dredge doesn't work when specifying glmer optimizer

Question

I am trying to use dredge from the R package MuMIn with a global binomial glmer model. I find that I need to specify the optimizer with control = glmerControl(optimizer="bobyqa") for convergence. However, when I go to use dredge, I get an error. If I reduce the number of predictors in the model, I can remove the bobyqa specification, get convergence, and use dredge. Any way I can get dredge to go with glmerControl(optimizer="bobyqa")?

test.glob=glmer(exploitpark~X + as.factor(Y) + Z + A + B + (1|ID), 
                family=binomial(), 
                glmerControl(optimizer="bobyqa"), data=df)
options(na.action = "na.fail")   #  prevent fitting models to different datasets
test.Set = dredge(test.glob, beta=c("partial.sd"), extra = c("R^2"))

Fixed term is "(Intercept)"

Error in glm.control(optimizer = c("bobyqa", "bobyqa"), calc.derivs = TRUE, : unused arguments (optimizer = c("bobyqa", "bobyqa"), calc.derivs = TRUE, use.last.params = FALSE, restart_edge = FALSE, boundary.tol = 1e-05, tolPwrss = 1e-07, compDev = TRUE, nAGQ0initStep = TRUE, checkControl = list(check.nobs.vs.rankZ = "ignore", check.nobs.vs.nlev = "stop", check.nlev.gtreq.5 = "ignore", check.nlev.gtr.1 = "stop", check.nobs.vs.nRE = "stop", check.rankX = "message+drop.cols", check.scaleX = "warning", check.formula.LHS = "stop", check.response.not.const = "stop"), checkConv = list(check.conv.grad = list( action = "warning", tol = 0.001, relTol = NULL), check.conv.singular = list(action = "message", tol = 1e-04), check.conv.hess = list(action = "warning", tol = 1e-06)), optCtrl = list())

Ben Bolker · Accepted Answer · 2018-12-19T18:14:14.643

tl;dr probably a bug in MuMIn::dredge() - I'm still digging - but it seems to work OK if you leave out the extra="R^2" specification.

reproducible example

set.seed(101)
dd <- data.frame(x1=rnorm(200),x2=rnorm(200),x3=rnorm(200),
                 f=factor(rep(1:10,each=20)),
                 n=50)
library(lme4)
dd$y <- simulate(~x1+x2+x3+(1|f),
                 family=binomial,
                 weights=dd$n,
                 newdata=dd,
                 newparams=list(beta=c(1,1,1,1),
                                theta=1))[[1]] 
## fit model
m0 <- glmer(y~x1+x2+x3+(1|f),
            family=binomial,
            weights=n,
            data=dd,
            na.action="na.fail")

now try glmer()+dredge(), with and without optimizer specification

library(MuMIn)
d0 <- dredge(m0)
m1 <- update(m0, control=glmerControl(optimizer="bobyqa"))
d1 <- dredge(m1)

These all work - so the problem must be with some of the optional arguments. Testing that:

d0B <- dredge(m0, beta=c("partial.sd"), extra = c("R^2")) ## works
d1B <- try(dredge(m1, beta=c("partial.sd"), extra = c("R^2"))) ## fails

Which of the extra arguments is the culprit?

d1C <- dredge(m1, beta=c("partial.sd"))  ## works
d1D <- try(dredge(m1, extra=c("R^2")))   ## fails

If you really, really want your R^2 values you could download/unpack the source code to the package, edit line 101 of R/r.squaredLR.R as indicated below (add cl$control to the list of elements that are set to NULL, and re-install the package ...

===================================================================
--- R/r.squaredLR.R (revision 443)
+++ R/r.squaredLR.R (working copy)
@@ -98,7 +98,7 @@
        if(formulaArgName != "formula")
            names(cl)[names(cl) == formulaArgName] <- "formula"
        cl$formula <- update(as.formula(cl$formula), . ~ 1)
-       cl$method <- cl$start <- cl$offset <- contrasts <- NULL
+       cl$method <- cl$start <- cl$offset <- cl$control <- contrasts <- NULL
    }
    cl <- cl[c(TRUE, names(cl)[-1L] %in% names(call2arg(cl)))]
    if(evaluate) eval(cl, envir = envir) else cl

Aha. Indeed, removing extra = c("R^2")) does the trick. Thank you. — ecologist1234, Dec 19 '18 at 18:14

score 2 · Answer 2 · answered Dec 20 '18 at 07:30

The issue is in r.squaredLR (implied by extra = "R^2"), which tries to fit a glm null model with glmer's argument control = glmerControl(optimizer="bobyqa"). (I will try to implement a solution for this bug in the forthcoming version of MuMIn.)

In case of glmer (or mixed models in general) it may be better to use r.squaredGLMM rather than r.squaredLR. So you would need to provide dredge with a function that extracts the R^2 vector from the result of r.squaredGLMM (which returns a matrix). For example:

# (following Ben Bolker's example above))
# Fit a null model with RE (use a non-exported function .nullFitRE or specify it by hand:
nullmodel <- MuMIn:::.nullFitRE(m1)
# the above step is not necessary, but avoids repeated re-fitting of the null model.

dredge(m1, beta="partial.sd", extra =list(R2 = function(x) {
    r.squaredGLMM(x, null = nullmodel)["delta", ]
}))

dredge doesn't work when specifying glmer optimizer

2 Answers2

reproducible example

now try glmer()+dredge(), with and without optimizer specification

Linked