Date: Sun, 15 Jul 2007 11:57:17 -0700
Reply-To: David L Cassell <davidlcassell@MSN.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: David L Cassell <davidlcassell@MSN.COM>
Subject: Re: Question re TSCSREG in SAS
Content-Type: text/plain; format=flowed
firstname.lastname@example.org wrote to me personally instead of to SAS-L:
>My name is Jonathan LaBerge, and I'm a Masters student in Finance
>using SAS for my thesis.
>I have a follow-up question of sorts in regards to the U of Georgia
>listserv discussion about R-Squared using PROC TSCSREG in SAS
>As I understood, your point to email@example.com was that the
>conventional OLS R-Squared measure reported by the TSCSREG procedure
>does not apply to any model run by it. I'm a little confused by this,
>because I was under the impression that a fixed effects model (either
>one or two way), which is what unplugged said he was running, simply
>replicated the OLS procedure with dummy variables for each effect
>class. I understood from the SAS help literature that this was true
>for the random effects models, but you specifically said in your post
>that it applied to the fixed effects models as well.
>I was originally running a two-way fixed effects model in TSCSREG
>simply out of laziness, i.e., avoiding typing out all the dummy
>variables manually in the PROC REG procedure. I was relying on the
>R-Squared value in the output as a comparable measure of fit when
>compared to OLS because of my assumptions mentioned above. But after
>reading your reply I went through the effort of running my model using
>the REG procedure, and both procedures gave me the exact same
>So, finally, my question: According to your post, the R-Squared value
>in TSCSREG cannot be trusted, but I am able to replicate the same
>result using OLS by way of PROC REG. Hence, I am inclined to continue
>to use TSCSREG for the sake of simplicity and to manually adjust the
>given R-Squared value for the number of predictors. What, if anything,
>is wrong with my reasoning?
>I really appreciate your time in answering. I can provide SAS program
>and dataset files if needed.
>Thanks in advance,
You are right in that the two-way-fixed model does match the OLS
regression point estimates. But it does not match the OLS regression
variance estimators, so the R-squared estimate is not going to be the
same as that for OLS regression. But that is what we would expect,
because the OLS fit does not take into account any temporal structure
or any cross-correlation behavior.
You can use a tool like PROC MIXED to get a covariance matrix that handles
these kinds of problems too, but that is not going to give you an OLS
regression fit either.
David L. Cassell
3115 NW Norwood Pl.
Corvallis OR 97330
Local listings, incredible imagery, and driving directions - all in one