Parameter estimation for MLM in
SAS:
GLIMMIX
Penalized quasi-likelihood
equations
User specifiesGandR
NLMIXED
Maximized approximation to
likelihood integrated over
random effects
User does not specifyGandR
User specifies variance
components ofGmatrix and
assumes an independentR
matrix (i.e.,R¼s^2 I)
Mixed models are flexible:
Layer random components
Handle nested clusters
Control for subject effects
Performance of mixed logistic
models not fully evaluated
There are various methods that can be used for
parameter estimation with mixed logistic mod-
els. The parameter estimates, obtained for the
Heartburn Relief data from the SAS procedure
GLIMMIX use an approach termed penalized
quasi-likelihood equations (Breslow and Clay-
ton, 1993; Wolfinger and O’Connell, 1993).
Alternatively, the SAS procedure NLMIXED
can also be used to run a mixed logistic
model. NLMIXED fits nonlinear mixed models
by maximizing an approximation to the likeli-
hood integrated over the random effects.
Unlike GLIMMIX, NLMIXED does not allow
the user to specify a correlation structure for
theGandRmatrices (SAS Institute, 2000).Instead, NLMIXED allows the user to specify
the individual variance components within the
Gmatrix, but assumes that theRmatrix has an
independent covariance structure (i.e. 0s on
the off-diagonals of theRmatrix).Mixed models offer great flexibility by allowing
the investigator to layer random components,
model clusters nested within clusters (i.e., per-
form hierarchical modeling), and control for
subject-specific effects. The use of mixedlinear
models is widespread in a variety of disciplines
because of this flexibility.Despite the appeal of mixedlogisticmodels,
their performance, particularly in terms of
numerical accuracy, has not yet been ade-
quately evaluated. In contrast, the GEE
approach has been thoroughly investigated,
and this is the reason for our emphasis on
that approach in the earlier chapters on corre-
lated data (Chaps. 14 and 15).Presentation: IV. The Generalized Linear Mixed Model Approach 587