Dear statalist users,
I am currently conducting a difference-in-difference approach and I need to empirically test the parallel trends assumption. Basically, our data covers 8 years and the time event (Post) starts in year 4 and we have one treatment variable (Treat). I was thinking of running the following model:
Y = a + β1 Treat*Year2+ β2 Treat *Year3+ β3 Treat *Post + control variables
From what I have read is that in a parallel trends assumption test, there should be a base period (e.g. Year1 in the model above) that should not be interacted with the treatment variable. This way, the coefficients of Treat*Year2 & Treat*Year3 would measure whether the difference between the base year (Year1) and each of the other pre-period years (Year2 & Year 3) is significant. An insignificant coefficient on β1 & β2 would imply that the parallel trends assumption holds. I just want to confirm that the above is accurate as I am not very familiar with this approach
Another question is related to the Treat*Post variable in the model above. In the original DiD model, Treat*Post is negative and significant at the 5% level. However, after running some preliminary analysis for the model above, the Treat*Post coefficient loses significance (p-value ~ 0.2). Should this be a concern? A colleague suggested that this might not be relevant as the Treat*Post coefficient in this case only measures the difference between the Post period and the base year and not the whole pre-period. Thus, Treat*Post in the model above does not capture the true difference-in-difference between the two periods.
If this is not truly a concern, do you think a better alternative is to just apply the model above for the pre-period years as the interpretation of Treat*Post might be a bit misleading as it differs from the original DiD model? In such a case, the model would simply be:
Y = a + β1 Treat*Year2+ β2 Treat *Year3+ control variables if year <Year4
Thanks a lot in advance for your help
Bilal
I am currently conducting a difference-in-difference approach and I need to empirically test the parallel trends assumption. Basically, our data covers 8 years and the time event (Post) starts in year 4 and we have one treatment variable (Treat). I was thinking of running the following model:
Y = a + β1 Treat*Year2+ β2 Treat *Year3+ β3 Treat *Post + control variables
From what I have read is that in a parallel trends assumption test, there should be a base period (e.g. Year1 in the model above) that should not be interacted with the treatment variable. This way, the coefficients of Treat*Year2 & Treat*Year3 would measure whether the difference between the base year (Year1) and each of the other pre-period years (Year2 & Year 3) is significant. An insignificant coefficient on β1 & β2 would imply that the parallel trends assumption holds. I just want to confirm that the above is accurate as I am not very familiar with this approach
Another question is related to the Treat*Post variable in the model above. In the original DiD model, Treat*Post is negative and significant at the 5% level. However, after running some preliminary analysis for the model above, the Treat*Post coefficient loses significance (p-value ~ 0.2). Should this be a concern? A colleague suggested that this might not be relevant as the Treat*Post coefficient in this case only measures the difference between the Post period and the base year and not the whole pre-period. Thus, Treat*Post in the model above does not capture the true difference-in-difference between the two periods.
If this is not truly a concern, do you think a better alternative is to just apply the model above for the pre-period years as the interpretation of Treat*Post might be a bit misleading as it differs from the original DiD model? In such a case, the model would simply be:
Y = a + β1 Treat*Year2+ β2 Treat *Year3+ control variables if year <Year4
Thanks a lot in advance for your help
Bilal
