Athey and Imbens have a big new paper out on nonparametric factor models [I was lucky enough to receive an advanced copy], which seeks to formulate an encompassing structure for causal inference with panel data that is comparable to the nonparametric (unconfoundedness, overlap) setup for cross-sectional data. As an occasional participant in this literature, I’ve found the over-reliance on functional-form assumptions (e.g. two-way FEs AKA parallel trends, or linear factor structure) frustrating, so attempts to avoid it are welcome. The estimators that fall out of this framework are not entirely obvious yet, but they take the form of nearest-neightbours-type estimators [pg 12]
Agarwal et al propose a similar linear antecedent to the above paper, and end up having a similar nn-type approach
Deaner et al propose a similar nonparametric, nonlinear factor structure and the implied estimators are, again, based on nearest-neighbours on a latent-factor scale.
Kolesar and Plagborg-Moller and Olea et al clarified LP/VARs a lot for me. Macroeconomists have a hard job and typically operate in a much more parametric structure than one would be comfortable with in microeconometrics. This lit helps outsiders like me understand the bias-variance tradeoff between LP and VARs as (loosely) comparable to IV v OLS.
Carrell et al’s Low-rank Thinning. The thinning approach to data compression is new to me [learned about it from another Lester Mackey paper] and this seems like a promising extension motivated by the old adage of “inside every big matrix, there is a smaller matrix, and your job is to find it”. Could be interesting to approach standard m-estimation with this approach, which would presumably entail constructing a surrogate loss using the thinned points.
quartz is a very promising looking static-site generator that renders markdown content. I liked the look of it enough to migrate my linklog from my previous barebones pandoc-powered version. I’ve been an indisciplined obsidian user over the last few years, and hope this migration serves as a commitment device to use its very cool linking / graph functions. For example, 2025-03-01 is the first linklog, and this link should be reflected in the graph that quartz generates.
pagedout 006 is a charmingly old-fashioned computer magazine. Single-page articles that run the gamut from esoteric cryptography to fun stuff you didn’t know about python.
Better Shell History Search : found on hackernews; I use ctrl-r on the terminal a lot and this improves it massively with the help of fzf.
gbnet gradient boosting implemented in pytorch. Unlikely to justify switching away from xgboost or lightgbm , but plugging into torch’s optimization ecosystem is nice either way.
History of animation through steamed hams the internet losing its mind and (the sparse remains of its sense of good) taste over mimicking Studio Ghibli with the new openai model’s (admittedly extremely impressive) image generation capabilities reminded me of this cool project where the steamed hams scene was animated in many different art styles.
Sheon Han’s article profile of Paul Ginsparg is great. arXiv is one of the best things about the internet (most of the content on this linklog catalogues it), and its history turns out to be quite interesting too.
shameless self-promotion
pyensmallen now has an EnsmallenEstimator class that implements GMM/M-estimation (wherein solves for parameters implicitly defined by g(Z;θ)=0 by root-finding or minimizing a quadratic form g′Wg). This uses jax for autodiff to calculate gradients automagically and is generally quite fast and functional.