iiinova — where your digital self keeps evolving

Reposted by Aidan Bindoff

Australian Museum Eureka Prizes @eurekaprizes.bsky.social · 29/07/2026

TAS Test @aidanbindoff.bsky.social from @utas.edu.au is a finalist in the Eureka Prize for Excellence in Interdisciplinary Scientific Research. Learn more: youtu.be/lcHGJQmcDjM #EurekaPrizes

Eureka Prize for Excellence in Interdisciplinary Scientific Research finalist team, TAS Test.

031

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

7/7 Drop-in package for the Stan/brms/loo ecosystem: github.com/ABindoff/rbloo Feedback welcome, especially on the edge cases - approximations are currently out of scope, falls back to plain PSIS-LOO.

github.com

GitHub - ABindoff/rbloo: Rao-Blackwellized leave-one-out cross-validation

Rao-Blackwellized leave-one-out cross-validation. Contribute to ABindoff/rbloo development by creating an account on GitHub.

010

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

6/7 And it's not cosmetic. In a model comparison, default PSIS-LOO calls it "decisive" evidence for the more complex model. The exact analysis (reproduced by RB-LOO) finds the two models indistinguishable. Wrong answer, high confidence.

100

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

5/7 On the brms epilepsy dataset's overdispersion stressor (97/236 folds failing), RB-LOO matches an 82-minute exact refit at ~zero cost. Moment matching, the current best fix, leaves over a third of those failures uncorrected.

100

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

4/7 The cure isn't new — marginalise the random effect instead of reweighting it. It's the standing recommendation in the loo docs. What's new here: closed-form/quadrature solutions for Gaussian, Bernoulli, binomial, Poisson random-intercept models, packaged as rb_loo(fit).

100

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

3/7 Turns out you can predict which folds will fail without forming a single importance weight. The Gelman–Pardoe pooling factor + structural leverage does it. In Gaussian LMMs it reduces to group size — a design-time diagnostic, AUC 0.96. GLMMs: one fit, AUC 0.81.

100

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

2/7 The problem: PSIS-LOO breaks on exactly the folds where a random-effect coordinate is data-driven and its group is small (think: singletons). Deleting the observation yanks the group's random effect around, weights get heavy-tailed, khat crosses 0.7.

100

Aidan Bindoff @aidanbindoff.bsky.social · 22/07/2026

1/7 Preprint's live: "Partial pooling predicts cross-validation reliability: a closed-form triage and Rao–Blackwellised cure for hierarchical LOO" arxiv.org/abs/2607.18836 PSIS-LOO fails predictably on hierarchical models. Here's the pattern, and the fix. 🧵

arxiv.org

Partial pooling predicts cross-validation reliability: a closed-form triage and Rao-Blackwellised cure for hierarchical LOO

For hierarchical models, Pareto-smoothed importance-sampling leave-one-out cross-validation (PSIS-LOO) fails on the folds where a random-effect coordinate is data-driven and its group is small. We sho...

120

Aidan Bindoff @aidanbindoff.bsky.social · 21/07/2026

This enables an a-priori two-level triage that recovers the exact LOO statistic while refitting only the few remaining folds where RB-LOO is unreliable. Methods paper to follow shortly.

000

Aidan Bindoff @aidanbindoff.bsky.social · 21/07/2026

Each observation's case-deletion influence is split into a pooling term, which governs where PSIS-LOO fails, and a variance-component term, which governs where RB-LOO's residual base importance sampling is strained.

110

Aidan Bindoff @aidanbindoff.bsky.social · 21/07/2026

If you've ever found yourself questioning the conclusions drawn from PSIS-LOO on a hierarchical model, or facing a multi-day exact LOO-CV, I think you'll appreciate my latest contribution to the #bayesian ecosystem - Rao-Blackwellised LOO with the rbloo package. www.github.com/ABindoff/rbloo

github.com

GitHub - ABindoff/rbloo: Rao-Blackwellized leave-one-out cross-validation

Rao-Blackwellized leave-one-out cross-validation. Contribute to ABindoff/rbloo development by creating an account on GitHub.

300

Aidan Bindoff @aidanbindoff.bsky.social · 23/06/2026

No but it looks a lot more user friendly than latex! On the rare occasion I first-author a paper I use latex, arXiv for example, insists.

010

Aidan Bindoff @aidanbindoff.bsky.social · 23/06/2026

If I need to submit in latex I'm not sure that typst can fill that gap yet? Looks good though

110

Reposted by Aidan Bindoff

Saloni @scientificdiscovery.dev · 19/06/2026

Very interesting talk by @mendelrandom.bsky.social on the apparent effect of shingles vaccines on dementia. Clear discontinuity from shingles vaccines on shingles, but not on dementia, in this much larger follow up in England, using similar analytic choices as the Wales study that showed a benefit.

RDD of cumulative risk of shingles over time, with a sudden drop when shingles vaccines are introduced.

RDD of cumulative risk of dementia over time, with no clear discontinuity when shingles vaccines are introduced.

511136

Aidan Bindoff @aidanbindoff.bsky.social · 19/06/2026

Understandable, he is not alone there

000

Aidan Bindoff @aidanbindoff.bsky.social · 19/06/2026

So if you have ever stared at divergent-transitions warnings and reparameterised by trial and error, there is now a number to compute instead. Would love to hear what people think. #Bayesian #MCMC #RStats #Stan #statistics

000

Aidan Bindoff @aidanbindoff.bsky.social · 19/06/2026

That leftover quantity is Gelman and Pardoe's "pooling factor". I show it is the exact weight for how much each coordinate should be centred vs non-centred, and derive it in closed form for every coordinate of a GLMM.

100

Aidan Bindoff @aidanbindoff.bsky.social · 19/06/2026

The idea: treat a hierarchical model as a fibre bundle, group parameters as fibres over the base space of population parameters. The natural connection on that bundle is flat everywhere. The geometry hides no obstruction to centring. What is left is purely statistical.

100

Aidan Bindoff @aidanbindoff.bsky.social · 19/06/2026

Building the smoothbp R package, I had an intuition about what causes funnel pathology in hierarchical models. I was wrong. But chasing that intuition led somewhere more fundamental to the centring vs non-centring decision than I had hoped. New paper: arxiv.org/abs/2606.19655

arxiv.org

A Flat Connection: The Pooling Factor and the Geometry of Centring in Hierarchical MCMC

Standard MCMC diagnostics ($\hat{R}$, effective sample size, divergence counts) detect whether a chain has mixed, but not why it has not. We ask whether the centring/non-centring obstruction in hierar...

161

Reposted by Aidan Bindoff

ArXiv Paperboy (Stat.ME+Econ.EM) @paperposterbot.bsky.social · 18/06/2026

arXiv📈🤖 smoothbp: Fast Bayesian Hierarchical Piecewise Regression with Smoothed Transitions and Spike-and-Slab Model Selection By Bindoff

Piecewise regression models are essential for identifying structural changes in longitudinal or spatial data across diverse scientific domains. While standard approaches often assume sharp, instantaneous transitions and single, non-hierarchical breakpoints, many real-world phenomena exhibit gradual, smoothed transitions that vary systematically across groups. We introduce smoothbp, an R package for fast, Bayesian hierarchical piecewise regression featuring logistic-smoothed transitions. By implementing a bespoke Metropolis-within-Gibbs sampler in Rust, smoothbp combines exact conjugate updates for linear terms with Hamiltonian Monte Carlo (HMC) transitions for non-linear location and sharpness parameters. smoothbp natively supports multiple change-points, random intercepts, random change-point timing, and structural covariates on all segment parameters. It also incorporates Kuo and Mallick (1998) spike-and-slab priors for automatic inference on the number of active breakpoints via the smoothbp_ss function. We document the sampler, validate parameter recovery and calibration through simulation-based calibration and interval-coverage studies, and contrast smoothbp against the existing software landscape across R, Python, Julia, and MATLAB, demonstrating its competitive efficiency against general-purpose probabilistic programming languages like brms and specialized packages like mcp.

021

Aidan Bindoff @aidanbindoff.bsky.social · 16/06/2026

Fantastic short article on how to understand uncertainty, apparent contradiction, and consensus in science

000

Aidan Bindoff @aidanbindoff.bsky.social · 14/06/2026

There's a classic story of Hilbert attending one of Von Neumann's lectures on Hilbert spaces, and Hilbert raising his hand at the end to ask "what is a Hilbert space?" - very relatable as I review my own work on fiber bundle connections in hierarchical models

000

Aidan Bindoff @aidanbindoff.bsky.social · 14/06/2026

Is anyone in my network an endorser for STAT.CO on arxiv.org? My co-authors typically submit elsewhere so I'm drawing a blank here #Rstats #Bayesian

010

Aidan Bindoff @aidanbindoff.bsky.social · 03/06/2026

smoothbp is now on CRAN! Fit hierarchical piecewise regression models with smoothed change-points, spike-and-slab regularisation, and a Rust-powered sampler — so it's fast. Great for data with structural breaks you want to localise without hard discontinuities (i.e differentiable) #rstats #CRAN

030

Aidan Bindoff @aidanbindoff.bsky.social · 28/05/2026

Almost... I'm hoping to get a preprint up next week because I need to submit a manuscript which uses the package. Thanks for your interest!

010

Aidan Bindoff @aidanbindoff.bsky.social · 28/05/2026

smoothbp has been submitted to CRAN! Hierarchical piecewise regression with smoothed change-points in #RStats - multi-breakpoint models, spike-and-slab regularisation for automatic breakpoint selection, and a fast MCMC sampler written in Rust under the hood. #rstats #statistics #bayesian

281

Aidan Bindoff @aidanbindoff.bsky.social · 15/05/2026

If you don't use your tokens they won't get the same budget next year

000

Aidan Bindoff @aidanbindoff.bsky.social · 15/05/2026

A pull request has been submitted to cline so that repro-mcp can be used effortlessly in VS Code, if you want to see this happen, go check out issue #10752 at GitHub.com/cline and vote it up or add your thoughts

github.com

Cline

Cline has 20 repositories available. Follow their code on GitHub.

000

Aidan Bindoff @aidanbindoff.bsky.social · 15/05/2026

I recently interviewed three PhD candidates with impressive publication histories in machine learning who couldn't describe where they might use AUC (let alone how or why)

000

Aidan Bindoff @aidanbindoff.bsky.social · 15/05/2026

If you're using agentic workflows (e.g. Claude) and value reproducibility and transparency, check out github.com/ABindoff/rep... - an MCP server that brings reproducibility logging to AI-assisted scientific computing

github.com

GitHub - ABindoff/repro-mcp

Contribute to ABindoff/repro-mcp development by creating an account on GitHub.

100

Aidan Bindoff @aidanbindoff.bsky.social · 01/04/2026

Unavoidable in R, might as well go all-in!

010

Aidan Bindoff @aidanbindoff.bsky.social · 29/03/2026

Ladies Cottage, Willow Court, Brigita Ozolins exhibition

010

Aidan Bindoff @aidanbindoff.bsky.social · 09/10/2025

Imagine mistakenly drinking cyanide laced Flavor-Aid, thinking it was cyanide laced Kool-Aid? They must have felt so silly when they realized they had been tricked into drinking Flavor-Aid!

000

Aidan Bindoff @aidanbindoff.bsky.social · 08/10/2025

Another day, another study using Shapiro-Wilk to assess "normality of data", and not reporting those results (which was fine, because the outcome was dichotomized anyway, because of course it was...)

000

Aidan Bindoff @aidanbindoff.bsky.social · 27/09/2025

I'm more scared of them than they are of me

120

Aidan Bindoff @aidanbindoff.bsky.social · 27/09/2025

I suggested my partner get something similar to stop her kids from coming into her room at night

110

Aidan Bindoff @aidanbindoff.bsky.social · 29/08/2025

We're all well aware that "correlation does not imply causation" but few heed this warning when it comes to ratios; indeed you can find them uncritically thrown about almost everywhere

000

Aidan Bindoff @aidanbindoff.bsky.social · 29/07/2025

You can have some of my painkillers, they make me very sleepy* *Side effects may include capture by a sleep paralysis demon

010

Aidan Bindoff @aidanbindoff.bsky.social · 29/07/2025

The vibe coding apps are marketing to me so hard rn. The thing is, I have to methodically specify what I need whether I write the functions or AI does it and happily I do that best by ...writing code.

020

Aidan Bindoff @aidanbindoff.bsky.social · 28/07/2025

Even a back of the envelope calculation that ignores most of the complexity should cause people designing (or reviewing) these studies some alarm

001

Aidan Bindoff @aidanbindoff.bsky.social · 09/07/2025

I appreciate that it's staggeringly brazen, but do we think Bondi erred when she claimed to have the Epstein list then later denied it exists? I think this administration is very comfortable with protecting power structures (beyond govt) that they are a part of, and want you to know you don't matter

000

Aidan Bindoff @aidanbindoff.bsky.social · 20/03/2025

They're not really "bears", they are marsupials

120

Aidan Bindoff @aidanbindoff.bsky.social · 03/03/2025

Greetings from the Hamilton Show (the original, not the musical), Tasmania

060