[R] Negative binomial models and censored observations

[Ben Bolker]

Achim Zeileis <Achim.Zeileis <at> uibk.ac.at> writes:

> 
> On Thu, 13 Mar 2014, Tim Marcella wrote:
> 
> > Hi,
> >
> > I am working with hurdle models in the pscl package to model zero 
> > inflated overdispersed count data and want to incorporate censored 
> > observations into the equation. 33% of the observed positive count data 
> > is right censored, i.e. subject lost to follow up during 
> the duration of 
> > the study. Can this be accounted for in the hurdle() function?
> 
> No, this is currently not supported. If the censoring points are fixed 
> (e.g., counts of "5" actually mean "5 or more") then using an ordinal 
> model might be an alternative to using a count model. However, if the 
> censoring points differ, then I wouldn't know of a package that provides 
> this out of the box...
> 
> > Thanks, Tim

  It's not quite out of the box, but if you defined a censoring level
it wouldn't be *too* hard to write a log-likelihood model for the 
truncated/censored case and using bbmle.  It would be much less efficient
and stable than pscl, but it should work ...

  Ben Bolker