Date: Thu, 7 Nov 2002 09:40:35 -0800
Reply-To: fred <xkrim3@HOTMAIL.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: fred <xkrim3@HOTMAIL.COM>
Subject: Re: Sum Horizontally Across 50,000 or more Variables - questions
from an Amateur
Content-Type: text/plain; charset=ISO-8859-1
Alright, here's my confession: I am 6 months and 100 lightyears from
being a well-grounded statistician. & possibly several times more from
being an effective programmer.
Actually, I was just moping around for an excuse to exercise my meagre
knowledge of ARRAYS and DO loops. But there is NO excuse for not
understanding basic statistical theory. You have just inspired me to
hit the books (seriously).
Cassell.David@EPAMAIL.EPA.GOV (David L. Cassell) wrote in message news:<OFC049CDD2.BBDF88C1-ON88256C69.email@example.com>...
> Roland "roland.rashleigh-berry" <roland.rashleigh-berry@NTLWORLD.COM>
> already gave an excellent technical reply. But I looked at your
> request and, well, winced. What you are attempting to do is bad
> for several reasons. You have a bad data structure which is
> handicapping your efforts. But your efforts are seriously misplaced.
> You (the original poster) wrote:
> > > I have written some code to determine the number of samples
> required before
> > > the actual mean converges to the expected mean.
> No, no, no! You cannot perform one realization of a stochastic process
> (or even a hundred) and come up with an understnading of the underlying
> ergodic behavior! It's a common mistake, though. I have even seen it
> done (horribly) in a peer-reviewed journal. (names omitted to protect
> the guilty...) You will only see when one chance behavior makes things
> look right, on one occasion. This is *not* what stochastic convergence
> is about. If you by chance get the sample mean to hit exactly the
> mean when you have only generated a sample of size 5, are you going to
> assume that you need only do n=5 for the rest of your life? If things
> go badly and you end up needing 100,000 samples, will you do that
> as well?
> Please, do NOT take this approach. Use statistical theory to tell you
> what sample size to use, or proper randomization, and not a single
> a process.