[R-sig-ME] prior specification in MCMCglmm

Celine Teplitsky teplitsky at mnhn.fr
Mon Apr 25 13:26:04 CEST 2011

Dear all,

I realise that Jarrod is doing field work, but I'm really hoping someone 
can answer my question while he's not around.

I am running animal models estimating covariances between life history 
traits, and I'm having trouble knowing which prior to use.

Thing is, if I use a prior as described on the Wam wiki site with 
V=PhenotypicVar/4 (as I have 3 random effects + residual), I have very 
nice results, with some significant genetic correlations between some 
life history traits.

However, one reviewer asked about prior sensitivity because CI were 
pretty large, so I went back to MCMCglmm course notes and saw that non 
informative prior were supposed to be V=diag(nbDimV)*0 and n=nbDimV-3. 
This led to an error message about G being ill conditioned, so I tried 
with diag(nbDimV)*0.001 and diag(nbDimV)*0.01 instead of diag(nbDimV)*0, 
and diag(nbDimV)*0.01 worked... But then I have the posterior of 
additive genetic variance collapsing on 0 for some trait. So my guess 
would be that I should use those latest priors, and believe my nice 
results did not exist. But as Hadfield et al paper and the Wam wiki 
website do not recommend those priors, I am a bit confused. Could 
someone help me figure out what would be the right thing to do?

All my apologies if this is a silly question, but I'm feeling a bit lost 

Thanks a lot in advance



Celine Teplitsky
Département Ecologie et Gestion de la Biodiversité UMR 7204
Unité Conservation des Espèces, Restauration et Suivi des Populations
Case Postale 51
55 rue Buffon 75005 Paris

Webpage : http://www2.mnhn.fr/cersp/spip.php?rubrique96
Fax : (33-1)-4079-3835
Phone: (33-1)-4079-3443

More information about the R-sig-mixed-models mailing list