Good thread here on why true Bayesian models aren’t as straightforward as just weighting predictions together

The null result penalty in economics: "Studies with null results are perceived to be less publishable, of lower quality, less important and less precisely estimated than studies with large and statistically significant results, even holding constant all other study features"

In certain parts of psychology, statistical modeling is an arcane art that requires intimate familiarity with contraptions known mostly by their acronyms (CLPM, RICLPM, ARTS, STARTS, LGCM...). Let's mix things up by putting substance first instead.

Two years ago, a single wind energy project in Germany with just three turbines required 36,000 pages of documentation. Then they passed permitting reform to cut red tape. Now Germany is a leader in clean energy deployment. Some lessons here for the United States...

Useful info: 1) Audit studies on racial discrimination in hiring suffer from publication bias. 2) There is still strong evidence of racial discrimination after adjusting for publication bias.

It feels under-discussed (at least on Twitter) that after three decades dodging every economic slowdown, Australia is about to have a lost decade.

"Findings unveil teacher bias in an essay grading task favoring girls and highbrow cultural capital, aligning with status characteristics and cultural capital theories." This month in our journal, Soumya Sankar Das

George Callas This whole line of research is completely ridiculous. The idea that you can add up the number of times "shall" appears in the CFR to quantify the "amount of regulation" and run regressions on it is something I'd come up with to make fun of economics research

Seeing Like A State is a brilliant book IMO, but it’s wild that anyone reads it as more than one side of an obvious dialectic. It’s ironically so uninterested in the complexity (and benefits) of the systems it critiques.

Got the exciting news that our paper "Zooming to Class" is out in print today! During the Pandemic, Department of Social Sciences at West Point allowed us to randomize cadets to online of F2F class sections. We found that Covid online learning had large negative effects on outcomes. aeaweb.org/articles?id=10…

"Amortize the cost of inference over time" is a phrase I regularly use in talks (to describe sequential/online Bayesian inverse planning) but I always am a little worried people think I'm talking about "amortized (variational) inference" instead.

Fallacious argument. A tax that kicks in at some wealth threshold applies to all Americans though. In this case, nobody's 100,000,001st dollar is excluded.

Genuinely not anti-AI but I think a lot about how much room there is to build better software today that has nothing to do with AI-enabled features. Faster, more stable, more secure, more useful... Most software isn't very good! We're leaving it all on the table!

Friday feature drop: Highlight text or code within an Artifact and quickly have Claude improve or explain the selection.

Interesting discussion of something we discussed in our (John Haltiwanger) Brookings paper last fall: that the average size of new firms dropped in the pandemic. But there are a couple of important nuances here.

🚨 New paper in the Journal of Development Econ. Are migrants from poor countries relatively high or low productivity workers? What happens to migration as incomes rise? Mariapia Mendola and I note: These are closely related questions with a counterintutive answer (thread)

"The Free Press did not publish corrections or clarifications, which only confirms my point that their vows to cover 'the world as it actually is' & pierce 'ideological narratives' ring hollow...that familiar, matter-of-fact, just-common-sense confidence of a heterodox pundit"

The "2024's best estimate" here is just the Lynn data again. A couple of race science bloggers averaged all the iterations of Lynn's data to make a "best estimate". This is like saying the data is valid because it correlates well with a copy and paste of itself