Chao Wang

Chao Wang

@excel-wang.bsky.social

Associate Professor in health and social care statistics at Kingston University. PhD in econometrics.

27 followers20 following91 posts

Chao Wang
@excel-wang.bsky.social

Study design does not necessarily dictate the level of evidence, as often depicted in "evidence pyramid". Context is important. See the example below (youtu.be/3pxudE0GNAo?...). Many people thought Study 2, a cross-sectional study, is the weakest. However Study 2 is arguably the strongest.

Chao Wang
@excel-wang.bsky.social

I asked various LLMs to rate the strength of the studies and here are the results. They all think Study 2 is the weakest, except Gemini Argon, which correctly pointed out "local ICB/CCG funding rules are largely exogenous to an individual's biological prognosis".

Chao Wang reposted

turgut keskintürk
@tkeskinturk.bsky.social

peter hull comparing TWFE and modern DIDs & it seems TWFE is OK. people will go "YOU SEE IT DIDN'T MATTER AFTER ALL" but regardless of whether TWFE is ultimately useful or not that all DiD chaos was a net good because it clarified the estimand. tinyurl.com/twfevsdid

Chao Wang
@excel-wang.bsky.social

If you think LLM is the way to achieve general intelligence (AGI?), think again. Using chess/Go as an example, LLMs are still far behind top human players, let alone top chess/Go AIs such as AlphaGo (which is not an LLM). x.com/i/grok/share...

x.comLLMs Poor at Chess and GoGeneral-purpose LLMs are mediocre at chess (roughly beginner to intermediate club level at best) and very weak at Go. They cannot beat professional human players or top specialized AIs like Stockfish,...
View full thread
Chao Wang
@excel-wang.bsky.social

LLMs are essentially next-token prediction. It tells us, given past game records, which move is most likely to happen. AlphaGo on the other hand, tells us, given the possible moves, which move has the best chance of winning.

Chao Wang
@excel-wang.bsky.social

Given the nature of LLMs, they also have limitations in other scenarios, such as driving. The human brain can write text, drive a car, and play games like chess or Go, demonstrating a more versatile kind of intelligence.

Chao Wang
@excel-wang.bsky.social

If you think LLM is the way to achieve general intelligence (AGI?), think again. Using chess/Go as an example, LLMs are still far behind top human players, let alone top chess/Go AIs such as AlphaGo (which is not an LLM). x.com/i/grok/share...

x.comLLMs Poor at Chess and GoGeneral-purpose LLMs are mediocre at chess (roughly beginner to intermediate club level at best) and very weak at Go. They cannot beat professional human players or top specialized AIs like Stockfish,...
Chao Wang
@excel-wang.bsky.social

Interesting article. It shows LLM cannot discover causal relationships purely based on the data without experts’ input.

Adam Kucharski@adamjkucharski.bsky.social

What can AI attempts to do causal inference tell us about the value of domain insights vs scale of raw data? Read more in my new post: kucharski.substack.com/p/confounded...

Chao Wang reposted

Jess Calarco
@jessicacalarco.com

Replying to @jessicacalarco.com

For example: When I teach qualitative methods, I ask my students which of these foods are sandwiches, and there's always lots of disagreement. So, we talk about how they can explain their criteria and try to persuade others to agree with them. But they can't objectively prove they're right.

Chao Wang reposted

Adam Kucharski
@adamjkucharski.bsky.social

Two fictional (presumably AI) stories in recent X posts, now repeated as a Google summary. No need to tell tall tales about your achievements when the AI ecosystem will now do it for you…

Chao Wang
@excel-wang.bsky.social

You will get a better chance with closed-source feature-rich (i.e. not heavily relying on third-party packages to do anything useful) software.

Olivia Guest · Ολίβια Γκεστ@olivia.science

this is important for those convinced by the reproducibility bros who I don't see for blessed reasons (they blocked me) > It is now more difficult to run a five-year-old Python script than a 40-year-old Fortran program, and even if it runs, it may not produce the same results as it did in the past.

Chao Wang reposted

Chao Wang
@excel-wang.bsky.social

Replying to @adamjkucharski.bsky.social

Tried a few models and both Gemini 3.5 Flash (Extended, screenshot1) and M365 Copilot GPT 5.5 (Think Deeper, s2) got the correct answer. Some reported Claude Opus 4.7 is also fine. Yes more advanced mode may not always be better, but they tend to be more correct than instant/quick/... mode.

Chao Wang reposted

Darren Dahly
@statsepi.bsky.social

We tried to tell y'all to stop calling everything "AI" many years ago and you just wouldn't listen and now the poor machine learners must also suffer alongside the statisticians 😜

This post is unavailable.
Chao Wang
@excel-wang.bsky.social

I asked ChatGPT how to calculate confidence intervals without researching and sourcing third-party R packages. Here’s the response.

Robert (Bob) Kubinec@rmkubinec.bsky.social

My #rstats cheat code for today is the binom.confint function in the binom package that will spit out *12* different ways of calculating a CI for a proportion. Also, this is why you use R for statistics... (and of course the correct CI method is bayes 😎)

Chao Wang reposted

Paolo Crosetto
@paolocrosetto.bsky.social

What is the most profitable industry in the world, this side of the law? Not oil, not IT, not pharma. It's *scientific publishing*. We call this the Drain of Scientific Publishing. Paper: arxiv.org/abs/2511.04820 Background: doi.org/10.1162/qss_... Thread @markhanson.fediscience.org.ap.brid.gy 👇

Mark A. Hanson@hansonmark.bsky.social

We wrote the Strain on scientific publishing to highlight the problems of time & trust. With a fantastic group of co-authors, we present The Drain of Scientific Publishing: a 🧵 1/n Drain: arxiv.org/abs/2511.04820 Strain: direct.mit.edu/qss/article/... Oligopoly: direct.mit.edu/qss/article/...

Chao Wang reposted

Compton Scatterbrained
@comptonscatter.bsky.social

Replying to @gdeejay.bsky.social

The future is bleak: AI "researchers" will send questionaires to AI "respondents" to collect data on what they "think" about Topic, poorly summarise the results and generate a paper for a mill. AI peers will review and leave comments. Other AI will cite the paper (and nonexistent ones). Wooooo.

Chao Wang reposted

Darren Dahly
@statsepi.bsky.social

Replying to @danlewer.bsky.social

I think it was a target trial emulation of difference in differences for a propensity score matched analysis with Bayesian borrowing of historical controls of digital twins based on synthetic data from the real world.

Chao Wang reposted

Peter Tennant
@pwgtennant.bsky.social

Once you truly understand causal inference theory, it's very hard to win an honest applied grant because: * Many questions can't be answered * Almost everything that sounds big and exciting is snake oil * Most reviewers won't understand * you're competing with the bandwagon-jumping cynics

Chao Wang
@excel-wang.bsky.social

More on “Simple method works equally well…” and do you really need SuperLearner? drwangstatsconsulting.wordpress.com/2025/06/17/m...

drwangstatsconsulting.wordpress.comMore on “Simple method works equally well…” and do you really need SuperLearner?I recently came across some interesting tutorials on TMLE. This reminds me of my previous post on comparing different estimators of treatment effects (see here). The default setting in the tmle R p…

Chao Wang reposted

Maarten van Smeden
@maartenvsmeden.bsky.social

Replying to @maartenvsmeden.bsky.social

If the goal is to describe or to explain, true high colinearity between important covariates may be a fact of life and difficult to overcome. Quick-fix solutions may result in bad confounding corrections. Bottom-line: quick fix solutions like PCA are often a poor man's solutions to colinearity

You are up to date.