Why does paramertic bootstrapping a mixed model with bootMer generate warnings?

Vote count:

0

What causes convergence warnings ("degenerate Hessian" et al) while doing parametric bootstrapping of models fit with R's glmer?

When I fit the initial model I am error free, but the bootstrapping can throw up to 50 warnings out of 500 bootstraps. These warnings seem to be correlated with highly biased bootstrap results. Is this most likely to occur with smaller sample sizes or when probabilities are relatively low or high? I'm modeling data from a 2x2 factorial design with probabilities for treatment combinations ranging from 0.3 to 0.8.

I've been trying to replicate this using the classic salamander mating data with only limited success (code below). Sometimes I can run 500 bootstraps and get 4ish errors, sometimes it runs clean. Reducing the sample size seems loosely correlated with errors, but seems most likely to generate the "some bootstrap runs failed" warning.

I'm interested practically in getting good results from my bootstrapping, but am also curious why this is happening. I can understand why classic non-parametric bootstrapping can be problematic with mixed effects models, especially if things are not balanced. But for parametric bootstrapping I don't totally understand how parameters from a well-behaved model could generate simulated data that couldn't then be fit by the same model structure.

Salamander data:

library(spaMM);data(salamander)

model data with reduced sample size (eg, use -c(1:100) to drop data from dataframe)

sala.redox <- glmer(Mate ~ TypeF + TypeM + (1|Cross), 
                   data = salamander[-c(1:100), ],
                   family = binomial)

parametric bootstrap

confint(sala.redox, method = "boot",
        nsim = 500, verbose = TRUE)

asked 1 min ago

N Brouwer

644

Why does paramertic bootstrapping a mixed model with bootMer generate warnings?