Date: Fri, 2 Jan 2004 12:24:48 -0800
Reply-To: David Reilly <dave@AUTOBOX.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: David Reilly <dave@AUTOBOX.COM>
Subject: Re: Test for Trend: SAS
Content-Type: text/plain; charset=ISO-8859-1
email@example.com (Sharon) wrote in message news:<firstname.lastname@example.org>...
> Hi there, two questions. I'm using SAS for my analysis.
> QUESTION ONE: I'm using generalized estimating equations (GEE) for my
> analysis and one of my variables has an obvious decrease in trend. Is
> there a way to do a "test for trend" using PROC GENMOD?
> QUESTION TWO: In the absence of GEE, I'd like to assume independence,
> just to get a sense of the trend (I know this is inappropriate, given
> the correlated data). I know I can use PROC FREQ to do a "test for
> trend" for unadjusted odds ratios but is there a way i can calculate
> "test for trend" for ADJUSTED odds ratios?
> Thanks in advance!
Tests for the difference in "TRENDS" between two time series require
that there is a common KIND OF TREND.
For example y(t)=y(t-1) + CONSTANT1 is one form of a TREND MODEL
a second example
y(t)= CONSTANT2 + CONSTANT3*TIME is another form of a
a third example
y(t)= CONSTANT4 + CONSTANT5*x(t) is yet another form of
a TREND MODEL .
Now let's assume that you have no causals (x's ) thus your problem is
to determine which of the first two TREND MODELS is appropriate.
To complicate matters consider the first kind of TREND
y(t)= y(t-1) + CONSTANT1 if t => to
otherwise y(t)= y(t-1) + CONSTANT2
thus the TREND depends on where you are in time.
Additionally consider models of the second kind
where y(t)= CONSTANT2 + CONSTANT3*TIME if t => to
otherwise y(t)= CONSTANT2
In order to test for a common trend one needs a common model form
across the two groups otherwise there is a de facto difference.
If one can assume a common model form then one estimates globally and
locally and performs an F Test on the differential SOS in order to
test the hypothesis of a COMMON TREND .
Hope this helps.
Automatic Forecasting Systems