[R-sig-ME] variance components models with zero estimates

[David Airey]

When a variance components mixed model is run in Stata, if some of the  
variance components are zero, the model may not converge, for rational  
reasons according to the manual entry. However, when the same model is  
run in SAS, the models with variance components that estimate to zero  
nonetheless converge. According to some SAS user friends, this is  
normal SAS behavior (I'm new to SAS as of yesterday). If I'm  
interested in looping through a set of such models, the SAS behavior  
is preferred. However, in Stata such models can be formulated as  
multilevel models that can dramatically reduce the dimension of the  
design matrix. The context where both behaviors is important is mixed  
models for gene set enrichment analysis, where there is a possibility  
of hundreds of models.

Does R lme4 handle variance components mixed models that have  
estimates of zero for some of the variance components like SAS or  
Stata? Is it possible to loop through variance components models when  
some of the variance components are zero? What is a suggested  
procedure for doing so in R? In Stata, I would probably use an EM only  
guess at which variance components were substantive, and then fit one  
of several models. What about R lme?

Cheers,

-Dave