Skip to content
Prev 1286 / 20628 Next

understanding log-likelihood/model fit

On Wed, Aug 20, 2008 at 2:24 PM, Andrew Robinson
<A.Robinson at ms.unimelb.edu.au> wrote:
I approach this from a slightly different point of view than does
Andrew.  I prefer to think of the process of estimating the fixed
effects parameters and the random effects as a penalized estimation
problem where the penalty is applied to the random effects only.  The
magnitude of the penalty depends on the (unconditional) variance
covariance matrix of the random effects.  When the variances are small
there is a large penalty.  When the variances are large there is a
small penalty on the size of the random effects.  The measure of model
complexity, which is related to the determinant of the conditional
variance of the random effects, given the data, has the opposite
behavior.  When the variance of the random effects is small the model
is considered simpler.  The simplest possible model on this scale is
one without any random effects at all, corresponding to a variance of
zero.  The larger the variance of the random effects the more complex
the model.

The maximum likelihood criterion seeks a balance between model
complexity and fidelity to the data.  Simple models fit the data less
well than do complex models.

The point to notice, however, is that the penalty on model complexity
applies to the random effects only, not to the fixed effects.  From
this point of view if the model can explain a certain pattern in the
data either with fixed-effects parameters or with random effects there
is an advantage in explaining it with the fixed effects.
Because the likelihood involves a penalty on the size of the variance
of the random effects.  Obtaining a similar fit (fidelity to the data)
with a much lower variance on the random effects (the penalty on model
complexity) is advantageous under the likelihood criterion.