Skip to content
Prev 58741 / 398502 Next

Some questions to GLMM

Jan Hattendorf wrote:
My guess is that your model is overparameterized.  Notice that the 
estimated correlations of the random effects are all near the extremes 
of -1 or +1.  I would start with a model that had fewer random effects 
terms.
It is a common misconception that a log-likelihood must be negative.  In 
fact, a log-likelihood can be positive when it is based on a probability 
density.  Probabilities cannot exceed one but probability densities can. 
  In this case the log-likelihood is a combination of the probability 
density for the random effects and the conditional probability of the 
observations.
There is an indication of overdispersion but I would not worry about 
that until I could get a handle on the random effects terms.
They fit the same model but with a different set of coefficients hence 
the estimated values are different.
The GLMM function evaluates the likelihood at the PQL estimates using 
the Laplacian approximation so the result will be different from that 
returned by glmmPQL.
No, it is an indication that the variance-covariance of the random 
effects is not positive definite.
You will need to reduce the number of random effects terms.
Same as above.