Jonas K. Lindeløv

I've made this cheat sheet and I think it's important. Most stats 101 tests are simple linear models - including "non-parametric" tests. It's so simple we should only teach regression. Avoid confusing students with a zoo of named tests. lindeloev.github.io/tests-as-linea… 1/n

Dear American friends. We agree that status quo in the Artcic is not an option. So let’s talk about how we can fix it - together. Lars Løkke Rasmussen Danish Foreign Minister

@Nate__Haines Yes, I think your title is fine. I'm surprised they managed to hijack the general term "Bayesian Optimization" for such a specific algorithm. Though I've used it for several manufacturing optimization problems and it works great!

@jonaslindeloev Good point! I used "Bayesian *design* optimization" purposefully, but can see how people could get confused. That said, the procedure is very Bayesian, so I think it makes sense to have that in the name 🤔

1/N Some New Years reading to share! In this post, we dive into Cronbach's alpha, Fisher info, KL divergence, and Bayes factors as measures of item informativeness. We then use these metrics to reduce a large 100 item pool down to just 15 items while maximizing information 🤖

@Nate__Haines Wiki article: en.wikipedia.org/wiki/Bayesian_… A much-used python module to do it: github.com/bayesian-optim… A great GUI implementation for practicioners: browniebee.io ("Brownie" because optimizing industrial baking-recipies is one application...)

@Nate__Haines Well done! Terminology note: "Bayesian Optimization" denotes a particular black-box algorithm for finding the combo of parameter values that maximizes a utility function. It's used a lot in Design Of Experiments (DOE). "Bayesian Design Optimization" sound like that but isn't.

Collection of insane and fun facts about SQLite. Let's go! SQLite is the most deployed and most used database. There are over one trillion (1000000000000 or a million million) SQLite databases in active use. It is maintained by three people. They don't allow outside contributions.

@mattansb BTW, we primarily relied on Bayesian inference, but we had added p-values after rejection from another journal for the use of "unconventional statistics".

@mattansb Rejection with reference to one reviewer whose main criticism was, "to my knowledge, an effect size less than 1 SD cannot be significant". I replied with two lines of R code demonstrating d=0.01 can be significant. But no reply. This was a very high-impact journal.

Hey #statsTwitter, what's that worst statistics-related comment/request you got from a reviewer or editor? Did you comply - did you do what they asked? If not, how did you push back?

NEW to {bayestestR} dev version - support for #marginaleffects! Take it for a test drive: remotes::install_github("easystats/bayestestR") @easystats4u #rstats

@SolomonKurz Left. At the right, it would be next to impossible to see extreme "low" lines.

2 ways of showing Bayesian uncertainty. Left is the posterior distribution with the mean line in bold, and the thinner lines 100 random HMC draws. The right is the same, but with semitransparent fill for each of the 100 draws. Thoughts? [#sec-Beta-binomial" target="_blank" rel="nofollow noopener">solomon.quarto.pub/sr2rstan/12.ht…] #RStats #rstan

@MicrosoftLoop Is math rendering coming too? Preferably accepting LaTeX as input.

Introducing Loop 2.0! We got a fresh new look 🔥🔥🔥 Beyond the new #MicrosoftLoop UI, we've made it easier to get to your meeting notes, favorites, recent, and more. #NewInLoop Check it out: msft.it/6018laHsr Here's what we've been brewing 👇

@JamesSteeleII @MatheusdPinto @AJThurston @AllenDowney @JASPStats For Christmas, I want a JASP plugin that does (G)LM, taking R's lm() syntax as model specification. Would be wonderful for teaching.

What if we taught #psychstats showing there is only one test, per @AllenDowney: allendowney.blogspot.com/2011/05/there-… Using the {infer} #rstats package: infer.tidymodels.org/articles/infer… Or at least leveraging #dataviz to illustrate its all OLS like @jonaslindeloev’s post here: lindeloev.github.io/tests-as-linea…

@AJThurston @AllenDowney Always a great point to make! I like to say that we should teach *statistical modeling*. Then testing becomes the easy part. For the {infer} package, that would be spending 75+% of the time on the specify() step.

@JamesSteeleII @AJThurston @AllenDowney I'd be eager to see them!

@statsepi #arithmeticmean

Gør mere godt for pengene ved at give effektivt. Se vores nye explainer-video og del!

@SatisfactoryAF December 25th, 2025

what date do you think 1.0 is releasing? if you guess correctly, you will officially be allowed to brag about it on the internet for getting it right.

@AmandaKMontoya @SJKim_Psych Thanks for the pointer to BUCCS! And I would have loved to be part of that class.

@SJKim_Psych @jonaslindeloev Jolynn Pek taught a power analysis class when I was in grad school hybridpower was a product of that class, and I think we either learned about Samantha’s method or I learned about it later. But that class got me really interested in power analysis.

There are no R packages to calculate sample size (/power) given uncertain population parmeters. I.e., like specifying distributions over mu and sigma in a t-test rather than point estimates. Must. Resist.

@fusaroli @lakens For now, just think t-test level complexity: mu and sigma. Then we can expand from there 😁 The thing I don't like about sensitivity is that it's just eye-balling/intuition-building. Exactly one *decision* has to be made in the end. One sample size; Run or don't run.

@jonaslindeloev @lakens Also do you really know eg the sd of random effects by participant or by stimulus? I do not and it has a huge impact on power. One could do a sensitivity analysis, but adding uncertainty does give more conservative estimates for sensitivity as well, so…

@MetinBulus I have not; thank you! This seems to align with the approach I had in mind. Do you have an evaluation of the merits of this approach? The package/github repo seems to have gone quite unnoticed.

@lakens Please say more :-) I'm often involved deciding whether to do a study or not, i.e., whether the required sample size surpasses the budget. Smallest eff of interest is absolute (non-standardized); We have imprecise info about true effect, impact of design differences, etc.

@jonaslindeloev One good reason for this is you should not do a power analysis if you are too uncertain about SD, and the effect size should be based on the smallest effect size of interest. So, in best practices there should be no need for such an R package.