[R-sig-ME] parallel option of the bootMer function: warnings for parallel=no, but not for parallel=snow

[Ben Bolker]

On 14-10-21 09:45 AM, Spronken, M. (Maitta) wrote:
> Dear all,
> 
> I have a question about the parallel option of the bootMer function,
> I hope you can help me with this. I am using R 3.0.1.
> 
> I ran a mixed model with the lmer function of the package lme4_1.1-7
> and tried to obtain confidence intervals with the bootMer function
> (in order to get the same results each time, I used set.seed(80)).
> 
> In order to divide the bootstrapping calculations over several cores,
> I used the parallel option of the bootMer function: 'parallel=snow,
> ncpus=7'. However, I also bootstrapped the model using 'parallel=no',
> in order to see whether it led to the same results.
> 
> The bootstrapping results were the same for 'parallel=snow' and
> 'parallel=no'. However, when using 'parallel=no', I got the following
> warnings: 1: In checkConv(attr(opt, "derivs"), opt$par, ctrl =
> control$checkConv,  : Model failed to converge: degenerate  Hessian
> with 1 negative eigenvalues 2: In checkConv(attr(opt, "derivs"),
> opt$par, ctrl = control$checkConv,  : Model failed to converge:
> degenerate  Hessian with 1 negative eigenvalues
> 
> When using 'parallel=snow', these warnings did not occur. Maybe also
> important, I do not get these warnings when running the original
> model.
> 
> This is a problem I have encountered with several models. Does
> someone know why using 'parallel=snow' does not lead to convergence
> warnings, while using 'parallel=no' does? Are these warnings
> problematic? Below you find the code I used.
> 
> Thanks in advance for your help!
> 
> 
> Code:
> 
> model_1 <- lmer(DV_continuous ~ factor_threelevels + (1 +
> factor_threelevels|userId), data=combined_data, control =
> lmerControl(optCtrl = list(maxfun = 1000000)))
> 
> FUN_bootMer <- function(fit) { return(fixef(fit)) }
> 
> set.seed(80)
> 
> boot_model_1a <- bootMer(model_1, FUN_bootMer, nsim = 1000, type =
> "parametric", parallel = "snow", ncpus =7)
> 
> set.seed(80)
> 
> boot_model_1b <- bootMer(model_1, FUN_bootMer, nsim = 1000, type =
> "parametric", parallel = "no")
> 
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 

  Possibly related to https://github.com/lme4/lme4/issues/231

  Do you get warnings with parallel="multicore"?

  I have a suspicion that some of the internal structures of `lme4` may
not be getting reset properly when using refit() (which is used
internally by confint/bootMer), and that new copies are getting made
(and thus reset) with parallel="snow" but not with parallel="no" ... ?

  Ben Bolker