@jessebruhn.bsky.social, Sai, and I are almost done with a paper that jointly estimates teacher VA and school VA. We're finding (in NC) school test effects of about .08 student SDs and that ~30% of that is due to variation in avg. teacher quality across schools.

It's not quite that, but I believe Bruce Sacerdote has a study using random adoption matching to get at it a bit.

"You either play hide and seek with me now, or pay for my psychologist in 5 years. Your choice daddy."

Isn't "fucking around at margins" more or less the definition of economics?

(Also, stop trying to guilt me to spend more time with my kids. They're good enough at that on their own, thank you very much.)

It's not just a question of "which matters more" from a causal inference sense but also "which can we use public policy to improve," no?

I've always wanted to meet the magical whiskey fairy. I'm ashamed to admit that I even started to doubt whether she actually existed, but you've given me hope Jesse.

I think a big part of it is that it's really hard to separate teacher effects from school effects, which you need to do if you really want to understand sorting. I'm working on a paper that tries to do that, in part to answer the sorting question, but it's a tricky empirical question.

4/ What makes me most excited about this work is that I think it can also be a useful building block for estimating effects in other complicated settings, such as averaging RCT and non-RCT estimates, RDs, judge IVs, etc.

So definitely let me know what you think of it, good or bad!

3/ Even more interesting, it turns out we can decompose the resulting posterior variance into that generated by traditional statistical uncertainty vs extrapolation uncertainty.

Much to my surprise, even in the huge Oregon Medicaid experience the statistical uncertainty seems to dominate.

2/ The basic idea is that by adding a Bayesian hierarchical model to traditional econometric models of selection, we can generate estimates/CIs of the ATE (or other estimands) that capture uncertainty in both the true values of the observed moments and uncertainty in how to extrapolate from these.

Isn't it partly a question of what you hold fixed and what is considered random when thinking about inference? E.g. if the size of spillovers depends on who is grouped, then it depends on treatment assignment and so cluster? If not, then they're fixed and so you don't need to cluster?

Oh wow, I hadn't even heard that!

See there's the difference between us: unlike you, I haven't gotten a grant in a while, but this was 100% me after putting my toddler to bed last night.