Mattan S. Ben-Shachar

1k followers! Calls for a re-introduction: Statistics lecturer, freelance stats consultant, & #rstats dev @easystats.bsky.social 📊 I try to help social scientists make better inferences from their data & communicate their findings 👨‍🏫 Hope to bring the #stats twitter/R community vibes over to bsky!

New post! Debunking the Google Gemini/Cloud commercial on the full moon and baseball stats. marcoshuerta.com/posts/full-m...

Okay, going to vote. See you on the other side. 🤞

Why it's important to consult a statistician before data collection, example 17445: If you collect data from 1000 patients, nested in 20 therapists, your effective sample size is much (much!) smaller than you think it is... 😏

People, if you wanna p-hack the shit out of your data, there's no need to pay a statistician - a subscription to the most capable LLM is much cheaper, without any of the pushback. Guess I'm firing a client today 🤷‍♂️

Your monthly reminder that if you're using ChatGPT to do your science, you're failing your colleagues, community, and all of us. Here, ChatGPT makes up data and presents it as the real thing — then apologizes when confronted.

Three different ways of presenting interactions in this little thread. I think "(non)parallel lines" is the most common, but I like the heatmap, especially when there isn't a focal predictor and a moderator and we're really looking to visualize a joint effect.

You know how the "No brown M&M's" clause in Van Halen's contracts works? (summary: https://en.wikipedia.org/wiki/Van_Halen_test) Well that's what demanding a simulation-based power analysis does for a complicated, e.g. multilevel model: forces you to understand that model properly before showtime.

I actually don’t think you can be a substantive expert without a certain amount of methodological competence. Otherwise you don’t really have a sense of how to interpret your data or the studies of others. It’s just vibes.

Very important for science that you plug a couple of numbers that you made up into the wrong fields in G*Power to get a number that does not change anything, because you’re going to tinker around until it’s the number you want 😌

Which famous movie, book, poem, etc, should Christopher Nolan adapt with no subtext? Entertaining answers only!

This was a great episode - both very introductory and nuanced.

Had this idea in lockdown, half-built it, forgot about it, remembered it again, though might as well finish it.

#statsmeme so I can find it later!

#dataviz #stats #rstats New blog post: Thinking about Penguin Variables Bill length & depth show Simpson's paradox. Re-expressed as size (L×D) and shape (L/D), it vanishes: same data, rotated 45°, a simpler story. friendly.github.io/blog/posts/2...

Just published open access in 𝘗𝘴𝘺𝘤𝘩𝘰𝘭𝘰𝘨𝘪𝘤𝘢𝘭 𝘔𝘦𝘵𝘩𝘰𝘥𝘴: “𝗪𝗼𝗿𝗸𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗖𝗶𝗿𝗰𝘂𝗹𝗮𝗿 𝗗𝗮𝘁𝗮: 𝗔 𝗧𝘂𝘁𝗼𝗿𝗶𝗮𝗹 𝗳𝗼𝗿 𝗖𝗼𝗴𝗻𝗶𝘁𝗶𝘃𝗲 𝗮𝗻𝗱 𝗕𝗲𝗵𝗮𝘃𝗶𝗼𝗿𝗮𝗹 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵” 🔗 doi.org/10.1037/met0... 🙎 Co-authors: @bayslab.org, @ivntmc.bsky.social, @cjungerius.bsky.social, Xiaolu Wang, Julie de Falco, Zahara Gironés, Adam Triabhall

🤠I'm excited to share our new paper, just published in PNAS, in which we examined more than ONE MILLION forecasts from Polymarket, Kalshi etc, The main finding is striking: Prediction markets and collective forecasting platforms are remarkably well calibrated.

🤫 {shiny} isn't the only "real" way to make an app in #rstats I've made the same website using {shiny} and {plumber2} for a made-up coffee company and deployed both on @ricochet.rs ricochet.rs/blog/ssr-in-r/ #databs

Why do I want access to all the books & articles? For training, testing & inference. But don't worry, it's 'transformative use'; what comes out will be very different from what I put in. It'll even grow the market for what I'm asking for. No, no, this isn't about LLMs. I'm talking about my students.

What about PhD students who confuse correlation and causation

Do you want to faithfully quantify false/true positive rates in a significance test? Model the entire data generating process. Faithfully calibrate an estimator? Model the entire data generating process. Design productive experiments? Model the proposed data generating process.

"If you don't open a book or talk to a statistician, you can experience the joy of discovery every day"

It's just one of those days...

If you haven’t watched this one yet, this video by the European Commission (!) is an absolute gem and worth the minute.

This is so bad. Why does rendering take 1min but running interactively in #Positron takes upward of 20 minutes??

Same five studies. Same treatment effect. One meta-analysis (raw): I² = 0%. Another meta-analysis (SMD): I² = 94%. How can both be right? 🤔 Find out in my latest blog post👇 mzloteanu.substack.com/p/standard-n... #Statistics #MetaAnalysis

This is what misalignment looks like

Is anyone else experiencing extreme variability in @quarto.org interactive run times? Like running `1` can take (within the same session) 0.001s to 1.2 MINUTES. Running in Positron, everything is up to date, and this isn't happening running in the same window just regular R...

Thanks! This inches us toward `glmmTMB` being a replacement for `gamm4` - which, with respect to its authors, I've found a headache to work with (huge model objects, lack of downstream methods, etc.). BUT `glmmTMB` still has issues with predictions, starting values, etc. github.com/glmmTMB/glmm...

🤯

"For observational designs, please describe association, prediction, or covariation in terms of statistical dependence. For example: 'we found that p(smoking, lung cancer) ≠ p(smoking)p(lung cancer)'".

Interviewer: Can you explain this gap in your CV? Statistician: Yeah, I was Missing At Random

How should we track human decisions in the kind of data science project that Alex envisions, and how can we ensure LLMs implement our decisions faithfully? One idea would be "Data Analysis Decision Records" What do you think? I'd love any feedback on #DADR ! vincentarelbundock.github.io/dadr

#rstats folks, what are some other must-use options()?

What does a sound data analysis workflow look like? In a first step: - Define your estimand (*quantity and unit of interest*) & research question - Map reality to your model (choose variables wisely) How the {modelbased}📦fits into this post-estimation workflow (second step): 👇 #rstats #easystats

Send your best work to PCI Psychology (psych.peercommunityin.org) where it will receive high quality & open peer review free of commercial interference. We review a variety of article types from a wide range of psychological topics (psych.peercommunityin.org/help/help_ge...).

One thing I find endlessly funny about the rise of AI is that it has revealed that there is a large number of people who believe that the reason they wrote essays at school or university is their teachers were desperately waiting for them to provide a definite answer to “Lear: what’s up with him?”

Psychologists: you should remove that outlier!

Remember folks, human peer review guarantees that at least 3-4 humans will read your paper!

no value judgment, but this got a laugh from me github.com/openai/math/...

Outside of the obvious open source problems with vibe coding, there is soooo much truth in the “now I have to read/review all of this stuff with my brain power when you used almost none at all” across all layers of work now. Feel this in my bones 😭