Hi,
I am estimating a difference-in-differences specification over a monthly panel dataset, and I want to test whether the pre-trends are parallel, so I take a similar approach used in Author (2003).
I estimate the following equation:
y = \beta_0 treated + \beta_1 after + \beta_2 interval x treated,
where treated is an indicator of whether the unit is treated, after is an indicator for the period in which the policy is applied, and interval is a categorical variable indicating the months before and after the treatment (policy is active) is applied, so -2 would be an indicator for 2 months prior to the treatment, 0 is the month in which the policy starter, +1 is the first month after the policy started.
Therefore, \beta_2 measure the difference in outcome (y) between treated and untreated units at different time periods.
Then I plot the estimates of \beta_2 (using as a baseline interval -1, which is set to be 0), and I get the plot I show below. As you can see in the pre-treatment period the difference between treated and untreated units is about 0.2 and significant, while in the treatment period there is a positive jump that I was expecting.
My question is whether I should be worried about the fact that in the pre-treatment the difference between treated and untreated units is not zero, as many papers using this test show.
Thanks for the help!
References
Autor, David H. 2003. Outsourcing at will: The contribution of unjust dismissal doctrine to the growth of employment outsourcing. Journal of labor economics 21(1) 1–42.
I am estimating a difference-in-differences specification over a monthly panel dataset, and I want to test whether the pre-trends are parallel, so I take a similar approach used in Author (2003).
I estimate the following equation:
y = \beta_0 treated + \beta_1 after + \beta_2 interval x treated,
where treated is an indicator of whether the unit is treated, after is an indicator for the period in which the policy is applied, and interval is a categorical variable indicating the months before and after the treatment (policy is active) is applied, so -2 would be an indicator for 2 months prior to the treatment, 0 is the month in which the policy starter, +1 is the first month after the policy started.
Therefore, \beta_2 measure the difference in outcome (y) between treated and untreated units at different time periods.
Then I plot the estimates of \beta_2 (using as a baseline interval -1, which is set to be 0), and I get the plot I show below. As you can see in the pre-treatment period the difference between treated and untreated units is about 0.2 and significant, while in the treatment period there is a positive jump that I was expecting.
My question is whether I should be worried about the fact that in the pre-treatment the difference between treated and untreated units is not zero, as many papers using this test show.
Thanks for the help!
References
Autor, David H. 2003. Outsourcing at will: The contribution of unjust dismissal doctrine to the growth of employment outsourcing. Journal of labor economics 21(1) 1–42.
Comment