Correct treatment of strictly exogenous variables with Xtabond2

[Frederik Federspiel]

Hi there, I'm trying to figure out how to correctly treat what I assume to be strictly exogenous variables using the Xtabond2 command:

My simplified code is as follows:

xtabond2 y L.y L.x1 x2 yr*, gmm(L.y L.x1, laglimits(1 5) collapse) iv(x2 yr*) twostep robust small

Should I add the eq(level) suboption in the iv-specification, if I believe x2 is strictly exogenous, and x1 endogenous?

[David Roodman]

I think "should" might be too strong because the identifying information being brought to bear is about the same either way. But yes, eq(level) is more standard. I think that's how it's done in the original Blundell-Bond paper. See the bbest.do file that accompanies xtabond2.

[Sebastian Kripfganz]

In the dynamic panel literature, "strictly exogenous" refers to the variables being uncorrelated with the idiosyncratic error component and any lag or lead of it. The eq(level) option imposes an even stronger assumption (as does the eq(both) option, which is the default) that those variables are also uncorrelated with the time-invariant unit-specific error component. This effectively becomes a "random-effects" assumption, which is much stronger than strict exogeneity as defined in my first sentence.

Blundell and Bond use first differences of the variables as instruments for the level model under the additional assumption that those first differences are uncorrelated with the unit-specific error component.