[R-sig-eco] GLS, GEE or LMM ??
Jens Oldeland
oldeland at gmx.de
Fri Apr 16 12:49:57 CEST 2010
Dear All,
I have run into a number of questions, and thus I hope you could help me
out. I am modelling the effect of oyster density and nutrients on the
bodyweight of mussels (population average).
Data was sampled at three different stations over 8 years, with values
measured in springtime once per year.
I was following Zuur et al 2009 Mixed Effects Models (wonderful book!),
but got lost at some points since different models lead to totally
different results.
a) the first question is about "centring data". Zuur suggest to center
parameters (p.334) if they are highly correlated with the intercept.
When I apply a lme (family=gaussian, random ~ 1 | bank, correlation =
corAR1(form = ~ daycount)) I have to center nearly all the values. When
I apply a GEE then there is no correlation at all (r=0.14).
Actually, centring the data leads to the same output at the end (for the
lme)
b) Choosing GEE, the effect of one parameter (salinity) is highly
significant, while using the LMM approach it is not, which would be
better for our interpretation...
But why? Is it because GEE should not be used on normally distributed
data? I know that GEE uses sandwich estimator and LMM uses ML. Which one
would be more "trustworthy" or conservative?
c) one last qeustion: negative AICs, which one is better. AIC: -10 or -5
? I have read contrasting statements. Is there any proof?? Does it hold
for BIC as well?
thank you in advance!
Jens
--
+++++++++++++++++++++++++++++++++++++++++
Dipl.Biol. Jens Oldeland
Biodiversity of Plants
Biocentre Klein Flottbek and Botanical Garden
University of Hamburg
Ohnhorststr. 18
22609 Hamburg,
Germany
Tel: 0049-(0)40-42816-407
Fax: 0049-(0)40-42816-543
Mail: Oldeland at botanik.uni-hamburg.de
Oldeland at gmx.de (for attachments > 2mb!!)
Skype: jens.oldeland
http://www.biologie.uni-hamburg.de/bzf/fbda005/fbda005.htm
+++++++++++++++++++++++++++++++++++++++++
More information about the R-sig-ecology
mailing list