Parentheses unbalanced - ivregress with interaction instrumented term

[Matthieu Lagarde]

Hello all,

I am trying to estimate the following equation where the variable S is instrumented by Z :
Log(y) on t t*S and a constant, where S is a binary variable potentially endogenous and that is why it is instrumented by Z.

To proceed, I would like to run a 2SLS by estimating first S_hat through a first stage : S on Z
And then replace S by S_hat, thus running this 2nd stage regression :
Log(y) on t t*S_hat and a constant.

To do so, I am trying to use the "ivregress 2sls" command :
ivregress 2sls y t c.t#(c.S=c.E), vce(robust)

Yet, I am getting the following error "parentheses unbalanced r(132);".
I looked at the documentation but could not manage to fix the problem.

Anyone can help ?
Thanks in advance !

Matthieu

[William Lisowski]

Welcome to Statalist, Matthieu.

My reading of help ivregress does not suggest that your factor variable interaction notation is permitted in the way you have used it.

Syntax

        ivregress estimator depvar [varlist1] (varlist2 = varlist_iv) [if] [in] [weight] [, options]

    varlist1 is the list of exogenous variables.

    varlist2 is the list of endogenous variables.

    varlist_iv is the list of exogenous variables used with varlist1 as instruments for varlist2.

It does not appear to support an interaction (#) connecting an exogenous variable and the parenthesized list defining the first stage models for the endogenous variables.

Thus, the misleading syntax error you receive is a result of your syntax being far enough from what the designers of ivregress expected to receive that the program did not handle it correctly - I'm guessing that when the command parsing reached the equal sign the program became totally confused.

Perhaps with this syntatical question out of the way someone more knowledgeable than I will address the more interesting methodological question of how to run two-stage least squares when an endogenous variable interacts with an exogenous variable.

[Matthieu Lagarde]

Hello William,

Thank you for your answer which helps a lot to clarify the problem. You are right, my syntax is too far away from the ivregress proper syntax. The question is thus: how to perform a 2SLS when an endogenous variable -that has to be instrumented- interacts with an exogenous variable? Maybe I should proceed "by hand" running both regression successively but I know the standard errors that I will get doing so will be false and I need to infer on the results I will get from this 2SLS regression. I am also more interested in finding a "direct" and neat way of doing it since I am sure it must exist.

Anyone has any clue ?

[Phil Bromiley]

You can create the interaction manually using generate and then include it in the variables to be instrumented.

[Matthieu Lagarde]

Hello Phil,

Thank you for the suggestion. However, I don't want to instrument the interaction itself, but only the right-hand side of the interaction. To be clear, in the following regression, log(y) on t t*S and a constant, I would like to instrument S on Z and not the whole interaction t*S on Z.

[William Lisowski]

The following discussion seems to support Phil's general recommendation that you instrument the interaction.

http://www.stata.com/statalist/archi.../msg01165.html

It references a text by Jeff Wooldridge; perhaps he and other methodologists have overlooked this discussion since the title suggests syntax problems rather than methodological issues.

A search of the Statalist archives, or more generally a Google search of the web, for

interact endogenous and exogenous variable

finds much more on this topic.