Instrumental Variables using dummy variables as instruments

[Kimara Saldanha]

Hi, I am using Stata 14 to run my analysis. My original regression is OLS and I want to run an IV regression for a potentially endogenous variable. However, the most relevant variables I can find are dummy variables. I was wondering whether dummies are ok to use and whether I would then have to run a different command in stata as much independent variable would now be binary (would I need to change IV regression to probit/logit - is this allowed?) Many thanks!

[Joao Santos Silva]

Dear Kimara,

You can stick to standard 2SLS. At most, you could consider a particular form of 2SLS if you wanted to instrument a dummy.

Best wishes,

Joao

[Eric Haavind-Berman]

Hi Kimara,

Joao's answer is absoutely correct. As an addition, there are a few pages related to your question in Angrist and Pischke's Mostly Harmless Econometrics in the section Forbidden Regressions (page 190-192) in case you are interested in digging a bit deeper into the question of dummy variables in IV regression.

Eric

[John Francois]

Hi all,

I have a question similar Kimara's. In my case, the IV is a dummy variable and there is one endogenous variable. When I use ivreg2 with country and time fixed effects, the IV variable is dropped and the model becomes unidentified. When I leave out the fixed year effects the model runs fine. So, my guess is ivreg2 is detecting collinearity between the year fixed effects and the IV. What is interesting is that the first-stage regression with year and country fixed effect runs fine and the IV variable doesn't get dropped. Please see the link below. Any help is appreciated.

. xi: ivreg2 lf15_64_f (fert = mcp) i.ctry i.year if regsmpl ==1, robust cluster(id)

i.ctry _Ictry_1-137 (_Ictry_1 for ctry==Afghanistan omitted)
i.year _Iyear_1990-2019 (naturally coded; _Iyear_1990 omitted)
Warning - collinearities detected
Vars dropped: _Ictry_4 _Ictry_35 _Ictry_49 _Ictry_64 _Ictry_66 _Ictry_78
_Ictry_82 _Ictry_128 mcp
equation not identified; must have at least as many instruments not in
the regression as there are instrumented variables

First stage regression results:

xi: reg fert mcp i.ctry i.year if regsmpl ==1, robust cluster(id)

Linear regression Number of obs = 3,863
F(28, 128) = .
Prob > F = .
R-squared = 0.9585
Root MSE = .34656

(Std. Err. adjusted for 129 clusters in id)
------------------------------------------------------------------------------
| Robust
fert | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mcp | .7631016 .0486615 15.68 0.000 6668165 .8593866
_Ictry_2 | -4.482033 6.92e-12 -6.5e+11 0.000 -4.482033 -4.482033
_Ictry_3 | -3.471633 6.91e-12 -5.0e+11 0.000 -3.471633 -3.4