Nathaniel Haines Profile picture
Paid to do p(b | a) p(a) p(a | b) = ——————— p(b) Manager, Data Science Research @LedgerInvesting
Dec 27, 2022 17 tweets 6 min read
1/n A thread demonstrating Stein's Paradox! Or, of how pooling of individual-level estimates toward groups means leads to estimates with lower mean squared error relative to approaches that disregard group-level information.

math.drexel.edu/~tolya/EfronMo… 2/n We are often in situations where we need to estimate some unobservable (latent) mean with only observed data. In such cases, using the sample mean as an estimate is standard practice. But in 1955, Charles Stein discovered that we could do better!

projecteuclid.org/ebooks/berkele…
Jan 29, 2022 7 tweets 4 min read
1/n It is ready! In this post, I walk through how to use renv, Docker, and GitHub Actions to automate a workflow that makes your R projects reproducible across time and space 🤖🤓.

The goal is to help you avoid the age-old problem of *dependency hell*:

haines-lab.com/post/2022-01-2… Image 2/n I first discuss renv, a super useful package management tool for R that allows you to install project-specific R packages. This means you don't have to worry updating a package for one project in a way that breaks another project—or, as I call it, computational whac-a-mole 🤖 Image
Aug 29, 2021 14 tweets 8 min read
1/n I felt like putting together a meta-thread on threads I have made in the past discussing statistical concepts including generative modeling, measurement error, penalized regression, and more. Mostly because this will make them easier for me to search for in the future 🤓 2/n First, a blog series on reinforcement learning in the context of laboratory behavioral tasks. The thread below describes the series, which covers basic principles and parameter estimation (including doing optimization, MAP estimation, & MCMC in R): Image
Aug 24, 2020 20 tweets 8 min read
1/n This project is finally updated! We argue that generative modeling can improve the quality of inference on behavioral data, and we use simulations and empirical data from the Stroop, Flanker, Posner Cueing, Delay Discounting Task, and IAT to show it.

2/n Recently, many papers have been published showing that traditional analyses lead to poor reliability for behavioral measures. We argue that there are both theoretical and statistical issues with traditional methods (e.g., mean contrasts) that are improved by generative models
Sep 6, 2019 7 tweets 3 min read
1/7 In 2017, Hedge, Powell, and Sumner showed that robust cognitive tasks are unreliable, which calls into question the use of behavioral tasks for studying individual differences. In this blog post, I show that this conclusion is misguided (haines-lab.com/post/thinking-…) 2/7 Specifically, Hedge et al. found that robust effects such as the Stroop effect have test-retest correlations in the range of .5 to .6 (r = .5 in the plot of their Stroop effects shown here), which severely impacts our ability to rank the performance of individual subjects.