In McChickenNuggets/blblm: Perform bag of little bootstrap for linear regression model

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(blblm)

Note: This package is an application of bag of little bootstrap(BLB) algorithm I learned from STA 141C (Big Data & High Performance Statistical Computing)

The bag of little bootstraps (BLB)

It is a procedure which incorporates features of both the bootstrap and subsampling to yield a robust, computationally efficient means of assessing the quality of estimators (i.e., construct confidence intervals )

DiagrammeR::grViz("blb.gv", height = 300)

Basically, the bag of little bootstraps = map_reduce + bootstrap. However, for each bootstrap, we sample $n$ from $b$ with replacement instead of sample $b$ from $b$ as in ordinary bootstrap.

• sample without replacement the sample $s$ times into sizes of $b$

• for each subsample

  • resample each until sample size is $n$, $r$ times
  • compute the bootstrap statistic (e,g., the mean of a variable, or cor between two variables) for each bootstrap sample
  • compute the statistic (e.g., confidence interval) from the bootstrap statistics

• take the average of the statistics

Intro on package blblm

• The main package realizes the application of BLB algorithm on Linear Regression Model and Logistic Regression model

• Provide two main functions

  • blblm() BLB on Linear Regression Model (Could specify Parallelization by setting Parallel = TRUE)
  • blbglm() BLB on Logistic Regression Model (Could specify Parallelization by setting Parallel = TRUE)

• Override Four functions

  • coef() Obtain the coefficients for objects
  • confint() Obtain the confidence intervals for coefficients
  • sigma() Obtain the sigma for objects (Could specify confidence level by setting confindence = TRUE)
  • predict() Predict values for new observation by given model (Could specify confidence level by setting confindence = TRUE)

Usage of blblm

On blblm

# Setting up the model for blblm
fitlm<-blblm(mpg ~ wt * hp, data = mtcars, m = 3, B = 100, Parallel = FALSE)

# Obtain coefficient
coef(fitlm)

# Obtain confidence Interval for coefficients
confint(fitlm, c("wt", "hp"))

# Obtain sigma 
sigma(fitlm)

# Obtain sigma with confidence intervals
sigma(fitlm, confidence = TRUE)

# Predict new observation
predict(fitlm, data.frame(wt = c(2.5, 3), hp = c(150, 170)))

# Predict new observation with confidence intervals
predict(fitlm, data.frame(wt = c(2.5, 3), hp = c(150, 170)), confidence = TRUE)

# Print
print(fitlm)

Turn warning = FALSE because we have a relatively small dataset Therefore for some observation, the algorithm won't converge and will have fitted probability as 0 or 1

ON blbglm

# Setting up the model for blblm
fitglm<-blbglm(Species ~ Sepal.Length * Sepal.Width, data = iris[1:100,], m = 3, B = 100, family = binomial, Parallel = FALSE)

# Obtain coefficient
coef(fitglm)

# Obtain confidence Interval for coefficients
confint(fitglm, c("Sepal.Length", "Sepal.Width"))

# Obtain sigma 
sigma(fitglm)

# Obtain sigma with confidence intervals
sigma(fitglm, confidence = TRUE)

# Predict new observation
predict(fitglm, data.frame(Sepal.Length = c(5.5, 5.1), Sepal.Width = c(2.3, 3.5)))

# Predict new observation with confidence intervals
predict(fitglm, data.frame(Sepal.Length = c(5.5, 5.1), Sepal.Width = c(2.3, 3.5)), confidence = TRUE)

# Print
print(fitglm)

Parallelization

library(future)
plan(multiprocess,workers=8)
options(future.rng.onMisuse = "ignore")

blblm

## We use system.time instead of bench::mark because the html will not give the output for bench::mark
system.time(fitlm1<-blblm(mpg ~ wt * hp, data = mtcars, m = 10, B = 5000, Parallel = FALSE))
system.time(fitlm2<-blblm(mpg ~ wt * hp, data = mtcars, m = 10, B = 5000, Parallel = TRUE))

blbglm

system.time(fitglm1<-blbglm(Species ~ Sepal.Length * Sepal.Width, data = iris[1:100,], m = 3, B = 5000, family = binomial, Parallel = FALSE))
system.time(fitglm2<-blbglm(Species ~ Sepal.Length * Sepal.Width, data = iris[1:100,], m = 3, B = 5000, family = binomial, Parallel = TRUE))

As you can see, when you turn on parallization, the package will significantly improved its efficiency.

McChickenNuggets/blblm documentation built on March 22, 2021, 1:45 p.m.