Combining regression summary outputs from multiple samples into a single dataframe in R

Question

I'm trying to combine multiple lm outputs into a data frame, for further calculations. I have a dataset of 1000 observations and 62 variables. The project is to randomly split the dataset 63/37, train the model, repeat this 1000 times and save the coefficients, the fitted values, and the r2 for all 1000 runs. So I'm doing most of that here (using mtcars ):

data("mtcars")
f <- function () {
  fit <- lm(mpg ~ ., data = mtcars, subset = sample <- sample.int(n = nrow(mtcars), size = floor(.63*nrow(mtcars)), replace = F))
  coef(fit)
}
output <- t(replicate(1000, f()))

I know I can get the rsq values with summary(fit)$r.squared and I can use predict() to get the fitted values. I'm just struggling with how to get them into the data frame with the saved coefficients.

Answer 1

The below should do

get_model <- function (input_data) {
    fit <- lm(mpg ~ ., 
              data = mtcars, 
              subset = sample <- sample.int(n = nrow(mtcars), 
                     size = floor(.63*nrow(mtcars)), replace = F)
             )      
    return(fit)
}

get_results <- function(lm_model){
    data <- data.frame()
    data <- rbind(data, coef(lm_model))
    data <- cbind(data, summary(lm_model)$r.squared)
    colnames(data) <- c(names(mtcars), "rsquared")
    return(data)
}


# running the above
input_data <- mtcars
general_df <- data.frame()

for(i in 1:1000){
    my_model   <- get_model(input_data)
    final_data <- get_results(my_model)
    general_df <- rbind(general_df, final_data)
}

Answer 2

You are very close:

library(tidyverse)
library(modelr)
data("mtcars")

get_data_lm <- function(data_df, testPCT = 0.37){

    data_resample <- modelr::crossv_mc(data_df, n = 1, test = testPCT)
    fit <- lm(mpg ~ ., data = as.data.frame(data_resample$train))

    stats <- c(coef(fit),
               "R2" = summary(fit)$r.squared,
               "AdjR2" = summary(fit)$adj.r.squared)
    pred_vals <- predict(fit, newdata = as.data.frame(data_resample$test))

    c(stats, pred_vals)

}

output <- t(replicate(1000, get_data_lm(mtcars)))

The only thing you needed to do is concatenate the other statistics and predicted values you want. Alternatively, you could use a parallel sapply() variant to make your simulation considerably faster.

Another comment: I use the crossv_mc() function from the modelr:: package to create one testing and training partition. However, I could have used n = 1000 outside the function instead; this would have created a resample data frame in my working environment for me to apply() a function over. See the modelr:: GitHub page for more info.