Martin Modrák

Make an effect size prediction!

@jamiecummins.bsky.social and I are replicating Balcetis & Dunning's (2010) "chocolate is more desirable than poop" (Cohen's d = 4.52)

Let us known in the replies what effect size you think we'll find. Details of the study in the thread below.

note that in the ggplot2 the default pointsize is not a constant anymore

GeomPoint$default_aes
#> Aesthetic mapping:
#> * `shape` -> `from_theme(pointshape)`
#> * `colour` -> `from_theme(colour %||% ink)`
#> * `fill` -> `from_theme(fill %||% NA)`
#> * `size` -> `from_theme(pointsize)`

Bayes Club A community of statisticians using Bayesian methods.

If you feel so inclined, I run a Mastodon instance bayes.club and try to make it a nice place and you can join if you like. I know some people have strong opinions on this and some have had bad experience on Mastodon. And that's all fair and I won't dispute that, there's a lot of work to do.

ELLIS PhD Program: Call for Applications 2025 The ELLIS mission is to create a diverse European network that promotes research excellence and advances breakthroughs in AI, as well as a pan-European PhD program to educate the next generation of AI...

I'm looking for a doctoral student with Bayesian background to work on Bayesian workflow and cross-validation (see my publication list users.aalto.fi/~ave/publica... for my recent work) at Aalto University.

Apply through the ELLIS PhD program (dl October 31) ellis.eu/news/ellis-p...

Nobody:
Line plots: These are her mountains and skies and she radiates
Through history's rivers of blood, she regenerates

One thing I find so strange about the P-curve is how once it was published, the burden of proof shifted to others to prove that it is bad (in ways that Simonsohn would care about, at least), rather than the authors, to prove that it works in general (for standard definitions of "works"). (21/x)

Simonsohn has now posted a blog response to our recent paper about the poor statistical properties of the P curve. @clintin.bsky.social and I are finishing up a less-technical paper that will serve as a response. But I wanted to address a meta-issue *around* this that may clarify some things. 1/x

Retraction Note: High replicability of newly discovered social-behavioural findings is achievable - Nature Human Behaviour Nature Human Behaviour - Retraction Note: High replicability of newly discovered social-behavioural findings is achievable

It's officially been a year since this paper was retracted. In the wake, the authors claimed (misleadingly) it was just a couple pesky sentences... a revise and resubmit.

Since then, radio silence with not even an update to the preprint.

www.nature.com/articles/s41...

Fewer people hit/spank/beat their kids and it is less socially acceptable to do so.

this might be a bit of a long shot, but is anyone here willing to share documents of an ERC (starting) grant? or any insights into the process, things you wished you knew before…? i would be very thankful as i think it is quite different from programs i am familiar with so far. ☀️

But apparently most of the Internet agrees that "population" estimates are "preferred for RCTs" but without any explanation why! Have I lost my mind? Or is this one of the situation where "we have always done things this way"? 3/3

they worry that for one outcome GLMMs provide "subject-specific" estimate and we should consider GEE to get "population" estimate. This seems weird to me, because the sites are a convenience sample and I really want to generalize to other sites and make decisions for individual patients 2/3

I am **NOW** writing a response to regulator on a first RCT I helped design and the plan is to use random intercepts for sites, so I ponder as well. The regulator doesn't mind the random effects (we plan fixed ef as sensitivity analysis), but, on a related note, 1/3

*cracks knuckles*

okay so you know how the media has a problem with taking snippets of scientific studies out of context and writing articles which mislead the public?

now imagine a computer program is doing that when you ask it questions about scientific literature but also it makes stuff up.

Release of CmdStan 2.37 We are very happy to announce that the 2.37.0 release of CmdStan is now available on Github! As usual, the release of CmdStan is accompanied by new releases of Stan Math, core Stan, and Stanc3. Thi…

Stan / CmdStan 2.37 release!
blog.mc-stan.org/2025/09/02/r...
- sum_to_zero_matrix type sums to zero across both rows and columns
- simplex and *_stochastic_matrix types should be now faster and more stable
- new functions exposing the built-in constraint implementations
1/2

A SQL query goes into a bar, walks up to 2 tables and says “Can I join you?”

Citogenesis

Obligatory: xkcd.com/978/
The LLM use could however increase efficiency of the process substantially

Destroy this myth. Please share if you have visited and researched in a library this past year! RT And give a shout out to your favorite library you have ever visited.

One Epidemic, Many Estimates (1EME) One Epidemic, Many Estimates (1EME)

📢 Interested in excess mortality methods, and want a challenge? I'm organising the "One Epidemic, Many Estimates" (1EME) project! Register *now* as a many analyst team (submissions due 15 March 2026), and then join us at LSE for a workshop on 21-22 May 2026! (1/n)

www.lse.ac.uk/Economic-His...

Great to see this happen and a good opportunity to caution against some pitfalls in many-analyst studies: www.pnas.org/doi/10.1073/...

Can't imagine a better team to pull this off, so do join

Models as Prediction Machines: How to Convert Confusing Coefficients into Clear Quantities Abstract Psychological researchers usually make sense of regression models by interpreting coefficient estimates directly. This works well enough for simple linear models, but is more challenging for more complex models with, for example, categorical variables, interactions, non-linearities, and hierarchical structures. Here, we introduce an alternative approach to making sense of statistical models. The central idea is to abstract away from the mechanics of estimation, and to treat models as “counterfactual prediction machines,” which are subsequently queried to estimate quantities and conduct tests that matter substantively. This workflow is model-agnostic; it can be applied in a consistent fashion to draw causal or descriptive inference from a wide range of models. We illustrate how to implement this workflow with the marginaleffects package, which supports over 100 different classes of models in R and Python, and present two worked examples. These examples show how the workflow can be applied across designs (e.g., observational study, randomized experiment) to answer different research questions (e.g., associations, causal effects, effect heterogeneity) while facing various challenges (e.g., controlling for confounders in a flexible manner, modelling ordinal outcomes, and interpreting non-linear models).

Figure illustrating model predictions. On the X-axis the predictor, annual gross income in Euro. On the Y-axis the outcome, predicted life satisfaction. A solid line marks the curve of predictions on which individual data points are marked as model-implied outcomes at incomes of interest. Comparing two such predictions gives us a comparison. We can also fit a tangent to the line of predictions, which illustrates the slope at any given point of the curve.

A figure illustrating various ways to include age as a predictor in a model. On the x-axis age (predictor), on the y-axis the outcome (model-implied importance of friends, including confidence intervals). Illustrated are 1. age as a categorical predictor, resultings in the predictions bouncing around a lot with wide confidence intervals 2. age as a linear predictor, which forces a straight line through the data points that has a very tight confidence band and 3. age splines, which lies somewhere in between as it smoothly follows the data but has more uncertainty than the straight line.

Ever stared at a table of regression coefficients & wondered what you're doing with your life?

Very excited to share this gentle introduction to another way of making sense of statistical models (w @vincentab.bsky.social)
Preprint: doi.org/10.31234/osf...
Website: j-rohrer.github.io/marginal-psy...

Israel hits Gaza hospital, killing at least 15 people Israeli strikes on Nasser Hospital in Gaza this morning have killed at least 15 people, including three journalists, one of whom worked for Reuters and one who worked for Al Jazeera.

Israeli strikes on Nasser Hospital in Gaza this morning have killed at least 15 people, including three journalists, one of whom worked for Reuters and one who worked for Al Jazeera

+1 I rarely get much stuff done when caring for the kids, except when the dishwasher needs unloading :-D

New paper with @richarddmorey.bsky.social now out in JASA, where we critically examine p-curve. Below is Richard’s excellent summary of the many poor statistical properties of p-curve (with link to paper). I wanted to add some conceptual issues that we also tackle in the paper.

Cover page for the manuscript: Morey, R. D., & Davis-Stober, C. P. (2025). On the poor statistical properties of the P-curve meta-analytic procedure. Journal of the American Statistical Association, 1–19. https://doi.org/10.1080/01621459.2025.2544397

Abstract for the paper: The P-curve (Simonsohn, Nelson, & Simmons, 2014; Simonsohn, Simmons, & Nelson, 2015) is a widely-used suite of meta-analytic tests advertised for detecting problems in sets of studies. They are based on nonparametric combinations of p values (e.g., Marden, 1985) across significant (p < .05) studies and are variously claimed to detect “evidential value”, “lack of evidential value”, and “left skew” in p values. We show that these tests do not have the properties ascribed to them. Moreover, they fail basic desiderata for tests, including admissibility and monotonicity. In light of these serious problems, we recommend against the use of the P-curve tests.

Paper drop, for anyone interested in #metascience, #statistics, or #metaanalysis! @clintin.bsky.social and I show in a new paper in JASA that the P-curve, a popular forensic meta-analysis method, has deeply undesirable statistical properties. www.tandfonline.com/doi/full/10.... 1/?

this is really good -- completely aligns with where ive come to myself on this after much difficulty and costly error

My experience also supports that. The workout/exercise stuff was explained to a hopeless ~22yo me by a friend and he was right: the effects of workout on your mental state and relationship to your body are both more important and easier to achieve than any noticeable change in your appearance.

So I think there is a niche of things like tool benchmarks where what you want to do is empirics in silico and this type of prereg can make sense. Though I agree that in a fashion typical to COS they assume a single type of simulation study they are familiar with is the only possible type..

And already at the time there was a long history of unfulfilled claims that the next game revolution will be in AI. Turns out that "smart" or "optimal" and "fun to play" only very rarely align. "smart" often meant hard to debug, hard to tweak to match design goals and expensive computation 2/2

It's just terminology: when I was involved in game AI (2010s), ~90% of industry-oriented articles self-described as AI were on navigation, finite state machines and behavior trees (kinda like decision trees). Making something that's fun to play with/against and eats little CPU is hard! 1/2