[R-pkgs] New versions of Matrix and lme4 packages for R-2.4.0
Douglas Bates
bates at stat.wisc.edu
Tue Oct 3 16:56:00 CEST 2006
Versions 0.9975-1 of the Matrix and lme4 packages will soon be available on
CRAN for use with R version 2.4.0 or later.
Purpose of the packages:
The Matrix package provides S4 classes and methods for sparse
and dense matrices. The lme4 package provides functions for fitting
and assessing linear or generalized linear mixed effects models (also
called multilevel models). Like the Matrix package, the lme4 package
uses S4 classes and methods.
Features of this release:
- These package versions are designed for use with release 2.4.0 or
later of R. Because of enhancements in S4 classes and methods in
R-2.4.0 the packages load much faster than before and the functions
can be organized differently. In particular, the lmer function for
fitting linear or generalized linear models is again in the lme4
package. The C code of lme4 now calls C code that is exported from Matrix
Matrix package improvements:
- The Matrix() constructor function has become more versatile and you
are advised to read the help page and try its examples.
- Reorganization and simplification of source code (source package is
about 1/3 the previous size). Much greater use of virtual classes
in method definitions.
- LU and QR decompositions for sparse matrices using the CSparse
library of C functions written by Tim Davis and documented in his
2006 SIAM book "Direct Methods for Sparse Linear Systems". New
function Cholesky (in addition to function chol) that allows the
full range of CHOLMOD library options for sparse Cholesky
decompositions.
- New classes to distinguish between sparse logical matrices and
sparse pattern matrices.
- NAs are now properly dealt with in sparse and dense matrices.
- Printing of sparse matrices now distinguishes structural zeros and
values that can be nonzero but happen to be equal to zero.
- New functions triu, tril and band patterned after those in Octave
and Matlab.
lme4 package improvements:
- Allow specification of nested random effects using terms of the form
(1|block/field)
- Model specifications without fixed effects, such as travel ~ (1|Rail),
are allowed. Previously the implicit intercept term needed to be
explicitly specified as travel ~ 1 + (1 | Rail).
- Error declared when the number of levels of a grouping factor is not
less than the number of observations.
- Default estimation method for GLMMs is now optimization of the
Laplace approximation to the log-likelihood
- Reorganization and simplification of source code. Allow for models,
such as multilevel item response theory models or models with
carry-over of teacher effects, that extend the linear or generalized
linear mixed model.
- Fixed bug, reported by Goran Brostrom, in calculation of standard
errors of coefficients in a binomial GLMM where the response is
given as a matrix.
On behalf of the authors Martin Maechler and Douglas Bates
More information about the R-packages
mailing list