I really dislike the idea that every contributor to a paper has to be perfectly fluent with every element of it (there are positive possibilities for load sharing / teamwork, not just free-riding). Otherwise yeah, improved mechanisms for assessing paper value are of course desperately needed...
Have a half finished paper on this comparing a fairly simple very fast approximation to 'proper' NUTS sampling, curious to chat when we meet in Malta :)
I would have said something more in the middle... (replace FIML with multiple imputation as needed)
As a different reviewer I hate it :)
You can find my fork here github.com/Lakens/Quart... but it is under extremely active development, and probably not fully documented, so maybe check in end of next week? It is functional, but setting it up might still be some effort without assistance at the moment.
For future me:
Username (optional) = Username (not optional)
Is there a 'git personal access token' region of the brain? Did I sustain damage there as a child?
Torn between viewing this as moral failure of the individual vs system.
I vaguely remember experiencing that this only held for some cases, and that in certain situations the increase in iterations per minute (for the computationally simpler state sampling) outweighed the efficiency loss - did you encounter anything similar?
Thoughts and prayers.
Yeah I mean I only asked the question so I can figure out which side of the flame war to be on, otherwise I don't think it matters either ;)
True, for a minimal definition of confidence region at least yeah. But if I anyway rely on asymptotic assumptions and use a multivariate-normal for confidence intervals (i.e. from hessian at max likelihood), I should just as well use it for confidence region samples / transforms, no?
Is it somehow philosophically non-frequentist if I sample from the frequentist confidence region and compute my quantities of interest? Asking for a friend.
or 3) stuff in between.
Nah I just thought wanting a decent abstract would have been less controversial 5 years ago 😅
As if we're all out there only trying to collect skills. Doing stuff is useful too.
cool. looking forward to the people who find it meaningful and worthwhile to create useful summaries using whatever good tools are at their disposal.
why have papers at all? isn't it just lazy? shouldn't we all be out there performing the experiments and gathering the data ourselves, or at least, paying assistants to do it?
Let's go back 5 years in time and imagine I'd still like a good summary. Maybe I'm not as skilled, maybe I'm less tolerant of sifting through too much mediocre work, maybe I'm lazy or busy.
Hatred for llm's creates some interesting logic. I'd just like a good summary, piecing through 40 pages to discover that eg. nothing very interesting was done is boring and wastes time. Summaries exist as a concept for a reason :)
Seriously, if only this were true! Many abstracts are more like teasers.
Sure yeah. Some domains are at least observationally rich enough with strong enough signal that it was clear well before ml. Curious how much further it can go.
That sounds like a 'not rich enough data' issue to me? what scenarios are there where this is a problem that better / more / different data can't resolve?
For some high enough value of 'well predicting', on a sufficiently rich dataset, they should converge at least! Interesting q to me is in which domains current available data is rich enough, and how we might determine that without explicit experiment...
as a re-evaluation / comparison / supplement I like it a lot, yeah. Feels like it makes the measurement problem even harder though, but will be interesting to see.
Why stop there? Why extract anything? Just put in the text and ask the llm what you want to know about the person / sample. (Bit sceptical here!)
#jobs #stellenangebote #GESISjobs #jobfairy #TrustedResearchEnvironments
Pursue your scientific career at GESIS!
youtu.be/zMY2zqYFIMg
Apply now for a job as a Associate in Trusted Research Environments (Salary group 11 TV-L, working time 39,5 h/week, unlimited contract)
gesis.jobs.personio....
It would be easier to catch if we hadn't let in so much human slop.
Hard agree on attending to the (far more robust) contemporaneous correlations. But considering directed cross-effects *in combination with* the contemporaneous correlation also leads to much more consistent (and typically much less causally exciting) inferences. Rough draft:
osf.io/preprints/ps...
The (RI)-CLPM (discrete-time models in general) gets moderation structure really wrong if we include actual exercise as a mediator -- no, fitness shouldn't take 2 years to be influenced by motivation just because you ran a yearly panel study, or 2 weeks because you ran a weekly study.