Skip to content
Prev 75780 / 398502 Next

GLMM - Am I trying the impossible?

It is not supported to call anova() on a glmmPQL fit.

For the glmmPQL fit you show, you have very large parameter estimates for 
a logistic and have partial separation (as you comment on for the control 
group): in that case PQL is not a reasonable method.

Try

fit <- glm(dead ~ Parasite * Bacteria, data=fish.data, family=binomial)
summary(fit)
anova(fit, test="Chisq")
fitted(fit)

and you have fitted values of zero (up to numerical tolerances).

This *is* discussed in MASS, around pp.198-9.

So there is little point in adding random efects for that model.  Now try

fit2 <- glmmPQL(dead ~ Parasite + Bacteria, random=~1|Tank,
                 family=binomial, data=fish.data)
summary(fit2)

Fixed effects: dead ~ Bacteria + Parasite
                 Value Std.Error  DF   t-value p-value
(Intercept) -4.243838 0.7519194 150 -5.644007  0.0000
Parasiteyes  1.264102 0.4205313   7  3.005964  0.0198
Bacteriayes  2.850640 0.7147180   7  3.988483  0.0053

which is pretty similar to the lmer fit you show.

I don't know what anova is doing for your lmer fit, but I do know that it 
should not be working with sums of squares as is being reported.
On Thu, 18 Aug 2005, Pedro J. Aphalo wrote:

            
I think you missed what glm was trying to tell you.