Message-ID: <85614903-5C4A-45AA-B1F7-A81417268A22@me.com>
Date: 2010-08-31T14:18:04Z
From: selling83 at me.com
Subject: ROC plots and predictions with glmm?
Dear List,
I am new to GLMM's and am trying to understand if a workflow i have used previously with GLM's will transfer to lmer.
I have a binary response variable - TRUE / FALSE and a series of 6 categorical factors with one interaction, and two random variables which account for phylogenetic structure.
I have run the model -
model1 <- lmer(response~1+(1|ord/fam) + a*b + c + d +e + f, family=binomial) which works fine.
Now i want to use the model to predict responses and am looking to validate the model with cross validation.
Using GLM my approach was
get predicted values using predict()
construct a ROC plot with ROCR or Epi.
cross validate with cv.glm
construct another ROC plot with cross validation lines to get a idea of variation etc.
I am thinking this is a common thing to do and people must want to do it with GLMM's? however with no predict function i don't know how this is possible?
Based on AIC criteria, the model that explains the data "best" for glm was also the best when using glmm, however i am unsure if this can justify using a glm and ignoring the random variables, even though i know they are important?
Any help would be greatly appreciated.
George