Measuring the strength of instruments using Weakivtest

[Tom Kemeny]

I'm interesting in estimating Montiel-Pflueger-style 'effective' F stats for a series of panel models I'm running. These are simple models with one endogenous regressor and one instrument. I started with xtivreg2, in the following form:

xtivreg2 y yr* (x=z), fe cluster(cbsa)

where yr* are year dummies, and cbsa captures the unit of observation (cities).

Since weakivtest only runs using ivreg2, I switched to an LSDV model, as follows:

tab cbsa, gen(city)
ivreg2 y yr* city* (x=z), robust cluster(cbsa) partial(city*)

I partialled out the city dummies as if I do not, I get errors saying "...estimated covariance matrix of moment conditions not of full rank...."

But when I now run weakivtest, it says:
Noconstant option is not supported by weakivtest.

Am I stuck or am I missing something? I can't get weakivtest to work with xtivreg2, and when I switch to LSDV-style, I have to partial out the constant to get any results, which also blocks weakivtest.

Another alternative, and I have no idea if it is a reasonable one, is to ignore the error message about the moment conditions, and push ahead with weakivtest. When I do that, though, I get a suspiciously large 'effective F' (ie over 300!).

Any thoughts?

Best,
Tom

[Joro Kolev]

The user written command -weakivtest- also works post -ivregress-, so you might try the latter.

I think you can safely ignore the error message about the moments as well, as this relates to the "second stage" and you are interested in the first stage basically.

[Tom Kemeny]

Thank you Joro - ivregress worked just fine, and ultimately I got a more reasonable effective F.