Need Help

[Saransh dhar]

I need guidance on whether I can run a staggered DID to analyse the effect of a treatment variable which is binary on an output variable, but my output variable is something which is generated at the end of the time period, like my period of study is from 2010 to 2023, and treatment is happening in different time periods for different entities and the output variable is something which is generated at the end of 2023. If yes, then it would be helpful if someone could briefly describe how to proceed or share any reference or document of the same.

[Clyde Schechter]

To use DID analysis you must have outcome assessments, at a minimum, both before and after treatment. Ideally you would have them at multiple times both before and after treatment. So, no, this data design is not suitable for a DID analysis.

What you have here is an intervention trial with no pre-intervention outcome measurements. So the only thing you can really do is a simple comparison of final outcomes in the treated and untreated groups. Now, since the intervention starts at different times in different units, but the outcome is always measured in 2023, you have an additional problem: different units will have been exposed to different durations of treatment. In general, this entails using a continuous treatment variable proportional to the expected treatment effect associated with the actual duration of treatment experienced by the unit. The specifics of doing that depend on how you expect the treatment effect to depend on the duration of treatment.

[Saransh dhar]

Cant I use DID to interpret the how treatment is affecting the outcome in comparison to the absence of treatment and if not DID can you mention any other technique which I can look upon for this?

[Clyde Schechter]

No. You can compare 2023 outcome in the presence of treatment to the outcome in the absence of treatment. But with no outcome assessments before 2023, DID is, by definition, impossible. DID means that you have the difference between the outcome values before and after treatment in each group, and then you evaluate the difference between those differences across groups. That's what DID means and what the initialism (?acronym) stands for: difference in differences. But your data design has only after-treatment outcome measures in the treated group, and only pre-treatment outcome measures in the untreated group. So you are missing the last D in DID.

All you can do is a straight treated vs untreated comparison. This is a weaker design because you have nothing to assure that the treated and untreated units were in any relevant way similar before treatment. The most you can do to shore this up is include lots of covariates, or a propensity score analysis, to reduce the influence of confounding variables (omitted variable bias).

If it is possible to obtain additional data from earlier years, you might be able to build up a data set that is suitable for DID analysis.