- Cosma Shalizi on (first linear, then nonparametric) instrumental variables from first principles.
- Dan Mackinlay on transfer learning / domain adaptation
- Francis Bach on Kernel Regression and the curse of dimensionality
- Barely a couple of weeks old, but already famous 4 parter on scientific fraud in social psych that beggars belief
- Scott Cunningham on Diff-in-Diffs vs Synth, with some Tom Friedman-esque metaphor mixing and distracting flights of fancy
- Extracts
Something about synth has always felt like an old sheep Dog. It doesn’t pretend to be a greyhound and make promises in part because it overshares.
But that’s like saying “Use a hammer to cook your dinner unless your dinner isn’t a nail, in which case use an oven.” It’s kind of odd to tell someone to try to use something for which it isn’t designed — you should instead go directly to the tool that is designed for that problem. When you need to hang a picture, you use a hammer. When you need to cook lasagna, you don’t, and you shouldn’t try unless you’re wanting to make a huge mess. When it’s dinner time, skip the hammer and go to the oven.
Charpentier's probabilistic foundations series
# econometrics
- https://freakonometrics.hypotheses.org/57649
- https://freakonometrics.hypotheses.org/57674
- https://freakonometrics.hypotheses.org/57693
- https://freakonometrics.hypotheses.org/57703
# machine learning
- https://freakonometrics.hypotheses.org/57705
- https://freakonometrics.hypotheses.org/57745
- https://freakonometrics.hypotheses.org/57782
- https://freakonometrics.hypotheses.org/57790
- https://freakonometrics.hypotheses.org/57813