I've been on vacation and so come late to this interesting discussion. Let
me add two minor points.
1. I have run into a lot of statements that "x is required" when
dealing
with pharma, and in particular wrt NDAs (new drug application). Almost all
were false. But I also understand a bit of where they are coming from. An
NDA costs millions, with the required trials, and is a huge submission in
terms of the sheer amount of paper. If a company's last NDA was successful,
there is a lot of draw to make the next one as much like it as possible, down
to the font size and the margins.
So I agree with Frank that "SAS is validated" is a false
requirement, yet
remain somewhat sympathetic to the person who said it.
2. We have been audited once on a portion of an analysis. The official said,
more or less, "show us how you got that answer" while pointing at a
part of
the report. That particular analysis happened to be done in SAS, but what
was important was our ability to quickly find the relevant code, not the
package.
3. If the other party is stubborn about SAS vs S, I have one example I like
to argue with. The S survival code has an extensive test suite, including
a set of small examples where I have worked out the correct results by hand.
Many of these latter are documented in the appendix of the Therneau and
Grambsch book. The SAS phreg procedure does not pass all the tests. (Cox
model + Efron approx for ties + deviance residuals for one. The size of
the error is numerically insignificant, a 1/n vs 1/(n-1) type of thing: but
it leads to slightly different robust standard errors).
