[R-sig-ME] Complete-separation-problem-resolution

Discussion:

C. AMAL D. GLELE

2018-07-26 13:32:53 UTC

Hi, dear all.
Please, How do I solve complete separation problem in binomial model built
with glmmTMB or glmer (or with any other mixed model builder)?
In advance, thanks foe your replies.
Kind regards,

[[alternative HTML version deleted]]

John Poe

2018-07-26 13:43:05 UTC

Permalink

The short answer is a bayesian hierarchical model. See this paper from
political analysis

http://www.carlislerainey.com/papers/separation.pdf

Post by C. AMAL D. GLELE
Hi, dear all.
Please, How do I solve complete separation problem in binomial model built
with glmmTMB or glmer (or with any other mixed model builder)?
In advance, thanks foe your replies.
Kind regards,
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

[[alternative HTML version deleted]]

Ben Bolker

2018-07-26 13:59:48 UTC

Permalink

See also

http://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#penalizationhandling-complete-separation

http://bbolker.github.io/mixedmodels-misc/ecostats_chap.html#digression-complete-separation

Post by John Poe
The short answer is a bayesian hierarchical model. See this paper from
political analysis
http://www.carlislerainey.com/papers/separation.pdf

[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

James Uanhoro

2018-07-26 13:47:04 UTC

Permalink

Check out the blme package. You can specify priors for fixed effects; this
is one way of handling the complete separation problem.
For choice of prior, Greenland and Mansournia (
https://www.ncbi.nlm.nih.gov/pubmed/26011599) suggest using log-F(1, 1)
over other common suggestions like normal, or Jeffrey's or t-priors. But
blme provides a quick solution that limits you to normal or t.

[[alternative HTML version deleted]]