Greg Stoddard

Hello friends! U Chicago Crime & Ed lab is looking to hire data scientists. If you want to work on data + public policy, we'd love to hear from you. Find out more at this link urbanlabs.uchicago.edu/attachments/ab… Plz DM me with questions and forward to anyone who might be interested.

Lots of people get worked up about citation styles but I simply don't care et al.

@arthur_spirling @MorucciMarco Love the paper! Any thoughts on the nature of prediction in ML (is this a cat or a dog) versus social science (will a conflict breakout in this region)? The former is almost perfectly predictable while the latter is an inherent forecast. Does that relate to instrinsic dimension?

Supervised machine learning was going to revolutionize prediction/inference for politics. What happened? New WP w @MorucciMarco uses "intrinsic dimension" to show it's v hard to beat simple (OLS, logit) models for "our" polisci tabular data. Ppr: arthurspirling.org/documents/Moru… (1/n)

@ChristophMolnar Well it's only a random forest if it comes from the ensemble region of France. Otherwise it's just sparkling bagged trees.

When you train a forest with sklearn's RandomForestRegressor, you don't get a random forest. You get bagged trees. At least when with the default parameter of max_features=1.0. Does anyone know why it was implemented this way?

@rajistics Thanks for making the video and for the kind words about our paper!

Data Scientist versus Data Analyst: Predicting Police Misconduct Good reminder of fundamentals in data science. I really like this paper because it includes many good data science practices (more than I could squeeze into the video). 🔍 We should always consider simpler rules-based approaches to using ML 📊 How much data is your ML Model using? 📈 How many features? (Do you need 800 features??) 🌍 Can you translate your ML performance into real-world impacts Great example for those of you teaching data science Thanks to @BeckerFriedman @HarrisPolicy @UChiUrbanLabs @gregstod

@LarsvanderLaan3 Super interesting work. I'm excited to try it in my applied work. Is there an example/guidance on how to use the package on RCT data with known treatment propensities?

4/n Experiments show find EP-Learner outperforms state-of-the-art including R-learner, causal forests, DR-learner, and T-learner! The gains are especially big for random forests and gradient boosted trees! We implement these learners in our R package github.com/Larsvanderlaan…

Introducing Efficient Plug-in Learning for HTEs! EP-learner solves two issues with orthogonal learning: 1.Extreme pseudo-outcomes lead to bound-violating predictions 2.Loss functions can be nonconvex We establish oracle bounds! (Not quasi) arxiv.org/pdf/2402.01972…

@StatsbyLopez @Stat_Ron @benbbaldwin @SeanfromSeabeck Really interesting! Where’s the chart from?

@benbbaldwin @SeanfromSeabeck Even easier to evaluate college talent with film. Why the NFL draft is always so efficient

@Jess_Hoel “For whom the nurse calls: Perceptions of missing work to care for sick kids”

Hey #EconTwitter, help me find a cute title for my new paper! It's about what parents *think* will happen when they miss work because a kid is sick. Abstract below. Sicko Kiddo: ???

@tommangan @Chris_arnade That’s literally the end of the book.

@Chris_arnade I felt like Moneyball had a great big glaring hole in the middle of it that he never talked about. EG: the tech is easily copied, and as soon as the big-city teams do it, the advantage to the A's evaporates. Since then I've respected his ability less and less.

As a Salomon Brothers bond trading alum, it warms my heart to see Michael Lewis finally living up to his full potential as the midwit salesman that couldn’t cut it as a trader & ends up ruined & fooled by a fast sweet talking customer

@peterbergman_ I'll take one please!

@PhDemetri There are a lot but I've found some of these papers to be helpful in thinking about the estimation of HTEs link.springer.com/article/10.100… openreview.net/pdf?id=FQLzQqG… and this library econml.azurewebsites.net

Looking for resources on heterogenous treatment effects. I found this and really like it stat.duke.edu/~fl35/teaching…

@ethayarajh Not clear this is true because of rich-get-richer dynamics. This famous paper (among others) shows that an early positive vote can lead to dramatically more upvotes later on. I'd say P(A>B) > .5 but certainly not 1. pubmed.ncbi.nlm.nih.gov/23929980/

They were inferred from the simple observation that if comment A was written after B but has a higher score despite getting less visibility, then ostensibly A > B. If A was written before B, then we can't conclude this -- the higher score could have come from more visibility!

📢 Models like #ChatGPT are trained on tons of human feedback. But collecting this costs $$$! That's why we're releasing the Stanford Human Preferences Dataset (🚢SHP), a collection of 385K *naturally occurring* *collective* human preferences over text. huggingface.co/datasets/stanf…

@AmandaCoston @anna_kawakami @HodaHeidari @d19fe8 @Haiyi_Zhu @satml_conf This is a great & thoughtful paper, I really enjoyed reading it. Congrats!

Thrilled that our work on evaluating the justifiability of decision-making algorithms got a best paper award at SATML and that @anna_kawakami presented the research! @HodaHeidari @d19fe8 @Haiyi_Zhu @satml_conf

This research is so important and ambitious. I’m proud to be colleagues with this amazing team.

Here's the story of a city's response to spiking gun deaths. Of READI Chicago—an ambitious effort to build & study a program of jobs & CBT for the men most likely to shoot or be shot. Of HUGE success by some measures (64% fewer shooting & homicide arrests!) & no impact by others.

Forthcoming in the AER: "Not Too Late: Improving Academic Outcomes among Adolescents" by Guryan, Ludwig, Bhatt, Cook, Davis, Dodge, Farkas, Fryer Jr., Mayer, Pollack, Steinberg, and Stoddard. aeaweb.org/articles?id=10…

@dggoldst But did it think fast or think slow?

Please note that there is a huge debate whether the first version is an error given conversational norms. The second one is more clearly an error and reflects current difficulties in getting these things to do mathematical reasoning.

GPT and the Linda Problem.

(please retweet) I am taking CS PhD students! If you are interested in ML, data science, health, or inequality, please apply to Cornell :) We have an island campus in New York City and a fantastic group of students I feel lucky to collaborate with every day. Join us!

@alexpghayes @george_berry Genuinely curious: does over or under sampling actually help for ranking? I have yet to find a case when it did, nor have I found an applied paper that showed it actually helped.