[R-sig-ME] Caution - Big changes in lme4 on Saturday
bates at stat.wisc.edu
Fri Jun 20 19:42:18 CEST 2008
On Fri, Jun 20, 2008 at 10:42 AM, hadley wickham <h.wickham at gmail.com> wrote:
>> Please be cautious about installing the new version if you depend on
>> using mcmcsamp in its current form. It would be a good idea to keep a
>> backup copy of the current CRAN version of the lme4 package in case
>> you find you want to back out the change.
> For those of you who don't already know about it, the installWithVers
> parameter to install.packages may be of use. This allows you to keep
> previous versions of the package around and refer to them by package
Thanks for the reminder, Hadley. Indeed, using installWithVers is recommended.
A few things to watch out for. Bin Dai is working on a Google Summer
of Code project incorporating the Adaptive Gauss-Hermite Quadrature
method for some types of generalized linear mixed models and nonlinear
mixed models. We realized that my previous stub code for AGQ where
the AGQ option would be specified with the `method' argument wouldn't
work. You must specify the number of quadrature points per axis.
Because PQL is no longer offered (nor is the Lindstrom-Bates
algorithm, with a name like that I didn't feel I could trust it :-)
the only distinction is Laplace versus AGQ. We therefore changed to
an argument nAGQ with a default of 1, which is the Laplace
approximation, in glmer and nlmer. In lmer the method argument was
only used to select ML or REML so we changed to a REML argument with a
default of TRUE. We will continue to honor a specification of method
= "ML" or method = "REML", because many published sources use it, but
you are encouraged to change to REML = FALSE if you want ML estimates.
Similarly the argument verbose = TRUE is preferred to the older form
control = list(msVerbose = TRUE). The ECME steps are no longer used
nor is the analytic gradient. The niterEM and gradient components of
the control list are still recognized but have no effect other than to
produce a warning that they are not active.
More information about the R-sig-mixed-models