Re: Interpreting the coefficient in a GLS binomial model

Serguei Kaniovski <kaniovsk@xxxxxxxxxx> wrote in
news:32470734.1172401192625.JavaMail.jakarta@xxxxxxxxxxxxxxxxxxxxxx:

Thanks a lot. Yes, this is what I have been looking for.

I see that truedough pointed out his error on the method of calculating
predicted proportions. I see that I made the same error. Should have been
1/(1+1/exp(-(2.983 + (-1.450829*0.5)))

I have another question, though. I show by other means that the votes
typically are highly positively correlation (c=0.6 on average).

The basic model has good fit:
R-sq.(adj)=0.854, Deviance explained=36.7%
BUT massive overdispersion
Residual deviance: 4267.0 on 106 degrees of freedom

The overdispersion cannot be remedied by regressing on LOG(index), and
using the quasibinomial family with a scale parameter for the
variance. The estimated Dispersion parameter for quasibinomial family
is also very large 39.52906.

I am guessing an error in the inputs to the procedure. How many "for" and
"against" votes did you have? How many predictor variables were in the
model?


QUESTION: Can overdispersion be due to correlation between the votes?
And, what can be done?

How could there be correlation between votes? Are you analyzing votes on
more than one candidate per voter? It might help to show the data layout
with, say, 10-15 cases and also the input to your statistical package.

--
David Winsemius
.