13.20 Penalized regression

  • dropping some controls
  • \(argmin_\beta \{\sum(Y - \hat{Y})^2 + \lambda F(\beta)\}\)
  • minmize sum of squared residuals AND make \(\beta\) function small
  • implementation: LASSO, ridge regression, elastic net regression (LASSO + ridge)
  • throw out variables that LASSO thinks are unimportant
  • watch out: standardize all variables
  • choose \(\lambda\) as you want; higher value –> toss out more variables