Difference in Differences with an endogenous shock?

[Nur Jahan]

I wonder how problematic it is to conduct a study using difference in differences when the shock (and the treatment) being studied is of an endogenous nature. Do I still need to fix the endogeneity problem of the shock or the treatment using instrumental variables ?

[Maxence Morlet]

If you're a microeconomist, unless you have a valid instrument as you suggested, this problem is not surmountable. Is treatment driven by the outcome, or is it driven by another variable, observed or unobserved and time-varying? Follow-up to the last question, do you have panel data?

The jist is that if you have instruments for the endogenous treatment, and you can argue their validity, feel free to use them! Their use is actually more than encouraged in this situation.

[Nur Jahan]

I am a microeconomist! So the treatment is driven by both other unobserved and time-varying variables and by the outcome as well. I do have an extensive panel data covering years before and after the treatment period. In fact, a DiD setting was ideal given the panel data and the research question I have but the endogeneity problem makes me wonder if I should use a DiD, an IV or maybe an IV in a DiD. What would you advise me to do in this case?

[Maxence Morlet]

Using a relevant but exogenous instrument for the treatment is pretty much your only hope.

Microeconomists, and more specifically labour economists, will never look past the fact that treatment is not exogenous. But using a TSLS IV with a Wald estimator (is your treatment binary?) may be a solution.

There is a community-contributed command you may want to check out, called

ivreghdfe

I think.

[Nur Jahan]

My treatment is continuous. Thank you very much for the suggestion.