I nonetheless bear in mind when the bacon decomposition clicked for me. I had labored via the paper backwards and forwards. The mixtape was underneath contract, and I knew I had a brief window to get on high of this paper that I may instantly inform was going to be influential. It was the weights. It was working via the Frisch-Waugh-Lovell steps and attending to these weighted 2x2s. That was the primary second.
However then the second second was a simulation that andrew baker wrote for his weblog which has staggered roll out with a panel of a thousand companies. No by no means handled, very dynamic therapy results, and the consequences had been recognized. I studied that simulation backwards and forwards.
And that’s how I study I believe. I study via pencil and paper, code, simulations, software, repeat. Doing issues. Notably easy simulations — issues which might be solely about one factor in order that I can see that one factor clear as day.
Nicely I haven’t had that occur but with Callaway, Goodman-Bacon and Sant’Anna’s new steady paper. I’ve learn it lots, and I really feel like I do know the parameters rather well, and I perceive the choice bias conceptually. However I haven’t had that ah hah expertise but. And I don’t know the brand new estimator tremendous effectively both.
So what I’ve determined to do is do a collection on right here the place I’ll be instructing myself steady did utilizing Claude code. I might be making an attempt to create the TWFE decomposition myself, and I’ll probably attempt to make a shiny app, in addition to Stata and R package deal for it too. Even when it doesn’t fairly work out, I believe it’ll assist me as a result of I believe I must know this paper inside and outside if solely as a result of steady therapies are widespread, and to be trustworthy they’re additionally enjoyable. It’s enjoyable to study new issues. So I’ll.
So I’m going to be doing that right here. After 38 Claude code entries, I believe I could stop numbering them altogether. And I’m going again to my authentic thought I had all alongside which is that I’m going to share about Claude code by utilizing it for issues. If that is useful for others, which I’m hoping studying to create econometrics packages might be, then I welcome you to come back do it with me!
I’ve different collection I’ll be doing too on Claude code. Certainly one of them is about utilizing the time stamps within the metadata of our information and folders to attempt to unearth forgotten occasions in our previous initiatives. However I’m going to start out right here with steady did, as I believe there’s lots of people that need assist studying it, and there’s lots of people which may wish to do these workouts with me.
You’ll nearly actually study this quicker than me! I hope if you happen to do, you’ll go away feedback although, and share with each other. You probably have an excellent simulation or one thing, put up it within the feedback. And I simply see it as us making an attempt to study this paper effectively and have enjoyable utilizing Claude code to do it. I’ve just a few steady dosage datasets we’ll use too.
Right here’s the paper. It’s conditionally accepted at AER. And there’s an R package deal too. I don’t know if the TWFE decomposition is in there although. I even have some concepts for covariates that I’m desirous to take us via. And we’re going to do simulations. It received’t be a category a lot because it’ll simply be me doing my factor, you doing all of your factor, after which we are able to chat on right here and no matter.
So perhaps begin studying the paper? Attempt to actually study the parameters in addition to you’ll be able to, and notably the notation. And the choice bias. The sturdy parallel tendencies half has to do with the estimator they’re going to suggest, so for now I’m pondering study the core constructing blocks, see if you happen to can’t study the ACRT rather well, after which let’s attempt to give attention to the TWFE weights. Thats the notation I’ve not liked but, but it surely’s as a result of I haven’t completed the shut research but.
Thanks! I’m enthusiastic about this. My objective is to fully grasp this paper backwards and forwards. I’m tremendous busy, so I’ll need to make stunning decks, tear down and construct up how I do it, and so forth. I haven’t determined one of the simplest ways to arrange every part however I’ll. At minimal I’ve my very own repo I’m engaged on for the TWFE decomposition that I’ll share however I believe you need to make your individual too tbh. And it’ll be enjoyable.
