“eteffects” command in the absences of instrumental variables

[Kaleb Jada]

Dear Statalist,

I would like to know if I can use "eteffects" command introduced in Stata 14 to estimate endogenous treatment effect model in the absence of instrument variable for the treatment assignment.

I am interested to see the impact of a treatment on a given outcome using observational data. As it happens, the treatment assignment is not random. In addition, I have reasons to believe that unobservable characteristics of my subjects affect both the treatment assignment and the outcome variable. One way I can think of addressing this issue would be to find instrumental variable/s for the treatment. Unfortunately, there is no such variable in my dataset. Is it possible to use the "eteffects" command in this case?

Kind regards,

[Phil Bromiley]

First, you'll increase your chances of a helpful answer if you follow the FAQ on asking questions - provide Stata code in code delimiters, Stata output, and sample data using dataex.

Others on this list are expert in this area, but I am not. However, they have not chosen to answer your question. With that caveat, I suspect it does not work. If you just have an intercept, then your error term looks just like the treatment (with a change in mean). I'm not even sure it would estimate - wouldn't it be co-linear with the treatment in the outcome equation? If you could get consistent estimates with no variable in the treatment equation, then we would never need to bother with other variables in the treatment equation.

I tried it with the sample used in the documentation:
use http://www.stata-press.com/data/r14/cattaneo2

eteffects (bweight i.prenatal1 i.mmarried mage i.fbaby) ///
(mbsmoke i.mmarried mage i.fbaby medu fedu)

This runs fine,but
eteffects (bweight i.prenatal1 i.mmarried mage i.fbaby) ///
(mbsmoke )

Doesn't run - not concave and not improving the criterion.

[FernandoRios]

Kaleb Jada Similar to what Phil already said.
The easy answer is Yes. Just like when using the command -heckman- you do not need an instrument to obtain the estimates. The identification in this case comes from the nonlinearity of the Inverse Mills Ratio.
However, you may find in many textbooks that in absence of an instrumental variable (a variable that explains the selection/treatment but not the outcome), the results may not be consistent, and will be hard to justify.
Hope this Helps
Fernando

[Kaleb Jada]

Phil and Fernando, thank you. If I have appropriate instrumental variable/s to explain the treatment assignment, why would I want to use eteffects command instead of ivregress then?

[Fissha Asmare]

Hi Kaleb Jada , I am also wondering if the etefefcts Stata command provides a consistent estimate without having a valid instrument. Did you figure it out?


Thanks!