Treatment effects (teffects) for change over time (pre- and post-treatment panel data)
Dear Statalist-users,
I have a question regarding the appropriate use of the teffects command. I hope someone cares to state an opinion on the matter – that would be a huge help! I haven’t been able to identify an answer to a similar question elsewhere.
I have a panel dataset containing data from a non-randomized study of a social intervention for children. I have 760 children for whom there are both pre- and post-treatment observations, meaning data from two time points in a wide format.
As the intervention was not controlled, I cannot simply take the difference in differences, as there are potential confounders to treatment that might bias the estimates of treatment effects. I have a list of observed socio-economic confounder-variables that need to be included in the models to hopefully balance the potential bias.
Thus, I have been looking at the teffects command – perhaps using propensity score matching - which seems suited for my purposes. The only problem is that what I wish to estimate is the difference in individual development over time for respondents – and not simply the difference in ATE from the post-treatment cross-section.
Thus, I have a number of ‘change variables’ that contain the change in observed outcomes between the two time points (T1 and T2), created by simply subtracting individual pre-treatment outcomes from post-treatment outcomes (T2 – T1 = change). What I am interested in is thus not just the difference between treated and untreated children at T2, but the difference in (positive or negative) development over time between the treated and the untreated.
My question is: Can I use the teffects (psmatch) command to estimate ATE for the change in outcomes? Does it sound feasible, or do I violate any assumptions here? I can’t seem to figure that out. I could use a fixed effects estimator, but I would like to control for a number of time-stable confounders, and I believe that wouldn’t be possible with a FE estimator.
Thankful for any advice you can provide me!
All the best,
Laerke
I have a question regarding the appropriate use of the teffects command. I hope someone cares to state an opinion on the matter – that would be a huge help! I haven’t been able to identify an answer to a similar question elsewhere.
I have a panel dataset containing data from a non-randomized study of a social intervention for children. I have 760 children for whom there are both pre- and post-treatment observations, meaning data from two time points in a wide format.
As the intervention was not controlled, I cannot simply take the difference in differences, as there are potential confounders to treatment that might bias the estimates of treatment effects. I have a list of observed socio-economic confounder-variables that need to be included in the models to hopefully balance the potential bias.
Thus, I have been looking at the teffects command – perhaps using propensity score matching - which seems suited for my purposes. The only problem is that what I wish to estimate is the difference in individual development over time for respondents – and not simply the difference in ATE from the post-treatment cross-section.
Thus, I have a number of ‘change variables’ that contain the change in observed outcomes between the two time points (T1 and T2), created by simply subtracting individual pre-treatment outcomes from post-treatment outcomes (T2 – T1 = change). What I am interested in is thus not just the difference between treated and untreated children at T2, but the difference in (positive or negative) development over time between the treated and the untreated.
My question is: Can I use the teffects (psmatch) command to estimate ATE for the change in outcomes? Does it sound feasible, or do I violate any assumptions here? I can’t seem to figure that out. I could use a fixed effects estimator, but I would like to control for a number of time-stable confounders, and I believe that wouldn’t be possible with a FE estimator.
Thankful for any advice you can provide me!
All the best,
Laerke
