Skip to content
Prev 12349 / 20628 Next

parallel MCMCglmm, RNGstreams, starting values & priors

Hi Ruben,

It might be hard to detect (near) ECPs with so many fixed effects (can  
you post the model summary (and give us the mean and standard  
deviation of any continuous covariates)). Also, the complementary  
log-log link (which is the za specification) is non-symmetric and runs  
into problems outside the range -35 to 3.5 so there may be a problem  
there, particularly if you use rcov=~trait:units and the Poisson part  
is highly over-dispersed.  You could try rcov=~idh(trait):units and  
fix the non-identifiable za residual variance to something smaller  
than 1 (say 0.5)  - it will mix slower but it will reduce the chance  
of over/underflow.

Cheers,

Jarrod




Quoting Ruben Arslan <rubenarslan at gmail.com> on Thu, 28 Aug 2014  
18:45:30 +0200: