Hi,
Centring will help reduce the correlation amongst covariates -- it is a good (and old) trick. A surer way is to use orthogonal polynomials. They are
slightly harder to interpret, but often this is immaterial. Try poly(x,3) and ?poly. Another option would be to ditch the idea of using a global
polynomial and use semi-parametric methods, such as B-splines (in library splines), or GAMMs (in library mgcv).
I fear that this is all fairly academic though. Cubics may be good enough -- check the model through diagnostics. Even though this is not trivial
for mixed models.
What concerns me a bit is the comment that the dependent variable is obtained from an NMDS. How is the uncertainty in the original data propagated
through the NMDS and the mixed model? At all? Are there any comments / opinions about this? Intrigued.
Scott
On 04/10/13 01:38, Zoltan Botta-Dukat wrote:
Dear Alexandre,
If your variables have only positive values, there will be strong correlation between linear, quadratic and cubic terms that leads to high VIF. You
can avoid it by centering the variables before calculating quadratic and cubic terms:
quadratic<-(x-mean(x))^2
cubic<-(x-mean(x))^3
Regards,
Zoltan
2013.10.03. 14:49 keltezéssel, Alexandre Fadigas de Souza írta:
Dear colleagues,
I am working on a Linear Mixed Model with nested structure, based on the book of Zuur et al. 2009. Mixed Effects Models and Extensions in
Ecology with R, and have a question to ask to those of you more experienced with this model family.
I am using several topographic variables (elevation, slope, convexity, facing) and one light-related variable (canopy openness) to explain
species abundance variation in 85 plots placed on 17 transects dispersed through a coastal vegetation complex on sandy soils in northeastern
Brazil. The dependent variable is axes of a floristic ordination (e.g., NMDS), with one separate model adjusted for each axis.
I will use two levels of spatial aggregation as random factors: the transects (5 plots per transect) and broad transect clusters (transect
clusters were logistically conditioned).
In the ecological literature, I have seen the suggestion of including quadratic and even cubic versions of the fixed effects variables (in this
case topographic and light) as a means to account for possible nonlinear effects of these variables on the dependent variable. One recent example is
Brunbjerg, A.K., Ejrnæs, R., Svenning, J.-C., 2012. Species sorting dominates plant metacommunity structure in coastal dunes. Acta Oecologica 39,
33–42.
However, the inclusion of such quadratic and/or cubic terms in the model
produces high colinearity levels (VIF > 25, sometimes even 38!).
Do you think this is a valid procedure, the inclusion of these polynomial
terms?
Thank you in advance for any ideas,
Alexandre
Dr. Alexandre F. Souza
Professor Adjunto II Departamento de Botanica, Ecologia e Zoologia Universidade Federal do Rio Grande do Norte (UFRN)
http://www.docente.ufrn.br/alexsouza Curriculo: lattes.cnpq.br/7844758818522706
_______________________________________________
R-sig-ecology mailing list
R-sig-ecology@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
--
Scott Foster
Computational Informatics
CSIRO
E scott.fos...@csiro.au T +61 3 6232 5178
Postal address: CSIRO Computational Informatics, GPO Box 1538, Hobart TAS 7001
Street Address: CSIRO Computational Informatics, Castray Esplanade, Hobart Tas
7001, Australia
www.csiro.au
_______________________________________________
R-sig-ecology mailing list
R-sig-ecology@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology