M.tareq
☆    

2019-07-21 21:09
(1711 d 19:54 ago)

Posting: # 20418
Views: 2,921
 

 REML, ML and MM different BE decision [RSABE / ABEL]

Dear all,

I've a question about the different estimation methods used for deriving BE in standard cross-over design.

Usually the BE is analysed using MM (Proc GLM in SAS) as per EMA (all factors as fixed effects) Q&A or REML as mentioned in FDA guidance.

Although when running a BE study and analyzing the data using MM/REML I get similar CI -we have a one drop out- ran the analysis with and without the subject and BE decision still the same. When using MM while including the drop-out using REML the BE decision changes.

the PK parameter of interest is the Cmax and it's outside the AL by 0.4

although when analyzing the data using ML, with and without the drop-out the BE is met in either cases.

I did some search, there's some paper indicating that the estimation using MM in unbalanced design might lead to biased variance estimates, although, in our case, the number of subjects on reference product is the same as the one on test product.

The drug is known to have low intra-subject variability (22%)

What are the assumption to use each method and when to do so ?

any insights on these discrepancies between the methods ? violation of assumptions?

Thanks

Mahmoud
Helmut
★★★
avatar
Homepage
Vienna, Austria,
2019-07-22 19:19
(1710 d 21:44 ago)

@ M.tareq
Posting: # 20426
Views: 2,258
 

 REML, ML and MM different BE decision

Dear Mahmoud,

not sure whatyou mean by “MM”. Methods of Moments? If yes, it gives only initial estimates for REML in a mixed-effects model. If the study is balanced, the optimizer should converge in the first iteration. In this case REML-estimates equal MoM-estimates.

❝ Although when running a BE study and analyzing the data using MM/REML I get similar CI -we have a one drop out- ran the analysis with and without the subject and BE decision still the same. When using MM while including the drop-out using REML the BE decision changes.


When you mean that the data was incomplete (i.e., the subject dropped out but you have still data of some periods) that’s quite possible. A mixed-effects model (including all data) tries to recover all information. However, in most cases you get pretty similar results by
  1. Mixed-effects model of all data
  2. Mixed-effects model of complete data (dropout excluded)
  3. Fixed-effects model like #2.

❝ I did some search, there's some paper indicating that the estimation using MM in unbalanced design might lead to biased variance estimates, …


Which one – Patterson/Jones?

❝ … although, in our case, the number of subjects on reference product is the same as the one on test product.


How is that possible when you had a dropout?

I would never try a mixed-effects model on the whole data. The residual variance should be the same (unless you run into convergence issues of the optimizer). Only the between-subject variance will be different. The latter is only “nice to know”. The EMA is clear in the BE-GL: Only subjects with at least one treatment of T and R should be included. If you want to submit to the FDA, exclude the subject as well. Not informative.

Dif-tor heh smusma 🖖🏼 Довге життя Україна! [image]
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
UA Flag
Activity
 Admin contact
22,957 posts in 4,819 threads, 1,638 registered users;
72 visitors (0 registered, 72 guests [including 7 identified bots]).
Forum time: 16:03 CET (Europe/Vienna)

Nothing shows a lack of mathematical education more
than an overly precise calculation.    Carl Friedrich Gauß

The Bioequivalence and Bioavailability Forum is hosted by
BEBAC Ing. Helmut Schütz
HTML5