conditions, e.g., the geometric mean ratio of the test to reference
335 drug in a PK BE study.
336
337 The protocol should include plans to minimize missing data. The trial protocol should
338 prospectively define anticipated causes of missing data, the corresponding statistical assumptions
339 about reasons for the missing data, and how missing data will be treated in the statistical
340 analysis. The treatment of missing data in the statistical analysis should be justified such that
341 valid statistical inferences can be made under the assumptions about the missing data
342 mechanism.
343
344 Statistical methods for handling missing data include complete case analysis, available case
345 analysis, weighting methods, imputation, and model-based approaches. For example, in a two-
346 way crossover study, a complete case analysis could be a general linear model as implemented in
347 SAS PROC GLM, which removes all subjects with any missing observations for any variables
348 included in the GLM model (i.e., removes subjects missing one or both periods). An available
349 case analysis could be done using SAS PROC MIXED, which uses all observed data (e.g., in a
350 two-way crossover study, uses all subjects with one or two complete periods of data).
351
352 Approaches for handling missing data and the statistical methods for the primary BE analysis
353 (e.g., GLM vs. MIXED) should be pre-specified in the study protocol or SAP. Depending on the
354 nature of the assumed or likely missing data mechanism, statistical methods from any of these
355 categories may be appropriate. The validity of a statistical approach to handle missing data
356 depends on a variety of factors, including, but not limited to, the mechanism for missingness, the
357 fraction of incomplete cases, the values that are missing, specifics of the analysis, and definition
358 of the estimand. Sensitivity analyses using alternative approaches may also be used in the
359 statistical analysis to address missing data. Sensitivity analyses should be pre-specified in the
360 trial protocol to evaluate the robustness of conclusions to deviations from the assumptions about
361 the missing data mechanism. The applicant should provide detailed information about reasons
362 for missing data and any observed intercurrent events.
363
364 For a particular drug product, if the PSG recommends certain approaches to handling missing
365 data, the applicants should refer to that PSG. Applicants may choose to contact the Agency via
366 the controlled correspondence, pre-ANDA meeting, pre-IND meeting, or pre-NDA meeting
367 pathway to discuss their proposed approach to handling missing data if such an approach is
368 different from what is recommended in the PSG or if the applicants have further questions.
369
370 3. Outlier Detection
371
372 Outlier data in BE studies are defined as subject data for one or more BA measures that are
373 discordant with corresponding data for that subject and/or for the rest of the subjects in a study.
374 Because BE studies are usually carried out as crossover studies, the most important type of
375 subject outlier is the within-subject outlier, when one subject or a few subjects differ notably
376 from the rest of the subjects with respect to a within-subject T-R comparison. The existence of a
10
377 subject outlier with no protocol violations and for which there are not bioanalytical errors could
378 indicate one of the following situations:
379
380 a. Product failure
381
382 Product failure could occur, for example, when a subject exhibits an unusually high or low
383 response to one or the other of the products because of a problem with the specific dosage unit
384 administered. This could occur, for example, with a sustained and/or delayed-release dosage
385 form exhibiting dose dumping or a dosage unit with a coating that inhibits dissolution.
386
387 b. Subject-by-formulation interaction
388
389 A subject-by-formulation interaction could occur when an individual is representative of subjects
390 present in the general population in low numbers, for whom the relative BA of the two products
391 is markedly different from that for most of the population, and for whom the two products are
392 not bioequivalent, even though they might be bioequivalent in most of the population. In the
393 case of product failure, the unusual response could be present for either the T or R product.
394 However, in the case of a subpopulation, even if the unusual response is observed on the R
395 product, there could still be concern about lack of bioequivalence of the two products. For these
396 reasons, applicants should not remove data from the statistical analysis of BE studies solely
397 because those data are identified as statistical outliers.
398
399 In general, outlier data (whether due to product failure, subject-by-formulation interaction, or
400 another cause) may only be removed from the BE statistical analysis if there is real-time
401 documentation demonstrating a protocol violation during the clinical and/or
402 analytical/experimental phase of the BE study. Applicants should include a prospective plan in
403 the BE study protocol for handling subjects (experimental outliers) in the BE statistical analysis.
404 Data from redosing studies are not considered valid evidence to support removal of outlier data
405 from the statistical analysis. All subject data should be submitted, with potential outliers flagged
406 with appropriate documentation as part of the submission. However, for a replicated PK BE
407 study, if reference-scaled average BE is used, the applicant should ensure that the calculated
408 intra-subject variability is not inflated due to extreme values or situations.
409
410 To characterize aberrant observations for exploratory or quality control purposes, the choice of
411 the appropriate technique depends on whether there are outlying subjects or outlying
412 observations, as well as on the study design.
413
414 C. Statistical Models
415
416 1. General Statistical Criteria for Bioequivalence
417
418 The general structure of a BE criterion is that a function (Θ) of population measures should be
419 demonstrated to be no greater than a specified value (θ). Using the terminology of statistical
420 hypothesis testing, this is accomplished by testing the hypothesis H0: Θ≥θ versus Ha: Θ<θ at a
11
421 desired level of significance, often 5%. Rejection of the null hypothesis H0 (i.e., demonstrating
422 that the estimate of Θ is statistically significantly less than θ) results in a conclusion of BE.
423
424 a. Use of confidence intervals to do two one-sided tests
425
426 In BE assessment we are frequently interested in testing whether a parameter (for example, the
427 difference of means for a T and R product for a specific endpoint) is contained within a defined
428 interval, call it [θ1, θ2]. The recommended method for doing such a test is the Two One-Sided
429 Tests Procedure. 22 A one-sided statistical test is carried out to determine whether the parameter
430 is ≥ θ1, and a second one-sided test is carried out to determine whether the parameter is ≤ θ2;
431 both tests are carried out at a level of significance α, which is usually 0.05. If both tests are
432 successful (that is, we reject the null hypothesis in both cases), we conclude that the parameter is
433 contained in [θ1, θ2].
434
435 These two one-sided tests are sometimes carried out by calculating a 100 (1-2α) % confidence
436 interval for the parameter and determining whether this confidence interval is completely
437 contained in the interval [θ1, θ2]. For this confidence interval method of carrying out the tests to
438 be valid, the confidence interval should be an equal tails confidence interval. If the lower and
439 upper confidence limits of the 100 (1-2α) % confidence interval are L1 and L2, respectively, then
440 the confidence interval is equal tails if L1, by itself, is at least a 100 (1-α) % lower confidence
441 bound for the parameter and L2, by itself, is at least a 100 (1-α) % upper confidence bound for
442 the parameter.
443
444 In some cases, there may not be general agreement as to the best choice of a particular statistical
445 testing methodology for carrying out the two one-sided tests (for example, if the parameter of
446 interest is the difference between the success probabilities for a T and R product for a binary
447 endpoint). In such cases, careful consideration should be given to the choice of statistical
448 methods for doing the two one-sided tests, which may or may not correspond to a confidence
449 interval method.
450
451 2. Statistical Information and Implementation of Criteria for PK Measures (AUC0-t,
452 AUC0-∞, and Cmax)
453
454 We recommend that applicants provide the following statistical information for AUC0-t,
455 AUC0-∞, and Cmax:
456
457 • Geometric means for the formulations tested
458 • Arithmetic means for the formulations tested
459 • Geometric mean ratios of Test vs. Reference and their corresponding 90% confidence
460 intervals or 95% upper confidence bounds (e.g., for highly variable drugs or narrow
461 therapeutic index drugs)
22
Schuirmann, D. J., 1987, A Comparison of the Two One-Sided Tests Procedure and the Power Approach for
Assessing the Equivalence of Average Bioavailability, Journal of Pharmacokinetics and Biopharmaceutics, 15(6):
657-680.
12
462
463 Recommended statistical information for other types of outcome measures is discussed in section
464 III: Specific Situations.
465
466 To facilitate BE comparisons, for crossover studies, the measures for each individual should be
467 displayed in parallel for the formulations tested. For each BE measure, the ratio of the individual
468 geometric mean of the T product to the individual geometric mean of the R product should be
469 tabulated side by side. The summary tables should indicate in which sequence each subject
470 received the product.
471
472 Statistical analyses of BE data are typically based on a statistical model for the logarithm of the
473 BA measures (e.g., AUC and Cmax). The model is a mixed-effects or two-stage linear model.
474 Each subject, j, theoretically provides a mean for the log-transformed BA measure for each
475 formulation, µTj and µRj for the T and R formulations, respectively. The model assumes that
476 these subject-specific means come from a distribution with population means µT and µR, and
477 between-subject variances σBT2 and σBR2, respectively. The model allows for a correlation, ρ,
478 between µTj and µRj. The subject-by-formulation interaction variance component, σD2, is related
479 to these parameters as follows:
480
481 σD2 = variance of (µTj - µRj)
482
483 = (σBT - σBR)2 + 2 (1-ρ)σBTσBR [23]
484
485 For a given subject, the observed data for the log-transformed BA measure are assumed to be
486 independent observations from distributions with means µTj and µRj, and within-subject variances
487 σWT2 and σWR2. The total variances for each formulation are defined as the sum of the within-
488 and between-subject components (i.e., σTT2 = σWT2 + σBT2 and σTR2 = σWR2 + σBR2). For analysis
489 of crossover studies, the means are given additional structure by the inclusion of period and
490 sequence effect terms.
491
492 The applicant may also consider prespecifying inclusion of important demographic and baseline
493 prognostic covariates in the statistical model for parallel studies. This sort of adjustment can
494 increase the precision and power of the statistical analysis and compensate for any lack of
495 balance between treatment groups with no inflation of Type 1 error.
496
497
498
499
23
Schall, R., and H. G. Luus, 1993, On Population and Individual Bioequivalence, Statistics in Medicine, 12(12):
1109-1124.
13
500 III. SPECIFIC SITUATIONS 24

501
502 A. In Vitro Bioequivalence and Population Bioequivalence
503
504 This section discusses statistical methods for assessment of in vitro BE, including population BE
505 (PBE), a similarity index (f2), statistical approaches respectively for in vitro release tests (IVRT),
506 in vitro permeation tests (IVPT) and in vitro abuse-deterrent formulations (ADF) comparative
507 studies, and a profile comparison approach based on Earth Mover’s Distance (EMD).
508
509 1. Population Bioequivalence
510
511 One of the recommended statistical approaches for evaluating in vitro BE is population BE
512 (PBE). To test for PBE, the null and alternative hypotheses are given as follows:
513 𝐻𝐻0 : 𝜃𝜃 ≥ 𝜃𝜃𝑃𝑃 vs. 𝐻𝐻𝑎𝑎 : 𝜃𝜃 < 𝜃𝜃𝑃𝑃
(μT −μR )2 +σ2T −σ2R (μT −μR )2 +σ2T −σ2R
514 where 𝜃𝜃 = if the estimated 𝜎𝜎𝑅𝑅 > 𝜎𝜎0 or 𝜃𝜃 = if the estimated
σ2R σ20
515 𝜎𝜎𝑅𝑅 ≤ 𝜎𝜎0 .
516 Here, 𝜇𝜇 𝑇𝑇 and 𝜇𝜇𝑅𝑅 are the population means, 𝜎𝜎𝑇𝑇2 and 𝜎𝜎𝑅𝑅2 are the population variances of the log-
517 transformed measure for T and R products, respectively; 𝜎𝜎02 is a regulatory constant for variance;
518 and 𝜃𝜃𝑃𝑃 is the PBE limit. The concept of PBE is to compare the difference of the T and R
519 products with that of the reference versus reference itself. This comparison can be denoted in
520 terms of the population difference ratio as follows:
E(YT − YR )2 (μT − μR )2 + σ2R + σ2T θ
521 � = � = � + 1.
E(YR − YR′ )2 2σ2R 2
522 The regulatory constant variance, 𝜎𝜎02 , is set based on the following considerations. Due to the
523 low variability of in vitro measurements, this guidance recommends that the ratio of geometric
524 means should fall within 0.90 and 1.11. As a result, an upper BE limit of 1.11 is recommended
525 for the average BE limit for in vitro data. Assuming σ2R = σ2T = 𝜎𝜎02 , μT − μR = ln 1.11 and the
526 maximum allowable limit for population difference ratio is 1.25, this leads to the recommended
527 choice of 𝜎𝜎02 = 0.01.
528
529 The determination of PBE limit, 𝜃𝜃𝑃𝑃 , is based on the consideration of average BE criterion and
530 the addition of variance terms to PBE criterion as the following form:
(μT − μR )2 + σ2T − σ2R Average BE limit + Variance term
531 = .
max{σ20 , σ2R } Scaled variance term
532
533 The FDA recommended allowance for the variance term is 0.01. This value may be adjusted
534 depending on the average BE limit for in vitro data based on further communication with the
535 Agency. Accordingly, the PBE limit, 𝜃𝜃𝑃𝑃 , is recommended as follows:
24
Some specific situations are addressed in the following subsections with specified choices of BE criteria. Further
discussion regarding these specified choices can be found in the guidances cited in those subsections.
14
(ln 1.11)2 + 0.01

536 𝜃𝜃𝑃𝑃 = = 2.089
0.01
537
538 A linearized form is recommended to use to test 𝐻𝐻0 : 𝜃𝜃 ≥ 𝜃𝜃𝑃𝑃 . That is, testing H0 : 𝜃𝜃 ≥ 𝜃𝜃𝑃𝑃 is
539 equivalent to testing H0 : γ ≥ 0 where γ = (𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2 + (σ2T − σ2R ) − 𝜃𝜃𝑃𝑃 𝜎𝜎𝑅𝑅2 if the estimated
540 𝜎𝜎𝑅𝑅 > 𝜎𝜎0 or γ = (𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2 + (σ2T − σ2R ) − 𝜃𝜃𝑃𝑃 𝜎𝜎02 if the estimated 𝜎𝜎𝑅𝑅 ≤ 𝜎𝜎0 . Here, γ1 =
541 (𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2, γ2 = 𝜎𝜎𝑇𝑇2 and γ3 = 𝜎𝜎𝑅𝑅2 + 𝜃𝜃𝑃𝑃 σ2R if the estimated 𝜎𝜎𝑅𝑅 > 𝜎𝜎0 or γ3 = 𝜎𝜎𝑅𝑅2 + 𝜃𝜃𝑃𝑃 σ20 if the
542 estimated 𝜎𝜎𝑅𝑅 ≤ 𝜎𝜎0 .
543 Suppose γ�𝑈𝑈 is a 95% upper confidence bound for γ. Then, PBE is supported if and only if γ�𝑈𝑈 ≤
544 0. Based on the work of Howe (1974) 25 and Ting et al. (1990) 26, an approximate 95% upper
545 confidence bound for γ is given as follows:
546 γ�𝑈𝑈 = γ�1 + γ�2 − γ�3 + �(γ�1 − γ�1 )2 + (γ�2 − γ�2 )2 + (γ�3 − γ�3 )2
547
548 where γ�1 , γ�2 , and γ�3 are point estimators of γ1 , γ2 , and γ3 , respectively; γ�1 and γ�2 are 95%
549 upper confidence bounds for γ1 and γ2 and γ�3 is a 95% lower confidence bound for 𝛾𝛾3. For
550 further detail, see, e.g., the draft PSGs for Budesonide suspension (September 2012) and
551 Fluticasone Propionate metered spray (June 2020). 27
552
553 2. Similarity Index (f2)
554
555 For a comparison of dissolution profiles, similarity is assessed using the similarity index, f2
556 (Shah et al., 1998), 28 as described in detail in the guidance for industry Immediate Release Solid
557 Oral Dosage Forms Scale-Up and Postapproval Changes: Chemistry, Manufacturing, and
558 Controls, In Vitro Dissolution Testing, and In Vivo Bioequivalence Documentation (November
559 1995). In particular, given that all profiles are conducted on a minimum of 12 individual dosage
560 units, 2 profiles are similar if the value of their similarity factor f2 is between 50 and 100.
561
562 3. In-Vitro Release Test
563
564 When an in-vitro release test (IVRT) is used to support a demonstration of BE for topical
565 dermatological drug products as part of an in vitro characterization-based BE approach, a two-
566 stage, nonparametric statistical approach is recommended, and described in the draft guidance
567 for industry In Vitro Release Test Studies for Topical Drug Products Submitted in ANDAs
568 (October 2022). 29 The statistical approach is the same as that used to assess the equivalence of
569 drug release rates for non-sterile semisolid dosage forms evaluated by a comparative IVRT study
570 in the context of certain postapproval changes; this is shown in detail in the guidance for industry
25
Howe, W.G., 1974, Approximate Confidence Limits of the Mean of X+Y Where X and Y are Two Tabled
Independent Random Variables, Journal of the American Statistical Association, 69:789-794.
26
Ting, N., R.K. Burdick, F. Graybill, S. Jeyaratnam, and T.F.C. Lu, 1990, Confidence Intervals on Linear
Combinations of Variance Components That Are Unrestricted in Sign, Journal of Statistical Computation and
Simulation, 35:135-143.
27
When final, these guidances will represent FDA’s current thinking on these topics.
28
Shah, V.P., Y. Tsong, P. Sathe, and J.P. Liu, 1998, In Vitro Dissolution Profile Comparison—Statistics and
Analysis of the Similarity Factor, f2, Pharmaceutical Research, 15(6):889-896.
29
When final, this guidance will represent FDA’s current thinking on this topic.
15
571 Nonsterile Semisolid Dosage Forms — Scale-Up and Postapproval Changes: Chemistry,
572 Manufacturing, and Controls; In Vitro Release Testing and In Vivo Bioequivalence
573 Documentation (May 1997).
574
575 The assessment of equivalence by an IVRT involves a comparison of the median in vitro drug
576 release rates of two formulations using a non-parametric statistical test which is resistant to
577 outliers that are expected to occur under the particular testing conditions.
578
579 4. In-Vitro Permeation Test
580
581 When an in-vitro permeation test (IVPT) is used to support a demonstration of BE for topical
582 dermatological drug products as part of an in vitro characterization-based BE approach, a mixed
583 scaled criterion is recommended, and described in detail in the draft guidance for industry In
584 Vitro Permeation Test Studies for Topical Drug Products Submitted in ANDAs (October 2022). 30
585 According to that methodology, a confidence interval is calculated for each of the endpoints, log-
586 transformed maximum flux (𝐽𝐽𝑚𝑚𝑚𝑚𝑚𝑚 ) and log-transformed total (cumulative) amount (AMT)
587 permeated. The permeation test is performed with excised skin sections from patients
588 undergoing a surgical procedure or from cadaver donors and the statistical test uses the within-
589 reference standard deviation, 𝑆𝑆𝑊𝑊𝑊𝑊 , as the threshold that prompts use of either the unscaled or
590 scaled confidence interval.
591
592 The mixed-scaled criterion uses the within-reference standard deviation as a threshold,
593 independently, for each endpoint. Specifically, for 𝐽𝐽𝑚𝑚𝑚𝑚𝑚𝑚 or log-transformed total (cumulative)
594 amount permeated, the reference-scaled average BE approach is used for the endpoint only if it
595 has a 𝑆𝑆𝑊𝑊𝑊𝑊 > 0.294. The regular ABE approach (refer to Schuirmann, 1987) 31 is used for the
596 endpoint with 𝑆𝑆𝑊𝑊𝑊𝑊 ≤ 0.294.
597
598 In the reference-scaled average BE approach, the hypotheses to be tested are:
599
(𝜇𝜇𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2
600 𝐻𝐻0 : 2 ≥ 𝜃𝜃
𝜎𝜎𝑊𝑊𝑊𝑊
(𝜇𝜇𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2
601 𝐻𝐻𝑎𝑎 : 2 < 𝜃𝜃
2
602 Here we determine the 100(1-α)% upper confidence bound for (𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2 − 𝜃𝜃𝜎𝜎𝑊𝑊𝑊𝑊
603 where:
604 - 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 = mean difference of T and R products
2
605 - 𝜎𝜎𝑊𝑊𝑊𝑊 = within-subject variance of R product
(ln (𝑚𝑚))2
606 - 𝜃𝜃 = , 𝑚𝑚 = 1.25, and 𝜎𝜎𝑊𝑊0 = 0.25 (regulatory constant)
(𝜎𝜎𝑊𝑊0 )2
607 For the T product to be bioequivalent to the R product, both of the following conditions must be
608 satisfied for each endpoint tested:
30
31
See footnote 22.
16
609
2
610 a. The 95% upper confidence bound for (𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 )2 − 𝜃𝜃𝜎𝜎𝑊𝑊𝑊𝑊 must be less than
611 or equal to zero (numbers should be kept to a minimum of four significant
612 figures for comparison).
613
614 b. The point estimate of the T/R geometric mean ratio must fall within the pre-
1
615 specified limits �𝑚𝑚 , 𝑚𝑚�, where m = 1.25.
616
617 In the case of the non-scaled approach, we calculate the 100(1-2α)% confidence interval for
618 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 as
619
𝑆𝑆𝐼𝐼2
620 𝐼𝐼.̅ ± 𝑡𝑡(1−𝛼𝛼),(𝑛𝑛−1) ∗ �
𝑛𝑛
621
622 where:
623 - 𝐼𝐼.̅ is the point estimate for the mean difference of T and R products
624 - 𝑆𝑆𝐼𝐼2 stimate of inter-donor variability
625 - 𝑡𝑡(1−𝛼𝛼),(𝑛𝑛−1) is the 100 (1 − 𝛼𝛼) percentile of the student’s t-distribution with (𝑛𝑛 − 1)
626 degrees of freedom
627 - n is the number of donors
628 - the value of α is usually set at 0.05
629
630 For the T product to be bioequivalent to the R product, the 100(1-2α)% confidence interval for
1
631 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 must be contained within the limits �𝑚𝑚 , 𝑚𝑚� in the original scale for each endpoint
632 tested, where m = 1.25.
633
634 5. Abuse-Deterrent Formulation Comparative Studies
635
636 An ADF is a formulation that has abuse-deterrent properties, which are defined as drug product
637 properties that are expected to meaningfully deter certain types of abuse, even if they do not fully
638 prevent abuse. 32 The general BE statistical considerations for in vitro ADF comparative studies
639 presented in this guidance align with the guidance for industry – Abuse-Deterrent Opioids —
640 Evaluation and Labeling33 and the guidance for industry – General Principles for Evaluating the
641 Abuse Deterrence of Generic Solid Oral Opioid Drug Products (November 2017). The potential
642 route of abuse (i.e., ingestion (oral route), injection (parenteral route), insufflation (nasal route), or
643 smoking (inhalation route)) and its relevance to ADF design feature(s) will determine how an
644 applicant should evaluate the abuse deterrence of the product utilizing a tier-based approach. To
645 support in vitro ADF comparative studies, the Agency recommends applicants provide
32
See the guidance for industry Abuse-Deterrent Opioids - Evaluation and Labeling (April 2015).
33
Ibid.
17
646 justification for the sample size, statistical test, and number of batches to assess the abuse-deterrent
647 properties and demonstrate consistency of abuse-deterrent performance throughout the drug
648 product shelf-life and lifecycle (i.e., postapproval changes). Applicants should consider a
649 standardized accept/reject criterion based on delta or confidence interval relevant to the abuse-
650 deterrent outcome. The Agency recommends the use of relevant statistics (e.g., sampling plans)
651 to support evaluation of abuse-deterrent properties.
652
653 For ANDA submissions, a non-inferiority approach should be taken when comparing T product
654 with R product to conclude that T product is no less abuse deterrent than R product. 34 The Agency
655 recommends inferential analyses to evaluate the abuse deterrence of T product versus R product.
656 In the analyses, a hierarchical set of null hypotheses serves as a gatekeeper for subsequent null
657 hypotheses, evaluating the abuse deterrence of T and R products under progressively more
658 challenging conditions. A hierarchical inferential approach is used to maintain a fixed family-wise
659 experiment Type I error rate. Typically, the acceptable Type I error probability (α) will be set at
660 5%.
661
662 6. Earth Mover’s Distance Based Profile Comparison Approach
663
664 EMD is a statistical metric that measures the discrepancy (distance) between distributions
665 without a prior assumption of the distribution. 35 The EMD has been recommended in a profile
666 comparison approach to assess equivalence of particle size distribution profile, 36 where the
667 profile exhibits complex distribution (i.e., multiple peaks) that cannot be accurately described by
668 some conventional descriptors (e.g., the D50 and SPAN). The EMD-based profile comparison
669 approach is briefly described as follows. To assess equivalence between the T and R product
670 formulations in the particle size distribution shape, an average profile of all R product samples
671 (i.e., R center) is calculated and serves as the reference profile to compute the distance between
672 an R or a T product sample to the R center using the EMD algorithm. After obtaining the profile
673 distances between each R product sample and the R product average (R – R center distance), and
674 the profile distances between each T product sample and the R product average (T – ‘R center’
675 distance), a statistical equivalence method, e.g., the PBE, is then applied to the two groups of
676 distances to indicate whether the T and R products are statistically equivalent in the particle size
677 distribution shape. For details, refer to Rubner et al. (2000). 37
678
679 Importantly, considering the increasingly emerging technologies and methods for in vitro BE
680 studies, applicants are encouraged to contact the Agency early to discuss their proposed study
681 designs and statistical methods via the controlled correspondence, pre-ANDA meeting, pre-IND
682 meeting, or pre-NDA meeting pathway. 38
683
34
Guidance for Industry Evaluating the Abuse Deterrence of Generic Solid Oral Opioid Drug Products (November
2017).
35
Rubner, Y., C. Tomasi, and L.J. Guibas, 2000, The Earth Mover’s Distance as a Metric for Image Retrieval,
International Journal of Computer Vision, 40(2):99-121.
36
Draft PSG for industry on Cyclosporine emulsion (October 2016). When final, this guidance will represent the
FDA’s current thinking on this topic.
37
See footnote 35.
38
18
684 B. Statistical Methods for Narrow Therapeutic Index and Highly Variable Drug
685 Products
686
687 1. Statistical Method for Narrow Therapeutic Index Drugs
688
689 If a drug is a narrow therapeutic index drug, a fully replicated cross-over design should be used.
690 The statistical analysis should be carried out using both the ABE and the reference-scaled
691 average BE tests for both AUC and Cmax.
692
693 The reference-scaled average BE is evaluated by testing the null hypothesis:
( 𝜇𝜇𝑇𝑇 −𝜇𝜇𝑅𝑅 )2
694 𝐻𝐻0 : 2 ≥ 𝜃𝜃
695 versus the alternative hypothesis:
( 𝜇𝜇𝑇𝑇 −𝜇𝜇𝑅𝑅 )2
696 𝐻𝐻𝑎𝑎 : 2 < 𝜃𝜃
697
698 where:
699 – 𝜇𝜇 𝑇𝑇 is the population average response of the log-transformed measure for the Test
700 formulation.
701 – 𝜇𝜇𝑅𝑅 is the population average response of the log-transformed measure for the
702 Reference formulation.
2
703 – 𝜎𝜎𝑊𝑊𝑊𝑊 is the population within subject variance of the Reference formulation.
[ln(∆)]2
704 – 𝜃𝜃 = 2 is the BE limit.
𝜎𝜎𝑊𝑊0
2
705 – ∆ and 𝜎𝜎𝑊𝑊0 are predetermined constants. Refer to the draft guidance for industry
706 Bioequivalence Studies With Pharmacokinetic Endpoints for Drugs Submitted
2 39
707 Under an ANDA (August 2021) for the values of ∆ and 𝜎𝜎𝑊𝑊0 .
708 Testing is usually done at α=0.05 and that rejection of the null hypothesis supports the
709 conclusion of bioequivalence.
710
711 Narrow therapeutic index BE studies should pass both the reference-scaled approach and the
712 unscaled average BE limits of 80.00 to 125.00%.
713
714 In addition, the test/reference ratio of the within-subject standard deviation should be evaluated.
715 The within-subject variability comparison of the T and R drug products is carried out by a one-
716 sided F test. The null hypothesis for this test is the following.
717
𝜎𝜎
718 𝐻𝐻0 : 𝜎𝜎 𝑊𝑊𝑊𝑊 ≥ 𝛿𝛿
𝑊𝑊𝑊𝑊
39
19
719
720 And the alternative hypothesis is:
721
𝜎𝜎
722 𝐻𝐻𝑎𝑎 : 𝜎𝜎 𝑊𝑊𝑊𝑊 < 𝛿𝛿
𝑊𝑊𝑊𝑊
723
724 where σWT is the within-subject standard deviation for the test product, σWR is the within-subject
725 standard deviation for the reference product and δ is the limit to declare the within-subject
726 variability of the test product is not greater than that of the reference product (refer to the draft
727 guidance for industry Bioequivalence Studies With Pharmacokinetic Endpoints for Drugs
728 Submitted Under an ANDA (August 2021) where 𝛿𝛿 was set to 2.5). 40
729
730 • The 100(1-α)% CI for σWT/σWR is given by
𝑠𝑠𝑤𝑤𝑤𝑤 /𝑠𝑠𝑤𝑤𝑤𝑤 𝑠𝑠𝑤𝑤𝑤𝑤 /𝑠𝑠𝑤𝑤𝑤𝑤

731 • � , �
�𝐹𝐹𝛼𝛼 (𝑣𝑣1 ,𝑣𝑣2 ) �𝐹𝐹1−𝛼𝛼 (𝑣𝑣1 ,𝑣𝑣2 )
2 2
732 Here, α=0.1, 𝐹𝐹𝛼𝛼 (𝑣𝑣1 , 𝑣𝑣2 ) and 𝐹𝐹1−𝛼𝛼 (𝑣𝑣1 , 𝑣𝑣2 ) are the values of the F-distribution with 𝑣𝑣1
2 2
733 (numerator) and 𝑣𝑣2 (denominator) degrees of freedom that has probability of α/2 and 1-
734 α/2 to its right, respectively.
735
736 2. Statistical Method for Highly Variable Drugs
737
738 If a drug is a high variable drug, a partial or fully replicated cross-over design should be used.
739 The statistical analysis should be carried out using the mixed scaling approach below for both
740 AUC and Cmax.
741
742 The mixed scaling approach:
743
744 If the estimated within-subject standard deviation of the RLD is < 0.294, the two one-sided test
745 procedure should be used to determine BE for the individual PK parameter. Otherwise, the
746 reference-scaled procedure should be used to determine BE for the individual PK parameter
747 together with a point estimate constraint for the estimated test/reference geometric mean ratio.
748
749 For the reference-scaled approach the upper BE limit for Test/Reference ratio of geometric
1
750 means is ∆ = 0.8 , the regulatory constant is 𝜎𝜎𝑤𝑤0 = 0.25 and the point estimate constraint is
751 80.00 to 125.00%.
752
753 Refer to the draft guidance for industry Bioequivalence Studies With Pharmacokinetic Endpoints
754 for Drugs Submitted Under an ANDA (August 2021) for further details. 41
755
40
41
20
756 C. Comparative Clinical Endpoint Bioequivalence Studies

757
758 For some products, the PSG may recommend an appropriately designed comparative clinical
759 endpoint BE study. In particular, a comparative clinical endpoint BE study is an option to be
760 considered for measuring BA or demonstrating BE of dosage forms intended to deliver the active
761 moiety locally, e.g., topical preparations for the skin, eye, and mucous membranes; oral dosage
762 forms not intended to be systemically absorbed, e.g., an antacid; bronchodilators administered by
763 oral inhalation.
764
765 In general, these studies will have a randomized, parallel group design, with three arms: test,
766 reference, and placebo/vehicle.
767
768 • A placebo/vehicle arm is recommended to demonstrate that the T product and R product
769 are active and to establish that the study is sufficiently sensitive to detect differences
770 between products at the lower end of the dose/response curve.
771
772 To establish BE, it is recommended that the following compound hypotheses (continuous
773 endpoint or dichotomous endpoint) be tested. Rejection of the null hypothesis supports the
774 conclusion of equivalence of the two products.
775
776 For a continuous endpoint:
777 The null hypothesis for this test is:
778
779 H0: μT /μR ≤ θ1 or μT /μR ≥ θ2
780
782 Ha: θ1 < μT /μR < θ2
783
784 where:
785 – μT = mean of the primary endpoint for the test group, and
786 – μR = mean of the primary endpoint for the reference group.
787
788 The null hypothesis, H0, is rejected with a Type I error (α) of 0.05 (two one-sided tests) if the
789 90% confidence interval for the ratio of the means between T and R products (μT /μR) is
790 contained within the interval [θ1, θ2].
791
792 For a dichotomous endpoint:
793 The null hypothesis for this test is:
794
795 H0: πT–πR≤ Δ1 or πT– πR ≥ Δ2
796
798 Ha: Δ1 < πT - πR < Δ2
799
21
800 where:
801 – πT= the success rate of the primary endpoint for the treatment group, and πR= the
802 success rate of the primary endpoint for the reference group.
803
804 The null hypothesis, H0, is rejected with a Type I error (α) of 0.05 (two one-sided tests) if the
805 estimated 90% confidence interval for the difference of the success rates between T and R
806 products (πT–πR) is contained within the interval [Δ1, Δ2].
807
808 • For continuous and binary endpoints, in order to demonstrate adequate study sensitivity,
809 the test product and reference product should both be statistically superior to placebo
810 (p<0.05) with regard to the primary endpoint.
811
812 • Refer to PSGs for comparative clinical endpoint BE study designs, definitions of study
813 populations, regulatory constant (e.g., equivalence interval limit), and analyses specific to
814 a given product.
815
816 D. Studies in Multiple Groups
817
818 There can be multiple sources of group 42 effects in BE studies. Sometimes, groups reflect
819 factors arising from study design and conduct. For example, a PK BE study can be carried out in
820 two or more clinical centers and the study may be considered a multi-group BE study. The
821 combination of multiple factors may complicate the designation of group. Therefore, sponsors
822 should minimize the group effect in a PK BE study as recommended below:
823
824 (1) Dose all groups at the same clinic unless multiple clinics are needed to enroll a
825 sufficient number of subjects.
826
827 (2) Recruit subjects from the same enrollment pool to achieve similar demographics
828 among groups.
829
830 (3) Recruit all subjects, and randomly assign them to group and treatment arm, at study
831 outset.
832
833 (4) Follow the same protocol criteria and procedures for all groups.
834
835 (5) When feasible (e.g., when healthy volunteers are enrolled), assign an equal sample
836 size to each group.
837
838 Bioequivalence should be determined based on the overall treatment effect in the whole study
839 population. In general, the assessment of BE in the whole study population should be done
840 without including the treatment and group interaction(s) term in the model, but applicants may
841 also use other pre-specified models, as appropriate (Fleiss 1986, Permutt 2003, Tsiatis et al.
42
In literature, the term group is sometimes referred to as subgroup.
22
842 2008). 43 The assessment of interaction between the treatment and group(s) is important,
843 especially if any of the first four study design criteria recommended above are not met and the
844 PK BE data are considered pivotal information for drug approval. If the interaction term of
845 group and treatment is significant (Alosh et al. 2015, Grizzle 1965), 44 heterogeneity of treatment
846 effect across groups should be carefully examined and interpreted with care. If the observed
847 treatment effect of the products varies greatly among the groups, vigorous attempts should be
848 made to find an explanation for the heterogeneity in terms of other features of trial management
849 or subject characteristics, which may suggest appropriate further analysis and interpretation.
850
851 It is important that statistical methods and models for the primary BE analysis are fully pre-
852 specified in the protocol or SAP (e.g., in an ANDA study, the applicant should pre-specify
853 detailed statistical criteria and models to be used if the interaction term of group and treatment is
854 applicable). In addition, the statistical model should reflect the multigroup nature of the study.
855 For example, if subjects are dosed in two groups in a crossover BE study, the model should
856 reflect the fact that the periods for the first group are different from the periods for the second
857 group, i.e., the period effect should be nested within the group effect.
858
859 When there are multiple centers with very few subjects in some centers and sponsors want to
860 combine centers in the analysis, any rules for combination should be pre-specified in the protocol
861 or SAP and a sensitivity analysis is recommended. More complicated scenarios may be
862 discussed with the appropriate CDER review division before submission.
863
864 E. Bioequivalence Statistics for Adhesion and Irritation Studies
865
866 In terms of the statistical method used in irritation, sensitization or/and adhesion studies for
867 Transdermal and Topical Delivery Systems, refer to the Statistical Consideration section in the
868 draft guidance for industry Assessing the Irritation and Sensitization Potential of Transdermal
869 and Topical Delivery Systems for ANDAs (October 2018) and the Considerations for Statistical
870 Analysis section in the draft guidance for industry Assessing Adhesion With Transdermal and
871 Topical Delivery Systems for ANDAs (October 2018). 45
872
873
874
43
Fleiss, J.L., 1986, Analysis of Data from Multiclinic Trials, Controlled Clinical Trials, 7(4):267-275;
Permutt, T., 2003, Probability Models and Computational Models for ANOVA in Multicenter Clinical Trials,
Journal of Biopharmaceutical Statistics, 13(3):495-505; Tsiatis, A.A., M. Davidian, M. Zhang, and X. Lu, 2008,
Covariate Adjustment for Two-Sample Treatment Comparisons in Randomized Clinical Trials: A Principled Yet
Flexible Approach, Statistics in Medicine, 27(23):4658-4677.
44
Alosh, M., K. Fritsch, M. Huque, K. Mahjoob, G. Pennello, M. Rothmann, E. Russek-Cohen, F. Smith, S. Wilson,
and L. Yue, 2015, Statistical Considerations on Subgroup Analysis in Clinical Trials, Statistics in Biopharmaceutical
Research, 7(4):286-303; Grizzle, J.E., 1965, The Two-Period Change-Over Design and Its Use in Clinical Trials,
Biometrics, 21(2):467-480.
45
See also the draft guidance for industry Assessment of Adhesion for Topical and Transdermal Systems Submitted in
New Drug Applications (July 2021). When final, these guidances will represent FDA’s current thinking on these
topics.
23
875 F. Dose Scale for Bioequivalence Assessment

876
877 In this method, the BE assessment is based on relative bioavailability of the test and reference
878 formulations at the site(s) of action. The relative bioavailability, F, is the ratio of the doses of
879 test and reference formulations that produce an equivalent PD response.
880
881 Generally, the F is estimated by fitting an Emax model that describes the within-study dose-
882 response relationship. Among available statistical methods for Emax model fitting, nonlinear
883 mixed effect (NLME) modeling is recommended, because the NLME modeling is capable of
884 characterizing between-subject variability and residual unexplained variability, and less sensitive
885 to aberrant observation and missing values.
886
887 For model fitting details, refer to the PSG on Orlistat oral capsule. 46
888
889 To determine BE, the 90% confidence interval for F can be estimated by a bootstrap procedure.
890 Each bootstrap estimation includes the calculation of F by fitting the selected model to a sample
891 dose-response data set, which is generated by resampling with replacement. To maintain the
892 correlation of observations within subject, resampling by subject (remaining observations from
893 all T and R treatment arms) is recommended rather than resampling by observations. The
894 Agency has also recommended using Efron's bias corrected and accelerated method to compute a
895 90% confidence interval for F. 47 Alternatively, the 90% confidence interval for F can be
896 estimated without a bootstrap procedure, directly from the point estimate of logF and its standard
897 error calculated using NLME modeling.
898
899 Given the complexity of dose scale analysis for comparative PD BE studies, applicants are
900 encouraged to contact the Agency early to discuss their proposed study designs and statistical
901 methods (e.g., alternative modeling approaches, impact of the missing data and the handling
902 strategy) via the controlled correspondence, pre-ANDA meeting, pre-IND meeting, or pre-NDA
903 meeting pathway. 48
904
905 G. Bioequivalence Studies Using Multiple References
906
907 In BE studies with more than two reference treatment arms (e.g., a three-period study including
908 two references, one from the European Union (EU) and another from the United States, or a
909 four-period study including test and reference in fed and fasted states), the BE determination
910 should be based on the comparison between the relevant test and reference products, using only
911 the data from those products. The BE analysis for this comparison should be conducted
912 excluding the data from the non-relevant treatment(s) — for example, in a BE study with a T
913 product, an EU reference product, and a U.S. reference product, the comparison of the T product
914 to the U.S. reference product should be based on an analysis excluding the data from the EU
915 reference. However, full data from the BE studies, including data comparing the T product that
46
Draft PSG for industry on Orlistat oral capsule (August 2021). When final, this guidance will represent FDA’s
current thinking on this topic.
47
Ibid.
48
24
916 is the subject of the application with non-U.S. reference products, should be submitted in the
917 application for completeness. The applicant may discuss the study design and statistical
918 approach with the appropriate CDER review division before study conduct.
919
920
25
921 V. APPENDICES
922
923 A. Choice of Specific Replicated Crossover Designs
924
925 Appendix A describes why FDA prefers replicated crossover designs with only two sequences,
926 and why the Agency recommends the specific designs described in section II.A.1.b of this
927 guidance.
928
929 1. Reasons Unrelated to Carryover Effects
930
931 Each unique combination of sequence and period in a replicated crossover design can be called a
932 cell of the design. For example, the two-sequence, four-period design recommended in section
933 II.A.1.b has eight cells. The four-sequence, four-period design below has 16 cells.
934
935 Period
936
937 1 2 3 4
938
939 1 T R R T
940
941 2 R T T R
942 Sequence
943 3 T T R R
944
945 4 R R T T
946
947 The total number of degrees-of-freedom attributable to comparisons among the cells is just the
948 number of cells minus one (unless there are cells with no observations).
949
950 The fixed effects that are usually included in the statistical analysis are sequence, period, and
951 treatment (i.e., formulation). The number of degrees-of-freedom attributable to each fixed effect
952 is generally equal to the number of levels of the effect, minus one. Thus, in the case of the two-
953 sequence, four-period design recommended in section V.A.1, there would be 2-1=1 degree-of-
954 freedom due to sequence, 4-1=3 degrees-of-freedom due to period, and 2-1=1 degree-of-freedom
955 due to treatment, for a total of 1+3+1=5 degrees-of-freedom due to the three fixed
956 effects. Because these 5 degrees-of-freedom do not account for all 7 degrees-of-freedom
957 attributable to the eight cells of the design, the fixed-effects model is not saturated. There could
958 be some controversy as to whether a fixed-effects model that accounts for more or all of the
959 degrees-of-freedom due to cells (i.e., a more saturated fixed-effects model) should be used. For
960 example, a sequence-by-period-by-treatment interaction effect might be included, which would
961 fully saturate the fixed-effects model.
962
963 If the replicated crossover design has only two sequences, use of only the three main effects
964 (sequence, period, and treatment) in the fixed-effects model or use of a more saturated model
965 makes little difference to the results of the analysis, provided there are no missing observations,
25
966 and the study is carried out in one group of subjects. The least squares point estimate of 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅
967 will be the same for the main-effects model and for the saturated model.
968
969 If the replicated crossover design has more than two sequences, these advantages are no longer
970 present. Main-effects models will generally produce different point estimates of 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 than
971 saturated models (unless the number of subjects in each sequence is equal), and there is no well-
972 accepted basis for choosing between these different estimates (though 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 from the
973 saturated model was determined to be appropriate for use in the reference-scaled average BE
974 assessment). Thus, use of designs with only two sequences minimizes or avoids certain
975 ambiguities due to specific choices of fixed effects to be included in the statistical model.
976
977 2. Reasons Related to Carryover Effects
978
979 One of the reasons to use the four-sequence, four-period design described above is that it is
980 thought to be optimal if carryover effects are included in the model.
981
982 Similarly, the two-sequence, three-period design is thought to be optimal among three-period
983 replicated crossover designs. Both of these designs are strongly balanced for carryover effects,
984 meaning that each treatment is preceded by each other treatment and itself an equal number of
985 times.
986
987 Period
988
989 1 2 3
990
991 1 T R R
992 Sequence
993 2 R T T
994
995 With these designs, no efficiency is lost by including simple first-order carryover effects in the
996 statistical model. However, if the possibility of carryover effects is to be considered in the
997 statistical analysis of BE studies, the possibility of direct-by-carryover interaction should also be
998 considered. If direct-by-carryover interaction is present in the statistical model, these favored
999 designs are no longer optimal. Indeed, the TRR/RTT design does not permit an unbiased within-
1000 subject estimate of 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 in the presence of general direct-by-carryover interaction.
1001
1002 The issue of whether a purely main-effects model or a more saturated model should be specified,
1003 as described in the previous section, also is affected by possible carryover effects. If carryover
1004 effects, including direct-by-carryover interaction, are included in the statistical model, these
1005 effects will be partially confounded with sequence-by-treatment interaction in four-sequence or
1006 six-sequence replicated crossover designs, but not in two-sequence designs.
1007
1008 In the case of the four-period and three-period designs recommended in section II.A.1.b, the
1009 estimate of 𝜇𝜇 𝑇𝑇 − 𝜇𝜇𝑅𝑅 , adjusted for first-order carryover effects, including direct-by-carryover
26
1010 interaction, is as efficient or more efficient than for any other two-treatment replicated crossover
1011 designs.
1012
1013 3. Two-Period Replicated Crossover Designs
1014
1015 For most drug products, two-period replicated crossover designs such as the Balaam design
1016 (which uses the sequences TR, RT, TT, and RR) should be avoided. However, the modified
1017 Balaam design (TR, RT, RR) may be useful for particular drug products (e.g., a long half-life
1018 drug for which a two-period study would be feasible, but a three-or-more-period study would
1019 not) when reference-scaled average BE is needed.
1020
1021 B. Rationale for Logarithmic Transformation of Pharmacokinetic Data
1022
1023 1. Clinical Rationale
1024
1025 The FDA Generic Drugs Advisory Committee recommended in 1991 that the primary comparison of
1026 interest in a BE study is the ratio, rather than the difference, between average PK parameter data from
1027 the T and R formulations. Using logarithmic transformation, the general linear statistical model
1028 employed in the analysis of BE data allows inferences about the difference between the two means on
1029 the log scale, which can then be retransformed into inferences about the ratio of the two averages
1030 (geometric means) on the original scale. Logarithmic transformation thus achieves a general
1031 comparison based on the ratio rather than the differences.
1032
1033 2. Pharmacokinetic Rationale
1034
1035 Westlake observed that a multiplicative model is postulated for PK measures in BA/BE studies (i.e.,
1036 AUC and Cmax, but not Tmax) (Westlake 1973 and 1988).49,50 Assuming that elimination of the drug is
1037 first order and only occurs from the central compartment, the following equation holds after an
1038 extravascular route of administration:
1039
1040 AUC0-∞= F*D/CL
1041
1042 = F*D/(V*Ke)
1043
1044 where F is the fraction absorbed, D is the administered dose, and F*D is the amount of drug absorbed.
1045 CL is the clearance of a given subject that is the product of the apparent volume of distribution (V) and
1046 the elimination rate constant (Ke). The use of AUC as a measure of the amount of drug absorbed
1047 involves a multiplicative term (CL) that might be regarded as a function of the subject. For this reason,
49
Westlake, W. J., 1973, The Design and Analysis of Comparative Blood-Level Trials, J. Swarbick, editor, Current
Concepts in the Pharmaceutical Sciences, Dosage Form Design and Bioavailability, Philadelphia: Lea and Febiger,
149-179.
50
Westlake, W. J., 1988, Bioavailability and Bioequivalence of Pharmaceutical Formulations, Biopharmaceutical
Statistics for Drug Development, 329-352.
27
1048 Westlake contends that the subject effect is not additive if the data are analyzed on the original scale of
1049 measurement.
1050
1051 Logarithmic transformation of the AUC data will bring the CL (i.e., V*Ke) term into the following
1052 equation in an additive fashion:
1053
1054 lnAUC0-∞ = ln F + ln D - ln V - ln Ke
1055
1056 Similar arguments were given for Cmax. The following equation applies for a drug exhibiting one
1057 compartmental characteristic:
1058
1059 Cmax = (F*D/V) * exp(-Ke*Tmax)
1060
1061 where again F, D and V are introduced into the model in a multiplicative manner. However, after
1062 logarithmic transformation, the equation becomes:
1063
1064 lnCmax = ln F + ln D - lnV – Ke*Tmax
1065
1066 Thus, log transformation of the Cmax data also results in the additive treatment of the V term.
1067
1068 C. SAS Program Statements for Average Bioequivalence Analysis of Replicated
1069 Crossover Studies
1070
1071 The following illustrates an example of program statements to run the unscaled average BE
1072 analysis using PROC MIXED in SAS version 9, with SEQ, SUBJ, PER, and TRT identifying
1073 sequence, subject, period, and treatment variables, respectively, and Y denoting the response
1074 measure (e.g., log (AUC), log (Cmax)) being analyzed:
1075
1076 PROC MIXED;
1077 CLASSES SEQ SUBJ PER TRT;
1078 MODEL Y = SEQ PER TRT/ DDFM=SATTERTH;
1079 RANDOM TRT/TYPE=FA0(2) SUB=SUBJ G;
1080 REPEATED/GRP=TRT SUB=SUBJ;
1081 ESTIMATE 'T vs. R' TRT 1 -1/CL ALPHA=0.1;
1082
1083 The Estimate statement assumes that the code for the test formulation precedes the code for the
1084 reference formulation in sort order (this would be the case, for example, if T were coded as 1 and
1085 R were coded as 2). If the R code precedes the T code in sort order, the coefficients in the
1086 Estimate statement would be changed to -1 1.
1087
1088 In the Random statement, TYPE=FA0(2) could possibly be replaced by TYPE=CSH or UNR.
1089
1090 In the Model statement, DDFM=SATTERTH could possibly be replaced by DDFM=KR2.
1091 However, the detailed model specification should be pre-specified in the protocol or SAP and
1092 data driven post hoc selection of the model is not allowed.
28
1093
1094 Additions and modifications to these statements can be made if the study is carried out in more
1095 than one group of subjects or other complicated scenarios. Alternative software could also be
1096 used if same results are generated as in PROC MIXED in SAS.
29

Fda On Statistical Approaches

Uploaded by

Copyright:

Available Formats

Fda On Statistical Approaches

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fda On Statistical Approaches

Uploaded by

Copyright:

Available Formats

What is the purpose of this document?

What statistical models are discussed for analyzing bioequivalence data?

Statistical Approaches

This guidance document is being distributed for comment purposes only.

U.S. Department of Health and Human Services

Additional copies are available from:

U.S. Department of Health and Human Services

1 Statistical Approaches to Establishing Bioequivalence

292 b. Data transformation for comparative pharmacodynamic and clinical

500 III. SPECIFIC SITUATIONS 24

(ln 1.11)2 + 0.01

𝑠𝑠𝑤𝑤𝑤𝑤 /𝑠𝑠𝑤𝑤𝑤𝑤 𝑠𝑠𝑤𝑤𝑤𝑤 /𝑠𝑠𝑤𝑤𝑤𝑤

756 C. Comparative Clinical Endpoint Bioequivalence Studies

875 F. Dose Scale for Bioequivalence Assessment

You might also like