We now review k-fold cross-validation. (a) Explain how k-fold cross-validation is implemented.
K-Fold Cross Validation is used to evaluate the performance of your predictive model. Previously we’ve learned about splitting our data into 1 train set and 1 test set. The K-Fold approach differs in that you divide your data into “k” equal size folds. Then the model is trained k amount of times. During each training, one fold is used as the validation set while the k-1 folds are the training set. Each time that the model is trained, performance metrics are calculated. The performance metrics are averaged to give the model’s overall performance scores (for example, accuracy).
Advantages relative to validation set approach: It can provide a better estimate of model performance because it’s training and validating on all of the data. For validation set approach, you could get very difference performance depending on how the data was split. Disadvantages relative to validation set approach: The performance metrics will vary depending on how many folds you choose. It is also computationally more “expensive” because you are training and evaluating multiple times. Advantages relative to LOOCV: It is less computationally “expensive” because it involves fewer iterations. Would be less biased than LOOCV because it averages over k-folds rather than being influenced by a single data point for each iteration. Disadvantages relative LOOCV: Again, the performance metrics could be more variable than LOOCV depending on how many folds you choose for k. LOOCV tends to have lower bias but higher variance compared to k-folds cv.
In Chapter 4, we used logistic regression to predict the probability of default using income and balance on the Default data set. We will now estimate the test error of this logistic regression model using the validation set approach. Do not forget to set a random seed before beginning your analysis.
library(ISLR2); library(corrplot); library(MASS); library(caret); library(car); library(dplyr); library(class);library(e1071);library(boot)
## corrplot 0.92 loaded
##
## Attaching package: 'MASS'
## The following object is masked from 'package:ISLR2':
##
## Boston
## Loading required package: ggplot2
## Loading required package: lattice
## Loading required package: carData
##
## Attaching package: 'dplyr'
## The following object is masked from 'package:car':
##
## recode
## The following object is masked from 'package:MASS':
##
## select
## The following objects are masked from 'package:stats':
##
## filter, lag
## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, union
##
## Attaching package: 'boot'
## The following object is masked from 'package:car':
##
## logit
## The following object is masked from 'package:lattice':
##
## melanoma
default = Default
str(default)
## 'data.frame': 10000 obs. of 4 variables:
## $ default: Factor w/ 2 levels "No","Yes": 1 1 1 1 1 1 1 1 1 1 ...
## $ student: Factor w/ 2 levels "No","Yes": 1 2 1 1 1 2 1 2 1 1 ...
## $ balance: num 730 817 1074 529 786 ...
## $ income : num 44362 12106 31767 35704 38463 ...
m5a = glm(formula = default ~ income + balance, data = default, family = binomial)
summary(m5a)
##
## Call:
## glm(formula = default ~ income + balance, family = binomial,
## data = default)
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -1.154e+01 4.348e-01 -26.545 < 2e-16 ***
## income 2.081e-05 4.985e-06 4.174 2.99e-05 ***
## balance 5.647e-03 2.274e-04 24.836 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2920.6 on 9999 degrees of freedom
## Residual deviance: 1579.0 on 9997 degrees of freedom
## AIC: 1585
##
## Number of Fisher Scoring iterations: 8
# i.
set.seed(42)
index = sample(nrow(default), 0.8*nrow(default), replace = F) # 80/20 split
default_train = default[index,]
default_val = default[-index,]
#ii.
m5b = glm(formula = default ~ income + balance, data = default_train, family = binomial)
#iii.
predprob_log_default = predict.glm(m5b, default_val, type = "response")
predclass_log_default = ifelse(predprob_log_default >= 0.5, "Yes", "No")
confusionMatrix(as.factor(predclass_log_default), as.factor(default_val$default), positive = "Yes")
## Confusion Matrix and Statistics
##
## Reference
## Prediction No Yes
## No 1932 44
## Yes 7 17
##
## Accuracy : 0.9745
## 95% CI : (0.9666, 0.981)
## No Information Rate : 0.9695
## P-Value [Acc > NIR] : 0.1061
##
## Kappa : 0.3895
##
## Mcnemar's Test P-Value : 4.631e-07
##
## Sensitivity : 0.2787
## Specificity : 0.9964
## Pos Pred Value : 0.7083
## Neg Pred Value : 0.9777
## Prevalence : 0.0305
## Detection Rate : 0.0085
## Detection Prevalence : 0.0120
## Balanced Accuracy : 0.6375
##
## 'Positive' Class : Yes
##
#iv. (1 - accuracy) is error
print(1-0.9745)
## [1] 0.0255
set.seed(42)
train_ctrl <- trainControl(method = "cv",
number = 3) #number is folds
m5c <- train(default ~ income + balance,
data = default,
method = "glm",
family = binomial,
trControl = train_ctrl)
print(m5c)
## Generalized Linear Model
##
## 10000 samples
## 2 predictor
## 2 classes: 'No', 'Yes'
##
## No pre-processing
## Resampling: Cross-Validated (3 fold)
## Summary of sample sizes: 6667, 6667, 6666
## Resampling results:
##
## Accuracy Kappa
## 0.9738002 0.4417961
print(m5c$results)
## parameter Accuracy Kappa AccuracySD KappaSD
## 1 none 0.9738002 0.4417961 0.002208342 0.05914816
Accuracy is 0.9738002 and error is 0.0261998. This is very similar to the previous method where error was 0.0255.
set.seed(42)
#dummy variable creation
default$student_dummy <- ifelse(default$student == "Yes", 1, 0)
default_train$student_dummy <- ifelse(default_train$student == "Yes", 1, 0)
default_val$student_dummy <- ifelse(default_val$student == "Yes", 1, 0)
m5d = glm(formula = default ~ income + balance + student_dummy, data = default_train, family = binomial)
predprob_log_default2 = predict.glm(m5d, default_val, type = "response")
predclass_log_default2 = ifelse(predprob_log_default2 >= 0.5, "Yes", "No")
confusionMatrix(as.factor(predclass_log_default2), as.factor(default_val$default), positive = "Yes")
## Confusion Matrix and Statistics
##
## Reference
## Prediction No Yes
## No 1932 42
## Yes 7 19
##
## Accuracy : 0.9755
## 95% CI : (0.9677, 0.9818)
## No Information Rate : 0.9695
## P-Value [Acc > NIR] : 0.06378
##
## Kappa : 0.4263
##
## Mcnemar's Test P-Value : 1.191e-06
##
## Sensitivity : 0.3115
## Specificity : 0.9964
## Pos Pred Value : 0.7308
## Neg Pred Value : 0.9787
## Prevalence : 0.0305
## Detection Rate : 0.0095
## Detection Prevalence : 0.0130
## Balanced Accuracy : 0.6539
##
## 'Positive' Class : Yes
##
# error
print(1-0.9755)
## [1] 0.0245
Including the dummy student variable decreased the test error by 0.001.
We continue to consider the use of a logistic regression model to predict the probability of default using income and balance on the Default data set. In particular, we will now compute estimates for the standard errors of the income and balance logistic regression coefficients in two different ways: (1) using the bootstrap, and (2) using the standard formula for computing the standard errors in the glm() function. Do not forget to set a random seed before beginning your analysis.
set.seed(42)
m6a = glm(formula = default ~ income + balance, data = default, family = binomial )
summary(m6a)
##
## Call:
## glm(formula = default ~ income + balance, family = binomial,
## data = default)
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -1.154e+01 4.348e-01 -26.545 < 2e-16 ***
## income 2.081e-05 4.985e-06 4.174 2.99e-05 ***
## balance 5.647e-03 2.274e-04 24.836 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2920.6 on 9999 degrees of freedom
## Residual deviance: 1579.0 on 9997 degrees of freedom
## AIC: 1585
##
## Number of Fisher Scoring iterations: 8
summary(m6a)$coefficients[, 2]
## (Intercept) income balance
## 4.347564e-01 4.985167e-06 2.273731e-04
Estimated standard errors:
(Intercept) income balance 4.347564e-01 4.985167e-06 2.273731e-04
boot.fn = function(data, index) {
# Subset data
data_subset = data[index, ]
# Fit model
model = glm(default ~ income + balance, data = data_subset, family = binomial)
# Return coefficient estimates
return(coef(model))
}
set.seed(42)
boot(data = default,
statistic = boot.fn,
R = 1000) # R is number of iterations
##
## ORDINARY NONPARAMETRIC BOOTSTRAP
##
##
## Call:
## boot(data = default, statistic = boot.fn, R = 1000)
##
##
## Bootstrap Statistics :
## original bias std. error
## t1* -1.154047e+01 -2.292405e-02 4.435269e-01
## t2* 2.080898e-05 2.737444e-08 5.073444e-06
## t3* 5.647103e-03 1.176249e-05 2.299133e-04
Standard errors: 4.435e-01, 5.07e-06,2.299e-04
The standard errors are very similar when you compare the two methods.
We will now consider the Boston housing data set, from the ISLR2 library.
boston = Boston
mu_hat = mean(boston$medv)
mu_hat
## [1] 22.53281
se_mu_hat = stats::sd(boston$medv)/(length(boston$medv)^(1/2))
se_mu_hat
## [1] 0.4088611
set.seed(42)
# bootstrapping function to calculate sample mean
boot_fn = function(data, indices) {
mean(data[indices])}
# bootstrap
boot_results =boot(boston$medv, boot_fn, R = 1000)
boot_results
##
## ORDINARY NONPARAMETRIC BOOTSTRAP
##
##
## Call:
## boot(data = boston$medv, statistic = boot_fn, R = 1000)
##
##
## Bootstrap Statistics :
## original bias std. error
## t1* 22.53281 0.02671186 0.4009216
SE from B is 0.4088611 and SE here is 0.4009216. They are very similar, with this error being slightly lower.
se_bootstrap <- sd(boot_results$t)
lower_bound <- mu_hat - 2 * se_bootstrap
upper_bound <- mu_hat + 2 * se_bootstrap
lower_bound
## [1] 21.73096
upper_bound
## [1] 23.33465
Based on bootstrap, the CE is 21.73 - 23.33.
t_test_result = t.test(Boston$medv)
t_test_result$conf.int
## [1] 21.72953 23.33608
## attr(,"conf.level")
## [1] 0.95
Based on t-test, the CE is 21.72953 - 23.33608. These results are basically the same as the bootstrap estimate.
mu_hat_median = median(boston$medv)
mu_hat_median
## [1] 21.2
set.seed(42)
# bootstrapping function to calculate sample median
boot_fn_median = function(data, indices) {
median(data[indices])}
boot_results_median <- boot(boston$medv, boot_fn_median, R = 1000)
boot_results_median
##
## ORDINARY NONPARAMETRIC BOOTSTRAP
##
##
## Call:
## boot(data = boston$medv, statistic = boot_fn_median, R = 1000)
##
##
## Bootstrap Statistics :
## original bias std. error
## t1* 21.2 0.0106 0.3661785
SE is 0.3661785.
mu_0.1_hat <- quantile(boston$medv, 0.1)
mu_0.1_hat
## 10%
## 12.75
set.seed(42)
boot_fn_10th_percentile <- function(data, indices) {
quantile(data[indices], 0.1)}
boot_results_10th_percentile <- boot(boston$medv, boot_fn_10th_percentile, R = 1000)
boot_results_10th_percentile
##
## ORDINARY NONPARAMETRIC BOOTSTRAP
##
##
## Call:
## boot(data = boston$medv, statistic = boot_fn_10th_percentile,
## R = 1000)
##
##
## Bootstrap Statistics :
## original bias std. error
## t1* 12.75 0.0215 0.4948966
Standard error estimate is 0.4948966 - slightly bigger than previous standard error.