Helmut ★★★ Vienna, Austria, 2021-08-27 17:51 (1200 d 00:30 ago) Posting: # 22550 Views: 5,966 |
|
Dear all (and esp. ElMaestro), with great interest I read two recent papers.1,2 On the author posted:
One of the conclusions drawn is that if we want to evaluate the data with a mixed model without running into model/convergence issues then we need to specify the covariance matrix \(\small{V }\) without using \(\small{V = ZGZ^t + R}\). Even so there is still no recipe for deriving the confidence interval. Hence, the options with the mixed model are:
I absolutely agree with this statement. However, what can we do? We need the bloody confidence interval. At the end of the day we are left out in the rain if we are dealing with a partial replicate design and the study has to be evaluated for average bioequivalence (according to the ’s wishful thinking if \(\small{s_\textrm{wR}<0.294}\)). To obtain the variance components which are estimable and the point estimate is just not enough. As long as we can’t get the standard error of the treatment difference and its associated degrees of freedom, we are at a loss. Hence, I repeat my mantra: Avoid the partial replicate design Of note: I used the supplementary script and got for
Pretty close to the modified script we discussed last year (given at the end). Much faster than mine.
########################### — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
ElMaestro ★★★ Denmark, 2021-08-28 00:18 (1199 d 18:04 ago) @ Helmut Posting: # 22552 Views: 4,633 |
|
Hi Hötzi, ❝ We need the bloody confidence interval. ❝ As long as we can’t get the standard error of the treatment difference and its associated degrees of freedom, we are at a loss. If we look at SAS documentation e.g. here and here, then
However, since FDA now threw the mixed model out for semi-replicate designs, interesting may in reality just mean academically attractive but possible not too useful at this moment. But who knows - until someone does it and explores it, it is not known if there is an advantageous property hidden somewhere. — Pass or fail! ElMaestro |
Helmut ★★★ Vienna, Austria, 2021-08-28 01:54 (1199 d 16:28 ago) @ ElMaestro Posting: # 22553 Views: 4,823 |
|
Hi ElMaestro, ❝ If we look at SAS documentation e.g. here and here, then ❝ 1. I do not understand the underlying math (just try and look into the spectral decomposition things mentioned as part of DDFM derivation; it sends my head spinning). Wooaah!! ❝ 2. Whatever SAS does, it seems to require that we work in V through Seems so. ❝ It would be interesting to see someone who has solid understanding of matrix theory and stats to work out the equations when the between and within-variability for T cannot be separated. Definitely not me… ❝ However, since FDA now threw the mixed model out for semi-replicate designs, … What on earth gives you this impression? I’m talking about ABE (Stats guidance Appendix E, Progesterone guidance page 8, ANDA guidance page 29): The following codes are an example of the determination of unscaled average BE for LAUCT with a partially replicate 3-way BE design: ❝ … interesting may in reality just mean academically attractive but possible not too useful at this moment. Nope. Talk to John about his data sets which didn’t converge. Given, FA0(1) did – with a warning – but I think there is a data set posted at Certara’s forum where nothing worked. Study done, no result. THX, .❝ But who knows - until someone does it and explores it, it is not known if there is an advantageous property hidden somewhere. Might be. — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
ElMaestro ★★★ Denmark, 2021-08-28 12:00 (1199 d 06:22 ago) @ Helmut Posting: # 22554 Views: 4,606 |
|
Hi Hötzi, ❝ ❝ However, since FDA now threw the mixed model out for semi-replicate designs, … ❝ ❝ What on earth gives you this impression? I’m talking about ABE (Stats guidance Appendix E, Progesterone guidance page 8, ANDA guidance page 29): Wow, that is actually right. Page 27: "PROC GLM should be used for partially replicate (3-way) BE studies" Page 29: "For PK parameters with a sWR < 0.294, use the unscaled average BE approach." (which implies PROC MIXED in their code) So, we must start to work out sWR using the equations. Once we got it, we know whether we need a mixed model or not for the evaluation of BE. We may need the mixed model for one metric, like AUCt and AUCinf, and the equations (ref scaled ABE) for another, like Cmax. I guess this is also the gist of the decision tree you gave. I wonder if this is truly their intention. The mixed model therefore seems to have been saved by the bell. This really emphasizes the need for a CI solution that allows V estimated without the existence of Z and R separately. I am curious now - we should make a little study on the type I error in case of missing ref values. — Pass or fail! ElMaestro |
Helmut ★★★ Vienna, Austria, 2021-08-28 12:38 (1199 d 05:44 ago) @ ElMaestro Posting: # 22555 Views: 4,677 |
|
Hi ElMaestro, ❝ ❝ What on earth gives you this impression? I’m talking about ABE (Stats guidance Appendix E, Progesterone guidance page 8, ANDA guidance page 29): ❝ ❝ Wow, that is actually right. ❝ Page 27: "PROC GLM should be used for partially replicate (3-way) BE studies" And Page 32:
❝ Page 29: "For PK parameters with a sWR < 0.294, use the unscaled average BE approach." (which implies PROC MIXED in their code) Now you got it. Some stuff there. ❝ So, we must start to work out sWR using the equations. Yep. At least that’s easy. ❝ Once we got it, we know whether we need a mixed model or not for the evaluation of BE. We may need the mixed model for one metric, like AUCt and AUCinf, and the equations (ref scaled ABE) for another, like Cmax. Correct. ❝ I guess this is also the gist of the decision tree you gave. I wonder if this is truly their intention. That’s the interpretation of all authors who assessed the Type I Error. Sent you Mehl. ❝ The mixed model therefore seems to have been saved by the bell. ? ❝ This really emphasizes the need for a CI solution that allows V estimated without the existence of Z and R separately. Happy that finally I conveyed the message. ❝ I am curious now - we should make a little study on the type I error in case of missing ref values. Oh dear, reaching for the stars! So far we have no clue how to come up with a solution for complete data. Appetizer. Since there were comments on about study costs, see there. — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
Mahmoud ★ Jordan, 2021-09-17 15:31 (1179 d 02:51 ago) @ Helmut Posting: # 22580 Views: 4,010 |
|
Dear all Dr Anders Fuglsang discussed in his paper Mitigation of the convergence issues associated with semi-replicated bioequivalence data. Pharmaceutical Statistics. 2021;1–3 used the proc mixed in RTR/TRR/RRT design in ABE I think that using proc mixed to fit RTR/TRR/RRT design is not suitable, because in most cases proc mixed produced Estimated G matrix is not positive definite. So proc GLM recommended by FDA and Emea is more appropriate in the Semi-replicated designs M.Youseef Edit: Post moved. [Helmut] |
Helmut ★★★ Vienna, Austria, 2021-09-17 17:15 (1179 d 01:06 ago) @ Mahmoud Posting: # 22581 Views: 4,062 |
|
Hi Mahmoud, ❝ I think that using proc mixed to fit RTR/TRR/RRT design is not suitable, because in most cases proc mixed produced Estimated G matrix is not positive definite. ❝ ❝ So proc GLM recommended by FDA and Emea is more appropriate in the Semi-replicated designs Please read the entire thread again. The FDA does recommend PROC MIXED for ABE in all guidances.That’s the problem. — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
Mahmoud ★ Jordan, 2021-09-17 17:24 (1179 d 00:58 ago) @ Helmut Posting: # 22582 Views: 4,103 |
|
❝ Please read the entire thread again. The FDA does recommend ❝ That’s the problem. Dear Dr Helmut Draft Guidance on Progesterone If SAS® is used for statistical analysis – PROC MIXED should be used for fully replicated (4-way) BE studies – PROC GLM should be used for partially replicated (3-way) BE studies or other software accomplishes same objectives Edit: Full quote removed. Please delete everything from the text of the original poster which is not necessary in understanding your answer; see also this post #5! [Helmut] |
Helmut ★★★ Vienna, Austria, 2021-09-17 19:11 (1178 d 23:11 ago) @ Mahmoud Posting: # 22584 Views: 4,247 |
|
Dear Mahmoud! ❝ Dear Dr Helmut Not a doctor. No intentions to become one. Told you that numerous times before. ❝ If SAS® is used for statistical analysis ❝ PROC MIXED should be used for fully replicated (4-way) BE studies ❝ PROC GLM should be used for partially replicated (3-way) BE studies Since you didn’t bother to read the entire thread as suggested – and not even the subject line of my post – would you mind to digest this one? Your quote refers to Reference-scaled Average Bioequivalence (\(\small{s_\textrm{wR}\geq 0.294}\)). However, for (conventional, unscaled) Average Bioequivalence (\(\small{s_\textrm{wR}<0.294}\)) PROC MIXED should be used regardless the design.— Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |