In dajmcdon/ubc-stat406-labs: Tutorials and labs for UBC Stat 406 in the 2020-2021 online year

library(knitr)
library(tidyverse)
opts_chunk$set(echo=FALSE, fig.align='center', fig.width=8, fig.height=8,
               cache=TRUE, autodep=TRUE, cache.comments=FALSE,
               message=FALSE, warning=FALSE)

Your goal is to use this data to predict economic mobility. Note that there are generally more observations than predictors.

mob = read.csv('mobility.csv')

1. Using glmnet, estimate 4 models: the linear model, ridge regression, the lasso, and the elastic net ($\alpha=.5$).

2. Plot the CV curves for each of the three regularized models (easy).

3. Use lambda.min to get a particular model for each of the regularized ones.

4. Plot the coefficients for each of the 4 models on one figure. What do you notice? Which features are most important?

library(glmnet)
linmod = lm(Mobility~.-ID-Name-State, data=mob, y=TRUE)
X = model.matrix(linmod)[,-1]
y = linmod$y
lasso = cv.glmnet(X, y)
ridge = cv.glmnet(X, y, alpha=0)
enet = cv.glmnet(X,y,alpha=.5)

par(mfrow=c(2,2))
plot(lasso)
plot(ridge)
plot(enet)
par(mfrow=c(1,1))

For enet and lasso, lambda.1se gives sparser models. For ridge, use lambda.min (more like GCV).

lasso1 = coef(lasso, 'lambda.min')
enet1 = coef(lasso, 'lambda.min')
ridge1 = coef(ridge, 'lambda.min')

ord = order(coef(linmod))
df = data.frame(lm = coef(linmod)[ord], lasso = lasso1[ord],
                elnet = enet1[ord], ridge = ridge1[ord])
df$var = rownames(df)
gather(df, key='method',value='estimate',-var) %>%
  ggplot(aes(y=var,x=estimate,color=method)) + geom_point() + 
  theme_minimal()

dajmcdon/ubc-stat406-labs documentation built on Aug. 18, 2020, 1:23 p.m.