[R-sig-ME] lme4 vs nlme?

Ben Bolker bbolker at gmail.com
Wed Mar 16 21:45:23 CET 2011

On 11-03-16 04:24 PM, Dominick Samperi wrote:
> Hello,
> I just read through Bates's excellent notes on lme4a and now I'm trying to
> compare what I learned with what is in nlme. The first thing I notice is that
> lme4a is not restricted to linear modeling (as its name and the lme4a package
> description might suggest).
> Second, I do not see the equivalent of pdMatrix, varIdent, etc. in lme4a.
> That is, I don't see how to specify different correlation structures for
> the random effect, or dependency information for the error terms, like
> heteroskedasticity, autocorrelation, etc.
> Are the latter missing because lme4a is not finished?
> What is the intended scope of lme4a and how does
> this compare with nlme?

  I have the impression from what Doug Bates (the primary author and 99%
heavy lifter) has said in the past on this list, nothing rules out
implementing equivalents of any of {pdClasses,varClasses,corClasses},
but there is a lot to work on before those come up.  Furthermore,
integrating these components with GLMMs will be challenging (if possible
at all).  Guessing wildly, I would say that an equivalent to pdClasses
would be easier/quicker than either of the other two.

  Ben Bolker

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