libaiyi ★ China, 20180626 12:00 (edited by Ohlbe on 20180626 12:24) Posting: # 18964 Views: 5,348 

Hi,all I am curious of the confidence interval decision of three treatment, three sequences, cross over BE study as below, if the goal is to prove the equivalence of any two of them (T1 and R or T2 and R). Will the CI still determined as 95% or 90%? Sequence/Period 1 2 3 I also want to know how to select the suitable design between latin squares and William design when there exists three treatment, R T1 T2. Thanks in advance. Edit: two posts merged. You can edit your message up to 24 hours after posting it [Ohlbe] 
Helmut ★★★ Vienna, Austria, 20180626 13:13 @ libaiyi Posting: # 18967 Views: 4,917 

Hi libaiyi, » […] the goal is to prove the equivalence of any two of them (T1 and R or T2 and R). Will the CI still determined as 95% or 90%? If this is a pivotal study (say you want to demonstrate BE of a capsule (T1) and a tablet (T2) to R which is either a tablet or a capsule) IMHO, you should employ Bonferroni’s 95% which preserves the familywise error rate at 1(10.05/2)^{2}=4.9375%. Reason: When both products are appoved, based on ABE a patient my switch from R to T1 or from R to T2. Slightly off topic: Another story would be one test and two references of different regions. If we ignore tourists, we have two different populations of patients. Then we don’t have to adjust α and go with the 90% CI. » I also want to know how to select the suitable design between latin squares and William design when there exists three treatment, R T1 T2. That’s a matter of taste. Williams’ designs are variancebalanced even for carryover (which is not part of the model). Hence, some people (see there) prefer them over Latin Squares. BTW, the EMA’s GL states the sixsequence Williams’ design for the story mentioned above. More important than the design is the evaluation (see this post). — Cheers, Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes 
ElMaestro ★★★ Belgium?, 20180626 22:20 @ libaiyi Posting: # 18969 Views: 4,860 

Hi libaiyi, » I am curious of the confidence interval decision of three treatment, three sequences, cross over BE study as below, if the goal is to prove the equivalence of any two of them (T1 and R or T2 and R). Will the CI still determined as 95% or 90%? Due to the little word highlighted in red above I am fairly sure I would not apply any other alpha than 5%, meaning I would go with the usual 90% CIs. — Le tits now. Best regards, ElMaestro 
Helmut ★★★ Vienna, Austria, 20180627 00:33 @ ElMaestro Posting: # 18970 Views: 4,860 

Hi ElMaestro, » » […] if the goal is to prove the equivalence of any two of them (T1 and R or T2 and R). » » Due to the little word highlighted in red above I am fairly sure I would not apply any other alpha than 5%, meaning I would go with the usual 90% CIs. Good point – I missed that! In the past I did not adjust because only one will be marketed (i.e., the entire patient’s risk lies with this product). However, some regulatory statisticians (you know, coffee break chats) seem to be more strict because one gets two chances of passing. Too lazy to crawl the supplementary material of your paper about pilot studies: Is such a case covered? — Cheers, Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes 
libaiyi ★ China, 20180627 08:25 @ Helmut Posting: # 18971 Views: 4,801 

Hi Helmut Sorry I am still a little confused, do you mean that if I try to test any of two tests and one reference (T1=R or T2=R), the CI should still be set as 90%, and if I want to test T1=T2=R, CI should be set as 95%? Thank you so much. 
ElMaestro ★★★ Belgium?, 20180627 08:29 @ Helmut Posting: # 18972 Views: 4,808 

Good morning Hötzi, » Good point – I missed that! In the past I did not adjust because only one will be marketed (i.e., the entire patient’s risk lies with this product). However, some regulatory statisticians (you know, coffee break chats) seem to be more strict because one gets two chances of passing. Too lazy to crawl the supplementary material of your paper about pilot studies: Is such a case covered? Whose paper? Some crackpot had one published paper and it covered the case of one product being tested in a pilot, then in a pivotal trial, depending on the figures. Rumours have it he is working on another manuscript dealing with scenarios like the one in this thread and that reviewers are way too tough. But then again, you cannot always trust rumours. Only trust google and wikipedia. Hi all, ok you'll think that I am nuts, but I think I will need to put it into reverse here: Let us say both R1 and R2 are (truly) not BE. For simplicity think of them as products with the exact same properties even though they formally may be called something different. Since we are happy if we approve one of them, applying alpha=5% could inflate the type I error (approval of a nonBE product) if the true relative performances are around the acceptance borders. We need that alpha correction. Sorry. Edit: Two posts merged. [Helmut] — Le tits now. Best regards, ElMaestro 
Helmut ★★★ Vienna, Austria, 20180627 11:45 @ ElMaestro Posting: # 18974 Views: 4,787 

Hi ElMaestro, » […] I think I will need to put it into reverse here: » Let us say both R1 and R2 are (truly) not BE… » » We need that alpha correction. Sorry. Haha, this time you answered a question not asked. You want to get both T1 and T2 approved. T1 = R ∧ T2 = R: ⇒ 95% CI Clearly needs an adjustment because of the simultaneous tests and both T1 and T2 will be marketed.You want to get either T1 or T2 approved. T1 = R ∨ T2 = R: ⇒ 95% CI In this case I argued for no adjustment (90% CI) in the past but nowadays I’m leaning towards an adjustment because which one will be marketed is part of a decision tree.— Cheers, Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes 
Ohlbe ★★★ France, 20180627 12:14 @ Helmut Posting: # 18975 Views: 4,804 

Dear all, » You want to get either T1 or T2 approved. T1 = R ∨ T2 = R: ⇒ 95% CI In this case I argued for no adjustment (90% CI) in the past but nowadays I’m leaning towards an adjustment because which one will be marketed is part of a decision tree.Isn't the common practice to first test T1 and T2 vs. R in a pilot study (and I will not get into a debate on whether alpha correction would be needed), select one of the two test formulations, and then test it in a pivotal study in the usual way ? How often do you see both test formulations directly tested in a pivotal study ? OK, maybe if the product has a very low variability and the number of subjects remains very low, it could make sense and save time to directly go for the pivotal. Any experience ? — Regards Ohlbe 
ElMaestro ★★★ Belgium?, 20180627 13:01 @ Ohlbe Posting: # 18977 Views: 4,827 

Hello Ohlbe » Isn't the common practice to first test T1 and T2 vs. R in a pilot study (and I will not get into a debate on whether alpha correction would be needed), select one of the two test formulations, and then test it in a pivotal study in the usual way ? How often do you see both test formulations directly tested in a pivotal study ? That's right. I have never seen an evaluation of two tests in a single study in which the applicant hoped for approval of both. » OK, maybe if the product has a very low variability and the number of subjects remains very low, it could make sense and save time to directly go for the pivotal. Any experience ? Yes absolutely. The majority of dossiers are being accepted without pilot trials. Whether to go for a pilot before a pivotal is dictated by many factors. First to file is a very significant one, but also involved is BCS class, budgets, previous failures ("You guys need to start showing some results!"), gut feeling and much more. However I get slightly worried when you are suddenly allowing the term sense to enter this discussion. My empirical observation is that that term has absolutely no place in the planning of BE studies. — Le tits now. Best regards, ElMaestro 
Ohlbe ★★★ France, 20180627 14:13 @ ElMaestro Posting: # 18978 Views: 4,730 

Dear ElMaestro, » » OK, maybe if the product has a very low variability and the number of subjects remains very low, it could make sense and save time to directly go for the pivotal. Any experience ? » Yes absolutely. The majority of dossiers are being accepted without pilot trials. Sure, when you only have one test formulation. But when you hesitate between two ? Would you directly go for a pivotal where both test formulations would be compared to the reference, and then file whichever one passes ? — Regards Ohlbe 
ElMaestro ★★★ Belgium?, 20180627 14:48 @ Ohlbe Posting: # 18979 Views: 4,733 

Hi Ohlbe, » Sure, when you only have one test formulation. But when you hesitate between two ? Would you directly go for a pivotal where both test formulations would be compared to the reference, and then file whichever one passes ? This is purely hypothetical for me. As far as I recall, I have not ever faced such a situation where two test formulations targeting the same authority were headson compared in a pivotal trial. I have faced it numerous times in pilot trials, but here the choice of formulation rarely is based on the CI but just on the observed GMR. If I were in a situation, hypothetically, where I had one pivotal trail with two passing T formulations against a single, and I only needed one approval, then I am not sure what I would base my choice on. It could be cost of goods, or a riskbased assessment of the manufacturing bottlenecks for the two T formulations. Not sure, I have a feeling this would not be sciencedriven but GuyinArmanisuitdriven. — Le tits now. Best regards, ElMaestro 
Ohlbe ★★★ France, 20180627 15:43 @ ElMaestro Posting: # 18981 Views: 4,737 

Hi ElMaestro, » This is purely hypothetical for me. As far as I recall, I have not ever faced such a situation where two test formulations targeting the same authority were headson compared in a pivotal trial. I have faced it numerous times in pilot trials, but here the choice of formulation rarely is based on the CI but just on the observed GMR. Precisely. But then, isn't this whole discussion on alpha correction purely hypothetical too ? Sorry if I'm spoiling a nice thread... — Regards Ohlbe 
ElMaestro ★★★ Belgium?, 20180627 22:13 @ Ohlbe Posting: # 18982 Views: 4,724 

Hi Ohlbe, » But then, isn't this whole discussion on alpha correction purely hypothetical too ? as I see it, it is a mild case of testing into compliance if we don't correct alpha: If T1 and T2 are both borderline bad (like GMR=0.8) then we will have increased prob >5% type I error) of approving one of them, i.e. approving an unapprovable product. — Le tits now. Best regards, ElMaestro 
jag009 ★★★ NJ, 20180629 20:53 @ Ohlbe Posting: # 19000 Views: 4,634 

Hi Olhbe » Isn't the common practice to first test T1 and T2 vs. R in a pilot study (and I will not get into a debate on whether alpha correction would be needed), select one of the two test formulations, and then test it in a pivotal study in the usual way ? How often do you see both test formulations directly tested in a pivotal study ? » » OK, maybe if the product has a very low variability and the number of subjects remains very low, it could make sense and save time to directly go for the pivotal. Any experience ? Damn. We did pivotals 4way crossover 3T vs R before! R&D said no time and material was cheap! I said "Ooooookay..." John 
ElMaestro ★★★ Belgium?, 20180629 22:11 @ jag009 Posting: # 19002 Views: 4,567 

Hello John, » Damn. We did pivotals 4way crossover 3T vs R before! R&D said no time and material was cheap! I said "Ooooookay..." How many of the T's turned out BE? I hope the CRO awarded you the title of "best client ever" afterwards. By the way and quite OT, I will be in Parsippany in July. That's still your whereabouts? Are you in office week of July 16 and does your coffee machine work ? — Le tits now. Best regards, ElMaestro 
jag009 ★★★ NJ, 20180703 17:38 @ ElMaestro Posting: # 19014 Views: 4,535 

Hi ElMaestro, » How many of the T's turned out BE? I hope the CRO awarded you the title of "best client ever" afterwards. Pivotals. The 2 4way 3T vs R studies turned out great as 2 out of 3 formulations made BE. I said great so which one are we going forward with and they said "the one which gives us the least headache in manufacturing. Lol We did this for 2 products. Haha » By the way and quite OT, I will be in Parsippany in July. That's still your whereabouts? Are you in office week of July 16 and does your coffee machine work ? You are about 3045 mins drive from me since I am down south (30 mile). I will be in the office 7/18 and after. Shoot me an email and I will give you my contact. Would love to meet you in person! Thanks John 
d_labes ★★★ Berlin, Germany, 20180628 19:48 (edited by d_labes on 20180628 23:33) @ Helmut Posting: # 18991 Views: 4,721 

Dear Helmut, dear other DiskuTanten, » You want to get both T1 and T2 approved. T1 = R ∧ T2 = R: ⇒ 95% CI Clearly needs an adjustment because of the simultaneous tests and both T1 and T2 will be marketed.I'm not quite sure if you are right here. Consider the analogy with the "simultaneous" tests of AUC and Cmax. I haven't seen any alpha correction for that case . The IUT (intersection union test) principle protects us. But can be very conservative. IMHO we don't need an alpha correction for your case above, if we combine the tests with 'and'. If we combine with 'or', e.g. any of the test products T1 or T2 is choosen for further development or for approval, than we have to split the familywise level of alpha = 0.05 and the two individual null hypotheses have to be tested at a comparisonwise type I error which is a fraction of alpha, e.g. according to the Bonferroni procedure. Means 95% CIs instead of 90% CIs in case of T1, T2 and R. Reference (wording of my post partly copied from that Reference): Hauschke D., Steinijans V.W., Pigeot I. Bioequivalence Studies in Drug Development: Methods and Applications. New York: Wiley; 2007. Chapter 7 Designs with more than two formulations / 7.4 Multiplicity Seems a case for sims if we don't trust in the Reference. Isn't it? Eventually we can missuse power.2TOST() if we consider for simplicity two simultaneous test of AUC_T1R and AUC_T2R as the two metrics, which are usually interpreted as two different metrics, e.g. AUC and Cmax, respectively.Then of course the question of the correlation parameter rho arises. How to choose? — Regards, Detlew 
Helmut ★★★ Vienna, Austria, 20180703 17:22 @ d_labes Posting: # 19013 Views: 4,496 

Dear Detlew, » IMHO we don't need an alpha correction for your case above, if we combine the tests with 'and'. THX for reminding me (not for the first time)… » Seems a case for sims if we don't trust in the Reference. Isn't it? In God we trust; all others must bring data. W. Edwards Deming » Eventually we can missuse power.2TOST() if we consider for simplicity two simultaneous test of AUC_T1R and AUC_T2R as the two metrics, which are usually interpreted as two different metrics, e.g. AUC and Cmax, respectively.» Then of course the question of the correlation parameter rho arises. How to choose? Not sure. Whereas two PK metrics are always correlated somehow, IMHO, here it may be different. One formulation doesn’t care how it behaves in relation to the other. Too lazy to browse through my pilot studies. I’m pretty confident that some kind of correlation will show up but I doubt that it will be causation. PK is relatively simple but the outcome of a PK study with different formulations depends on an awful lot of manufacturing / process parameters. I think that the safe way would be subjects simulations. Normally for the TIE we use a true GMR of one of the BElimits and add variability. How to do that in a higher order design? Say for one subject I will have R=1, T1=1.25±se, T2=1.25±se. Now what? — Cheers, Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes 
d_labes ★★★ Berlin, Germany, 20180704 13:28 @ Helmut Posting: # 19020 Views: 4,505 

Dear Detlew, » » IMHO we don't need an alpha correction for your case above, if we combine the tests with 'and'. » THX for reminding me (not for the first time)… Correct. But old belives die hard » » Eventually we can missuse power.2TOST() ...» » Then of course the question of the correlation parameter rho arises. How to choose? » Not sure. Whereas two PK metrics are always correlated somehow, IMHO, here it may be different. One formulation doesn’t care how it behaves in relation to the other... Sounds for me like rho=0 . — Regards, Detlew 