1) GREAT Vids!!! 2) Is a matrix completion my sole causal option if my IT/treatment has a dichotomous pattern that looks like 0;1;0;1;1;0;1;1;1;0 (suggesting compounding impacts) over a 10-yr period? I thought I'd be doing a DID until seeing my effect wasn't a one-n-done like a policy might be. Instead, I have staggered effects bc another site might have the first effect in year 7, then nothing again until year 10, PLUS what looks like a compounding effect on about 1500 sites annually over a decade. 3) Any suggestions?
There are a few other approaches that allow for "non-monotonic treatment", Ie treatment that turns off and on. I haven't tried them myself so I can't vouch for any one in particular though. But that's the search term I'd use.
@@NickHuntingtonKlein TY, I'll do that! TBH, I've looked at using a DID with decomposition to show the weight differences in just-effected vs already effected groups vs their untreated counterparts, but IDK. This is all so complicated TBH. I'm now wondering if a fixed effects analysis is my safe route bc it feels like I'm diving repeatedly into the "How the Pros Do It" portion of your chapters! 😙Any thoughts?
@@Tee_Cee_ Fixed effects won't make use of the "control group" element of the design. Using something that takes that into account will probably give you a more plausible estimate. Something to note is that what you want to do here needs to keep in mind what you think the fadeout of the effect is. If you expect the treatment to have a lasting effect of any kind, then in a treatment pattern of 1, 0, 1, etc., the unit is still being treated to some degree in the 0 period. THinking about this again you may instead want to look at this paper www.nber.org/system/files/working_papers/w29873/w29873.pdf and the associated software packages did_multiplegt (stata) and DIDmultiplegt (R)