IVLASSO Weak identification F stats

[Ivan Tzintzun]

Hello,

I am using high-dimensional ML techniques to find optimal instruments.
After using IVLASSO command, including the option idstats, I obtain the following results:

Weak identification F stats (i.i.d.):
Optimal Lasso IV(s): 0.02
Optimal Post-Lasso IV(s): 0.02
Full IV set: 6.50

Weak identification F stats (robust):
Optimal Lasso IV(s): 9.33
Optimal Post-Lasso IV(s): 9.33
Full IV set: 4384.46

What is the main difference between the Optimal Lasso and Full IV set F-stats ?
Which would be the relevant F-stat to be reported (e.g. Optimal Lasso IV) ?
Is it strange the robust option yields considerably higher F-stats, for instance when compaing the Full IV set ?

Thanks !

[Mark Schaffer]

Ivan,

The optimal Lasso IV in effect combines the selected instruments for that endogenous regressor and combines them into an optimal IV (similar to how you could do 2SLS by hand but use the fitted value of the endogenous regressor as an IV). ivlasso does this separately for every endogenous regressor.

The "Full IV set" is the union of all selected instruments for all endogenous regressors.

Regarding your results - can you share more of the output? The results are very odd. The iid estimation using the optimal instruments is basically unidentified. How many endogenous regressors do you have? Do you have any instruments that are dummy variables or are you using a cluster-robust setup? You could be getting strange results with the robust case if e.g. singleton dummies are being selected as instruments.