[R-sig-ME] mixed mutlinomial regression for count data with, overdisperion & zero-inflation

[dave fournier]

Hi,

I ran your data outside of R.  It was immediately clear that the version 
of glmmADMB I had
in R was missing a feature.  the additional feature was the rescaling of 
badly scaled parameters
in the model.  Aftewr a simple rescaling I ran your data according to 
your model hypothesis
using both ZI and non ZI and Neg Bin type 1 and type 2 overdispersion.  
the model converged easily
and produced the following results.  the Objective function value is the 
negative log likelihood so smaller is better.
clearly zero inflation is indicated and type 2 fits a lot better than 
type 1.





Type 2
# Number of parameters = 33  Objective function value = 5698.40 Maximum 
gradient component = 7.96141e-05
# pz:
  0.0000
# Number of parameters = 34  Objective function value = 5672.69 Maximum 
gradient component = 4.96526e-05
# pz:
0.118535625347
Type 1
# Number of parameters = 33  Objective function value = 5954.95 Maximum 
gradient component = 9.43931e-05
# pz:
  0.0000
# Number of parameters = 34  Objective function value = 5936.85 Maximum 
gradient component = 7.44436e-05
# pz:
0.0684308821563