[R-sig-ME] AIC and BIC with ML and REML
d.rizopoulos at erasmusmc.nl
Tue Mar 31 09:21:04 CEST 2009
Gorjanc Gregor wrote:
>> In a way this question relates to the earlier discussion on whether to
>> regard the random effects as parameters or as unobserved random
>> variables. The difference between -2 * l + 2 * p and -2 * l + 2 * (p
>> + q) can be considered to be a question of how many parameters there
>> are in the model. In particular, do the random effects count as
>> parameters? I had an interesting discussion with Georges Monette
>> about this a few days ago and both of us feel the saying the random
>> effects don't affect the parameter count is underestimating the
>> complexity of the model but saying they should add q to the number of
>> parameters is overestimating the complexity.
> Thanks for this comment. I will take this as a yes to: "Are there are several
> definitions of AIC and BIC lurking around?".
> DIC "solves" the issue of effective number of parameters to some extent
> though it also has its own problems
there is also the conditional AIC of Vaida and Blanchard (Biometrika,
2005, 351--370) that calculates the effective degrees of freedom for the
random effects. Both these authors and the DIC authors distinguish
between the focus of inference (or better the focus of prediction) in
mixed models, i.e., either the marginal log-likelihood or the
conditional on the random effects log-likelihood.
>> I don't know a good answer to the question of how to "count" the
>> number of parameters in a mixed model (but I also don't feel that AIC
>> or BIC should be taken too seriously - these quantities are, at best,
>> a guide for model comparisons).
> R-sig-mixed-models at r-project.org mailing list
Department of Biostatistics
Erasmus University Medical Center
Address: PO Box 2040, 3000 CA Rotterdam, the Netherlands
More information about the R-sig-mixed-models