Goodman-Bacon Decomposition #1232
Conversation
✅ Deploy Preview for unique-scone-f4eb67 ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
Hi @arutyunovv nice job! I haven't done a detailed review but I believe we have some articles about why TWFE may fail in staggered settings, @srosh2000 or @valerievossen maybe you can point out some relevant ones for Victor to cross reference? |
|
Hi @arutyunovv, It reads super clear! I added only a few small comments.
|
|
Hi @valerievossen and @shrabasteebanerjee, Thanks a lot for your comments!
|
|
@arutyunovv I was thinking the same, could be a bit shorter. What about "Uncovering Staggered DiD Bias with the Goodman-Bacon Decomposition"? |
There was a problem hiding this comment.
Great job @arutyunovv! Added some minor suggestions but otherwise its good to go! :)
| {{<katex>}} | ||
| {{</katex>}} | ||
|
|
||
| The problem with negative weights is that they can bias our final estimate by altering the direction (sign) of the effect. Let’s suppose we have three comparisons with coefficients of 0.5, 1, and 5 and weights of 0.8, 0.8, and –0.6, respectively. Our total effect is then: $0.5*0.8 + 1*0.8 + 5*(-0.6) = -1.8$. We therefore obtain a negative treatment effect estimate even though each of the individual comparisons yields a positive effect of the treatment! |
There was a problem hiding this comment.
the total effect equation is not rendering properly when locally hosted for some reason..
|
|
||
| The decomposition also allows us to see the weight attached to each type of comparison. This way, we are able to see how our overall TWFE estimate is constructed and whether it is prone to the limitations and biases discussed above. In practice, computing the Goodman-Bacon decomposition is very straightforward with the `bacondecomp` [package](https://github.com/evanjflack/bacondecomp) in R. | ||
|
|
||
| ## An example in R |
There was a problem hiding this comment.
We discussed earlier about keeping the example consistent throughout the article so that people can connect the code to the intuition but if you feel like the divorce dataset is not straightforward to explain the intuition then lets keep it as is. Either is fine by me :)
|
|
||
| The Goodman-Bacon decomposition is a diagnostic tool which allows us to see whether our TWFE estimator is reliant on forbidden comparisons or negative weights. The decomposition breaks down the estimator into three types of comparisons: | ||
|
|
||
| 1. Treated vs. never treated: this is equivalent to the classic 2x2 DiD comparison. We are only comparing treated to untreated units. |
There was a problem hiding this comment.
Very minor: Good to render the bullet headers (e.g. treated vs never treated) in bold or italics
|
Hi @srosh2000, Thanks for your comments! I've now made the changes you suggested - let me know whether you think there is anything else to change still. @shrabasteebanerjee I'm also tagging you so that you can have a final look before merging. Thanks! Best, |
Hi @valerievossen, @srosh2000, and @alexandervossen,
I hope you've all had a great weekend!
I am opening a pull request for my topic on the Goodman-Bacon decomposition. I also went into some detail on why the TWFE estimator may fail in settings with staggered treatment timing, so it could be worthwhile to change the title of the article to reflect that.
Please let me know what improvements you think could be made and also what you think about the title issue. Thank you!
Best,
Victor