Lightnews — Scholar-powered news

Rachel Leah Childers

@donskerclass.bsky.social

2.3K followers 360 following 510 posts

Econometrics, Statistics, Computational Economics, etc University of Zurich http://donskerclass.github.io 🇺🇲 in 🇨🇭. 🏳️‍⚧️

donskerclass.github.io

Posts Media Videos Starter Packs

Pinned

Rachel Leah Childers @donskerclass.bsky.social · Jan 14

For the Spring semester, I am restarting my free weekly open office-hours for anyone in the world with Econometrics questions. Wednesdays 10-12AM Eastern or by appointment; sign up and drop by!

Details and sign up at donskerclass.github.io/OfficeHours....

7 22

Rachel Leah Childers @donskerclass.bsky.social · 9h

Our guarantees are asymptotic based on GMM theory, just like indirect inference. We examined finite sample results in simulation and found quite good performance in the low-non-simulated data regime, relative to existing methods. Some methods based on sample splitting perform badly in small samples.

Rachel Leah Childers @donskerclass.bsky.social · 11h

Ex post yes, you get valid inference in such a setting. Ex ante IDK. The number of parameter strengths which you would have to conjecture is something like m(m-1) moment covariances, which is a huge improvement over ∞ if specifying relations directly on the text. Maybe there's a pseudo-R^2 summary?

Rachel Leah Childers @donskerclass.bsky.social · 11h

Working on this project to understand how LLM simulation data could contribute to drawing valid scientific conclusions from real world data, led by @yewonbyun.bsky.social with Shantanu Gupta was a great learning experience, and I'm happy to see it forthcoming @neuripsconf.bsky.social 2025!

Valid Inference with Imperfect Synthetic Data

Predictions and generations from large language models are increasingly being explored as an aid in limited data regimes, such as in computational social science and human subjects research. While pri...

arxiv.org

Rachel Leah Childers @donskerclass.bsky.social · 11h

Valid inference w/ LLM-simulated data:

1. Take subsample of texts, extract variables
2. Construct moments identifying param on those variables
3. Ask LLM to simulate variables on sample & remaining texts
4. Use same moments w/ simulated variables
5. Combine moments, estimate jointly w/ 2-step GMM

Emily Byun @yewonbyun.bsky.social · 11h

💡Can we trust synthetic data for statistical inference?

We show that synthetic data (e.g., LLM simulations) can significantly improve the performance of inference tasks. The key intuition lies in the interactions between the moment residuals of synthetic data and those of real data

3 3 19

Rachel Leah Childers @donskerclass.bsky.social · 16h

Everybody wants to join me at Universität Zürich now, huh.
Herzlich willkommen in der Schweiz, Esther and Abhijit!

Department of Economics @Universität Zürich 🇨🇭 @econ.uzh.ch · 18h

We are thrilled to welcome Esther Duflo and Abhijit Banerjee as the new Lemann Foundation Professors to our department, beginning in the summer of 2026.

Rachel Leah Childers @donskerclass.bsky.social · 4d

Dynamic investment under rational expectations: Yale preparing press materials in advance for Sveriges Riksbank season.

yaleeconomics.bsky.social @yaleeconomics.bsky.social · 4d

To celebrate Professor Steve Berry and his lifetime of accomplishments, we interviewed many colleagues and students and wrote about the lasting impact of his work, in the words of those who know him best: tobin.yale.edu/news/250930/...

The Lasting Impact of Steve Berry’s Work — Through the Eyes of Colleagues and Students

tobin.yale.edu

1 1 4

Rachel Leah Childers @donskerclass.bsky.social · 4d

I’m still partial to heavy-tailed priors in many applications, which can approximate sparsity, but existing evidence and plausible models don’t support many exact 0s. doi.org/10.3982/ECTA...

Economic Predictions with Big Data: The Illusion of Sparsity - The Econometric Society

doi.org

1 2 5

Rachel Leah Childers @donskerclass.bsky.social · 6d

Disagree about people not proposing competing ideas to KL divergence. I was once flown out for a talk specifically so that Florian Gunsilius could spend a day evangelizing Wasserstein distances to me. I'm not yet a card-carrying member of the Optimal Transport cult, but compelling points were made.

1 2

Rachel Leah Childers @donskerclass.bsky.social · 6d

L^∞ is just sup norm, which is cursed because it may not be measurable or well-defined and can be a pain to bound.
Most of my PhD thesis was about extending results from Frobenius norm to operator norm (the L^∞ of Schatten norms). Frobenius makes perturbation very easy if you can show it's finite.

Rachel Leah Childers @donskerclass.bsky.social · 6d

I've been at this a long time....

Tweet from Rachel Leah Childers, @donskerclass

Favorite Banach spaces, ranked (partial order): W^k_2, Cameron-Martin space (any), L^2, L^∞, L^p (2<p<∞), W^k_p,..., L^1, L^p (1<p<2)

9:18 PM, August 27, 2015

Ranking does not necessarily reflect current views of the author. Apologies due to L^1 and their friends and loved ones....

Rachel Leah Childers @donskerclass.bsky.social · 6d

Drop your favorite distances. Lately I've been thinking about Fisher-Rao divergence (not a distance), Kernel Stein Discrepancy, L^∞ (cursed, evil), etc.

3 5

Rachel Leah Childers @donskerclass.bsky.social · 8d

For today's meeting of the UZH ML+Macro Finance reading group, I read about Forward-Backward Stochastic Differential Equations. Among other things, these allow for a stochastic version of Pontryagin's maximum principle for optimal control along a path.
Some notes I made explaining how they work:

donskerclass.github.io

2 1

Rachel Leah Childers @donskerclass.bsky.social · 9d

This post by Nan Jiang, nanjiang.cs.illinois.edu/2025/09/29/p... who knows more about RL theory than I ever will, on gaps in standard justifications for Policy Gradient, reassures me as I go through the PG lectures in interactive-learning-algos.github.io that its use is indeed a bit weird.

Office meme from Nan Jiang blog post
Frame 1: Two cards reading "policy gradient with 100 tweaks" and "policy gradient with another different 100 tweaks", respectively. "Corporate needs you to find the differences between this picture and this picture."
Frame 2: Pam Office: "They're the same picture"

Rachel Leah Childers @donskerclass.bsky.social · 12d

I remember feeling like I needed to hide it even though it would be years until I was self-aware enough to know why.

Pictured: my entire self-justification for watching every show or movie about trans people I could find as a teen:

marge simpson from the simpsons is holding a potato and saying i just think they 're neat

Alt: Pictured: Marge Simpson from the Simpsons is holding a potato and saying "I just think they're neat"

media.tenor.com

1 1

Rachel Leah Childers @donskerclass.bsky.social · 13d

Eh, ok fine.
Hi @gabbiegibson.bsky.social! There is a large crowd of women here who saw a clip from your documentary and seem to like you quite a lot.

2 4

Rachel Leah Childers @donskerclass.bsky.social · 13d

I found her @, and she seems to be doing fine, but it seems invasive to tag her even to let her know how many people think she's incredibly cool...

1 8

Rachel Leah Childers @donskerclass.bsky.social · 13d

It was pretty much literally just that. It was a documentary where they followed 4 random college students and had them talk about being trans. 2005 was a different world… en.m.wikipedia.org/wiki/TransGe...

TransGeneration - Wikipedia

en.m.wikipedia.org

2 1 22

Rachel Leah Childers @donskerclass.bsky.social · 14d

Core memory unlocked! I had this show DVRed as a teenager and watched it when my parents weren't home. I had totally forgotten about the dancing bit but can assure you that DDR was huge with the trans girls I knew around 2003-2004...

3 19

Rachel Leah Childers @donskerclass.bsky.social · 15d

Sounds pretty standard. The author list on that new book looks very promising! You didn't hear it from me, but these dynamic methods have been much more empirically successful for social-insurance style questions (your forté anyway) than for business cycles, so I'd focus more on those chapters.

Rachel Leah Childers @donskerclass.bsky.social · 15d

Nice! What textbook/materials are you using? I’ve found Stachurski johnstachurski.net/edtc to be much cleaner and simpler than the Stokey-Lucas/Ljungqvist-Sargent tomes I used around that era, but there are a lot of good intros nowadays. Is there a particular macro topic you’re targeting?

1 3

Rachel Leah Childers @donskerclass.bsky.social · 17d

On Halley, I'm reminded of Cosma's old post on whether Copernicus should have used AIC or the LASSO bactra.org/weblog/922.h...

Ockham Workshop, Day 2

bactra.org

1 2

Rachel Leah Childers @donskerclass.bsky.social · 17d

Agreed, though the word "parsimonious" is just as slippery as "overfitting". I like to operationalize it using entropy integrals, but for some reason this makes philosophers of science's eyelids twitch.

Rachel Leah Childers @donskerclass.bsky.social · 17d

In the context of probability theory, these are distinct based on whether we maintain the assumption of a common distribution; Ben, not a fan of probability models, clearly does not like that distinction www.argmin.net/p/thou-shalt...
I'm ok with it but teach them as views of the same phenomenon:

Sample Selection and External Validity

donskerclass.github.io

2 3

Rachel Leah Childers @donskerclass.bsky.social · 17d

There's nothing wrong with benchmarks, but taking them as the only paradigm is a bit limiting.
How temporally unstable are the laws of physics? Basic chemistry and biology? Those are models too. Maybe they'll stop working tomorrow too: it is time t+1, after all.

emmie malone 🏳️‍⚧️ @emmiemalone.bsky.social · 18d

I’ve been training my whole life for this moment

A screenshot of an article published in popular science. It reads: “First known wild 'grue jay' hybrid spotted in Texas. Green and blue jays are crossing paths as temperatures rise.”

1 3

Rachel Leah Childers @donskerclass.bsky.social · 17d

Dataset shift is just external validity, but underspecification refers specifically to multiple indistinguishable predictors. For that to matter, you need some setting that isn't your train-test set to distinguish, which is an external validity problem, but defined in terms of a model class.

1 2