Choosing a best model when you have several with similar xR2

[Chris Bowman-Prideaux]

I am interested in understanding what variables affect the cover of a plant species. My question general is what environmental, land management, and species cover/density variables best predict the cover and density of the species of interest. In my analysis I have 80+ variables. The analysis yielded a series of 3 variable models. Two variables occur in all of the top 100 models with the greatest xR2. The third variable differs between species cover variables and land management variables.

What I would like to do is discuss two models that have very similar xR2, but address somewhat different aspects of the community/environment/land management interactions. Can I present two models in a paper that come from the same NPMR analysis? They would address two more specific questions
1) what combination of species variables explains the target species cover?
2) what combination of land management variables explains the target species cover?

Or can I only present one model from a single NPMR analysis?

[Bruce McCune]

I'd say that you should feel free to present contrasting models, as you suggest. It is often very interesting to compare one group of variables against another. For example, one might ask how well one can predict a species occurrence based on geographic coordinates alone (easting, northing, elevation) vs a model with habitat variables but no geographic coordinates. I find it easiest to do this kind of comparison by setting up different predictor matrices, each with a different subset of predictors. They can be mutually exclusive subsets or partially overlapping.

Bruce McCune

--
You received this message because you are subscribed to the Google Groups "HyperNiche and NPMR" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hyperniche+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Chris Bowman-Prideaux]

Bruce, 

So I shouldn't use 2 models from the same analysis using all 80 variables? I should break down the variables in to two sets of variables with at least some differences in the predictor matrix. Correct? 

If so, how many variables need to be different?

The original reason for putting all the variables into a single matrix was to look for patterns between land management variables and species variables. If I separate them and use a species only matrix and a land management only matrix I loose a rather interesting result.

Or should I do the analysis with a predictor matrix with species only and then a second analysis with species and land management variables in the predictor matrix?

Chris

[Bruce McCune]

Chris, the idea of two completely separate sets of variables was just an example. I guess my general point is that you can structure the model selection and predictor sets in any way that makes sense to you. And you can examine and report on more than one specific model from the same group of models. Keep in mind that predictors that aren't selected in a particular model have nothing to do with the qualities of that model EXCEPT that during the model fitting phase they are competing with other predictors and combinations of predictors for inclusion in the model.

As I understand it, you already have two models with an interesting contrast between them, so no problem with presenting and discussing both of them.

Bruce McCune