Tom Oliver 03/13/09 9:36 am >>>
Dear Emmanuel & Helplist,
Thanks for your response. So would a valid approach be to test for
phylogenetic autocorrelation beforehand (e.g. Using Gittleman and
Kot's method implemented by the gearymoran function in R package
ade4) and if there is significant autocorrelation, then to use the
GEE method to get the regression slopes and significance of the
explanatory variables? In my case there is no significant
phylogenetic autocorrelation in the data (using the gearymoran
function). Does this then justify using a straight forward
non-phylogenetic regression? I am slightly confused that if we
cannot reject the null hypothesis of phylogenetic autocorrelation
(p>0.05), does this mean that we can really 'accept' the null
hypothesis and use non-phylogenetic methods?
I guess that you'd prefer this relationship to be significant ;)
Since you seem to have a continuous response, I suggest you switch to
GLS and compare different correlation structures with AIC.
EP
Thanks for all your help,
Tom
Emmanuel Paradis <emmanuel.para...@mpl.ird.fr> 03/03/09 3:25 pm >>>
Actually, I have already observed the same thing with simulated data.
When you shuffle your data (independently for both vectors), you
cancel the (potential) relationship that exists between the variables,
but you also remove the covariance among species. So using GEEs is no
more valid because you now assume a wrong covariance structure.
Instead, you should use a "comb" tree. The real test of the type I
error rate of compar.gee is to simulate two independent sets of
variables with a multivariate distribution which variance-covariance
is given by the tree (alternatively you can simulate the characters
along the tree, but it's tedious).
In practice, a solution to this lack of robustness of the GEE approach
may be to test for phylogenetic autocorrelation beforehand.
Alternatively, if your response is continuous you can use GLS where
the parameter(s) of the correlation structure may be estimated, and
different correlation structures can be compared (not sure about
similar comparisons with GEEs).
EP
--
This message (and any attachments) is for the recipien...{{dropped:13}}
_______________________________________________
R-sig-phylo mailing list
R-sig-phylo@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-phylo
