[R-sig-ME] [R] Missing interaction effect in binomial GLMM with lmer

[David Duffy]

On Sun, 14 Feb 2010, Ben Bolker wrote:

>
>  I don't think this is necessarily a mixed model question (as you
> ...  Is it not just a reasonable conclusion
> that there is not a significant difference between (the difference
> between 29/30 and 16/30) and (the difference between 29/30 and 5/30)?
>

What he said ;)

Perhaps the OP is confused by the fact that the subgroup analysis gives a 
"signficant" association in one subgroup and not the other, but the 
difference in association strength between the two groups is not 
significant -- this is quite common.

It seems that the maternal random effect variance can be set to zero, eg

library(glmmML)
m1 <- glmmML(Hatched ~ Origin + Treat, cluster=Female, data=hatch.frame)

Scale parameter in mixing distribution:  0.6484 gaussian
Std. Error:                              0.5779

m2 <- glm(Hatched ~ Origin + Treat, family=binomial,  data=hatch.frame)
LRTS <- m1$deviance-m2$deviance

so the results of the fixed effects GLM are a reasonable thing to look at,
and as Ben Bolker pointed out reduce to testing equality of odds ratios in 
two 2x2 tables.

  t1 <- with(hatch.frame,as.data.frame(table(Origin,Treat,Hatched)))
  library(exactLoglinTest)
  mcexact(Freq ~ Origin+Treat+Hatched+Treat:Hatched+Origin:Hatched,
          data=t1)

is one approach to an exact test of the significance of the interaction 
term, I think.  When I dug up my copy of Donald and Gart's IC2x2 program, 
their exact test for interaction P=0.52.

-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v