Date:         Tue, 26 Jan 2010 00:42:04 -0800
Reply-To:     wolfgang <cmat.wolfgang@GMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         wolfgang <cmat.wolfgang@GMAIL.COM>
Organization: http://groups.google.com
Subject:      Re: Is Regression Using Proc IML Faster?
Comments: To: sas-l@uga.edu
Content-Type: text/plain; charset=ISO-8859-1

Hi,
I cannot believe that you could be faster using IML code than PROC
REG.
As was already said here, the main problem with IML is that you must
load the matrix into core.
[1] Now, PROC REG computes the X'X matrix by running observationswise
through the data set
and X must never be in code. That means PROC REG accumulates a lot of
error (the condition
of X'X is squared the condition of X)  by computing X'X, but it should
be faster than IML when doing
the same in two steps, loading the matrix ad computing X'X.
[2] After X'X is computed, PROC REG is using sweep, which is of course
much slower than Cholesky.
Here, IML could be faster when Cholesky is used. However, using
Cholesky on a ill conditiones X'X
may get you in trouble.
[3] So, if X is ill conditioned, I would use either PROC ORTGOREG or
IML.
But if the condition of X is okay, I would use PROC REG.

When computing benchmarks, the condition of X should be one factor.
Another factor would be the form of the matrix, i.e. is Nobs >> nvar
or is Nobs not
much larger than nvar. These are two very different situations. Also
running very many
small problems or just a few very large problems should be a factor
etc.

Wolfgang