Reply from Helmut, 2014-11-14 14:11:
Hi Lucas,
» Hi Helmut, how are you?!
THX; reasonably good. :smoke:
» Did you validated all workflow templates that you use on WinNonlin…
Good question! In the past I compared the WinNonlin’s output obtained from datasets to the results given in the literature (see here --> 10914). I survived one inspection in 2004.
» …or there wasn't need for that?
Software has to be validated, right? Unfortunately all these datasets were balanced crossovers. We recently published a paper comparing eight datasets evaluated by five different software packages – EquivTest/PK, Kinetica 5.0.10, SAS 9.2, Phoenix/WinNonlin 6.3.0.395, and R 3.0.2 (see here --> 13653). We obtained identical results – with the exception of Kinetica, which screwed up all imbalanced datasets (PE & CI wrong). Thermo Fisher Scientific will correct the bug (see here --> 13826).
Amazingly one of the reviewers of the manuscript asked “Was the software used in the analyses validated?” which we answered “For the purpose of bioequivalence the softwares were not validated beyond installation and where possible automatic IQ/OQ/PQ. The point of the paper is that it is virtually impossible to do in absence of a range of datasets, and the paper makes a proposal on which datasets such a validation could be based.”
Last Wednesday another paper dealing with 2-group parallel designs was accepted by the AAPS Journal. In PHX/WNL one needs a workaround to adjust for unequal variances (i.e., Welch Satterthwaite approximation) and the software is limited to 1,000 subjects/group (not an issue in BE). You can make an educated guess about how Kinetica performed.
We “validated” PHX/WNL’s templates for FDA’s RSABE and EMA’s ABEL (see here --> 13748) by comparing its results with SAS’. Still some work to be done. EMA’s fully replicated dataset is imbalanced (nRTRT ≠ nTRTR) and incomplete (periods missing) whereas the partial replicate is both balanced and complete. We learned (from examples John posted here and others in Pharsight’s Extranet) that FDA’s ABE code for the partial replicate fails to converge sometimes due to the overspecified model (both in PHX/WNL and SAS). It seems that FA0(1) instead of FA0(2) could do the job. CSH (suggested in FDA’s guidance) did not work for these nasty datasets.