@neurostats.org
AI in Bio & Health & Therapeutic Development Bio: https://linktr.ee/mnarayan Substack: https://blog.neurostats.org Peek into my brain: notes.manjarinarayan.org Previously @dynotx @StanfordMed PhD@RiceU_ECE | BS@ECEILLINOIS 🧪🧮⚕️🧬🧠🖥🤖📈✍️🩺👩📈📉
I think the problem here is that most people who do epidemiological research - including many who might call themselves 'epidemiologists' - have not actually had any epidemiology training. I wish it was a protected title. With an exam that would require proving you understand residual confounding!
Otherwise, I'm more considered about pre-validation issues — correcting for chronological age without nested cross-validation in biomarker prediction models.
The nonlinear thing becomes an issue if you want to make a claim that residuals are independent of age.
For example, current evidence is that LLMs rely too much on interpolation rather than discovering underlying invariant causal abstractions.
📌
Same goes for structure determines function concepts
I think Cox is earnest. It is scary to listen to him because he is right that the AI presentation of "facts" is largely based on what is popular, but you can also steer it to selectively give you evidence that confirms your own worldview.
This appears to be a very problematic weaponization of casual reasoning that feels 'rigorous' because it involves a selective presentation of evidence. Some very excellent researchers have looked into the problems from air pollution.
#CausalSky #MetaSky
www.theguardian.com/technology/n...
📌
What year was this picture taken in?
Exactly
> The people who care about a topic enough to research it will often have ideological motivations. The ideological influence is there. That doesn’t mean it’s bad science.
Every few years there is a new article like this, but not much will change unless scientists stop relying on GraphPad Prism that doesn't offer any of this and have something they can use more easily.
LLM powered workflows might change this going forward.
📌
The grass is not greener, I'm afraid.
It's a niche for sure. But it exists and it's underutilized and could be further developed
📌
Unless you are using inferential clustering methods that are designed to severely test any hypotheses against the baseline that "how often would you find clusters if the data was generated by these high dimensional null patterns or if data were corrupted by structured measurement error"
Can you demonstrate the extent and degree of different kinds of sequential selection biases from a #CausalSky #StatSky #MedSky perspective in public clinical trial datasets?
DM me if you are interested.
For instance in TrialBench?
github.com/ML2Health/ML...
the Mesoscopic Integrated Neuroimaging Data (MIND) platform will soon be opening also for outside researchers leveraging our new 15.2 T 🐁 MRI, SHIELD clearing and fully upgraded lightsheet microscope plus computational resources developed by @neuroak.bsky.social to handle large datasets.
📌
📌
I also think most of the authors were not solid neuroimaging researchers. It was like a fun project. Not particularly good science.
Neurobullshit is alive and well in education research. Reminds me of the good old days of @neuroskeptic.bsky.social on this 10+ years ago
📌
Also, longitudinal studies don't fix causality problems. Fortunately, they help you discover entirely new kinds of confounding.
📌
If you use a model dialogically, it’s effectively guiding you through the free-writing/revision cycle composition teachers recommend.
I get why this mode of use isn’t very visible. All the rhetoric around AI has been that it’s “generative” and does the work for you. Things like the “Dear Sydney” +
I feel vindicated
cc:
@jackiesullivan.bsky.social @kordinglab.bsky.social nglab.bsky.social and Paul Cisek.
