Date:         Mon, 15 Jul 1996 07:49:07 -0500
Reply-To:     "Christopher Alan Strickland (407) 726-7351"
              <STRICKLA@JS-JTF.AF.MIL>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         "Christopher Alan Strickland (407) 726-7351"
              <STRICKLA@JS-JTF.AF.MIL>
Subject:      Re: trying to get R-squared to 9 decimal places
In-Reply-To:  <01I6ZYX3R7AQ0001CP@js-jtf.af.mil>

In article <4rs8e2$p8m@news.wco.com>, doliov@wco.com says...
>>
>>Dean Nelson (denelson@facstaff.wisc.edu) wrote:
>>:Does anyone know how to get R-squared from an OLS model to 9 decimal places?
>>:
>>your post piqued my interest.  just as a matter of curiosity, why would
>>you want the ninth decimal place of an R-square value?
 
>Another problem has cropped up in this solution.  The model he is using has
>68 independent variables.
 
How about eliminating a number of variables before doing the all possible
regressions.  Sounds like some academic overkill here.  There has got to be
severe problems with co-linearity which will invalidate the all possible
regressions.  Before even attempting to do this the number of independent
variables needs to be pared down big time.  One approach, though time consuming
would to be drop the variable with the lowest t-value and continue until all
variables in the model are significant.
 
Secondly, the R-squared is going to be high just from the number of independent
variables.  If you want to use an R-squared, use the adjusted R-squared.
 
Bottom line: A lot more homework needs to be done before even attempting to do a
all possible regressions on this model.  The co-linearity along is going to
destroy any reliability.
 
+------------------------------------------+
| Chris Strickland                         |
| http://www.teg.saic.com/mote/people.html |
| strickla@js-jtf.af.mil                   |
+------------------------------------------+