Try-catch slowing down simulation of linear mixed effects

Question

I am simulating linear mixed effects in R, but sometimes the model does not converge. Therefore, I wrap it in a try function so the simulation keeps running even if the model fails. However, my simulation is very slow, and I think it's because of my code. I wrap the model in try and if it works, then I do the model again, which I know will eat up my time. There has to be a better way to do this. Any ideas?

Here is some sample code:

set.seed(16420)
y=rnorm(10000,100,15)
x=rnorm(10000,100,15)
t=rep(seq(1,100,1),100)
i=sort(rep(seq(1,100,1),100))
my.data=data.frame(i,t,x,y)
if(!inherits(try(lme(y~x+t,random=~x+t|i,data=my.data),
                 silent=T),"try-error"))
{
  m1=lme(y~x+t,random=~x+t|i,data=my.data)
}
summary(m1)

The title "Trying and not having to do it again" is probably too vague to get much attention for the question. I don't know enough about the question to retitle it for you, but you might want to get more descriptive — camille, Oct 23 '18 at 14:14

score 4 · Accepted Answer · answered Oct 26 '18 at 14:55

A simple first step you could take to speed up your code (by almost 100%) is to not run the lme() model twice; assign the result of try(lme(...)) to m1 and then test to see if it inherits an error; compare your code from above (placed into a function) to a function with this approach:

sim_slow <- function() {
    y=rnorm(10000,100,15)
    x=rnorm(10000,100,15)
    t=rep(seq(1,100,1),100)
    i=sort(rep(seq(1,100,1),100))
    my.data=data.frame(i,t,x,y)
    if(!inherits(try(nlme::lme(y~x+t,random=~x+t|i,data=my.data), silent=T),"try-error"))
    {
        return(nlme::lme(y~x+t,random=~x+t|i,data=my.data))
    }
    return(NULL)
}

sim_fast <- function(no_warn = TRUE) {
    if ( no_warn ) { # If you also don't want the warnings
        warn_option <- getOption("warn")
        options(warn = -1)
        on.exit(options(warn = warn_option))
    }
    y <- rnorm(10000, 100, 15)
    x <- rnorm(10000, 100, 15)
    t <- rep(seq(1, 100,1), 100)
    i <- sort(rep(seq(1, 100, 1), 100))
    my.data <- data.frame(i, t, x, y)
    m1 <- try(nlme::lme(y ~ x + t, random = ~x + t|i, data = my.data), silent = TRUE)
    return("if"(inherits(m1, "try-error"), NULL, m1))
}

^{Created on 2018-10-26 by the reprex package (v0.2.1)}

Now let's see if there's a speed difference:

set.seed(16420)
system.time(slow_sims <- replicate(10, sim_slow()))
#> Warning in logLik.lmeStructInt(lmeSt, lmePars): Singular precision matrix
#> in level -1, block 1

#> Warning in logLik.lmeStructInt(lmeSt, lmePars): Singular precision matrix
#> in level -1, block 1

#> (Many similar warnings are omitted here for space)

#>    user  system elapsed 
#>  19.612   0.008  19.621
set.seed(16420)
system.time(fast_sims <- replicate(10, sim_fast()))
#>    user  system elapsed 
#>  11.704   0.000  11.703
sum(sapply(slow_sims, is.null))
#> [1] 3
sum(sapply(fast_sims, is.null))
#> [1] 3
all.equal(slow_sims, fast_sims)
#> [1] TRUE

^{Created on 2018-10-26 by the reprex package (v0.2.1)}

Try-catch slowing down simulation of linear mixed effects

1 Answers1