[R-sig-eco] quasi-complete seperation in logistic regression

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 05/19/2011 02:41 PM, Stolen, D Eric (KSC-IHA-4400)[Innovative Health
Applications LLC] wrote:
> Hello; I am working on a logistic regression model in which I have
> quasi-complete separation on an explanatory variable (see table
> below).  The response variable is Success of parrot reintroductions,
> and one of the explanatory variables is PredThreat, a 3 category
> variable designating the level of predator threat to the population.
> When I fit the univariate logistic regression model Success ~
> PredThreat, I get a huge standard error, which I believe is an
> indication of the optimization algorithm failing due to
> quasi-complete separation.  I am testing a variety of models using
> information-theoretic model selection to judge which variables are
> important to reintroduction success.  My question concerns what to do
> to about PredThreat, since it appears to be an informative variable.
> First I'm wondering if I can trust the AIC value calculated from the
> model with PredThreat?  Second, to get at an effect size and also to
> include it in multivariate models, I thought of treating PredThreat
> ! as a continuous variable.  When I do that in the univariate model,
> I get a more reasonable parameter estimate and standard error, but a
> much lower AIC.  I'd really appreciate any insights in how to deal
> with this problem.

  You might look into the brglm, logistf, and arm packages which all
offer options for bias-reduced (or Bayesian) GLMs that should (?) do a
better job with separation ... ?
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAk3Vjj4ACgkQc5UpGjwzenOvrgCePLXdZ9hny71Suy4MHOvfRS3e
r0gAn2Iw7XQziXKHEIPpIeZhhmNZSkCx
=8Fty
-----END PGP SIGNATURE-----