Tessera

Pepijn Vink

@pepijnvink.bsky.social

PhD Candidate Methods & Statistics @ Utrecht University | he/him | intensive longitudinal data, Hidden Markov Models, & Bayes

18 Posts44 Follower 75 Folgt

Pepijn Vink @pepijnvink.bsky.social · 6d ago

Apparently my university now offers a summer school course on family constellations and I have some thoughts about this...

Rafe Meager (they/them)@economeager.bsky.social · last wk.

we have got to put that into a causal inference lecture

Museum of Twitter@museum-of-twitter.bsky.social · last wk.

@RuthePhoenix

Took the batteries out of the carbon monoxide alarm because the loud beeping was giving me a headache and making me feel sick and dizzy.

tj mahr 🤘@tjmahr.com · last wk.

interesting how “correlation is not causation” shifted from a warning about spurious associations to something incurious people say to refute things

Pepijn Vink @pepijnvink.bsky.social · 2w ago

So apparently I’m very good at guessing predictor importance. Thanks @eikofried.bsky.social for the great prize! #SAA2026

Pepijn Vink @pepijnvink.bsky.social · 2w ago

SAA 2026 kicks of at a gorgeous location! @saa2026.bsky.social

Pepijn Vink @pepijnvink.bsky.social · last mo.

💡 New preprint! Cross-lagged panel models only work if stable traits (genes, upbringing) aren't secretly confounding your results. Can including a latent variable save the day? We tested 6 SEM models to find out when it does — and when it doesn't. Link: osf.io/preprints/psyarxiv/sdjry_v2

Edo Navot @edonavot.bsky.social · last mo.

When the panel model sits cross-legged, can we still identify the latent class of men who prefer butts?

Pepijn Vink @pepijnvink.bsky.social · last mo.

Eiko Fried @eikofried.bsky.social · last mo.

Join our Many-Analysts Simulation Study—yes, you read that correctly—accepted as Stage 1 Registered Report in (AMPPS) @psychscience.bsky.social. We're trying to find out how researchers' simulation design choices shape methodological conclusions. Further info: osf.io/8vcxh/files/...

Pepijn Vink @pepijnvink.bsky.social · last mo.

Bayesian Workflow arriving the week of a deadline and the week before a conference could not be worse timing

Pepijn Vink @pepijnvink.bsky.social · last mo.

As a man who also uses mixture models, I am very sorry for this abomination of a model.

Julia M. Rohrer@dingdingpeng.the100.ci · last mo.

Saved you a click

samuel mehr @mehr.nz · last mo.

inspired by @russpoldrack.org's AI policy for his lab, last night I drafted up this AI policy and came up with an initial set of use case suggestions for my lab

General principles
Here are a few key principles that govern our use of AI in scientific work.
AI tools are not necessarily useful. Despite what you might have heard in OpenAI's advertising, using AI tools is the wrong choice for many tasks. Alternative and often more old-fashioned tools will do a better job for quite a few specific things (see this instagram feed for a variety of examples); one should never assume that an AI tool is the default "best thing to try". Moreover, because verifying AI results can be time-consuming, there are plenty of cases where you'll spend less time learning how to do X (i.e., by working on it yourself) rather than having Claude do X and then spending weeks to painstakinly confirm that it worked.
AI usage does not absolve you of your responsibility for your own science. We are responsible for the science that we produce, both from an ethical standpoint and a practical one. Any errors in your work are your responsibility, any security risks caused by AI usage (see below) are your responsiblity, and so on. Just as you must debug the outputs of your analysis code, the outputs of any/all AI usage must be carefully verified.
AI usage requires disclosure. The decision to use (or not use) AI tools for any aspect of the scientific process is one that requires discussion with all leaders of a project. This is both for scientific reasons (i.e., we ourselves must know what our project invovles) and practical ones: many scientific societies and jouranls require the disclosure of AI usage.
AI tools may present security risks. Many commercially available AI tools, especially free ones, do not provide substantive protections regarding inputs provided by users. Put another way, you don't know what will happen to the text, data, images, or anything else you show a third-party AI tool. This may present substantial security risks.
AI is not for scientific writing. Papers are the primary end product of our science. Scientific writing distills our thinkin…

How to handle AI usage
Like many things in our lab, it's best to be vocal. You should always feel free to raise the possibility of AI usage on a project. For any potential use case, chat with your labmates, your collaborators, and Sam about it. Discussion with external-to-lab collaborators can be particularly important given the issue of disclosure, mentioned above; you may not be aware of what AI tools collaborators are using, unless you've asked.
You should also be skeptical of the utility of AI for a specific task (not least because of the billions of dollars being thrown at selling AI subscriptions to universities and other institutions). One thing to consider might be whether the use case is something that would actually be useful for you to master yourself. For example, a common web element that turns up in tons of online experiments would probably be worth your figuring out how to code up from scratch, since you'll use it often in online experiments and figuring it out may help you come up with new ideas for how to design a game. In contrast, an esoteric web element that you'll use for one experiment and never again might be a better fit to farm out to Claude.
With the above General Principles in mind, we do not have any blanket policy about where AI should and shouldn't be used in research, with the clear exception of drafting text for manuscripts, grant proposals, posters, reports, or any other text-based scientific product: AI tools should never be used for writing, i.e., no copying and pasting of text directly from an LLM is appropriate, without explicit discussion and agreement prior to generating the text.
Other use cases are worth considering, however, and should be considered freely. Some good and bad examples follow, based on the principles above (we should populate more examples as they come up).

Some potential examples of AI usage
AI as a cheap web developer. If you're coding up a gamified experiment it can be quite difficult to learn a whole bunch of web development from scratch. Claude will be better at this than you, and that's fine: we are not a web development lab, we're a cognitive science one, so it's not essential that you become a super-proficient dev. Vibe-coding web stuff is a pretty decent use case. (But you must very carefully ensure that whatever AI tools code up for you behave exactly as you expect them to, e.g., if you've vibe-coded a specific form of randomized experiment, you'll need to check that the randomization is being carried out appropriately).
AI as a code reviewer. You might try using an AI tool to figure out why your CSS isn't displaying properly or to help debug code in general. (But you must never provide participant data to a third-party LLM, as this would be a data security risk; and you must always double- and triple-check the tool's output, just as you would with a human code reviewer or if you were trying out tips from Stack Exchange).
AI as a search engine. You might try using an AI tool to discover types of models or analysis approaches that you hadn't heard of before, or to turn up obscure research papers. (But you'll need to verify the output carefully with trusted sources, to ensure accuracy; and you should never, ever treat AI-generated search results as an objective and accurate summary, as they are far from unbiased and are often inaccurate, especially when it comes to scientific questions — try asking Gemini whether music makes you smarter).
AI as a sketch pad. You might try using an AI tool to generate ideas for what to put in a scientific figure, or come up with creative different layouts, which you then use as inspiration for figures you create yourself. (But you should not provide data to an LLM and ask for figures to be produced for you, since this often leads to laughably bad outcomes; it also would not be r…

Some bad examples of AI usage
AI as a data analyst. You should not provide a dataset to an AI tool (this is a security risk) and you should not ask an AI tool to write analysis code for you from scratch. Analyses done without statistical understanding are worse than no analyses at all. While it's OK to use an AI tool to help debug your analysis code, your development as a statistical analyst is part of your scientific development, so it's a bad idea to farm this out.
AI as a literature reviewer. You should not use AI to generate reference lists or summarize whole fields of literature for you. AI tools notoriously cite papers that do not exist, and they recapitulate well-known biases in science (e.g., citing more male authors than female ones).
AI as an illustrator. We are fundamentally opposed to the practice of AI models appropriating human artists' work to produce generative imagery and so we do not support its use for scientific work. The reputational risks associated with embarrassing errors in generative imagery are also substantial. We are lucky to have grants available to us that can be used to pay illustrators and graphic designers for work, and we regularly hire folks like this. Talk to Sam if you have a design need on a project that you'd like us to fund; it's a much better call than using AI.

Mattan S. Ben-Shachar @mattansb.msbstats.info · 3mo ago

I can't be the only one who hears this, right? #statsmeme

Daniel Lakens @lakens.bsky.social · 3mo ago

New blog post: Evaluating Dr. Cuddy’s Claim that the Debunking of Power Posing is a Myth. daniellakens.blogspot.com/2026/05/eval... On an AI generated description of a non-existent study, incorrectly citing findings from studies, and the importance of scientific criticism.

Evaluating Dr. Cuddy’s Claim that the Debunking of Power Posing is a Myth

In this blog post I will analyse the arguments that Dr. Amy Cuddy provided in a blog post “The "Power Posing Was Debunked" Myth: What the Re...

daniellakens.blogspot.com

Leonid Schneider @forbetterscience.bsky.social · 5mo ago

Oh nothing, just a peer reviewed paper my colleagues found... doi.org/10.1016/j.ma... "1 mL of the mass killing of an ethnic group was opposed to 20 mL of the skin sample and unprotected to light for 7 min." Even AI knows what's wrong here, but @elsevierconnect.bsky.social doesn't.

Pepijn Vink @pepijnvink.bsky.social · 6mo ago

Me every time I get a .name_repair error

Julia M. Rohrer@dingdingpeng.the100.ci · 6mo ago

I just hate tibbles so much. (don't @ me)

Pepijn Vink @pepijnvink.bsky.social · 6mo ago

I was very confused when I read this until I realised that this was, in fact, not about statistics

Jason Cohen@asmartbear.com · 6mo ago

Bootstrapping needs high margins and quick payback periods so you can grow without raising money. (2/2)

madeline odent @oldenoughtosay.com · 6mo ago

The drunk uncle theory. You don’t argue with the casually homophobic uncle at Thanksgiving dinner to change his mind; you argue so that the closeted cousin at the kids table knows there’s safe people and better possibilities out there

Marie Le Conte@youngvulgarian.marieleconte.com · 6mo ago

agree with this (hah) but also think a particular mistake the left made for a long time online, and still makes to an extent, is failing to understand that the person whose mind you may actually change is the one reading the argument you're having, not the one you're arguing with

Julia M. Rohrer @dingdingpeng.the100.ci · 7mo ago

they used what now to measure testosterone

Ruben C. Arslan@ruben.the100.ci · 7mo ago

New meta finds no correlation between testosterone and risk preference, buuuuut testosterone was mainly measured using 2D:4D (nonsense) or saliva (still bad) and we know from Frey et al. 2017 that the risk measures don't tap into a coherent preference. www.sciencedirect.com/science/arti...

violin plot of results by testosterone measure

Pepijn Vink @pepijnvink.bsky.social · 7mo ago

Title case is based purely on vibes

Julia M. Rohrer@dingdingpeng.the100.ci · 7mo ago

As a non-native speaker who could never figure out how title case works (I *just* learned the last word is always capitalized? wtf?) I don’t have my objections.

Dr. Casey Fiesler @cfiesler.bsky.social · 8mo ago

The ACM Digital Library, where a LOT of computing-related research is published (I'd say at least 75% of my own publications), is now not only providing (without consent of the authors and without opt-in by readers) AI-generated summaries of papers, but they appear as the *default* over abstracts.

Screenshot of a paper entry:
Fictional Failures and Real-World Lessons: Ethical Speculation Through Design Fiction on Emotional Support Conversational AI
Authors: Faye Kollig, Jessica Pater, Fayika Farhat Nova, Casey Fiesler
(There are tabs with "abstract" and "summary" and "summary" is selected.)

Jan Vanhove @janhove.bsky.social · 8mo ago

A recent study purports to have found that multilingualism protects against accelerated ageing. I've taken a closer look at it, and it doesn't look good. New blog post: "Does multilingualism really protect against accelerated ageing? Some critical comments" janhove.github.io/posts/2025-1...

Positive trend between monolingualism and accelerated ageing.

Negative trend between per capita GDP and accelerated ageing.

Pepijn Vink @pepijnvink.bsky.social · 9mo ago

Donald Rubin being in the Epstein files wasn't on my bingo card for this week

Julia M. Rohrer @dingdingpeng.the100.ci · 9mo ago

"For instance, randomized controlled trials could explicitly manipulate multilingualism"

a woman wearing a blue and white floral dress stands in front of a sign that says natural beauty

Alt: Jennifer Lawrence nodding and doing a thumbs up

media.tenor.com

Casper Albers 🟥@casperalbers.nl · 9mo ago

Doing non-causal inference (and being explicit about it), yet using a causal word as second word in the title. If you pay Nature € 10.690, they will publish this in Nature Ageing. I can tell you what I think of that for free. www.nature.com/articles/s43...

screenshot from the paper, stating that no causal claims (like they did in the title) should be made.

Hadley Wickham @hadley.nz · 10mo ago

Do you teach #rstats? Do your students complain about how lame and old-fashioned dplyr is? Don't worry: I have the solution for you: github.com/hadley/genzp.... genzplyr is dplyr, but bussin fr fr no cap.

GitHub - hadley/genzplyr: dplyr but make it bussin fr fr no cap

dplyr but make it bussin fr fr no cap. Contribute to hadley/genzplyr development by creating an account on GitHub.

github.com

apache/@apache.be · 10mo ago

Belgian AI scientists are advocating *against* the use of AI in academia. “If independent thinking is no longer encouraged at university, where would it?” apache.be/2025/10/24/b...

Belgian AI scientists resist the use of AI in academia

Several AI scientists have published an open letter calling for a ban on AI use by students.

apache.be

Maarten van Smeden @maartenvsmeden.bsky.social · last yr.

Depending which methods guru you ask every analytical task is “essentially” a missing data problem, a causal inference problem, a Bayesian problem, a regression problem or a machine learning problem