Apple Vision Presents: Why is TWFE biased in some event study graphs but not others?

This is a video I did for a substack. In it, I try to explain why sometimes a two-way fixed effects event study plot is biased, and other times it isn't, by focusing on two papers. One of them was in the American Economic Review published in 2022 entitled “Social Media and Mental Health” by Luca Braghieri, Ro’ee Levy, and Alexey Makarin. And the second one is as of this posting a working paper by Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond entitled “Generative AI at Work”. Both of them are differential timing difference-in-differences studies where the treatment (Facebook in the first; generative AI chatbot in the second) gets adopted by different groups of units (schools; workers) at different dates. In the Facebook example, the two-way fixed effects estimates are very different (and not statistically significant) from the robust estimators, suggesting that the estimates are perhaps biased. But in the generative AI example, the two-way fixed effects estimates are nearly the same as the robust methods. Why? I try to explain what could be going on, both in the substack below, but also using my Apple Vision Pro. I wanted people to see me sticking my arms through the event study coefficients as I'm curious if watching me move between the various papers and plots was helpful. Warning, though -- because the Apple Vision is mounted to my head, the picture moves when I move. So it can be shaky, and for it not to be shaky, I would have to change my entire body language, and the concentration it takes is easy to lose when you start concentrating on the papers. Anyway, I thought some of you might want to see this new explainer of Sun and Abraham's decomposition also, with an applied example.
causalinf.subs...