[R-sig-phylo] lambda for single traits and regression

Isabella Capellini isabella.capellini at durham.ac.uk
Fri Jan 23 12:15:35 CET 2009

Dear all,

some time ago we discussed testing for the phylogenetic signal in the  
data with Pagel's lambda and marginally touched the issue of  
estimating lambda while simultaneous;y running a regression.

So here is my question. As Marguerite pointed out (see below), it may  
happen that both your dependent and independent trait have a high  
lambda, but the lambda for both of them together is lower (or  
different). How is this possible theoretically? if X has lambda=1, y  
has lambda=1, why isn't the lambda of the regression not 1 too? This  
happened to me a few times and I really cannot figure out why (sorry!)

On a related issue, say that your X and Y have both lambda=1  
individually; what should you do while testing their regression? what  
is your current thinking about estimating lambda while testing for the  
regression? should this be done or not? It seems to me that while some  
people do estimate lambda value of the 2 traits together in the  
regression, others check the lambda for each trait separately and if  
they do find a strong phylogenetic signal they proceed without further  
estimating lambda during the regression analysis (which should imply  
they use a lambda=1, right?).


On 11 Apr 2008, at 20:45, Marguerite Butler wrote:
> Just wanted to add a couple of more points:
> 1) From what you describe, I think you are fitting Pagel's lambda to
> univariate models. But then you are comparing bivariate models. These
> are not the same, in the sense that the best lambda for each variable
> may not be the best one for them jointly.

Isabella Capellini, PhD
Research Associate

Evolutionary Anthropology Research Group
Department of Anthropology
Durham University
Dawson Building
South Road, Durham
DH1 3LE (U.K.)

phone: +44 (0)191 3343298
fax: +44 (0)191 3341618



More information about the R-sig-phylo mailing list