Skip to content

Question on significancy of terms

3 messages · Leonel Lopez, Emmanuel Curis, Geoff Brookshire

Original
1 day later
Emmanuel Curis
# 
Hello Leonel,

I think your problem is not specific of GLMM. The (fixed-part) models
a*b = a+b+a:b and b+a:b will lead to similar results, but expressed
with a different set of coefficients.

Assume a and b are two-levels factors (a1,a2 and b1,b2) for sake of
simplicity : the first model will have coefficients
 - associated to a   : for a2
 - associated to   b : for b2
 - associated to a:b : for a2:b2

The second model will have coefficients
 - associated to   b : for b2
 - associated to a:b : for a1:b2 and a2:b2

(you can check this also with the simple lm function, by the way).

So they have the same number of coefficients and, I think, express the
same linear relationship but in a different basis.

So the likelyhood will be the same...

I never tested that with GLMM, but I guess it is the same problem as
with lm and lmer...

Best regards,
On Tue, Sep 11, 2012 at 08:29:03PM +0100, Leonel Lopez wrote:
? Hi lme4 users:
? I am new to mixed models in R and started using lme4 and apart of all I?m not an statistician.
? I am developing a GLMM model like the one below which contains the independent effect and interaction of two factors, plus the repeated measurement effect on individuals.
? Y~a*b+(1/Ind)I am using theglmer( function, family=poisson.?This model has three terms (a+b+a:ab). If I want to test the significance of the three terms I am using the?likelihood ratio test (anova) comparing the model?with the term and without the term
? Let say for example:
? m1<-glmer(y~a*b+(1/ind), family=poisson)
? m2<-glmer(y~a+b+(1/ind),?family=poisson)
? anova(m1,m2)
? but, how to test the significancy of only "a" or only "b"
? For example to test a:
? 
? m1<-glmer(y~a*b+(1/ind),?family=poisson)
? m3<-glmer(y~b+a:b+(1/ind),?family=poisson)
? anova(m1,m3)
? 
? This seems to be a incorrect model development as the LRT gave exacttly the same values for the two models.
? 
? 
? I am only considering transform my variable, get a normal distribution and conduct the analysis with the lmer() function.
? 
? I?ll be very grateful for your help!!
? 
? Cheers
? 
? Leo
? 	[[alternative HTML version deleted]]
? 

? _______________________________________________
? R-sig-mixed-models at r-project.org mailing list
? https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
Geoff Brookshire
# 
Hi there,

I asked a similar question about a year ago:
https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q3/006690.html

To test for the significance of a, you could either use pvals.mcmc (in
the languageR package) to get an MCMC p-value, or do Wald chi-square
tests of model fits without the interaction term like so:

m1 <- glmer(y ~ a + b + (1|ind), family=poisson)
m2 <- glmer(y ~ a + (1|ind), family=poisson)
anova(m1, m2)

(Make sure you use the pipe character | for random effects.)

cheers,
Geoff
On Tue, Sep 11, 2012 at 3:29 PM, Leonel Lopez <llopezt2004 at yahoo.co.uk> wrote: