In ipalvis/BLBLogistic: Bag of Little Bootstrap for Logistic Regression Model

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

This package builds a logistic regression model using R's built-in "glm" function, utilizes BLB to subsample and resample the data, and retrieves both estimates and intervals of regression coefficients, sigmas, and predictions when applicable. The following is a walkthrough of the package's utilization, using 1) a provided dataset "bank-additional-clean.csv" that is a cleaned version of the dataset found here: https://archive.ics.uci.edu/ml/datasets/bank+marketing and 2) a 1992 questionnaire regarding the telecommution of City of San Diego Employees.

Note: Imported dataset cannot include any NA values.

First we load the package:

library(BLBLogistic)

We read the clean .csv file and assign it to our data variable of interest. We build the model using the package's blbglm function, that takes in five parameters:

1. formula: The formula that we wish to regress the data upon.

2. data: PATH of the dataset to take in.

3. m: The number of partitions ("bags") you wish to make in your future bootstrap data (default to 2)

4. B: The number of bootstrap resamplings you wish to make of each partition in m (default to 10)

5. parallel: Boolean indicating whether you want to utilize parallelization (for larger datasets) (default to FALSE)

bank <- read.csv("bank_data_clean.csv")
bankdata <- blbglm(y ~ age + job + marital + housing + contact + month + day_of_week + campaign + pdays + previous + poutcome + cons.price.idx + cons.conf.idx + euribor3m + nr.employed + y, data = bank, m = 3, B = 10, parallel = TRUE)

telcom = read.csv("tele_data_clean.csv")
teledata <- blbglm(C3H17M ~ OVERTIME + EFACT9 + EFACT6 + MANCONST + JOBCONST
                 + TECHCONS + CSO9FT2, data = telcom, m = 3, B = 10)

We then extract coefficient estimates for each bootstrap in each bag, and generate a mean confidence interval for said estimates.

• NOTE: confint() takes two additional parameters:

1) parm to specify which specific estimate CI's you wish to retrieve

2) level to specify quantile lengths for confidence intervals

coef(bankdata)
confint(bankdata, parm = c("age", "marital", "nr.employed"), level = 0.99)

coef(teledata)
confint(teledata)

We can then extract sigma from said estimates, and generate a singular confidence interval for sigma. (sigma takes the same parameters as confit from above)

sigma(bankdata, confidence = TRUE, level = 0.99)

sigma(teledata, confidence = TRUE)

We generate a probability prediction/prediction interval for each estimate, taking in up to four parameters:

1. object: PATH of the dataset to read.

2. newdata: A data frame of equal length of object, to replace fitted values for wanted predicted estimates.

3. confidence: A boolean toggle indicating output of prediction interval (default to FALSE)

4. level: The quantile of which you wish you prediction to be generated upon (default to 0.95 or 95%)

x_bank <- data.frame(age = 40, job = 2, marital = 3, housing = 1, contact = 1, month = 7, day_of_week = 4, campaign = 2, pdays = 999, previous = 0, poutcome = 2, cons.price.idx = 93.918, cons.conf.idx = -41.8, euribor3m = 4.864, nr.employed = 5228.1, y = 0)
predict(bankdata, x_bank, confidence = TRUE, level = 0.99)

x_tele <- data.frame(OVERTIME = c(4, 5), EFACT9 = c(0.5, 0.6), EFACT6 = c(1,0), MANCONST = c(1, 1), JOBCONST = c(1, 0), TECHCONS = c(0, 1), CSO9FT2 = c(0.5, 1))
predict(teledata, x_tele, confidence = TRUE)

The updated print function takes in the blbglm data, and displays the blbglm model/sampling, coefficient estimates of the BLB samples, the sigma, and returns the requested subsample m and bootstrap size B.

print(bankdata)
print(teledata)

ipalvis/BLBLogistic documentation built on March 21, 2020, 12:19 a.m.