reg_logreg: Reguarized logistic regression

In andybega/demspaces: Functions for the Democratic Spaces Project

Description

Standardized interface for self-tuning regularized logistic regression.

Usage

reg_logreg(x, y, folds = 5, alpha_n = 3, cost = "mse", lambda = "1se")

Arguments

x	Data frame with features.
y	Binary vector indicating outcome event.
folds	Number of folds to use for CV tuning
alpha_n	Number of alpha values to sample for CV tuning
cost	Cost measure to use, see glmnet::cv.glmnet()
lambda	Decision rule to pick lambda, one of "min", "1se", "0.5se"

Details

Tuning is performed using cross-validation with glmnet::cv.glmnet(). Both lambda and alpha values are tuned. The lambda values are left to the model default and a uniform grid of alpha values is used. The lambda value is picked with glmnet::cv.glmnet()'s more robust 1se value (i.e. not the absolute minimum, but closest value within 1 SD of the minimum value). Then the globally optimum alpha value is picked.

See Also

Other Other base models: logistic_reg_featx(), logistic_reg()

Examples

library(modeldata)
data(credit_data)
credit_data <- credit_data[complete.cases(credit_data), ]

mdl <- reg_logreg(credit_data[, setdiff(colnames(credit_data), "Status")],
                    credit_data$Status,
                    folds = 5, alpha_n = 4)
# plots to review tuning results
plot(mdl)
plot(mdl, "alpha")
plot(mdl, "lambda")
preds <- predict(mdl, new_data = credit_data)
head(preds)

andybega/demspaces documentation built on April 18, 2021, 11:05 p.m.