if you have a specific non-linear function in mind then brms has a special “non-linear syntax” that is well described in the docs

depends on what you mean by non-linear, but if you want something like a GAM then it’s about as simple as:

bf(y ~ s(x), phi ~ s(x), family=“beta”)

a particularly fun example of this is that the term “sensitivity” now means both P(true positive) and also z(P(true positive)) - z(P(false positive)), which really helps with explaining ROC curves

moral of the story is that just about anything can be used as a black box, which isn’t inherently bad. but this is also why we need experts to spot the “landlord special”/causal salad regressions for us

this might sound like a knock on this particular professor, but it’s also something we do all the time. most people don’t have a deep causal understanding of how e.g. their fridge, electric, or bicycle work. we usually don’t even notice it until pointed out by a handyman/mechanic

I once worked with a senior professor with a habit for calling anything they didn’t understand a “black box”. it wasn’t until they used it to describe GLMMs that I realized the habit was less an exercise in intellectual humility than an excuse to avoid statistical thinking

sounds like a job for @singmann.bsky.social

singmann.org/topics/mpt-m...

and sometimes it’s “literally anything with computational models, except you have to call it AI or else it doesn’t count”

the thing I don’t understand about this trend is that “AI” is being prioritized without anybody seeming to know what it is. sometimes it’s “do OpenAI’s job for them but for less money”, sometimes it’s “expose people to AI slop and see what happens”,