Best of #econtwitter - Week of January 30, 2022 [4/3] special metrics edition
Because there were already three parts to last week’s newsletter (part one, two, three), broken out here is a high concentration of good metrics content from last week:
![Twitter avatar for @theJohnBonney](https://substackcdn.com/image/twitter_name/w_96/theJohnBonney.jpg)
Hi there #econtwitter—you may have noticed my new paper with @cblandhol, Magne, and Alex: "When is TSLS Actually LATE?" I am stoked about this paper and wanted to briefly summarize our findings. (1/n)
Paper here: papers.ssrn.com/sol3/papers.cf…
![Image](https://substackcdn.com/image/fetch/w_600,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fpbs.substack.com%2Fmedia%2FFKNFAiOVgAAUGWm.png)
![Twitter avatar for @theJohnBonney](https://substackcdn.com/image/twitter_name/w_96/theJohnBonney.jpg)
Our findings? Without additional parametric assumptions, there is precisely one 2SLS specification that is *guaranteed* to be a positively-weighted average of LATEs. Remember that SW specification? The one that no one ever uses? Yep, that’s the only one. Oof. (9/n)
![Twitter avatar for @theJohnBonney](https://substackcdn.com/image/twitter_name/w_96/theJohnBonney.jpg)
Where does this leave us? Well, unless you are using the SW approach (and we both know you aren’t), then under the usual IV assumptions alone, 2SLS is *not* LATE. You must also assume that you have rich covariates and that your 1st stage is monotonicity-correct. (13/n)
![Twitter avatar for @matt_blackwell](https://substackcdn.com/image/twitter_name/w_96/matt_blackwell.jpg)
Kudos to everyone who is figuring out each case, but maybe just make your working hypothesis that if you’re using least squares with anything other than a fully saturated discrete covariates, you may not be getting what you expect.
![Twitter avatar for @agoodmanbacon](https://substackcdn.com/image/twitter_name/w_96/agoodmanbacon.jpg)
Every day I stray farther from the view that controlling for things using a regression is a good idea.
---
![Twitter avatar for @instrumenthull](https://substackcdn.com/image/twitter_name/w_96/instrumenthull.jpg)
Ok, so I come bearing good news for ~93% of you: esp. those bootstraping complex models (e.g. w/many FEs)
Instead of resampling, which can be seen as reweighting by a random integer W that may be zero, you can reweight by a random non-zero non-integer W
![Twitter avatar for @instrumenthull](https://substackcdn.com/image/twitter_name/w_40/instrumenthull.jpg)
Peter Hull @instrumenthull
^thread, lots of discussion in the replies/QTs, truly the best of econtwitter, including:
![Twitter avatar for @grant_mcdermott](https://substackcdn.com/image/twitter_name/w_96/grant_mcdermott.jpg)
Okay, here’s my quick attempt at making this a fast, reusable function. gist.github.com/grantmcdermott…
I _think_ it’s working, but please kick the tyres (examples at the bottom). If all looks good then I might port to a package.
![Twitter avatar for @instrumenthull](https://substackcdn.com/image/twitter_name/w_40/instrumenthull.jpg)
Peter Hull @instrumenthull
![Twitter avatar for @paulgp](https://substackcdn.com/image/twitter_name/w_96/paulgp.jpg)
Here's a nice write-up by Bryan Graham on this method here:
bryangraham.github.io/econometrics/d…
also Chamberlain and Imbens (2000):
nber.org/papers/t0200
(ungated older version here: scholar.harvard.edu/chamberlain/fi…)
![Image](https://substackcdn.com/image/fetch/w_600,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fpbs.substack.com%2Fmedia%2FFKYL0cfWUAADkgZ.png)
![Twitter avatar for @instrumenthull](https://substackcdn.com/image/twitter_name/w_40/instrumenthull.jpg)
Peter Hull @instrumenthull
---
![Twitter avatar for @jiafengchen42](https://substackcdn.com/image/twitter_name/w_96/jiafengchen42.jpg)
Excited to share a new preprint with @DRitzwoller on semiparametric estimation (read, *machine learning* estimation) of long-term treatment effects, such as those identified by a latent unconfoundedness or a surrogacy assumption (1/n)
arxiv.org/abs/2107.14405
![Twitter avatar for @causalinf](https://substackcdn.com/image/twitter_name/w_96/causalinf.jpg)
Couple things I'm learning today:
1) Limited information maximum likelihood (LIML) originates in the Cowles Commission work on simultaneous equation modeling and instrumental variable methods in the 1940s and 50s by Haavelmo, Koopmans and others
1/n