Timeline for Difference-in-difference model with mediators: Estimating the effect of different elements of a policy
Current License: CC BY-SA 3.0
6 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 21, 2017 at 15:51 | comment | added | Julian Schuessler | @Peter OPs first question implies D potentially affects Y through channels other than the specific M. Then D can not be a valid IV. Certainly also not conditional on the other Ms, if these and Y are also driven by unobserved confounders. Then, even conditional on the other Ms, D and Y will be connected because one conditions on colliders. | |
| Feb 21, 2017 at 14:57 | comment | added | Peter | @JulianSchuessler, that is why I started my reply with "For the second question,..." | |
| Feb 20, 2017 at 11:50 | comment | added | Julian Schuessler | This does not make any sense. OP explicitly asks how to check whether M mediates the effect of D on Y. Your approach ASSUMES it COMPLETELY mediates, and provides an altnerative approach to identify the ATE of D, which the OP assumes is identified anyways via DiD! | |
| Feb 13, 2017 at 21:04 | comment | added | Peter | @greg Slapping myself as I forgot the exclusion restriction. I agree that you could control for the other observed Ms, but you'd be assuming there are no unobserved Ms. | |
| Feb 13, 2017 at 17:41 | comment | added | greg | I thought about IV but the problem is the exclusion restriction -- i.e. the assumption that the only link between D and Y goes through M_j. I could control for the other measured M, which partly addresses the problem, but relies on the assumption that there are no other unobserved M. Correct? | |
| Feb 13, 2017 at 15:23 | history | answered | Peter | CC BY-SA 3.0 |