LASSO (LARS) - SELECTION OF MANY NOT SIGNIFICANT VARIABLES

[nickvlachog]

Hellow,
I am using LARS so as to select from a pool of about 80 regressors (independent variables) the best ones in explaining my dependent variable.

LASSO selects almost 70 out of 80 variables (using the Cp criterion), but when I use all these variables in estimating my final model specification almost half of those regressors (selected by LARS) appear to be insignificant (p-value >> 5%)!!!

Is there something I can do when using LASSO so as not to end up in a model with so many insignificant variables?

Or, do you have any other ideas on how can I overcome this issue with my dataset?

Thank you in advance,
Nikos

[Carlo Lazzaro]

Nick:
model maqullage should be discouraged.
That said, and without knowing the reason you are after selecting "the best predictors" for your dependent variable, you may be interested in taking a look at -help stepwise- with the usual warning that there are very good reasons to avoid that approach (i.e., overfitting and subsequent poor out of-sample prediction) (please, see a related thread at http://www.statalist.org/forums/foru...ise-regression).

[ben earnhart]

I'll go on the assumption you're doing something where data-mining techniques *are* appropriate, a purely predictive exercise where causality and theory are moot. Most people around here have backgrounds that make data mining anathema, but I took a class in it from B-school here, and for some purposes (marketing, credit card fraud detection, etc.), it's potentially OK so long as the results are validated on new datasets. Anyway a few possibilities:

• Are they anywhere near .05 or .10? Near-significant variables can have predictive value. The magic .05 cut-off is an arbitrary construct. Something totally insignificant is one thing, but if it's close to significance, it may pass the Cp criterion.
• What happens to the other regressors when you remove the non-significant ones? Do significance levels go up or down? It's possible LASSO includes stuff like supressor effects, or sets of dummies where only one out of five is actually significant.
• Have you tried CHAID (try -findit chaid- to get it). LASSO seems a blunt instrument compared to CHAID, which is another method of data mining.

Of course, depending on what you're really up to, data mining may not be appropriate anyway.

[nickvlachog]

Regarding the issue you raised, I am providing below some further clarifications:
- I think that data mining techniques are appropriate in my study, since the current literature selects arbitrarily the regressors (accounting ratios) while in reality there is a wide range of variables that can be employed but not given any special attention.
- almost half of the variable selected have a p-value above 30% (highly insignificant)
- when i remove the non significant regresors the remaining are not materially affected (slight increase in significance). The R-squared of the model almost remains unchanged (decrease by about 0.1%)
- I will also try CHAID.

Thank you in advance for your valuable help.
Nikos,