On Sep 7, 10:42 am, jim4s...@YAHOO.CO.UK (Jim Groeneveld) wrote:
> Well, that is not so difficult. Just look at the median, not the mean (you
> do it nonparametric don't you? Though in the output of NPAR1WAY I see means)
> values of your effects; are these in the expected direction? If these are
> not included in the standard output it must be possible to caugh them up in
> some other way. And you know the direction that you expect. But maybe the
> PROC could be more clear about this. Yet it is possible to interprete the
> output correctly.
Thanks for your help. It seems that there is indeed sufficient
information for me to figure out what is being tested, and how to
interpret the one-sided tests, if only I examine enough of the output
and understand the convoluted way SAS has set this up.
I guess this now morphs into a rant about why can't SAS make these
things clearer. The PROC prints a one-sided test result, but never
tells us what the alternative hypothesis is that is being tested. They
could have programmed the output to say "One-sided test for Ha: mean
(feature) > mean (check)" but they didn't do so.
They write in great detail in the documentation how the test statistic
z is computed, but nowhere in that write-up do they discuss exactly
what hypothesis z is testing. You have to know enough to use other
outputs from the procedure to know what hypothesis is being tested,
and it may not be the hypothesis you want to test. This is exactly the
opposite of statistical theory, where you state the alternative
hypothesis and then compute the test. I have this funny little idea in
my head that SAS ought to mirror common statistical practices...
PROC NPAR1WAY isn't the only SAS procedure that does not clearly state
the hypothesis being tested, in either the output or in the
documentation. When it is difficult for an experienced statistician
such as myself to figure out what is being tested, imagine how
beginners must feel. And it is so unnecessary ... another 5 hours of
programming and testing would have made the PROC NPAR1WAY outputs much
Seems like a letter of complaint to SAS Institute is in order here,
complete with examples where this happens in other PROCs. I will get
And finally, I have similar issues with MATLAB as well, so I'm not
just picking on SAS. They too are guilty of not explaining things well
enough so an experienced statistician knows what is going on. At one
point, I wrote to The Mathworks complaining about an example in their
documentation, which I said was the most incomplete and misleading
example I had ever seen in software documentation.
To both SAS and MATLAB, I say that five more hours of their time would
have saved me (and many others) ten hours of my time.
paige\dot\miller at kodak\dot\com