The models brmultinom
handles are also known as
baseline-category logit models (see, Agresti, 2002, Section 7.1),
because they model the log-odds of every category against a
baseline category. The user can control which baseline (or
reference) category is used via the ref
. By default
brmultinom
uses the first category as reference.
The maximum likelihood estimates for the parameters of
baseline-category logit models have infinite components with
positive probability, which can result in problems in their
estimation and the use of inferential procedures (e.g. Wad
tests). Albert and Andreson (1984) have categorised the possible
data patterns for such models into the exclusive and exhaustive
categories of complete separation, quasi-complete separation and
overlap, and showed that infinite maximum likelihood estimates
result when complete or quasi-complete separation occurs.
The adjusted score approach to bias reduction that
brmultinom
implements (type = "AS_mean"
) is an
alternative to maximum likelihood that results in estimates with
smaller asymptotic bias that are also *always* finite, even in
cases of complete or quasi-complete separation.
brmultinom
is a wrapper of brglmFit
that fits
multinomial logit regression models through the 'Poisson trick' (see, for
example, Palmgren, 1981; Kosmidis & Firth, 2011).
The implementation relies on the construction of an 'extended'
model matrix for the log-linear model and constraints on the sums
of the Poisson means. Specifically, a log-linear model is fitted on
a Kronecker product
(https://en.wikipedia.org/wiki/Kronecker_product) of the
original model matrix X
implied by the formula, augmented by
nrow(X)
dummy variables.
The extended model matrix is sparse, and the Matrix package
is used for its effective storage.
While brmultinom
can be used for analyses using
multinomial regression models, the current implementation is more
of a 'proof of concept' and is not expected to scale well with
either of nrow(X)
, ncol(X)
or the number of levels in
the categorical response.