Matthijs Hollanders @matthollanders.bsky.social · 15/09/2026Do you know if it’s bad for the longevity when this happens? My fans are constantly going when running MCMC on my M5 Pro MBP. 110
Matthijs Hollanders @matthollanders.bsky.social · 15/09/2026I don’t know enough about frequentist stats to understand. 100
Matthijs Hollanders @matthollanders.bsky.social · 15/09/2026Also the prior basically vanishing with lots of observations unless it’s very narrow (you already made up your mind). 010
Matthijs Hollanders @matthollanders.bsky.social · 15/09/2026Bayes is a fresh of breath air. Infinitely more intuitive and scalable. Write the Stan program and see how well the optimiser recovers your hierarchical effects. 😁 000
Matthijs Hollanders @matthollanders.bsky.social · 13/09/2026I was talking about most principled, not the simplest! 000
Matthijs Hollanders @matthollanders.bsky.social · 13/09/2026are you sure brms doesn’t do fixed effects without a shared scale by default, and random effects without sum to zero? 120
Matthijs Hollanders @matthollanders.bsky.social · 13/09/2026It’s not much of a hassle, and it seems reasonable that if you include a categorical predictor, that the effects share a common scale. 100
Matthijs Hollanders @matthollanders.bsky.social · 13/09/2026You identify the intercept. The interpretation is now the mean across your levels. 100
Matthijs Hollanders @matthollanders.bsky.social · 13/09/2026In Stan or pymc, create a sum to zero vector and give it a normal prior. 000
Matthijs Hollanders @matthollanders.bsky.social · 13/09/2026I remember a tweet from @rmcelreath.bsky.social saying he sometimes modeled categorical effects as hierarchical (I assume drawn from a normal with shared scale). I reckon modeling them this way, with a sum-to-zero constraint to retain a nice intercept, *has* to be the most principled strategy. 130
Matthijs Hollanders @matthollanders.bsky.social · 10/09/2026Great time to be looking for some tiger snakes. 010
Matthijs Hollanders @matthollanders.bsky.social · 08/09/2026Data stickers that are science enthusiasts 000
Reposted by Matthijs HollandersJarno Vanhatalo @jarnovanhatalo.bsky.social · 08/09/2026During the many long term ecological change projects in @rececochange.bsky.social, we have concretely learned how important long term data are. Here some thoughts on the topic 👇 Long-Term Community Data Reveal Ecological and Evolutionary Phenomena - www.annualreviews.org/content/jour...annualreviews.orgLong-Term Community Data Reveal Ecological and Evolutionary PhenomenaResolving the ecological and evolutionary processes affecting biodiversity requires long-term community data. By separating short-term variability from directional change and by revealing lags spannin... 0116
Matthijs Hollanders @matthollanders.bsky.social · 08/09/2026How about those of us that feel ethically conflicted using these carbon fiends 031
Matthijs Hollanders @matthollanders.bsky.social · 03/09/2026The strongest evidence that I’ve seen for that we live in a simulation is the apparent number of people that know how to use regex. 150
Matthijs Hollanders @matthollanders.bsky.social · 30/08/2026Stan in the hands of stats GOATs but still inclusive 👏🏼 010
Matthijs Hollanders @matthollanders.bsky.social · 29/08/2026Wow what are the chances I also saw a full moonrise last night 000
Matthijs Hollanders @matthollanders.bsky.social · 29/08/2026Epic few days finding wildlife in the Top End. 020
Matthijs Hollanders @matthollanders.bsky.social · 23/08/2026Just watched your talk, it was great. The recovery of the population mean, is there a parallel with kriging/making predictions for unobserved locations with a GP? 110
Matthijs Hollanders @matthollanders.bsky.social · 22/08/2026It would be the one place where the imposter syndrome isn’t imagined. 011
Matthijs Hollanders @matthollanders.bsky.social · 22/08/2026One day I’ll be at StanCon. Not presenting that’s for sure though. 110
Matthijs Hollanders @matthollanders.bsky.social · 15/08/2026Not to be that guy but did you organise an inspection prior to purchase? 100
Matthijs Hollanders @matthollanders.bsky.social · 14/08/2026I mostly use it for math/stats concepts. I’ll look for ones that do that! 010
Matthijs Hollanders @matthollanders.bsky.social · 14/08/2026Do you have tips? It's overwhelming for a noob. I've got a Pro chip with 64 gb ram; using qwen3.6-35b-a3b now. What do you mean with hooking up infra to API? 100
Matthijs Hollanders @matthollanders.bsky.social · 14/08/2026Set it up yesterday, seems to work well! thanks 100
Matthijs Hollanders @matthollanders.bsky.social · 14/08/2026That looks good. But you could easily do this with your other approach too right? Just set alpha[j] = tau[j], or as i do, alpha[1] = gamma, alpha[2:J] = mu * tau[1:(J-1)], where tau is scaled to ahve geometric mean 1, mu is estimated, and gamma is estimated as well as an initial pop. correction. 110
Matthijs Hollanders @matthollanders.bsky.social · 12/08/2026I do it here too, estimating either the Dirichlet concentration (1/alpha) or logistic-normal scale. Seems to work well. mhollanders.github.io/occARU/artic...mhollanders.github.ioThe occARU model 110
Matthijs Hollanders @matthollanders.bsky.social · 10/08/2026alpha=1 also gets quite informative with large number of surveys. 010
Matthijs Hollanders @matthollanders.bsky.social · 10/08/2026Have you tried estimating alpha? The sparsity of the entry simplex is identified. 110
Matthijs Hollanders @matthollanders.bsky.social · 05/08/2026See these are exactly the things a layperson like me doesn’t know. 110