Issue with Omitted Cohort Values in CSDID

[Alexa Smith]

Hello, everyone!

I’m working on an analysis using the csdid command in Stata to estimate ATETs across multiple cohorts. While running my model, I encountered an issue where values for certain cohorts are omitted in the output. Specifically, these omissions occur for certain time periods, and I suspect the issue might be related to covariates. What I've done so far:

• Ran the model with and without covariates to test their impact. Omissions are less frequent when covariates are excluded.
• Checked for missing data. Some of the covariates have lots of missing values or lack sufficient variation within specific cohort.
• Ran a VIF test to check for multicollinearity and dropped concerning variables.

Is there a way to determine which specific paired observations were kept or dropped when computing the ATET for each cohort? I’m particularly interested in identifying these observations to better understand why some cohorts are omitted.

Does anyone have recommendations for identifying problematic covariates (e.g., lack of variation) that might lead to cohort omissions in csdid?

Thank you very much!

[FernandoRios]

There is no direct way of doing it
What you may want to do is identify your cohorts for each year (typically a tabulation of tvar and gvar)
This gives you the "sample" sizes for each group.
Then look into the cohorts you think are incorrectly omitted, and try to look only into those records to run the regressions you are interested in.
Within that sample, And the potential control for that group, you can check if you can run a simple regression with all your covariates. If any comes collinear, you cant include it in CSDID. Or at least you should be very careful of doing it.

Again, the goal is to look into the smallest group, because that is the driving element to determine model specification (or that is how I think about it)