|Date: ||Mon, 30 Jun 1997 10:58:10 -0500|
|Reply-To: ||Brian Smith <B.SMITH@LILLY.COM>|
|Sender: ||"SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>|
|From: ||Brian Smith <B.SMITH@LILLY.COM>|
|Organization: ||Eli Lilly and Company|
|Subject: ||Re: Singular Matrix Problem|
|Content-Type: ||text/plain; charset=us-ascii|
> In article <5p12id$12pa@ns5-1.CC.Lehigh.EDU>, drt2@Lehigh.EDU writes:
> >Subject: Singular Matrix Problem
> >From: drt2@Lehigh.EDU
> >Date: 27 Jun 1997 14:59:25 -0400
> >Sorry in advance if this is a "statistics" question vs. a SAS question
> but I
> >am desperate for some help.
> >I am doing an analysis using several continuous and discrete variables. I
> >using Proc GLM with the /solution option to get the beta estimates. I get
> >note that says" The X'X matrix has been found to be singular and a
> >inverse was used to solve the normal equations. Estimates ... are biased
> >are not unique estimates of the parameters."
> >I understand what it is saying but, and here's the SAS question: What can
> >to "fix" this problem? I did a correlation analysis but there does not
> >to be any significant correlation between the variables in the model.
> >anyone help me with this?
> >While I'm abusing this group, can anyone tell me the "best" way to
> >the Beta Coefficients? Is it better to calculate them with all variables
> >the model or alone?
> >Thanks in advance.
> the fix for your problem is hard to specify without knowing the nature of
> your predictor variables. Doing simple pairwise correlations between
> variables will not necessarily point you toward the problem. It is often
> the case that some one variable is a linear combination of two or more
> other variables. In this case, none of the paiwise correlations are
> likely to be high. This often happens when dummy coding categorical
> variables and you forget to "leave out" the comparison level.
> If you describe your variables in a little more detail someone may be able
> to provide a little more guidance. This would also help with answering
> your second question.
> Good luck,
Since he says he has both continuous and discrete variables, I imagine
the only problem is he overparamerterization due to having the intercept
in the model. More than likely he can eliminate the message with the
noint option in the model statement. But, in reality it is not a
To the second question about beta coefficients. The answer is it
depends on what you want to find out. Find yourself a good statistician
to work through all the relative points.
Brian Smith, PhD