31:50 Why are states that switched before the data started having a high weight, and not a low weight? Also, it is bad to use already treated states as controls, right? So it would be a bad idea to include these states.
Two quick questions: 1. If I am filtering my sample population to better capture the target demographic of the policy that I am evaluating, is this considered conditioning the parallel trends assumption on a set of pre-treatment covariates? 2. I am a bit confused about post-treatment parallel trends. If one of my treatment groups is negatively impacted by the policy, whereas the rest are positively impacted, is this a volition of this assumption?
Thank you for this. However, I must appoint that the estimator does fail when we have a common year for the treatment intervention for all treated units. Secondly, the explanation of how to define gvar is really poor in both the code and the video
Brilliant, thank you so much for sharing this Taylor, and than you Kirill for the amazing work. This has helped me tremendously with my master's thesis.
Hello, thank you for the video. I've been trying to replicate this methodology but I'm having issues with defining the gvar and get an error message when I run the code. I currently have a dummy for that variable, indicating when a treatment was implemented. Instead of having a value of 1 or 0 for when the treatment was implemented, how should my variable be coded in my dataset? I would really appreciate any advice! Thank you so much!
Hi Petra, hope you solved your issue. If not, the group variable is an id for assign each entry to a group, it isn't a dummy. For example: Say i have 100 entries, each entry is a individual. The first 50 represents the group that's treated in 2004. The last 50 represents the group that's treated in 2005. In this case, we would have two groups: group 1, all the entries that's treated in 2004 and group 2, all the entries that's treated in 2005. Hope it could help you or anyone with the same issue.
It's been a little while since I looked at it, but I believe the code I had used can be found here: github.com/taylorjwright/did_reading_group/blob/main/callaway_santanna/callaway_santanna_castle.R You can also see some examples from Pedro and Brant in their vignette: bcallaway11.github.io/did/articles/did-basics.html#other-features-of-the-did-package
@@taylorwright3880 Hi, the code doesn't work at line 76-80. The error is Error in pre_process_did(yname = yname, tname = tname, idname = idname, : data[, tname] must be numeric
Thanks for this presentation - it complements well your paper. I have one question, does the 'not yet treated' group include cases that are 'never treated'? Also, a slight typo on slide 20, missing a '['.
Hey Taylor, got my notifications on for your channel! Such rich content that you`re sharing with us on these videos. Really helpful for researchers trying to improve and get updated with applied econometrics. Thank you and hoping for more of these.
You're saving my life right now. I've been stuck with my masters thesis and I couldn't find any good explanations in pervious literature on how this actually works in practically. Thank you so much for explaining this in a way that is comprehensible to mere mortals. Thanks, Taylor and thanks Andrew for coming up clutch.
This is a really impressive talk! Focuses absolutely on the essential arguments and insights. And - depressingly rare for an economist - Andrew has really crafted this presentation. Thanks so much Taylor for this video series.
Can anyone confirm that the "Partial out fixed effects" step at 12:17 only works when you've observed data for every unit at every time point (i.e., fully balanced)?