[R-sig-ME] How to deal with Hessian error if the model only contains binomial (0, 1) fixed factors?

[Ben Bolker]

ONKELINX, Thierry <Thierry.ONKELINX at ...> writes:

> 
> Dear Julia,
 
> Note that the smallest category of the response (samemate == 0) has
> 12 observations. According to a rule of thumb, you need 10
> observations in each response category per parameter in the
> model. Hence a model with one (if you're lucky two) parameters is
> doable.
 
> Complete separation appears when for a combination of covariates all
> responses are 0 or 1. Adding complexity to the model, increases the
> probability of complete separation because you have more possible
> combination and a lower number of observations per
> combination. Adding a random effect with 59 levels to a dataset of
> 66 observations guarantees complete separation...
 
> The separation of the simple glm(samemate ~ success) is not that
> bad. But glm(samemate ~ success + type) will give (quasi-)complete
> separation. The glm model gives implicit warnings (see the last
> model in your first email): type has a very large effect size and a
> huge standard error.

There is a short example of how to deal with complete separation
by setting weakly informative priors at http://rpubs.com/bbolker/glmmchapter
(search for "complete separation"), via the blme or MCMCglmm packages,
but I would second Thierry's opinion that you may be overcomplicating
things.

  Ben Bolker