[R-sig-ME] Random effect variance = zero

[Ben Bolker]

Ben Bolker <bbolker at ...> writes:

> 
> On 14-08-14 06:54 AM, Marco Plebani wrote:
> > Dear list members,
> > 
> > Package blme has been suggested for fixing issues with random effect
> > variance = zero in other occasions, but I do not understand the
> > rationale behind it. What does blme that lme4 does not? In which way
> > do the two approaches differ? In particular: - what is the prior
> > information that blme is using, and - how comes that blme still
> > estimates parameter values and assign p-values to them? According to
> > my (very limited) knowledge of bayesian stats the outcome of the
> > analysis should be an updated distribution of the possible parameter
> > values.
> > 
> > The available documentation about blme is limited and/or I could not
> > find it. I realize that my question on blme hides another, much
> > broader, on how bayesian stats work; regarding the latter, a
> > suggestion of a good, practice-oriented reference book would be
> > appreciated.
> > 
> > Thank you in advance,
> > 
> > Marco
> 
>  (Started writing this before Doug's comments, which I agree are [as
> usual] thoughtful and sensible but think represent one point of view.)
> 
>   For a start, there's a paper that describes the approach in detail:
> Chung, Yeojin and Rabe-Hesketh, Sophia and Dorie, Vincent and Gelman,
> Andrew and Liu, Jingche. "A Nondegenerate Penalized Likelihood Estimator
> for Variance Parameters in Multilevel Models". Psychometrika
> doi:10.1007/s11336-013-9328-2
> 
>   As for the p-values; I would say that in this context they're not
> particularly philosophically coherent but do still represent a rough
> measure of strength of evidence ...
> 
>   cheers
>     Ben Bolker
> 


  PS Andrew Gelman has a thoughtful post on this subject (the fact
that in many practical cases there's insufficient information in the
data to determine the among-group variance) from a Bayesian perspective
(a "stalwart" Bayesian view, using informative priors, rather than
a "cringing" Bayesian view that tries to come up with a weakly
informative prior that fixes (?) the problem (?) of zero variances ...)

http://andrewgelman.com/2014/08/11/
   discussion-sander-greenland-posterior-predictive-checks/
(broken URL to make Gmane happy)