martin ★★ Austria, 20180608 09:15 (1062 d 15:23 ago) Posting: # 18867 Views: 2,788 

Dear all, I am facing an interesting BE study and would be happy to get the forum members thoughts on the analysis strategy Background: one reference (R) and two test items (T1, T2) with different doses per body weight stratum Body weight stratum 1: x mg for T1, T2 and R for subjects < z kg Body weight stratum 2: y mg for T1, T2 and R for subjects > z kg Study design: Williams’ design (3treatment, 3period, 6sequence) with stratified randomization according to body weight stratum (<z kg and > z kg) to ensure balance regarding this strata. Questions: 1) Can the BE assessment be based on all data combined or does this require a separate analysis per body weight stratum 2) If a combined assessment is the way to go how should the model look like a) Ignoring body weight strata and use classical model (FDA): fixed effects for period, sequence, treatment and random effect for subject nested in sequence. b) Including body weight strata as mentioned in ICH E9 (i.e. factors on which randomization has been stratified should be accounted for later in the analysis): fixed effects for period, sequence, treatment, body weight strata and random effect for subject nested in sequence. Best regards and looking forward to an interesting discussion Martin Edit: Category changed; see also this post #1. [Helmut] 
Helmut ★★★ Vienna, Austria, 20180608 11:04 (1062 d 13:34 ago) @ martin Posting: # 18868 Views: 2,446 

Hi Martin, » Body weight stratum 1: x mg for T1, T2 and R for subjects < z kg » Body weight stratum 2: y mg for T1, T2 and R for subjects > z kg » Study design: Williams’ design (3treatment, 3period, 6sequence) with stratified randomization according to body weight stratum (<z kg and > z kg) to ensure balance regarding this strata. Modifying this code:
» 1) Can the BE assessment be based on all data combined or does this require a separate analysis per body weight stratum Strictly speaking BE is a comparison of at the same molar dose. Hence, separate? » 2) If a combined assessment is the way to go how should the model look like » a) Ignoring body weight strata and use classical model (FDA): fixed effects for period, sequence, treatment and random effect for subject nested in sequence. » b) Including body weight strata as mentioned in ICH E9 (i.e. factors on which randomization has been stratified should be accounted for later in the analysis): fixed effects for period, sequence, treatment, body weight strata and random effect for subject nested in sequence. If you go with a) what is the purpose of the stratification? In b) the stratum is a betweengroup term anyhow and I guess that the residual error will be the same as in a). — Diftor heh smusma 🖖 Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes 
d_labes ★★★ Berlin, Germany, 20180608 17:04 (1062 d 07:34 ago) @ martin Posting: # 18875 Views: 2,393 

Dear Martin, » ... » Questions: » » 1) Can the BE assessment be based on all data combined or does this require a separate analysis per body weight stratum » » 2) If a combined assessment is the way to go how should the model look like » » a) Ignoring body weight strata and use classical model (FDA): fixed effects for period, sequence, treatment and random effect for subject nested in sequence. » » b) Including body weight strata as mentioned in ICH E9 (i.e. factors on which randomization has been stratified should be accounted for later in the analysis): fixed effects for period, sequence, treatment, body weight strata and random effect for subject nested in sequence. Additionally to what Helmut wrote: If you will use a Proc GLM or lm() for the combined assessment you will face a confounding between subject effects and bodyweightstrata effects and the analysis code will show you the finger. Something like df=0 will happen, at least for type III tests of effects.You have to include at least a treatment by bodyweightstrata interaction to get meaningfull results. I'm not really sure if this confounding is also an issue if you plan to use Proc MIXED or lme()/lmer() . Make a example data set and try it.I suggest that you modify the FDA code for logistic groups (see f.i. this post) accordingly (change group to bodyweightstrata and drop Period(nested within Group)). And read Helmut's lectures about "MultiGroup Studies in BE. To pool or not to pool?". All the criticism regarding the group effects apply also to bodyweightstrata effects. The treatment effect (diff in the log domain, ratio on the original scale) in xover studies is determined intrasubject. Thus all subject characteristics constant over the study can not influence it really. If you are interested in having a look at BE assesment (ratio & CI) for the different bodyweightstrata you have to go with 1) anyway. BTW: "factors on which randomization has been stratified should be accounted for" should read the other way round: randomize stratified for factors planned to account for in the analysis. — Regards, Detlew 
zizou ★ Plzeň, Czech Republic, 20180608 21:11 (1062 d 03:27 ago) (edited by zizou on 20180609 03:21) @ martin Posting: # 18876 Views: 2,365 

Dear Martin, additionally to what Helmut and Detlew wrote: » 1) Can the BE assessment be based on all data combined or does this require a separate analysis per body weight stratum Maybe easier way would be to separate the analysis as 2 studies:  for subjects with weight <z kg,  for subjects with weight >z kg. Pro: standard evaluation (2x) Con: approx. double number of subjects required (than in option 2)) » 2) If a combined assessment is the way to go how should the model look like ... I would be inspired by FDA model with Groups (as Detlew pointed). Then it is an interesting question which factors should be included in the model: Maybe:
Possible option would be to use dose correction and then evaluation using the standard model. Though I realized that there is the same dose correction for T1,T2,R in the half of the subjects and similarly in the second half ... so the ratios will stay unchanged regardless the correction. Pro: Less subjects required in total (if you are not interested in separate analysis) Con: model , varied questions may be arised Best regards, zizou 