Good afternoon,
I am using a synthetic difference-in-differences (SDID) estimator to evaluate the impact of a shock on a very volatile country. I have 1 treated unit, 33 units in my donor pool and 65 years of data (1960-2024) with 3 post-treatment periods. The "sdid" command produces ATTs that are reasonable and statistically significant. The trend graphs look relatively similar and the placebo tests and leave-one-out donor tests all produce desirable results.
The problem is when I inspect the time weights (lambda), it appears that the algorithm places full weight on the period right before the treatment period and ignores the rest. I have tried restricting the time period in different ways (e.g. shorter pre-treatment window, restricting start date to a period after a prior volatile period, etc.) but the issue remains. My question is, is the estimation still valid or should I be concerned? If so, is there a solution to this?
I would appreciate any insight or resources on this.
Thank you,
Naureen.
I am using a synthetic difference-in-differences (SDID) estimator to evaluate the impact of a shock on a very volatile country. I have 1 treated unit, 33 units in my donor pool and 65 years of data (1960-2024) with 3 post-treatment periods. The "sdid" command produces ATTs that are reasonable and statistically significant. The trend graphs look relatively similar and the placebo tests and leave-one-out donor tests all produce desirable results.
The problem is when I inspect the time weights (lambda), it appears that the algorithm places full weight on the period right before the treatment period and ignores the rest. I have tried restricting the time period in different ways (e.g. shorter pre-treatment window, restricting start date to a period after a prior volatile period, etc.) but the issue remains. My question is, is the estimation still valid or should I be concerned? If so, is there a solution to this?
I would appreciate any insight or resources on this.
Thank you,
Naureen.
