R BootStrap with Strata

Question

I learned how to do bootstrap regression. Here is my example and Code.

   library(boot)
    sampledata = data.frame(y = sample(0:20, r = T, size = 5000),
                            x1 = runif(5000),
                            x2 = runif(5000),
                            group1 = sample(1:3, r = T, size = 5000),
                            group2 = sample(0:1, r = T, size = 5000))


# function to obtain R-Squared from the data
rsq <- function(formula, data, indices) {
  d <- data[indices,] 
  fit <- lm(formula, data=d)
  return(coef(fit))
}

# bootstrapping with 250 replications
results <- boot(data=sampledata, statistic=rsq,
   R=250, formula=y~x1+x2)
df <- data.frame(results$t)
names(df) <- names(results$t0)

However if you want to do this process for subgroups of your data how can you make it more automated? This is what I hope to do but this is very tiresome and I have many more group categories but this is basically the hopeful result, basically I subset the sampledata into 6 groups and then use all those data sets separate in 6 different bootstrap models and then combine the results in DATA. I see boot has strata feature but am not getting it to function. Also are you supposed to subset your data to group and then do the bootstrap sample? Or get a bootstrap sample and then subset your data? Any way, I am wondering how can it be possible to make this more automatic where I can code the program to do the bootstrap estimate for all groups separately, where the models are stratified.

sampledata1 = subset(sampledata, group1 == 1 & group2 == 0)
sampledata2 = subset(sampledata, group1 == 2 & group2 == 0)
sampledata3 = subset(sampledata, group1 == 3 & group2 == 0)
sampledata4 = subset(sampledata, group1 == 1 & group2 == 1)
sampledata5 = subset(sampledata, group1 == 2 & group2 == 1)
sampledata6 = subset(sampledata, group1 == 3 & group2 == 1)



# bootstrapping with 250 replications
results <- boot(data=sampledata1, statistic=rsq,
   R=250, formula=y~x1+x2)
df1 <- data.frame(results$t)
names(df1) <- names(results$t0)
df1$group1 = 1
df1$group2 = 0

# bootstrapping with 250 replications
results <- boot(data=sampledata2, statistic=rsq,
   R=250, formula=y~x1+x2)
df2 <- data.frame(results$t)
names(df2) <- names(results$t0)
df2$group1 = 2
df2$group2 = 0

# bootstrapping with 250 replications
results <- boot(data=sampledata3, statistic=rsq,
   R=250, formula=y~x1+x2)
df3 <- data.frame(results$t)
names(df3) <- names(results$t0)
df3$group1 = 3
df3$group2 = 0

# bootstrapping with 250 replications
results <- boot(data=sampledata4, statistic=rsq,
   R=250, formula=y~x1+x2)
df4 <- data.frame(results$t)
names(df4) <- names(results$t0)
df4$group1 = 1
df4$group2 = 1

# bootstrapping with 250 replications
results <- boot(data=sampledata5, statistic=rsq,
   R=250, formula=y~x1+x2)
df5 <- data.frame(results$t)
names(df5) <- names(results$t0)
df5$group1 = 2
df5$group2 = 1

# bootstrapping with 250 replications
results <- boot(data=sampledata6, statistic=rsq,
   R=250, formula=y~x1+x2)
df6 <- data.frame(results$t)
names(df6) <- names(results$t0)
df6$group1 = 3
df6$group2 = 1

DATA = rbind(df1, df2, df3, df4, df5, df6)

---

rsq <- function(formula, data, indices) {
  dataSUB <- data %>% filter(GROUP1 == group1, GROUP2 == group2)
  d <- dataSUB[indices,] 
  fit <- lm(formula, data=d)
  return(coef(fit))
}

Answer 1

This should do the trick:

set1 <- c(1,2,3)
set2 <- c(0,1)
for(i in set1) {
    for(j in set2) {
        sampleDataSet = subset(sampledata, group1 == i & group2 == j)
        results <- boot(data=sampleDataSet, statistic=rsq, R=250, formula=y~x1+x2)
        df <- data.frame(results$t)
        names(df) <- names(results$t0)
        df$group1 = i
        df$group2 = j
        assign(paste("df",i+(j*length(set1)),sep=""), df)
    }
}

Note: if you plan to do this with larger groups, assuming 0 is always in set2 you can just update set1 and set2 and the naming scheme should continue to work.

I hope this helps!

Answer 2

Using data.table may not enhance the time this would take to run. However here is a similar alternative with data.table and an assisting function to generalize the process:

library(data.table)

setDT(sampledata)

boot_group <- function(sel_group1, sel_group2){
  sampledata1  <- sampledata[group1 == sel_group1 & group2 == sel_group2]
  results <- boot(data=sampledata1, statistic=rsq,
                  R=250, formula=y~x1+x2)
  df1 <- data.table(results$t)
  names(df1) <- names(results$t0)
  df1[, group1 := sel_group1]
  df1[, group2 := sel_group2]
  return(df1)
}

DATA <- data.table()
groups1 <- sampledata[, unique(group1)]
groups1 <- sort(groups1)
groups2 <- sampledata[, unique(group2)]
groups2 <- sort(groups2)

for (g2 in groups2) {
  for (g1 in groups1) {
    subdf <- boot_group(g1, g2)
    DATA <- rbind(DATA, subdf)
  }
}