In my paper, I have estimated a triple difference-in-difference model where the outcome variable is a dummy variable indicating if the given individual is employed or not. Let's say that policy that I am evaluating happened in the capital of country $S$, in year 2000. As treated group I choose the capital of country $S$, as control I choose all the other cities in country $S$. The model I am estimating is

$$y = \alpha + \beta\text{State} + \gamma\text{Post} + \omega\text{Capital} +\theta(\text{Post}\times\text{Capital})+ \rho(\text{Capital}\times\text{State})+ \\ \psi(\text{Post}\times\text{State}) + \delta(\text{State}\times\text{Capital}\times\text{Post}) + \epsilon$$

Assume that the outcome variable is a dummy that indicates if the individual $i$ is employed or not. If I estimate the regression model, and I get a coefficient of 0.06 on the triple interaction variable $\delta$, is it correct to say that the probability to be employed of an individual living in the capital of country $S$ in the post-treatment period (treated group) increases by 6 percentages points with respect to a same individual living in the control group (any of the other urban areas of country S) ?

  • $\begingroup$ Welcome to CV paulo morales. A few suggestions to improve your question: (1) "get a[n estimated] coefficient of 0.06" for which coefficient?? ($\beta$? $\gamma$? etc.), and (2) "the probability to be employed of an individual living in the country S in the post-treatment period increases by 6 percentages points?" Please clarify "increases" relative to which group? You can edit your question to make such improvements by clicking the "edit" link in the lower left. $\endgroup$ – Alexis Jan 24 '19 at 20:53
  • 1
    $\begingroup$ Please register your account (you can find information on how to do this in the My Account section of our help center), then you will be able to edit & comment on your own question. $\endgroup$ – gung - Reinstate Monica Jan 24 '19 at 21:11

I think it is easiest to think of $\delta$ as the time change in employment rate for the capital city residents, minus the change in employment for residents in the control cities (all non-capital), and also minus the change in employment for the non-capital city residents in target state S.

The hope is that this controls for two kinds of potentially confounding trends: country-wide trends in employment (that would have nothing to do with the policy, like national economic booms/busts) and changes in employment of other people living in state S (possibly due to other events, like local economic booms/busts).

Assuming the assumptions have been met, the interpretation is a counterfactual one. This is the expected effect of the policy on the average capital resident, relative to his employment without the policy.

This post contains a nice graph of the intuition.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.