[R] geepack(geese) vs. glmmPQL

[Dieter Menne]

Dear R-List,

I have a data set (available on request) of observations from
a 2x2 conditions psychological experiment, with count of
correct/wrong decisions, and vp as the subject id.

 $ vp      : Factor w/ 10 levels "1","10","2","3",..: 1 1 1 1 1 1 1 1 2 2
...
 $ bed     : Factor w/ 2 levels "Hz15","Hz1": 1 1 1 1 2 2 2 2 1 1 ...
 $ task    : Factor w/ 2 levels "n","p": 1 1 2 2 1 1 2 2 1 1 ...
 $ decision: Factor w/ 2 levels "corr","wrong": 1 2 1 2 1 2 1 2 1 2 ...
 $ count   : int  20 14 25 26 5 29 27 7 7 10 ...

Analyzed by glmmPQL, the result is meaningful and "as expected"

--------------------------------------------
gep<-glmmPQL(decision~bed+task+bed*task,data=mag3,random=~1|vp,
             weights=count, family=binomial )

Fixed effects: decision ~ bed + task + bed * task
              Value Std.Error DF t-value p-value
(Intercept)  -0.666     0.504 65   -1.32  0.1910
bedHz1        1.187     0.746 65    1.59  0.1163
taskp         1.359     0.663 65    2.05  0.0443
bedHz1:taskp -2.328     0.986 65   -2.36  0.0213

I first had tried with geese (from geepack), but
the results were totally off.
Is this a convergence or my error in specifying
the model?

------------------------
geese(decision~bed+task+bed*task,data=mag3,id=vp,
             weights=count, family=binomial)

 Coefficients:
              estimate   san.se wald p
(Intercept)   3.36e+16 6.48e+14 2696 0
bedHz1        3.14e+16 9.54e+14 1081 0
taskp         3.54e+16 5.50e+14 4155 0
bedHz1:taskp -6.18e+16 1.14e+15 2934 0

Scale Model:
 Scale Link:                identity

 Estimated Scale Parameters:
            estimate   san.se     wald p
(Intercept) 2.31e+15 1.36e+38 2.91e-46 1

Correlation Model:
 Correlation Structure:     independence

Returned Error Value:    1
Number of clusters:   10   Maximum cluster size: 8


------------

Dieter Menne