Papers I Actually Reread

Marcus Ellison

@latecheckoutmark

5Items

0Likes

Most academic papers are like bad first dates: too long, weirdly proud of themselves, and you’re relieved when it’s over. These eight are the opposite. They’ve all wrecked my brain in a good way, either by giving me a new stat toy, changing how I think about causality, or just making me stare at a figure like it’s the last shot of a playoff game. If you’re a data person who likes your theory with a side of real-world stakes, this is the little reading rotation I keep coming back to when I should probably be sleeping.

See all 5 items Follow Marcus EllisonOpen in app

The list

Showing 1–5 of 5

Papers I Actually Reread — 5 items
Name	Authors	Year	Field	One-line gist	Why I care	Best stat or figure	Difficulty	Where to read	My take
1The Probable Error of a Mean	William S. Gosset (as “Student”)	1908	Statistics / Inference	Derives the t-distribution so you can do inference with tiny samples instead of needing a mountain of data.	Any time I’m A/B testing with laughably small n, this is the ghost in the machine bailing me out.	The whole idea that you can get sane uncertainty estimates from like a dozen observations still feels like cheating.	Medium if you’ve survived one real stats class; otherwise bring coffee.	Biometrika archives, most stats textbooks, or free scans online.	It’s the OG “do more with less data” paper—basically Moneyball for sample sizes.
2Regression Discontinuity Designs in Economics	David S. Lee and Thomas Lemieux	2010	Econometrics / Causal inference	A tour of regression discontinuity designs and how to treat arbitrary cutoffs like nature’s randomized trials.	If you work with messy policy data, this is the closest you get to sorcery without p-hacking.	Those jump-at-the-cutoff plots where one side of the line looks like it’s living in a different universe.	Medium-hard; the intuition lands fast, the assumptions take a few re-reads.	Journal of Economic Literature; also floating around as a PDF from various universities.	This is the paper that made me see every score cutoff—test scores, credit scores, you name it—as a potential natural experiment.
3The Use of Multiple Measurements in Taxonomic Problems	R. A. Fisher	1936	Statistics / Pattern recognition	Introduces linear discriminant analysis using iris flowers, basically the grandparent of half the classifiers we still use.	It’s a reminder that a simple linear boundary, done right, can hang with the fancy deep models on the right problem.	The separation of iris species in low-dimensional space—still wild how clean it looks for a real dataset.	Medium; algebra-heavy but conceptually pretty clean.	Annals of Eugenics archives or any ML history reader; PDFs are everywhere.	Feels like watching black-and-white game tape of a legend and realizing the fundamentals still work in today’s league.
4Random Forests	Leo Breiman	2001	Machine learning / Ensemble methods	Shows how averaging lots of noisy decision trees with randomness baked in gives you a shockingly strong predictor.	Any time I need a baseline model that punches above its weight, this is still the first jersey off the bench.	The error vs. number-of-trees plots that flatten out like a good defensive rotation—diminishing returns, but steady.	Medium; readable for practitioners, the theory parts get spicy but skimmable.	Machine Learning journal or the author’s reprints page online.	It’s the paper behind half the Kaggle gold medals and more than a few production systems nobody brags about but everyone trusts.
5Object-Level Representation of Images	Jitendra Malik, Serge Belongie, Thomas Leung, Jianbo Shi	1999	Computer vision	Pushes beyond raw pixels to represent images in terms of objects and segments, laying groundwork for modern vision models.	It’s a reminder that all the fancy deep nets are still chasing the core idea: see scenes as objects, not noise.	The segmentation examples where messy real-world scenes suddenly break into clean, meaningful regions.	Medium-hard; you’ll need some comfort with both stats and vision basics.	Papers from the IEEE or CVPR proceedings; PDFs are widely mirrored.	Reading it now feels like seeing the early sketches for the stuff your phone casually does every time you open the camera.