Skip to content

lme4 and deviance

2 messages · Nadège Jacot, Douglas Bates

Original
Nadège Jacot
# 
Dear list,

I'm trying to understand why the deviance function returns -2 log likelihood in lme4 and not the "true" deviance as with lm().

The scaled deviance is defined as the difference of -2 log likelihood between the model of interest and the saturated model but what is the saturated mixed model? For balanced data, some authors (e.g. Hoffman 2014, Longitudinal Analysis: Modeling Within-Person Fluctuation and Change) define the saturated mixed model as a model with saturated means (one fixed effect for each time point) and unstructured variance but such a model is not available for unbalanced data.

And if we can compute the scaled deviance, what is the deviance?

Thanks in advance for any hint.

Nad?ge Jacot
6 days later
Douglas Bates
# 
I share your concern about how to define the scaled deviance.  The trick,
as you say, is deciding what the saturated mixed model is.  I don't know of
a good way of defining it.

It might be best to just refer to the log-likelihood and not derive
anything called "deviance".
On Thu, May 21, 2015 at 9:09 AM Nad?ge Jacot <Nadege.Jacot at unige.ch> wrote: