I have uploaded a new version (0.30-2) of glmmML to CRAN today.
This is a rather extensive upgrade, mostly internal. Adaptive
Gauss-Hermite quadrature (GHQ) is now used for the evaluation of the
integrals in the log likelihood function. The user can choose the number
of points (default is 16), I _think_ that choosing 1 point will result
in a Laplace approximation. The integrals in the score and hessian
are evaluated by the QUADPACK function 'Rdqagi' which is the C code
behind the R function 'integrate'. This specific combination of the two
methods seems to work best. (I often get _exactly_ (up to seven digits)
the same value with the two methods, but in some extreme cases one may
fail and not the other.)
New components in the output from 'glmmML' are 'posterior.means'
and
posterior.modes'. The modes are found by using 'vmmin' (behind
R's
'optim') on the integrands in the GHQ, the means by numerical
integration. Usually, they do not differ much.
A special problem is situations where the random effects variance is
very small or zero. I may happen that glmmML is unable to get the
likelihood value above the value given by 'glm' on the corresponding
model with no clustering. In such a case zero variance is reported, with
a standard error that is NA. A warning is also given. If a test of the
hypothesis that sigma = 0 is on the wish list, a p-value can be
estimated by bootstrapping, see the input parameter 'boot'. The only
option now is a parametric bootstrap; I have removed the 'conditional'
approach.
As usual, comments, and error and bug reports are welcome.
G?ran
--
G?ran Brostr?m
_______________________________________________
R-packages mailing list
R-packages at stat.math.ethz.ch
https://stat.ethz.ch/mailman/listinfo/r-packages
