parallel computing in R , implementing bootstrap - r

Question

I'm currently trying to compute model estimators using the BLB bootstrap , and would like to do so parallel. my code works fine when I'm not doing it parallel. the problem when I'm computing in parallel is that the results I get from each core contains NA values. I don't understand how I get NA values while the Iris Data set's values don't contain NA at all.
here is the code that I'm using :
library(doParallel)
library(itertools)
num_of_cores <- detectCores()
cl <- makePSOCKcluster(num_of_cores)
registerDoParallel(cl)
attach(iris)
data <- iris
coeftmp <- data.frame()
system.time(
r <- foreach(dat = isplitRows(data, chunks=num_of_cores),
.combine = cbind) %dopar% {
BLBsize = round(nrow(dat)^0.6)
for (i in 1:400){
set.seed(i)
# sampling B(n) data points from the original data set without replacement
sample_BOFN <- dat[sample(nrow(dat), size = BLBsize, replace = FALSE), ]
# sampling from the subsample with replacment
sample_bootstrap <- sample_BOFN[sample(nrow(sample_BOFN), size = nrow(sample_BOFN), replace = TRUE), ]
bootstrapModel <- glm(sample_bootstrap$Petal.Width ~ Petal.Length + Sepal.Length + Sepal.Width, data = sample_bootstrap)
coeftmp <- rbind(coeftmp, bootstrapModel$coefficients)
}
#calculating the estimators of the model with mean
colMeans(coeftmp)
})

Answer 1

I think you're going to have to go through a few iterations of the debugger on this to solve it. But you're getting NAsfrom this line
bootstrapModel <- glm(sample_bootstrap$Petal.Width ~ Petal.Length + Sepal.Length + Sepal.Width, data = sample_bootstrap)
I am guessing that you get a singularity from one of your sample_bootstraps, since a singularity would give you an NA coefficient. But it's possible something else is causing this error, though it's definitely from this line of code.... you'll need to step through the debugger to isolate it.
... ie, this is not a complete answer. But this should allow you to solve your own problem:
You can see this by investigating:
r2 <- foreach(dat = isplitRows(data, chunks=1)) %dopar% {
BLBsize = round(nrow(dat)^0.6)
for (i in 1:400){
set.seed(i)
# sampling B(n) data points from the original data set without replacement
sample_BOFN <- dat[sample(nrow(dat), size = BLBsize, replace = FALSE), ]
# sampling from the subsample with replacment
sample_bootstrap <- sample_BOFN[sample(nrow(sample_BOFN), size = nrow(sample_BOFN), replace = TRUE), ]
bootstrapModel <- glm(sample_bootstrap$Petal.Width ~ Petal.Length + Sepal.Length + Sepal.Width, data = sample_bootstrap)
coeftmp <- rbind(coeftmp, bootstrapModel$coefficients)
}
#calculating the estimators of the model with mean
# return a list, not just the colMeans -- for debugging purposes
return(list(coeftmp= coeftmp, result= colMeans(coeftmp)))
}
sum(is.na(r2[[1]][[1]])) # no missing coefficients with 1 core
r <- foreach(dat = isplitRows(data, chunks=num_of_cores)) %dopar% {
BLBsize = round(nrow(dat)^0.6)
for (i in 1:400){
set.seed(i)
# sampling B(n) data points from the original data set without replacement
sample_BOFN <- dat[sample(nrow(dat), size = BLBsize, replace = FALSE), ]
# sampling from the subsample with replacment
sample_bootstrap <- sample_BOFN[sample(nrow(sample_BOFN), size = nrow(sample_BOFN), replace = TRUE), ]
bootstrapModel <- glm(sample_bootstrap$Petal.Width ~ Petal.Length + Sepal.Length + Sepal.Width, data = sample_bootstrap)
coeftmp <- rbind(coeftmp, bootstrapModel$coefficients)
}
#calculating the estimators of the model with mean
# return a list, not just the colMeans -- for debugging purposes
return(list(coeftmp= coeftmp, result= colMeans(coeftmp)))
}
# lots of missing values in your coeftmp results.
lapply(r, function(l) {sum(is.na(l[[1]]))})