Matthijs Hollanders

IG

Make everything the same shade of black and that’s my theme.

I can’t stand that online journals can’t format equations properly.

Have you read ai-2027.com?

How about the rise of Stan tho

McElreath uses Mac!

AI 2027 A research-backed AI scenario forecast.

What do people think of this?

ai-2027.com

Is it at least the least of two evils?

Non collapsibility? Time invariance? I know for a fact the latter isn’t a problem with hazard ratios.

@jordannafa.bsky.social

Yeah, but working with rates is a bit more intuitive and comparing two rates I think is more intuitive than comparing log odds. I wonder why it isn’t used more.

Hey #stats people, what do we think about interpreting coefficients in cloglog binomial regressions? Since everyone hates odds ratios I wonder if interpreting hazard ratios is easier.

I’m quitting.

Why do we peer review for free?

Seen.

Hasn't it been like this for a while with R, Python, and Julia?

Iconic interaction.

Looks great can’t wait to read.

Change your tattoo plans to this.

Im confused too. I think often Dirichlet and Gaussian processes are termed non-parametric, which doesn’t make sense because you model them with parametric assumptions (eg MVN for GPs).

Bottom signal.

I need some help with parameterising a latent simplex, so a set of latent probabilities that sum to 1. Can anyone have a yarn?

Hazard rates FTW 🎉

This thing slaps.

brms is one of the most amazing R packages ever. Nevertheless I can't get myself to use it for my own work because nothing beats coding up your own models where you have ultimate control and insight into what's going on. There remains a "black box" element unless you are deeply familiar with it.

Working on a multisite/multistate/robust design Jolly-Seber model in Stan where I need to marginalise (1) the entry occasion of each individual and (2) the sub-area within each site where each individual enters. 😮‍💨

It’s all just latent variable models right?

But are you buying?

It’s literally impossible.

Hey Aki, that’s really cool. But wouldn’t the level of impact on the posterior be different from dataset to dataset? I’d imagine that with SBC you’re simulating a bunch of datasets and looking at the overall behaviour.