0

I have just spent a decent amount of time learning how to do a stacked adaptive elastic net regression on several multiply-imputed datasets using the saenet package.

I understand that recommended practice is to use the value of lambda within one standard error of the minimum cross validation error. When I use the lambda value that yields the minimum CV error itself several of the predictors are non-zero. However when I use the 1SE value of lambda, all the predictors are zero.

Is there any empirical reason forbidding me from choosing the model with the minimum CV error rather than 1SE from minimum?

llewmills
  • 2,151
  • 17
  • 35
  • 1
    No, you absolutely don't have to pick the lambda that obeys the one-standard error rule. It's just that in a situation like this, it looks like the trivial model is competitive with the one with minimum CV error. – Stephan Kolassa Jul 15 '23 at 22:10

0 Answers0