Testing the Suitability of an Instrumental Variable in a Random Effects Model

[Louisa Roberts]

Dear Listserv—

I’m trying to test the suitability of an instrument with a random effects (or mixed effects) model.

I have one instrumental variable (iv) and one endogenous predictor variable (ngo). I’m using Stata 13; my N is about 290; and my cases (country-years) are clustered by country.

My Questions:

(1) Here is the code I’ve been using: xtivreg hojust year relz l.gdpc1000 (l.ngo=l.iv), re. I want to test whether the predictor variable “ngo” is endogenous, as theory would lead me to expect. Is there any way to do this? (I tried the estat endogenous postestimation command, but it apparently only works after ivregress, not after xtivreg.)

(2) Can you suggest any tests in Stata of whether I have a weak instrument?

(3) Is there any way to use an instrumental variable with the mixed command? My main models actually use the mixed command (not xtreg) and include three random effects. (Here is my code: mixed hojust year relz l.gdpc1000 || country: year, covariance(unstructured).) The xtivreg command basically works but isn’t ideal, because it only allows me to include a single random effect for country.

Thanks very much!

Louisa

[Mark Schaffer]

On (2), xtoverid, available from SSC in the usual way, gives you some options. There's a noi (=noisily) option that people have used in the past to get more diagnostics. See e.g. the threads here:

http://www.stata.com/statalist/archi.../msg00766.html

or here:

http://www.stata.com/statalist/archi.../msg00610.html

[Louisa Roberts]

Thanks very much Mark! I'll definitely use that. Does anyone else have thoughts on my questions #1 or #3 above?

[Ariel Karlinsky]

I would also love an answer to question (1) as I'm in a similar situation.
frankly I don't really understand why there are so few post-estimation tests available for xtivreg as opposed to ivregress

[Louisa Roberts]

Ariel, this is a second-best solution to the problem identified in #1 above, but one thing I've done is run ivregress, clustering by country. So, my model looks like this: ivregress 2sls hojust yr relz l.gdpc1000 (l.ngo=iv), vce(cluster country). The results aren't identifical to the results from my xtivreg model (above), but it does allow me to use the estat endogenous post-estimation command.

[Ariel Karlinsky]

Since the results are not identical this seems to suggest to me that this method is flawed... Your'e using tests that are probably(for some reason unknown to me) are unvalid with XT models, and just like you can't report a t stat for a coefficent from a reduced model as the t stat for a coefficent from a full model, I doubt that it's good practice to report various IV tests not from the model your'e actually running...

[Ariel Karlinsky]

Since the results are not identical this seems to suggest to me that this method is flawed... Your'e using tests that are probably(for some reason unknown to me) are unvalid with XT models, and just like you can't report a t stat for a coefficent from a reduced model as the t stat for a coefficent from a full model, I doubt that it's good practice to report various IV tests not from the model your'e actually running and presenting...