Bioequivalence and Bioavailability Forum

Hi again,

❝ Yes, how much is possible and what are the efforts? Example: Phoenix’s validation suite costs more than the software itself. If you test all components (NCA, BE, IVIVC, PK, NLME/PopPK,…) estimated runtime is about 40 hours. ‘Validation’ should be done everytime to OS is altered (not only service packs). This would include Hotfixes/Patches which are distributed by M$ monthly. Your productivity would go down by 25%. In a larger company hotfixes are centrally distributed, which means that you can’t even retreat to another machine in the meantime. Anyway I would call everything you obtain from a SW vendor (how extensive it every might be) not validation but qualification. It’s up to us to find out the flaws.

I once attended a course on WNL modeling near Paris where 20 computers were available to the attendees. WNL was installed but malfunctioned on each and every one of them. The course director was rather perplexed. It turned out to be a conflict between some hardware driver and WNL that was the problem. Which meant that the conflicting hardware had to be deactivated on all units (was not essential, iirc it was a token ring i/o driver. Yup back in those days!). So to rule out unpleasant regulatory interrogations I guess all validations should be done in presence and in absence of any other piece of software or running exe/driver on the target machine, plus all their possible interactions.

"But you didn't validate when Minesweeper, MS Excel and Adobe Suite was running and while your printer driver and RSS reader was turned off. I'll give you a 483, sir."

❝ Cough. (at least one)

Still active?

Hi Helmut!

❝ BTW, @Detlew: How did you discover the existence of Owen’s Q in SAS?

Mentioned above the Analyst application in SAS before V9.1 allowed the sample size estimation for the paired design and the parallel group design. Within that application (designed as some sort of "point and click") the user had the possibility to look behind the scenes and get the code the really used.

And oh wonder there where code lines like
   ...
   df=n-1
   t1=tinv(1-alpha,df);
   P1=OwenQ(t1,d1,0,R,df);
   P2=OwenQ(-t1,d2,0,R,df);
   power=P2-P1;
   ...
Remembering having heard the name Owen sometimes ago, getting knowledge of the formulas behind that AlGore-Rhythms it was not too hard -but costs nevertheless some sweat- for me amateur to figure out what the arguments of the OwensQ where .

Regarding your validation question:
I would say in principle it could be validated what we do with such undocumented functions, features or whatever if we could show the results are what is expected.

I have seen all around in the software industry usage of undocumented functions, features, data structures, especially in using features of the Betriebssystem Windoofs. Maybe the well-known stability of Progs under Kleinweich (measured as blue screen probability) was partly due to that.

The problem with such undocumented features/functions is that they may change in future versions, maintenance releases or even in hot-/bug-fixes without any notice. In the best case our use of them will then throw errors (or blue screens). But according to McMurphy's laws I expect we get wrong results without noticing them and are surprised if others do.

I had decided to use OwenQ because at that time I had no other opportunity to do my sample size planning the right way within "The validated power ...". Thus the alternatives were "use or die".
Meanwhile R and the author of PowerTOST came to rescue .
Now Owen's Q function is documented, the implementation code can be inspected by every body, but is it validated? What does this mean?
Of course the author of PowerTOST has checked the results of using it in sample size estimation as far as possible (regarding time, money and other resources like literature tables, access to other power/sample size software and and ...). Really! Nevertheless during the life cycle of PowerTOST up to now cases arose (extremal with respect to the usual use, see NEWS file of PowerTOST) where the implementation lead to implausible or erroneous results. What does this mean with respect to validation? Does this mean we have to validate every of our application cases because we don't know if this is an extremal? If yes how? Extremal cases are usually not tabulated. Other software may use other algorithms/approximations which by themselves call for validation/qualification. Hier beisst sich die Katze in den Schwanz. Two pices of software each giving 2+2=5 (a well known result) validate one another.

I meanwhile think the whole software validation fairy tale is a Chimera invented by the Greeks having nothing to do (the Greeks are nowadays responsible for all) .

Validated - Has been tried once before.
Guernsey McPearson's Drug Development Dictionary