I am trying to select out variables using a correlation matrix as the "subset" argument with dredge() in MuMIn (R).
My problem is precisely as described in this unresolved post: I run a model fm1*, and use dredge to test all combinations of variables. To exclude some combinations I use a subset matrix sub1 following the method described in dredge.subset demo.
> sub1
X1 X2 X3 X4 X5 X6 X7 X8 X9 X10
X1 NA NA NA NA NA NA NA NA NA NA
X2 TRUE NA NA NA NA NA NA NA NA NA
X3 FALSE FALSE NA NA NA NA NA NA NA NA
X4 FALSE FALSE TRUE NA NA NA NA NA NA NA
X5 FALSE FALSE FALSE FALSE NA NA NA NA NA NA
X6 FALSE FALSE FALSE FALSE TRUE NA NA NA NA NA
X7 TRUE FALSE TRUE FALSE FALSE TRUE NA NA NA NA
X8 FALSE TRUE TRUE TRUE TRUE FALSE TRUE NA NA NA
X9 TRUE TRUE TRUE TRUE FALSE TRUE FALSE FALSE NA NA
X10 TRUE TRUE TRUE TRUE TRUE FALSE FALSE FALSE TRUE NA
It works fine with up to 9 variables; with more I get an error message:
> form
N ~ X1 + X2 + X3 + X4 + X5 + X6 + X7 + X8 + X9 + X10
> fm1<-glm(form,data=dfmod)
> ms1<-pdredge(fm1,subset=sub1,cluster=clust)
Warning message: In pdredge(fm1, subset = sub1, cluster = clust) :
non-missing values exist outside the lower triangle of 'subset'
This is not true, as shown by:
any(!is.na(sub1[!lower.tri(sub1)]))
[1] FALSE
In addition to the warning, the model selection table includes sets of variables that were not allowed by the correlation matrix.
I did find a workaround by converting the matrix sub1 into a logical expression, which then works fine as subsetting condition. But it would be interesting to understand what's going on with the matrix.
*note that the same happens with glm and glmer
