Test for instrument validity.

[Jane Wilson]

Hi all, I am estimating the following using 2SLS.

xtivreg lemp lunion lunemployment lyouth i.year (lfertility = lagfertility), fe vce(cluster id)

I was hoping someone would be able to advise me as to whether there are tests I can run to check that my instrumental variable is valid? Thank you.

[Carlo Lazzaro]

Jane:
I would take a look at the user-written command -xtoverid- (type -search xtoverid- from within Stata to spot and install it).
Please note that, being a bit old-fashioned, -xtoverid- does not allow -fvvarlist- notation (see -help xi:- in this respect).

[Joao Santos Silva]

Dear Jane Wilson,

Adding to the excellent advice provided by Carlo, please note that the test of the validity of over-identifying restrictions is just that; not a test of instrument validity. Instrument validity is an identifying assumption and as such cannot be tested.

Best wishes,

Joao

[River Huang]

Dear Jane, Please consider (ssc install) xtivreg2.

[Jane Wilson]

Dear all, thank you for your help so far.

When running xtoverid I obtain the following:

xtoverid

Test of overidentifying restrictions: 
Cross-section time-series model: xtivreg fe  robust cluster(id)
Sargan-Hansen statistic   0.000  (equation exactly identified)

I am unsure how to proceed from here, as examples I have looked at do not share the same output. I would appreciate any help with this!

Thanks

[Carlo Lazzaro]

Jane:
-xtoverid- output mirrors the evidence that you have one endogenous regressor and one instrument only.

[Jane Wilson]

Thank you Carlo! One more question, xtoverid is not producing output when I include i.year. When i remove this term, it does produce output. Is there any way of getting an output including i.year? Thanks

[Carlo Lazzaro]

Jane:
being a bit old-fashioned, -xtoverid- does not allow -fvvarlist- notation.
The usual trick is to prefix your code by -xi:-:

xi: xtivreg lemp lunion lunemployment lyouth i.year (lfertility = lagfertility), fe vce(cluster id)
xtoverid

[Jane Wilson]

Hi Carlo, for reason reason when I run that code, it still doesn't work. Do you know anything that may help resolve this issue?

[Carlo Lazzaro]

Jane:
as you know (and as reminded by the FAQ) stating the something (code, command, function) does not work as expected (and our expectation can be far different from the way the code, command, function actually works as per our instructions) is not helpful.
Please share what you typed an what Stata gave you back (via CODe delimiters). Thanks.

[Jane Wilson]

Carlo,

I typed:

xi: xtivreg lemp lunion lunemployment lyouth i.year (lfertility = lagfertility lagfertility2), fe vce(cluster id)

Stata gave back:

  xtoverid
o. operator not allowed

[Carlo Lazzaro]

Jane:
-xtoverid- does not allow o.operator (a reminder of an omitted variable).
The only chance is to re-estimate the miodel without the omitted variable (this can be pretty unsatisfying under -fe- specification and time-invariant predictors).

[Jane Wilson]

Thank you for your help so far Carlo. Would there be any other way to test the validity of the instruments so that I could include year fixed effects?

[Carlo Lazzaro]

Jane:
as Joao Santos Silva helpfully pointed out in #2, you cannot test the validity of your instruments, but the validity of over-identifying restrictions.
That said, I'm not aware of othet test that can accomplish what you're after.
I still believe that the best option at hand is to re-run your model without the omitted variable (that would be omitted anyhow) and then call -xtoverid-.

[Jane Wilson]

Thank you for all your help Carlo!