s-news
[Top] [All Lists]

Looking for help with lme correlation structure

To: s-news@lists.biostat.wustl.edu
Subject: Looking for help with lme correlation structure
From: Andrew Robinson <andrewr@uidaho.edu>
Date: Thu, 22 Aug 2002 21:53:52 -0700
Hi all Splus-gurus,

I'm trying to fit a mixed-effects model with the following structure: I have 6 regions, within which I have 90 sites (not exactly 15 per region, but close), within each of which I have 2 blocks, within each of which I have 3 plots. For each plot I have a set of diameters of a varying number of trees. I've fit the 3-parameter Weibull function to each of these plots, therefore I have three parameters per plot, for a total of 540 triplets in the structure described above. I have numerous potential predictor variables (eg average height, basal area, etc.) that I'd like to use to try to predict these three parameters.

I've constructed a scatterplot matrix of the parameters within the plots, which reveals considerable correlation between the parameters (especially location and scale). I'd like to construct a model that includes such a correlation structure in the predictive portion. That is, it's important that if I predict a triplet of three parameters for any given plot, that the predictions evince something like that correlation structure. Such a correlation structure is not produced from models that predict the three parameters separately.

So, what I'd like to do is predict the three parameters simultaneously for each plot using the predictor variables and the appropriate random effects, including a correlation structure and an appropriate variance structure. I have a model statement like

A.overall <- lme(parameter ~
               TPA * height * ba * model,
                weights=varIdent(form=~1|model),
               correlation = corSymm(form=~1|region/site/block/plot),
               random=~1|region/site/block/plot,
                data=fitbasis)

where model is a 3-level factor distinguishing between the fitted parameters of the Weibull distribution (location, shape, and scale).

So, when I fit this model, I get a correlation structure like this:

 Correlation:
       1      2
2 -0.783
3 -0.998  0.767

So, the within-plot residuals have that correlation structure. Now, how do I induce the model to apply that correlation structure to the within-plot fitted values? I guess I want the BLUPS to be correlated in this way within the plots. Can it be done?

Thanks for any and all ideas,

Andrew

Andrew Robinson                  Phone: 208-885-7115
Department of Forest Resources   Fax:   208-885-6226
University of Idaho              E: andrewr@uidaho.edu
Po Box 441133                    WWW: http://www.uidaho.edu/~andrewr
Moscow, ID 83843                 and: http://www.biometrics.uidaho.edu/
No statement above may be construed to necessarily represent my employer.


<Prev in Thread] Current Thread [Next in Thread>
  • Looking for help with lme correlation structure, Andrew Robinson <=