Date:         Thu, 17 Jan 2008 14:29:15 -0600
Reply-To:     "data _null_," <datanull@GMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "data _null_," <datanull@GMAIL.COM>
Subject:      Re: Weird result in PROC GLM
Comments: To: Peter Flom <peterflomconsulting@mindspring.com>
In-Reply-To:  <7407591.1200594296930.JavaMail.root@mswamui-chipeau.atl.sa.earthlink.net>
Content-Type: text/plain; charset=ISO-8859-1

Perhaps this example will help.  The means are depend on the coding 0
vs -1.  In the REG the intercepts are different.  The intercept and
mean are related, yes?

data test;
   do coding = -1,0;
      do year = 1989 to 2007;
         LateYear =ifn(year>1997,1,coding);
         earlyYear=ifn(year<1998,1,coding);
         output;
         end;
      end;
   run;
proc print;
   run;
proc means;
   class coding;
   run;
proc reg noprint outest=est;
   by coding notsorted;
   late:model year = lateYear;
   early:model year = earlyYear;
   run;
proc print;
   run;


On Jan 17, 2008 12:24 PM, Peter Flom <peterflomconsulting@mindspring.com> wrote:
> Hello
>
> We have a data set with 3 variables: Year, percent HIV and research estimate.
>
> At the request of our boss, we converted year into a dichotomous variable (not my choice of analysis...)
>
> My colleague did this two ways:
>
> 1) Lateyear = 1 if year > 1997, else 0
> and
> 2) earlyyear = 1 if year < 1998, else 0
>
> as a learning exercise we  then ran both
>
> PROC GLM;
>  MODEL RESEARCH_ESTIMATE = PERCENT_HIV LATEYEAR PERCENT_HIV*LATEYEAR;
> RUN;
>
> and
>
> PROC GLM;
>  MODEL RESEARCH_ESTIMATE = PERCENT_HIV EARLYYEAR PERCENT_HIV*EARLYYEAR;
> RUN;
>
> I expected these to be identical, except for change of sign.  Indeed, most of the results are identical, R-square, type I SS, F value etc..  But the TYPE III SS results for PERCENT_HIV are different by a factor of 8; the parameter estimate for the intercept is different, and the parameter estimate for PERCENT_HIV is different by a small amount, and the SE for the intercept and PERCENT_HIV are different.
>
> Any ideas how this could happen?
>
> Thanks
>
> Peter
>