Date: Wed, 2 Jan 2008 16:14:55 -0500
Reply-To: Peter Flom <email@example.com>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject: Re: A conceptual question about regression
Content-Type: text/plain; charset=UTF-8
In general, I agree.
There might be some exceptions. Suppose that both variables are measured repeatedly, but you only have data on certain time points. Further suppose that both variables are highly autocorrelated.
Then you might want to do what the OP suggested, since the variables could be acting as surrogates....
Also, sometimes cause and effect are not so clear....let's say you had data on economic conditions and elections....which causes which?
>From: Mary <mlhoward@AVALON.NET>
>Sent: Jan 2, 2008 3:56 PM
>Subject: Re: A conceptual question about regression
>It doesn't really seem quite right to me. For instance, I do opthalmology research and the presence of drusen (inflammatory spots) almost always occurs before the occurance of the disease AMD, and a certain GENE SNP, a factor H SNP, would occur at birth (since it is inherited or an abnormality in the gene occurs at birth. So in this example would we want to have a model saying that we want to predict the factor H abnomality given independent variables of having AMD or not and having drusen or not?
>In this example, it doesn't seem quite right to me- the dependent variable would seem to be the final event, not the initial event.
>Thus a model
>model amd_disease= b0 + b1*factorh_snp + b2*drusen_present;
>makes more sense to me than a model like
>model factorh_snp = b0 + b1*drusen_present + b2*amd_disease;
>So I'd agree with you theoritically, it seems backward to use the event that first occurs. Perhaps you could let us know what your variables are to see if there might be a situation that makes more sense.
> ----- Original Message -----
> From: Tommy Xie
> To: SAS-L@LISTSERV.UGA.EDU
> Sent: Wednesday, January 02, 2008 2:12 PM
> Subject: A conceptual question about regression
> Hi all,
> I'd appreciate if anyone can help me answer this question. If variable
> X takes place before Y and Z, which may indicates some sort of causal
> effect, is it O.K. to run a regression like X=B1*Y+B2*Z? I suspect
> that the regression has a wrong direction.
> Thanks in advance!