[R-sig-ME] glm model with all zeros for one of the factor level

[Ben Bolker]

  Agreed. I added a section to the glmm FAQ giving guidance on how to to
this in the GLMM case:
http://bbolker.github.io/mixedmodels-misc/ecostats_chap.html#digression-complete-separation


On 17-12-06 09:36 AM, Paul Johnson wrote:
> (This question is about GLMs rather than mixed models in R.)
> 
> I recommend reading up on separation in logistic regression, where
> the proportion in any of the categories formed by the fixed effects
> is exactly 1 or 0, so that a maximum likelihood estimate of the log
> odds doesn't exists. The logistf package is the simplest way of
> dealing with this in R.
> 
> Good luck, Paul
> 
> 
> Sent from BlueMail<http://www.bluemail.me/r?b=11327> On 5 Dec 2017,
> at 16:00, Juan Pablo Edwards Molina
> <edwardsmolina at gmail.com<mailto:edwardsmolina at gmail.com>> wrote:
> 
> Dear List members:
> 
> I performed two independent experiments (CRD) to test if a whitefly 
> has preference to infect with a virus: potatos, tomatos or peppers 
> (target hosts, TH), wether if the virus was obtained from potato or 
> tomato (source hosts, SH). So I released 100 white flyes (previously 
> infected with the virus from one or other SH) inside cages
> containing 10 plants of each TH (30 total). This is how the data
> looks like:
> 
> exp  SH    TH    cage    tot  posit
> 
> 1     tom    tom     1       10     4 1     tom    bat      1
> 10     3 1     tom    pep     1       10     0 1     bat    tom
> 2       10     1 1     bat    bat       2       10     2 1     bat
> pep      2       10     0
> 
> 2     tom    tom     3       10     6 2     tom    bat      3
> 10     4 2     tom    pep     3       10     0 2     bat    tom
> 4       10     4 2     bat    bat       4       10     0 2     bat
> pep      4       10     0
> 
> The issue I found here  is that pepper was not infected at all, 
> however it was infected in another experiment without chance of TH 
> choice: i.e. I released infectious whiteflies inside cages
> containing the same pepper genotyope and they present the typical
> virus disease symptoms.
> 
> So, how should I consider modeling this data? Zero-inflated negative
> binomial using the total plants as offset? Hurdle-model? Should I
> remove the pepper level for the model?
> 
> Any help would be really helpful.
> 
> Juan Edwards
> 
> ________________________________
> 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 
> [[alternative HTML version deleted]]
> 
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>