[R] R and clinical studies

[Cody_Hamilton at Edwards.com]

I agree that most problems arise in the data management / file derivation
phase.  From my reading of 21 CFR 11, it appears that this document focuses
primarily on data management (as well as on software directly involved in a
medical device) rather than on validation of statistical functions.  I
believe this point has been made previously on the R-help list.

With regards to validating functions, I have often wondered how one can
validate a function when one cannot see what it is doing.  You could
certainly compare calculations from one package to the same calculations
from another package, but then you must purchase (ouch!) and know how to
properly use two software packages instead of one.  And I suppose they
could both be wrong!  Is not peer-review the best form of validation?  . .
. I suspect I may be "preaching to the choir" here.

I would love nothing more than to migrate our stat group over to R from
SAS.  Based on my experience with R/Splus, the language seems more
extendable, flexible, and has much better graphics (as has been pointed out
many times on this list).  It also has available the many contributions of
generous R users.  However, it has been hard to win pure SAS users onto R
(even if it saves the company money!).  One can't send the biostat group
off to R training like one would to SAS classes.  Learning R requires
initiative from the user (which is not necessarily a bad thing).  I
considered encouraging the purchase of Splus as an intermediate step
(hoping that its proprietary nature would soothe fears regarding open
source software), but that option was not as cheap as I thought.

Regards,
    -Cody




Delphine Fontaine wrote:
> Thanks for your answer which was very helpfull. I have another question:
>
> I have read in this document
> (http://cran.r-project.org/doc/manuals/R-intro.pdf) that most of the
> programs written in R are ephemeral and that new releases are not
> always compatible with previous releases. What I would like to know is
> if R functions are already validated and if not, what should we do to
> validate a R function ?
>

In the sense in which most persons use the term 'validate', it means to
show with one or more datasets that the function is capable of producing
the right answer.  It doesn't mean that it produces the right answer for
every dataset although we hope it does.  [As an aside, most errors are
in the data manipulation phase, not in the analysis phase.]  So I think
that instead of validating functions we should spend more effort on
validating analyses [and validating analysis file derivation].  Pivotal
analyses can be re-done a variety of ways, in R or in separate
programmable packages such as Stata.

--
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.