Forum: help

LME4 and SAS GLIMMIX Differences [ Reply ]
By: Bob Kuhn on 2011-03-07 03:58

This email refers the example on page 11 of the PROC GLIMMIX documentation,

see http://support.sas.com/rnd/app/papers/glimmix.pdf

The SAS code for running this problem is:

proc glimmix data=multicenter;
class center group;
model sideeffect/n = group / solution;
random intercept / subject=center;

Note that in the SAS code, the form of the dependant variable causes SAS to correctly assume that the family is binomial.

The R code that I used for the same problem is:

noside=n-sideeffect
d=cbind(sideeffect,noside)
m1=glmer(d ~ group+(1|center),family=binomial,dat)

Note that the data contains info on 30 centers, 15 in each group.

Partial SAS output follows:

Solutions for Fixed Effects
Standard
Effect group Estimate Error DF t Value Pr > |t|
Intercept -0.8071 0.2514 14 -3.21 0.0063
group A -0.4896 0.2034 14 -2.41 0.0305
group B 0 . . . .

The partial output from my R run is given below.

Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.3370 0.2597 -5.148 2.63e-07
groupB 0.4995 0.2038 2.451 0.0142

I have two questions:

1. In the test for GROUPS there is a slight difference between the absolute value of the Z value (2.41 vs 2.451). Is this difference caused by an error in my R code for LME4 (I checked that the data was read in correctly) or the fact the two programs use a different algorithm?

2. The biggest difference in ths P value between GLIMMIX and LME4 is caused by the fact that GLIMMEX uses the t distribution with 14 df while LME4 uses the normal approximation. For the z value of 2.451 the Normal approximation P value is 0.0142 while at 14 df, the t value is 0.0280. Do you have any comments on which P value is a better approximation?