Storing results from simulated merMods in data frame - r

Question

Updated:
I'm trying to examine the variability in my parameter estimates from a merMod object by simulating known data and running the model 100 times. I'd like the result to be a data frame that looks like the following:
| simulation | intercept | est.x1 | est.x2 |
| ---------- | --------- | ------ | ------ |
| sim_study1 |.09 |.75 |.25 |
| sim_study2 |.10 |.72 |.21 |
| sim_study3 |NA |NA |NA |
My code to generate multilevel data with a random intercept and 2 predictors is:
# note. this code block runs as expected, and if I run a lmer() call
# on a simulated data set I get values that one would expect.
gen_fake <- function(i, j){
school <- rep(1:j)
person <- rep(1:i) # students nested in schools
# parameters
mu_a_true <- 0.10 # real intercept
sigma_a_true <- 0.10 # varince of intercept
sigma_y_true <- 0.40
b1_true <- .75
b2_true <- .25
# random intercept for schools
a_true <- rnorm(j, mu_a_true, sigma_a_true)
# random data for predictors
x1 <- rnorm(i, 0, 1)
x2 <- rnorm(i, 0, 1)
# outcome
y <- rnorm(i, a_true[school] + b1_true*x1 + b2_true*x2, sigma_y_true)
return (data.frame(y, person, school, x1, x2))
}
I'm attempting to conduct a 100 simulations of a model, while generating new data each time. Note, I'm trying to implement tryCatch within the loop because with more complex models, where the model might not terminate normally, I'd like value returned in the table to be NA for parameters.
My code for this is as follows:
# create an empty data frame with names of parameters (there's probably
# a slicker way to do this within the loop where I can match names from
# the model call)
sim_results <- data.frame(matrix(nrow=100, ncol=3,
dimnames=list(c(),
c("intercept",
"est.x1", "est.x2"))),
stringsAsFactors=F)
# load library for analysis
library(lme4)
# conduct 100 simulations of the model generating fake data for each run
sim_study <- function (i, j, n.sims){
for (sim in 1:n.sims){
fake_dat <- gen_fake(i, j)
tryCatch({
lmer_sim <- lmer(y ~ x1 + x2 + (1|school), data = fake_dat)
}, error = function(e){
return(NA)
}) #return previous value of fm if error
estimates <- rbind(fixef(lmer_sim))
}
sim_results[sim,] <- estimates
}
# run the simulation study
sim_study (1000,5,100)
The issue I am having is that the function only returns 1 row and it isn’t populating the empty data frame I made:
(Intercept) x1 x2
[1,] 0.09659339 0.7746392 0.2325391
I'm unsure of the issue. Finally, any feedback you might have for how to make this work faster would also be appreciated, as I'd like to learn more about that issue. Thanks for any assistance.

Answer 1

This may be a bit of a forehead-slapper, but I think you just misplaced the loop brackets? This works for me:
sim_study <- function (i, j, n.sims){
for (sim in 1:n.sims) {
if (sim %% 10 == 0 ) cat(".\n") ## print progress
fake_dat <- gen_fake(i, j)
tryCatch({
lmer_sim <- lmer(y ~ x1 + x2 + (1|school),
data = fake_dat)
}, error = function(e){
return(rep(NA,3)) ## return vector of correct length
}) #return previous value of fm if error
estimates <- rbind(fixef(lmer_sim))
sim_results[sim,] <- estimates
}
return(sim_results)
}
A few more points:
I'm not sure whether the tryCatch() logic works, since I didn't hit any errors (but I think it ought to be modified to return an object with the current length, as above)
you could replace some of your gen_fake (not the generation of the predictors, but the generation of the response with the built-in ?simulate.merMod(), but I don't think it would actually work any better (or worse)
speeding this up significantly would be a bit of work/hacky. There is a refit() function that works quickly if only the predictor variable has changed, but it doesn't hold in this case. You could use the tricks specified here ...