Hi I'm using glmer (lme4) to analyse binomial data. I have used observation level to account for overdispersion. My problem is that a referee asked to get wald t or F instead of wald Z mentioning Bolker's review in TREE 2008. Can you suggest a way to get t or F instead of z in the output? this is the code I'm using now: y=cbind(data$Success,data$Failure) glmer.r<-glmer(y~ treat + (1|obs), data=data, family=binomial(logit), weights =total.size) summary(glmer.r) Many thanks in advance, Clelia
glmer z values binomial data
2 messages · Clelia Gasparini, Ben Bolker
Clelia Gasparini <clelia.gasparini at ...> writes:
Hi I'm using glmer (lme4) to analyse binomial data. I have used observation level to account for overdispersion. My problem is that a referee asked to get wald t or F instead of wald Z mentioning Bolker's review in TREE 2008. Can you suggest a way to get t or F instead of z in the output? this is the code I'm using now: y=cbind(data$Success,data$Failure) glmer.r<-glmer(y~ treat + (1|obs), data=data, family=binomial(logit), weights =total.size) summary(glmer.r) Many thanks in advance, Clelia
The Bolker 2008 t/F suggestion is based on using penalized quasi-likelihood (PQL) to account for overdispersion; lme4 no longer does PQL, because we decided that we don't really understand/aren't comfortable with the properties of PQL for GLMMs. * if you can decide on the degrees of freedom (see http://glmm.wikidot.com/faq#df , and scroll down to "DF alternatives"), then you can take the Z scores, reinterpret them as t scores, and use 2*pt(abs(scores),df=df,lower.tail=FALSE) to get p-values. * including a per-observation random effect should account for overdispersion anyway (see the "overdispersion" section of the FAQ) * I'm a little puzzled that you have specified the response as a two-column matrix *and* have separately specified the weights: normally you would do one or the other ... * the model specification you have above has no grouping variable/ random effect term other than the observation-level RE, which makes it look just like a regular binomial GLM + overdispersion. Therefore you might be better off/have an easier time with glm(y~ treat , data=data, family=quasibinomial) or a beta-binomial model.