Date: Thu, 5 Jan 2006 13:07:47 -0800
Reply-To: Steve <sczapka@HOTMAIL.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Steve <sczapka@HOTMAIL.COM>
Subject: Re: Estimate Statements with continuous variables
Content-Type: text/plain; charset="iso-8859-1"
I probably should have added to my original listing that I would be
running the analysis on a sub-set of the data only containing data from
seedlings that were protected by shelters. I realize that the analysis
would be confounded by control seedlings (shelter diameter of zero).
Dale McLerran wrote:
> --- Steve <sczapka@HOTMAIL.COM> wrote:
> > Greetings.
> > I am analyzing tree seedling growth in response to various treatments
> > using PROC MIXED. Treatments include the use of tree shelters to
> > protect the seedlings, herbicide to control competing vegetation, and
> > different species. I now want to include tree shelter diameter (a
> > continuous variable) as a covariate. However, I'm not exactly sure
> > how
> > to further analyze a significant interaction between for example tree
> > shelter use and shelter diameter.
> > As an example, say my model is the following....
> > PROC MIXED;
> > CLASS shelter herbicide species;
> > MODEL growth = shelter herbicide species diameter
> > shelter*herbicide shelter*species
> > shelter*diameter;
> > In the event of a significant shelter*diameter interaction, how can I
> > determine where the differences exist? In ther past I've only
> > compared
> > two categorical variables (e.g., shelter*herbicide). In that
> > instance
> > I used estimated statements to determine where the differences lied.
> > For example, say I had a significant shelter*herbicide interaction.
> > If
> > I wanted to compare sheltered versus unsheltered seedlings when
> > Herbicide was used or not, I could write two estimated statements
> > like...
> > ESTIMATE "shelter v. unsheltered, herbicide" shelter 1 -1
> > shelter*herbicide 1 0 -1 0;
> > ESTIMATE "shelter v. unsheltered, no herbicide" shelter 1 -1
> > shelter*herbcide 0 1 0 -1;
> > In this instance SAS gives me a p-value I can use to determine
> > significant differences between the different levels.
> > Someone told me I can do the same sort of thing with a continous
> > variable by "comparing effect (beta) estimates for the different
> > levels
> > of the categorical variable", but I'm not sure how to do this in SAS.
> > Can anyone offer advice as to how this may be accomplished?
> > Thank you.
> > Steve Czapka
> I believe there is a conceptual problem here. You indicate
> that the shelter variable is an indicator of whether the
> seedlings are sheltered or unsheltered. If the seedlings
> are unsheltered, then you must set the shelter diameter to
> zero, right? Note that the shelter diameter cannot vary for
> unsheltered seedlings. If the seedlings are sheltered, then
> the shelter diameter is nonzero and can vary.
> Thus, the shelter diameters for sheltered and unsheltered
> seedlings are mutually exclusive of each other. Further,
> there is NO SLOPE FOR THE SHELTER DIAMETER EFFECT when the
> seedlings are unsheltered. Since there can be no slope for
> the shelter diameter effect, then how can you estimate and
> interpret an interaction effect?
> The estimate could only be obtained by extrapolating the
> shelter diameter effect for sheltered trees to zero diameter.
> Depicted graphically, you might have an effect something
> like the following:
> | G
> | G
> | G
> | G
> 0 1 2 3 4
> Shelter diameter
> Here, the upper case G represents mean value given the
> fitted model. Note that the upper case G at shelter
> diameter zero is for an unsheltered tree. The other three
> upper case G values are mean growth for sheltered trees
> with diameters 1, 2, 3, and 4. The lower case g is an
> extrapolation of the effect of shelter diameter in sheltered
> trees to a shelter diameter of zero. The difference between
> extrapolated zero shelter diameter for a sheltered tree and
> the observed zero shelter diameter for an unsheltered tree
> is the only basis for an interaction effect. This difference
> may be "significant", but it seems to me completely
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: dmclerra@NO_SPAMfhcrc.org
> Ph: (206) 667-2926
> Fax: (206) 667-5977
> Yahoo! for Good - Make a difference this year.