Fitting Mixed-Effects Models Using Efficient EM-type Algorithms
The Journal of
Computational and Graphical Statistic
David A. van Dyk
Department of Statistics, Harvard University, Cambridge, MA
02138, U.S.A.
In recent years numerous advances in EM methodology have
lead to algorithms which can be very efficient when compared
with both their EM predecessors and other numerical
methods (e.g., algorithms based on Newton-Raphson). In this
paper we combine several of these new methods to develop
a set of mode-finding algorithms for the popular
mixed-effects model which are both fast and more reliable than
standard algorithms such as proc mixed in SAS.
We present efficient
algorithms for Maximum Likelihood (ML), Restricted Maximum
Likelihood (REML), and computing posterior modes (with conjugate
proper and improper priors). These algorithms
are not only useful in their own right, but also illustrate
how parameter expansion, conditional data augmentation, and
the ECME algorithm can be used in conjunction to form efficient
algorithms. In particular, we illustrate a difficulty in
using the typically very efficient PXEM (parameter-expanded
EM) for posterior calculations, but show how algorithms based
on conditional data-augmentation can be used. Finally, we
present a result that extends Hobert and Casella's (1996)
result on the propriety of the posterior for the mixed-effects
model under an improper prior, an important concern in Bayesian
analysis involving these models that when not properly understood
has lead to difficulties in several
applications.
Key Words: EM algorithm, ECME algorithm, Gaussian hierarchical
models, Posterior inference, PXEM algorithm, Random-effects models,
REML, Variance-component models, working parameters.
Return to David van Dyk's homepage.