CODEWITHSUNDEEP
rlme4mixed-models
isDotR
isDotR

Reputation: 1063

Setting rhoend parameter with lme4

I am running a lmer model, with verbose = 2L, as in the following simple example:

library(lme4)
myData <- data.frame(Y = rnorm(100), 
                     Group = sample(LETTERS[1:2], 100, replace = TRUE))
LMER <- lmer(Y ~ (1 | Group), data = myData, verbose = 2L)

The verbose output looks like this:

npt = 3 , n =  1 
rhobeg =  0.2 , rhoend =  2e-07 
   0.020:   6:      293.717;0.200000 
  0.0020:   8:      293.684;0.166601 
 0.00020:  12:      293.683;0.160047 
 2.0e-05:  14:      293.683;0.160260 
 2.0e-06:  15:      293.683;0.160260 
 2.0e-07:  17:      293.683;0.160254 
At return
 20:     293.68311: 0.160254

I would like to change the rhoend parameter, with the idea of reducing the amount of time it takes to fit, presumably at the expense of getting less precise estimates.

How can I re-write my lmer() call to alter the rhoend parameter?

Upvotes: 2

Views: 237

Answers (1)

Ben Bolker
Ben Bolker

Reputation: 226087

Like this:

See ?lmerControl: you need to put the rhoend setting inside a list called optCtrl, to be passed to the optimizer.

LMER2 <- lmer(Y ~ (1 | Group), data = myData, 
  verbose = 2L, control=lmerControl(optCtrl=list(rhoend=1e-5)))
npt = 3 , n =  1 
rhobeg =  0.03090896 , rhoend =  1e-05 
start par. =  0.1545448 fn =  314.382 
rho:   0.0031 eval:   3 fn:      314.382 par:0.154545 
rho:  0.00031 eval:   5 fn:      314.380 par:0.165471 
rho:  5.6e-05 eval:   7 fn:      314.380 par:0.164620 
rho:  1.0e-05 eval:   9 fn:      314.380 par:0.164573 
At return
eval:  10 fn:      314.37951 par: 0.164571

You could also consider using control=lmerControl(optimizer="nloptwrap") in the latest (1.1-7) version of lme4 -- by default it uses a different implementation of the BOBYQA optimizer. See the examples in ?nloptwrap to see that you can change the tolerances for changes in the parameters (xtol_abs) or the response (ftol_abs). The default tolerances for nloptwrap are somewhat milder (= faster run times) than those for the default optimizer.

By the way, my answers are quite different from yours because we picked different random numbers. Best to use set.seed(101) (or some other arbitrary integer of your choice) for reproducibility.

Upvotes: 4

Related Questions