Skip to content
Prev 16293 / 20628 Next

Estimates for groups within fixed effects & also: enough info for estimating variance? MCMCglmm

Hi,

1/ It is not a good idea to remove the intercept in this instance. It is 
forcing the regressions to go through the origin so that when Zdbh and 
Znnd are zero the probability of being in either class is 50:50 (0 on 
the logit scale). The random effects are not deviations from the first 
species but deviations from the average species. Only with fixed effects 
are the contrasts (usually) with the first level. If you want random 
slopes you need to replace ~species with us(Znnd):species but with only 
23 species and a categorical response you will probably not get very 
precise estimates.

2/ i.i.d species effects are often hard to separate from 
phylogenetically correlated species effects, and the phylogenetic 
heritability (aka Pagel's lambda) is often poorly estimated. This is 
particularly the case with categorical response variables and you have 
the added difficulty that numerical issues are often encountered if 
there is support for heritabilities close to one. Presumably there is 
variation in the outcome within species? If you do encounter numerical 
problems then you can use trunc=TRUE in the call to MCMCglmm which keeps 
the latent variable from going into extreme-probability territory or you 
can assume the heritability=1 by using MCMCgmmRAM:

http://jarrod.bio.ed.ac.uk/MCMCglmmRAM_2.24.tar.gz

If you have a lot of replication within species (?) my feeling is that 
you can probably get away with fitting both terms.

Cheers,

Jarrod
On 05/03/2018 17:08, Drager, Andrea Pilar wrote: