Skip to content
Back to formatted view

Raw Message

Message-ID: <alpine.OSX.2.00.1101111608590.7311@qatna.eawag.wroot.emp-eaw.ch>
Date: 2011-01-11T15:15:45Z
From: sovo0815 at gmail.com
Subject: Catch separability problems in a (multinom) regression

So far I have "learned" that a binary logit model can either not 
converge or show linear separability (two different warnings). And 
so far I assume that this is also the case for multinomial models. 
For these models, if such errors occur, the coefficients are not 
very useful. In a loop (bootstrapping XYZ repetitions), I can 
catch the convergence problem via the <model>$convergence variable 
(set to 1). But how can I catch cases of separability?

library(nnet)
y <- factor(rep(LETTERS[1:3], each=20))
a <- c(rep(1:4, each=10), rep(5:6, 10))
b <- c(rep(1:2, 20), rep(3:4, each=10))
mod <- multinom(y ~ a + b, model=T)


-- 
S?ren Vogel, sovo0815 at gmail.com, http://sovo0815.wordpress.com/