[R] inconsistency on p-value calculation of anova for quasi binomial

[Ulrich Halekoh]

Hej,

providing the dispersion parameter estimate to 
the anova function
for a quasibinomial fit results in two different
ways to calculate the p-value for the same statistic.


In the following example I test for the interaction effect.

In the versions  (a1 and a2) the p-value is based on the
F_1_17 distribution,
in the version (a3) it is calculated via the normal.

I think  anova should behave (by default) in all versions either
 in the one or the other 
way.


#example based on the orobanche data set provoded with the dispmod-package
library(dispmod)
data(orobanche)
orobanche$y<-with(orobanche,cbind(germinated,seeds-germinated))
g1<-glm(y~host+variety,family=quasibinomial,data=orobanche)
g2<-glm(y~host+variety+host:variety,family=quasibinomial,data=orobanche)

a1<-anova(g2,g1,test='F',dispersion=summary(g2)$dispersion)
a2<-anova(g2,test='F')

a3<-anova(g2,test='F',dispersion=summary(g2)$dispersion)



ulrich

R 1.9.0
Windows 2000


==============================================================
Ulrich Halekoh,  PhD                         Phone: +45 8999 1825
Biometry Research Unit                       Fax:   +45 8999 1300
Danish Institute of Agricultural Sciences    E-mail: ulrich.halekoh at agrsci.dk
Research Centre Foulum, DK-8830 Tjele, Denmark