R help - Jul 2003 - How to use quasibinomial?

Dear all,

I've got some questions, probably due to misunderstandings on my behalf,
related
to fitting overdispersed binomial data using glm().

1. I can't seem to get the correct p-values from anova.glm() for the F-tests
when
supplying the dispersion argument and having fitted the model using
family=quasibinomial. Actually the p-values for the F-tests seems identical to
the
p-values for the Chi-squared tests. When not supplying the dispersion argument,
i.e. when anova.glm() uses the default scaled Pearson statistic from
family=quasibinomial, both tests returns the p-values I'd expect. What am I
doing
wrong here and how can I make it work?
> fit.1<-glm(y/n~host*variety,family=quasibinomial,weights=n)
>
dscale<-sum(residuals(fit.1,type="deviance")^2/fit.1$df.residual)
> dscale
[1] 1.957517
> anova(fit.1,test="F",dispersion=dscale)
Analysis of Deviance Table
Model: quasibinomial, link: logit
Response: y/n
Terms added sequentially (first to last)
             Df Deviance Resid. Df Resid. Dev       F    Pr(>F)
NULL                            20     98.719
host          1   55.969        19     42.751 28.5916 8.937e-08
variety       1    3.065        18     39.686  1.5657    0.2108
host:variety  1    6.408        17     33.278  3.2736    0.0704

I expected:
> 1-pf(3.2736,1,17)
[1] 0.08812074
> anova(fit.1,test="Chisq",dispersion=dscale)
Analysis of Deviance Table
Model: quasibinomial, link: logit
Response: y/n
Terms added sequentially (first to last)
             Df Deviance Resid. Df Resid. Dev P(>|Chi|)
NULL                            20     98.719
host          1   55.969        19     42.751 8.937e-08
variety       1    3.065        18     39.686     0.211
host:variety  1    6.408        17     33.278     0.070

As expected:
> 1-pchisq(6.408/dscale,1)
[1] 0.07040576

2. When using summary.glm() on a glm object fitted using family=quasibinomial
the
reported tests are t-tests. Why?

Thanks,
Henric

I think I can now answer my own questions:

Henric Nilsson (hennil at statisticon.se) wrote*:
>1. I can't seem to get the correct p-values from anova.glm() for the
F-tests when
>supplying the dispersion argument and having fitted the model using
>family=quasibinomial.
Using family=quasibinomial does exactly what it says: It fits a quasi-likelihood
model with mean=mu and variance=mu(1-mu). Hence, supplying the dispersion
argument
results in R treating this as a known value, i.e. not estimated from data, in
which
case the F-test shouldn't be used and therefore returns the same p-value as
the Chi-
squared test.
>2. When using summary.glm() on a glm object fitted using
family=quasibinomial the
>reported tests are t-tests. Why?
This makes sense since we're estimating the dispersion parameter. Can't
really
remember what confused me here...

//Henric