[R] Zero-inflated neg.bin. model and pscl package
Ben Bolker
bolker at zoo.ufl.edu
Sun Dec 4 03:43:12 CET 2005
Simone Vincenzi <svincenz <at> nemo.unipr.it> writes:
> I don't understand from the sentence in the pscl guide "Zero-inflated count
> models are a type of two-component mixture model, with a component for zero
> counts, and the other component for the positive counts" if:
> a)to get true estimate of the relative mean abundance, the model multiply
> the relative mean abundance at a site by the probability that the relative
> mean abundance at a site is generated through a negative binomial
> distribution, as proposed by Lambert (Technometrics, 1992). By using this
> kind of mixture model, zeros arise from one or two processes and their
> related covariates.
> b) we have two independent models, where the first part is a binary outcome
> model and the second one is a negative binomial model, assuming that zeros
> arise from a single process and set of covariates, as proposed by Dobbie and
> Welsh (Austr.N.Z.J.Stat., 2001)
>
> Thanks
>
> Simone Vincenzi
> PhD student in Ecology, University of Parma, Italy
>
From your descriptions and a quick look at the two papers you cite,
my main conclusion is that I don't really think these are actually
different models -- just different descriptions of the same statistical
model.
Digging into the code in the package and looking at
?zeroinfl shows that the package is fitting a binomial
model for the probability of a structural zero and
a Poisson or negative binomial for the result otherwise ...
you can specify covariates for both models.
As a minor point: both references you cite actually focus
on ZI Poisson (not NB) regression models, although Dobbie and
Welsh do allow for overdispersion ...
hope that helps
Ben Bolker
More information about the R-help
mailing list