Paul Bürkner

the logging is done in rstan. so a fix if needed will have to be there I assume.

there is not unfortunately. I didn't have time to look into it anymore.

Image of a graduating PhD student in the trending Studio Ghibli style.

I defended my PhD last week ✨

Huge thanks to:
• My supervisors @paulbuerkner.com @stefanradev.bsky.social @avehtari.bsky.social 👥
• The committee @ststaab.bsky.social @mniepert.bsky.social 📝
• The institutions @ellis.eu @unistuttgart.bsky.social @aalto.fi 🏫
• My wonderful collaborators 🧡

#PhDone 🎓

can you post a reprex on github?

What advice do folks have for organising projects that will be deployed to production? How do you organise your directories? What do you do if you're deploying multiple "things" (e.g. an app and an api) from the same project?

Amortized inference for finite mixture models ✨

The amortized approximator from BayesFlow closely matches the results of expensive-but-trustworthy HMC with Stan.

Check out the preprint and code by @kucharssim.bsky.social and @paulbuerkner.com👇

Finite mixture models are useful when data comes from multiple latent processes.

BayesFlow allows:
• Approximating the joint posterior of model parameters and mixture indicators
• Inferences for independent and dependent mixtures
• Amortization for fast and accurate estimation

📄 Preprint
💻 Code

Title: Posterior SBC: Simulation-Based Calibration Checking Conditional on Data Authors: Teemu Säilynoja, Marvin Schmitt, Paul Bürkner, Aki Vehtari Abstract: Simulation-based calibration checking (SBC) refers to the validation of an inference algorithm and model implementation through repeated inference on data simulated from a generative model. In the original and commonly used approach, the generative model uses parameters drawn from the prior, and thus the approach is testing whether the inference works for simulated data generated with parameter values plausible under that prior. This approach is natural and desirable when we want to test whether the inference works for a wide range of datasets we might observe. However, after observing data, we are interested in answering whether the inference works conditional on that particular data. In this paper, we propose posterior SBC and demonstrate how it can be used to validate the inference conditionally on observed data. We illustrate the utility of posterior SBC in three case studies: (1) A simple multilevel model; (2) a model that is governed by differential equations; and (3) a joint integrative neuroscience model which is approximated via amortized Bayesian inference with neural networks.

If you know simulation based calibration checking (SBC), you will enjoy our new paper "Posterior SBC: Simulation-Based Calibration Checking Conditional on Data" with Teemu Säilynoja, @marvinschmitt.com and @paulbuerkner.com
arxiv.org/abs/2502.03279 1/7

A study with 5M+ data points explores the link between cognitive parameters and socioeconomic outcomes: The stability of processing speed was the strongest predictor.

BayesFlow facilitated efficient inference for complex decision-making models, scaling Bayesian workflows to big data.

🔗Paper

cool idea! I will think about how to achieve something like this. can you open an issue on GitHub so I don't forget aboht it?

Join us this Thursday for a talk on efficient mixture and multilevel models with neural networks by @paulbuerkner.com at the new @approxbayesseminar.bsky.social!

Paul Bürkner (TU Dortmund University), will give our next talk. This will be about "Amortized Mixture and Multilevel Models", and is scheduled on Thursday the 30th January at 11am. To receive the link to join, sign up at listserv.csv.warwick...

Paul Bürkner (@paulbuerkner.com) will talk about amortized Bayesian multilevel models in the next Approximate Bayes Seminar on January 30 ⭐️

Sign up to the seminar’s mailing list below to get the meeting link 👇

Hochschulen und Forschungsinstitutionen verlassen Plattform X - Gemeinsam für Vielfalt, Freiheit und Wissenschaft

More than 60 German universities and research outfits are announcing that they will end their activities on twitter.

Including my alma mater, the University of Münster.

HT @thereallorenzmeyer.bsky.social nachrichten.idw-online.de/2025/01/10/h...

what are your best tips to fit shifted lognormal models (in #brms / Stan)? I'm using:
- checking the long tails (few long RTs make the tail estimation unwieldy)
- low initial values for ndt
- careful prior checks
- pathfinder estimation of initial values
still with increasing data, chains get stuck

I think this should be documented in the brms_families vignette. perhaps you can double check if the information you are looking for is indeed there.

happy to work with you on that if we find the time :-)

indeed, I saw it at StanCon but I am not sure anymore how production ready the method was.

something like this, yes. but ensuring the positive definiteness of arbitrary constraint correlation matrices is not trivial. so there may need to be some restrictions of what correlation patterns are allowed.

I already thought about this. a complete SEM syntax in brms would support selective error correlations by generalizing set_rescor()

OK, here is a very rough draft of a tutorial for #Bayesian #SEM using #brms for #rstats. It needs work, polish, has a lot of questions in it, and I need to add a references section. But, I think a lot of folk will find this useful, so.... jebyrnes.github.io/bayesian_sem... (use issues for comments!)

A ternary plot visualizing data points classified into three groups: CN (Cognitively Normal), MCI (Mild Cognitive Impairment), and AD (Alzheimer's Disease dementia). The triangular axes represent predicted probability corresponding to each diagnosis category, with the corners labeled CN (top), MCI (bottom left), and AD (bottom right). Each point is colored according to its real diagnosis group (blue for CN, green for MCI and red for AD). Trajectories of predicted probabilities pertaining to the same subject are denoted with arrows connecting the points. Shaded regions in blue, green, and orange further highlight distinct areas of the plot associated with CN, MCI, and AD, respectively. A legend on the right identifies the diagnosis categories by color.

Ternary plots to represent data in a simplex, yay or nay? #stats #statistics #neuroscience

Writing is thinking.

It’s not a part of the process that can be skipped; it’s the entire point.

1️⃣ An agent-based model simulates a dynamic population of professional speed climbers.
2️⃣ BayesFlow handles amortized parameter estimation in the SBI setting.

📣 Shoutout to @masonyoungblood.bsky.social & @sampassmore.bsky.social

📄 Preprint: osf.io/preprints/ps...
💻 Code: github.com/masonyoungbl...

Neural superstatistics are a framework for probabilistic models with time-varying parameters:

⋅ Joint estimation of stationary and time-varying parameters
⋅ Amortized parameter inference and model comparison
⋅ Multi-horizon predictions and leave-future-out CV

📄 Paper 1
📄 Paper 2
💻 BayesFlow Code

I think the link you cited points to the wrong paper.

yeah indeed it seems we don't have it yet. but perhaps may be worthwhile to implement? I will ask on the Stan forums.

I don't know. didn't check yet

I am always looking for count data distributions that can handle both under and overdispersion without being a computational nightmare. PRs are welcome :)

New report highlights the scientific impact of open source software Two of the scientists who won this year’s Nobel Prize for cracking the code of proteins’ intricate structures relied, in part, on a series of computing

“We don’t value software, data, and methods in the same way we value papers, even though those resources empower millions of scientists” 💯

www.statnews.com/sponsor/2024...