GLMM- Is it possible to limit variable combinations? - r

Question

Let's assume I want to see all possible variable's combinations of a GLMM (using lme4) but I don't want to consider two variables at the same time in a model. How do I do that? For instance, I want to consider 3 fixed effects and 3 random effects, but I don't want any of the random or fixed effects to be considered at the same time in a model. If I construct the model this way:
model1 <- glmer(x~var1+var2+var3+(1|var4)+(1|var5)+(1|var6),
data=data1)
and I use MuMIn::dredge() function (to perform model averaging later), I will get all possible combinations between them, but I don't want (1|var4) to be in the same model as (1|var5).
So, is it possible to limit model combinations? This way I would avoid unnecessary models and save computing time.

Answer 1

Just in case anyone is looking for an updated solution to this... You can now do this easier with the subset argument of dredge:
#full model
model1 <- glmer(x~var1+var2+var3+(1|var4)+(1|var5)+(1|var6),data=data1)
#exclude models containing both (1|var4) & (1|var5) at the same time
dredge(model1, subset = !((1|var4) && (1|var5)))

Answer 2

I don't know how to do this within MuMIn::dredge() (but see my attempts below).
set.seed(101)
dd <- data.frame(x=rnorm(1000),
var1=rnorm(1000),
var2=rnorm(1000),
var3=rnorm(1000),
var4=sample(factor(sample(1:20,size=1000,replace=TRUE))),
var5=sample(factor(sample(1:20,size=1000,replace=TRUE))),
var6=sample(factor(sample(1:20,size=1000,replace=TRUE))))
library(lme4)
m0 <- lmer(x~var1+var2+var3+(1|var4)+(1|var5)+(1|var6),dd,REML=FALSE,
na.action=na.fail)
If we try to use the m.lim argument it subsets only the fixed effects, but leaves in all the random effect terms:
dredge(m0,m.lim=c(0,1))
## Model selection table
## (Intrc) var1 var2 var3 df logLik AICc delta weight
## 1 0.02350 5 -1417.485 2845.0 0.00 0.412
## 3 0.02389 -0.03256 6 -1416.981 2846.0 1.02 0.248
## 5 0.02327 0.02168 6 -1417.254 2846.6 1.56 0.189
## 2 0.02349 -0.002981 6 -1417.480 2847.0 2.02 0.151
## Models ranked by AICc(x)
## Random terms (all models):
## ‘1 | var4’, ‘1 | var5’, ‘1 | var6’
Following demo(dredge.subset), I tried this as an example:
dredge(m0,
subset=expression(!( (var1 && var2) || ((1|var4) && (1|var5)))))
but got
Error in dredge(m0, subset = expression(!((var1 && var2) || ((1 | var4) && :
unrecognized names in 'subset' expression: "var4" and "var5"
I can't find any documentation on how to do dredging/model averaging with MuMIn::dredge() across models with different random effects (indeed, I'm not convinced this is a good idea). If you wanted to fit all models with exactly one fixed-effect and exactly one random-effect term, you could do it as follows:
Set up all combinations:
fvars <- paste0("var",1:3)
gvars <- paste0("(1|var",4:6,")")
combs <- as.matrix(expand.grid(fvars,gvars))
Now fit them:
mList <- list()
for (i in 1:nrow(combs)) {
mList[[i]] <- update(m0,
formula=reformulate(combs[i,],response="x"))
}
Now you can use lapply or sapply to operate on the elements of the list, e.g.:
lapply(mList,formula)
## [[1]]
## x ~ var1 + (1 | var4)
##
## [[2]]
## x ~ var2 + (1 | var4)
##
## [[3]]
## x ~ var3 + (1 | var4)
##
## [[4]]
## x ~ var1 + (1 | var5)
## ... et cetera ...
bbmle::AICtab(mList,weights=TRUE)
## dAIC df weight
## model5 0.0 4 0.344
## model6 0.5 4 0.262
## model4 1.0 4 0.213
## model8 4.1 4 0.044
## ... et cetera ...
... but you'll have to work a bit harder to do model averaging. You might try r-sig-mixed-models#r-project.org, r-sig-ecology#r-project.org, or e-mail the maintainer of MuMIn (maintainer("MuMIn")) ...