nlcom after elastic net?

[Kjell Weyde]

Hi!
I have used the user-written cvlasso (Ahrens, A., Hansen, C.B., Schaffer, M.E. 2018. cvlasso: Program for cross-validation using lasso, square-root lasso, elastic net, adaptive lasso and post-OLS estimators. http://ideas.repec.org/c/boc/bocode/s458458.html) to simplify a model containing several interaction terms. For simplicity, let's say that the model is reduced to:

outcome = a + bX1 + cX2  dX3 + eX1X2 + fX1X3

I would like to understand more about the results, i.e., what the interaction terms mean. Normally, I would have used nlcom to do this, but in my case, I don't have the e(V), only e(b) and e(se). Is there any way to use nlcom when you dont have e(V)? Or is there any similar Stata command that can do the job?

Best,
Kjell

[Achim Ahrens]

cvlasso implements K-fold and rolling cross-validation which is aimed at out-of-sample prediction. Cross-validation is not good at selecting the true model and I would advise against interpreting your coefficients. The predictors are chosen because they help predicting the outcome-- that does not mean that the selected predictors have a meaningful causal effect on the outcome.

Please have a look at our working paper that explains these issues in more detail: http://ftp.iza.org/dp12081.pdf