[R-sig-ME] lme4_0.999375-1: not clear what works and what doesn't
bates at stat.wisc.edu
Thu Jan 24 15:15:21 CET 2008
On Jan 23, 2008 8:59 PM, Michael Kubovy <kubovy at virginia.edu> wrote:
> Dear Mixed modelers,
> I realize that the package is in transition; I see that some examples
> have a comment #not run, which I take to mean "won't run for
> now" (e.g., mcmcsamp() ).
Yes. I still need to work out some details on mcmcsamp. They may be
easy to do or they may not - I can't tell yet.
> But I'm confused about some functions that don't work:
> What seems not to work:
> 1. simulate() "Error: 'simulate' is not implemented yet";
Indeed. The innards of the simulate function are closely tied to the
representation of the model. This has stabilized, relative to recent
major upheavals, but it is not yet completely stable. There isn't
much purpose in devoting a lot of energy to functions like simulate()
and, as shown below, plot() until the representation and the
optimization are set.
> 2. The examples in lme4::lmer that call nlmer() all say "Warning
> message: In mer_finalize(ans, verbose) : false convergence (8)"
That, and the handling of the weights argument, are the main hangups
at present. Updating weights for the residuals is slightly different
from updating the weighted cross products of model matrices and that
part of the code apparently still has some "infelicities". I suspect
the problem is in the logic of updating the model matrix weighted
> 3. plot(fm2 <- lmer(Reaction ~ Days + (1|Subject) + (0+Days|Subject),
> data = sleepstudy)))
> complains: "Error in as.double(y) : cannot coerce to vector"
To summarize. There are known problems in
use of weights argument for LMMs and GLMMs
Collateral damage from changing the class representation occurs in
simulate, plot and possibly other extractors.
More information about the R-sig-mixed-models