
PhD student @MITEcon «- @UChiUrbanLabs «- @CivisAnalytics Late PhD potpourri: behavioral + methods. Putting the yo in RStudio.
Boston, MA
Joined June 2009
- Tweets910
- Following846
- Followers519
- Likes993
Stay tuned for my eventual tirade on the lasso in econ.
It's a cool tool! But I'd bet:
1) We'd be way more critical of lasso stated as a Laplace prior on coefficients.
2) Cross-validation is secretly doing a lot of work.
3) Lasso's popular because GBTs aren't in Stata.
This Post is from an account that no longer exists. Learn more
1
595
Advik Shreekumar retweeted
I’m thrilled to announce Conformal Risk Control: a way to bound quantities other than coverage with conformal prediction.
arxiv.org/abs/2208.02814
Check out the worked examples in CV and NLP!
The best part is: it’s exactly the same algorithm as split conformal prediction🤯🧵1/5
Retweeting b/c I want to test two hypotheses based on regression:
1) Is A less sensitive to treatment than B? H0: |b1| < |b2|
2) Does treatment push A and B the same way? H0: sign(b1) = sign(b3)
I'm frustrated b/c this is hard with my econ toolkit, but easy with Bayes.
(1/n)
Apparently there was a series of Bayesian statistics conferences in Spain starting in the late 1970s, and they had a tradition of cabaret songs about how great Bayes' theorem is
1
2
I appreciate that econ tries to make few assumptions in regression*.
But it's not costless! Makes it hard to answer questions like:
— Is A less sensitive to treatment than B?
— Does treatment push A and B the same way?
* the assumptions get shifted to the asymptotics.
(n/n)
1
Advik Shreekumar retweeted
Here's a 1-minute video showing--in the words of visualization hero @EdwardTufte--the "compared to what" factor for today's amazing @NASAWebb image. Thanks to the @WWTelescope for making this possible, and @ADavidWeigel. I only wish @NASA and @POTUS had shown the image this way!
Replying to @AlyssaAGoodman @NASAWebb
And you can zoom on the first one (SMACS 0723) interactively here in @WWTelescope: web.wwtassets.org/specials/2…
Advik Shreekumar retweeted
NEW PREPRINT
The increasingly popular class imbalance approaches (such as SMOTE) for risk prediction modeling: they are likely to do more harm than good
arxiv.org/abs/2202.09101
Wordle's been a fun reminder of how many fantastic 5-letter words we have: angst, thyme, inure, grist, haunt, spurn, glyph, psalm, foist ...
I love how they look, feel to say, evoke so much despite being short.
Any crowd favorites?
1
9
To my fellow young researchers: learn from the past.
Use leave-one-out cross-validation.
Important point. Young researchers have mostly missed out on the opportunities we had back in the 1980s and 1990s to start a program running and then claim to be "working" for days on end while it ran. #multitasking
5














