[R-sig-ME] MCMCglmm update
Jarrod Hadfield
j.hadfield at ed.ac.uk
Wed Oct 7 07:39:18 CEST 2015
Hi,
MCMCglmm has been updated to version 2.22. A lot of minor annoying
bugs have been fixed, but as far as I am aware no major bugs have been
found. Quite a bit of new functionality has been added:
1) Antedependence structures.
Structured antedependence models can now be fitted using the new
variance structure ante[]. The suffix [] takes a number, giving the
order of the antedependence model (e.g ante1 and ante2 give first and
second order antedependence models), and the number can be prefixed by
a 'c' to hold all regression coefficients of
the same order equal. The number can also be suffixed by a
'v' to hold all innovation variances equal. For example,
antec2v has 3 parameters: a constant innovation variance, and two
constant regression coefficients (one 1-lagged, and one 2-lagged).
Priors for antedependence structures allow priors to be placed
directly on the regression parameters via a beta.mu (a vector of prior
means) and a beta.V (a matrix of prior variances) element to the prior
list
2) Path analysis.
Path analysis could be performed previously using the sir function,
but it was cumbersome and did not work if all response variables were
not Gaussian and completely observed. The path function is less
flexible than the sir function, but it is easier to use and works with
non-Gaussian data. Paths are allowed between observations within the
same residual block, and path(cause, effect, k) specifies which of
the k variables affect each other. For example, if a three-response
model was fitted then
cbind(a,b,c)~trait+path(1,2,3)+path(1,3,3), rcov=~us(trait):units
then states that a[i] determines b[i] and c[i].
3) Simulate
A simulate method now exists and can be used to simulate observations
from a model defined by a MCMCglmm object.
4) Predict
The predict method is now more complete and accepts new data
5) Random effect - residual correlations
Random effect - residual correlations can now be fitted by specifying
covu=TRUE in the prior specification for the residual structure. The
set of residuals defined by this structure are allowed to covary with
the random effects specified by the final random effect structure. If
the residual (co)variance matrix is of dimension n, and the final
random effect (co)variance matrix is of dimension m, then the residual
prior specification must be of dimension n+m. The final random effect
(co)variance matrix should not have a prior specification.
6) Random effect Bradley-Terry models
Bradley-Terry models without random effects could already be fitted in
previous versions by simply taking the difference between the two
opponents predictors (and potentially fixing the intercept at zero if
no order effects were modelled). Random effects can now be fitted
using the multimembership model formulation mm(opponent1-opponent2),
which now allows a `-' as well as the traditional `+'.
Cheers,
Jarrod
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
More information about the R-sig-mixed-models
mailing list