Guido Biele

I think a "rich-get-richer" effect could be stronger if one does not provide additional info. But that would in my view also be bad practice. Asking an LLM to help with a review by using a RAG system with the relevant literature could well make the effect weaker.

Genau für diese Art von Text, der in erster Linie für Verwaltungszwecke verfasst wird, sind m. E. LLMs nützliche und effiziente Werkzeuge, bei denen, aus meiner Sicht, die Vorteile ggü den Nachteilen der Nutzung überwiegen.

Generally agree. An neat use of ordinal regressions is to use them as a semi-parametric model for weird (multimodal) VAS data. If one splits the original scale in bins, one can with some effort calculate marginal effects on the original scale, with resolution depending on # of bins.

Typos in the post are virtue signals, indicating that no LLM was used in its writing 😀

The last few days a couple of texts made the round here that equate using AI in the research process with doing bad research. I think these are 2 orthogonal things: Using AI can make research worse and it can make it better, depent on how it is used.

I don't see how the post effectively criticises target population estimands. It shows that if you have a target population estimand in mind, but your study isn't designed to estimate it, you'll pay with a high variance of your estimate.

OTH, “Our analysis should take into account that children grow around 7.5 cm per year and rarely less than 1 cm or more than 15 cm per year.” sounds entirely reasonable and illustrates that priors often reflect not just subjective beliefs but objective information that can robustify inference.

One reason for the lack of appreciation might be that Bayesian methods are often taught in a very abstract way, just as in the book at the start of the thread. Teaching the formula “posterior = prior × likelihood” tends to focus attention on so-called subjective beliefs as priors.

I agree that one can use Bayes for many reasons and that it is useful in relatively simple situations. However, my (admittedly limited) observation is that the ability to fit complex models attracts more people to Bayesian estimation than improved inference in simpler cases.

It works usually well if one uses t-tests or anovas to analyse simple experiments. But one shouldn't generalise from that to the goals/usefulness of Bayesian or Frequentist methods in general.

It seems to be a common misunderstanding that the goal of applied Bayesian inference is to let the prior influence the inference. I think a more common goal is to be able to reliably fit complex models that are hard/impossible with Frequentist methods.

Did they look at both parental income and education? In Norwegian data the association of child outcomes with parental education is clearly stronger than with parental income.

To let LLMs produce less common and therefore likely also more creative text, one can always set the temperature parameter.
But there are trade offs. Less common might also be more likely wrong.

Sometimes psychologists collaborate with causal inference experts and use it for their research, here the probability of benefit or harm of ADHD medication for school outcomes: arxiv.org/abs/2502.10049
All math by @johandh2o.bsky.social

YouTube video by 3Blue1Brown Transformers (how LLMs work) explained visually | DL5

Understanding the transformer architecture helps explain why LLMs are so powerful—and when they’re likely to fail.
The Andrej Karpathy video goes into less detail about transformers than would be helpful. Here's a quick and accessible intro from 3Blue1Brown. www.youtube.com/watch?v=wjZo...

I believe large language models (LLMs) can be very useful and can be used to support learning effectively.
However, this plot from a recent study (www.nature.com/articles/s41...), which praises the impact of LLMs on learning outcomes, doesn't exactly inspire confidence.

Easily only for models with identity link function (also careful with treatment-interactions). Otherwise one might have to code up a model in e.g. Stan, omit priors on coefficients involving the treatment & put a prior on the marginal effect (calculated in the transformed parameters block).

The article talks about non-collapsibility, but doesn't seem to use the term. I think its a helpful term, because it puts a name to a key problem with OR, which is there even if the OR is not interpreted as a RR.

Links:
documents1.worldbank.org/curated/en/0...
assets-eu.researchsquare.com/files/rs-424...
arxiv.org/pdf/2409.090...
arxiv.org/pdf/2402.09809

Would be great to have that on CRAN 😀

Ist natürlich auch mit Vorsicht zu geniessen. Eine wichtige Schwäche ist, dass es alle Aussagen von Autoren unkritisch akzeptiert.

Elicit: The AI Research Assistant Use AI to search, summarize, extract data from, and chat with over 125 million papers. Used by over 2 million researchers in academia and industry.

Genau, wenn schon, dann besser elicit.com :-)

Anyhow, I am sympathetic to focusing on selection bias, because I think quasi-experimental designs, which are often possible only in selected samples, focus so much on internal validity, that it easy to overlook that they might exchange confounding bias with selection bias.

Using missingnes graphs provide an additional way to define internal and external validity: violations of internal validity are due to backdoor paths that do not involve missingness indicator nodes (or their children) and violations of external validity involve those. (Just made this up ...)

Agreed that one usually cares about some population, and not about a specific sample. On the other hand, (some?) Trial statistician will say that an RCT is designed to estimate the SATE and no more.
I think it is therefore useful to be explicit about the estimand before one starts discussing bias.

Agreed that selection bias receives too little attention.
Still, the problem only materialises if the estimand is the effect in a target population.
Then, a sensible thing to do is to draw a missingnes graph (m-DAG) and check if (conditional) exchangebility is violated (any backdoor paths?).

synthpop: Generating Synthetic Versions of Sensitive Microdata for Statistical Disclosure Control A tool for producing synthetic versions of microdata containing confidential information so that they are safe to be released to users for exploratory analysis. The key objective of generating synthet...

cran.r-project.org/web/packages...
Though it is in my experience not easy to generate synthetic data that capture all relevant aspects of the data, especially if some binary variables are very unevenly distributed (like rare diagnoses).

Yann LeCunYann LeCun VP & Chief AI Scientist at Meta in a linkedin post: Hey Europe, you want a vibrant tech industry, right? The US seems set on destroying its public research funding system. Many US-based scientists are looking for a Plan B. You may have an opportunity to attract some of the best scientists in the world. Scientists will go where they have the means to be the most creative and productive. Here are the criteria that attract them: 1- access to top students and junior collaborators. 2- access to research funding with little administrative overhead. 3- good compensation (comparable with top universities in the US, Switzerland, Canada). 4- freedom to do research on what they think is most promising. 5- access to research facilities (e.g. computing infrastructure, etc). 6- ability to collaborate/consult with industry and startups. 7- moderate teaching and administrative duties. They will seek the best trade-off between these criteria in academia, public research, or industry. European academia rates high on 1 and 4, low on 2 (even if you can get an ERC grant), 5, 6 and 7, and *very* low on 3. European industry rates low on almost every criterion, particularly on 4, but also on 3 and 5 compared to top US industry labs. To attract the best scientific and technological talents, make science and technology research professions attractive. It's pretty straightforward.

Sometimes it's not only *what* is being said but *by whom* it is being said.

www.linkedin.com/feed/update/...

Causal inference is hard.
Here you can see how we, really @tvarnetperez.bsky.social, do our best to give a good answer to an important question for those with ADHD: Does longterm medication improve learning outcomes?

Does anyone know of good resources for staying updated on large language model developments (ChatGPT, Gemini, Claude) relevant to quantitative researchers? I'm looking for feeds/accounts that describe experiences with models or highlight new releases (including betas) and other research updates.