I am running a Diff-in-Diff analysis about the triggering of a policy that once triggered bans a certain action for 6 months.

I have run the analysis considering only pre and post period, including in the post-period also the observations that are after the 6-month period, that is, when the ban was actually released.

Now I want to run a robustness check and I want to consider the fact that after the 6-month period the ban is gone.

How can I do this?

I thought I could simply change my EVENT vector, which had a dummy equal $0$ for pre-treatment and $1$ for post-treatment. Could I just change this and do: $0$ for pre-treatment, $1$ for post-treatment, $0$ for released-treatment.

It seems wrong to me. isn't it? should I then change also the treatment vector?


1 Answer 1


The choice depends on whether you expect a lingering effect in the released-from-treatment period. This expectation can come from theory or domain expertise.

If that is sensible in your setting, then you should have two separate post dummies and two treated-post interactions. The two interaction coefficients will allow the effect to vary in the two post-treatment regimes.

The coding scheme you propose would be suitable if you expect the effect to go away immediately (like a light switch being turned off).

  • $\begingroup$ Yeah I expect lingering effects. so you mean to have an event dummy vector = 1 for the dates after the event and 0 for dates before the event. Then, create another event dummy vector = 1 for all dates before the release and 0 after the release. Then, multiply the treatment vector for both event vectors, generating two treatment*event vectors. Then, run everything in one regression. is that right? if that's the case, is this a robustness check? it looks more like an original regression: the main regression that I should run and that should be later tested by another different robustness check. $\endgroup$
    – Mining
    Mar 2, 2022 at 0:02
  • 1
    $\begingroup$ If you have monthly data, I would use post_treatment dummy (with pattern ...0,0,1,1,1,1,1,1,0,0,...) and post_release dummy (...0,0,0,0,0,0,0,0,1,1,1,1,1,1,...) for pre, during, and post months. If you expect a lingering effect, then this would be your main specification and not a robustness check. A potential robustness check would use time dummies and their interactions with treated instead of post dummies. This would add even more flexibility: a separate effect for every period. This could let you test parallel trends in the pre-period, and allow the treatment effect to vary over time. $\endgroup$
    – dimitriy
    Mar 2, 2022 at 0:17
  • $\begingroup$ @Mining To complement the answer above, see the last equation in my answer here. A second interaction is warranted if you suspect a lingering policy effect. $\endgroup$ Mar 2, 2022 at 5:55
  • $\begingroup$ @dimitriy to make sure i understood. I do a post treatment dummy in which 0 is pre treatment, 1 is during treatment, and 0 is after release. then I do a post release dummy in which 0 is pre treatment, 0 is during treatment, and 1 is after release. Then I multiply the treatment (1 for treated and 0 for control) for the vectors explained above and I get 2 main diff in diff coefficients: one for the during treatment and one for the after release. right? $\endgroup$
    – Mining
    Mar 2, 2022 at 12:52
  • $\begingroup$ That is exactly right. $\endgroup$
    – dimitriy
    Mar 2, 2022 at 14:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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