Hacker News Reader

Comment by tptacek | Hacker News Reader

I estimate, conservatively, that about 90% of papers published outside the STEM fields can be categorized as Cargo Cult Science.

I like the idea that he calls this a "conservative" estimate. That sounds much better than "factoid I pulled out of my butt". An unfathomably vast number of papers are published in every mainstream field of academia, STEM and not, every year. The reason we only have "conservative" "estimates" is that nobody can possibly read all of them.

Pick a field and take it down if you can. But trying to build a case against all of social science using the example of someone notoriously elevated despite a lackluster career as a publishing academic doesn't get you anywhere!

View on HN

DARPA is doing a relatively large scale effort to see if social sciences papers replicate, and how easy it is to look at a paper and guess if it will replicate

the preliminary estimation phase predicts like 50% rates of replication

let's be real, the average social science paper is not good

edit: there are lots of garbage papers in every field, the culture of academia is deeply flawed, the low standards are just even lower in some fields

The problem is that replication studies aren't done hard enough.

And by replication studies I mean "take this paper, try to replicate, write a paper on it".

Every singular paper can be wrong, that's why you're supposed to replicate them - that's at least what was taught in psychology, which many consider "social science of less rigour" - well, it was probably the course most hardcore about experimental design in my education and ability to replicate a study...

Are they not done or are they done and nobody wants to publish them?

Nobody CAN publish them. An actual criterion for most top journals is literally "novelty". Reviewers will immediately levee "this has been done before" as a damning critique in peer review. There is no replication because you can't currently share replications, let alone sustain a career on them.

My experience of "representative" studies in psychology (university of Berne) was that they'd come up with something, then gather a group of "volunteer" psychology students - 80% women - who need to participate at x studies as requirement to pass, publish it as representative serious research, and call it a day.

>The problem is that replication studies aren't done hard enough.

>And by replication studies I mean "take this paper, try to replicate, write a paper on it".

You don't necessarily need to perform a straight-up "replication study" in order to do some useful sanity checking on published results. It can be done as part of another novel study that builds upon prior work. That's basically what happened with the fraudulent UCLA political persuasion study several years ago [1]. Some Berkeley researchers wanted to extend the ideas in that earlier published paper, reused similar methods under the assumption that the UCLA folks were correct, and couldn't get anything reasonable within the ballpark of what was published.

[1] https://fivethirtyeight.com/features/how-two-grad-students-u...

Yes there’s a serious issue, but thrust of the previous comment was that it’s not productive to criticize poor research using poor research.

Conservatively 90%? Bullshit. The real numbers are already terrible so why use unintentional hyperbole to make the argument it just weakens it.

I don't think there's any objective way to say that a majority of papers are bullshit, even if they are. Failure to replicate is "easy" to show, but doesn't say anything about the utility or quality of the science done, beyond the absolute basics of a paper describing what a scientist did in detail and not lying about the results. A "sky is blue" paper will replicate, but I'm not sure how you'd find a "fair" way to call it useless.

Anyway, jumping from "50% of papers will probably fail an absolute bare minimum test" to "90% of papers are garbage" is subjectively reasonable to me.

If 90% seems shocking to someone, I suggest they skim a bunch of papers and develop their own feel for how bad the situation is.

I read a lot of history and some linguistics, it seems really fine to me. Linguistic especially, they work on _a lot_ of data, you don't have the "tested on 15 subjects" from biology or medecine effect.

Even if i agreed with your "subjectively reasonable" idea, the issue is that non-STEM science encapsulate a lot of different fields.

what's non-STEM science? the incentives of publishing and funding systems are similar across most fields (publish papers in the cool journals, get as many citations as possible, leverage that for more funding), I would be astonished if there were fields substantially exempt from bad science.

I feel like my position is fair for any field publishing papers with falsifiable claims based on evidence?

Possibly you and the OP would be better served to say “I wouldn’t be surprised if some claim about x%” is true. Or, “I would guess it’s true that”.

Back to the main subject Vertasium has a related video:

Is Most Published Research Wrong https://youtu.be/42QuXLucH3Q?si=p9M0ZOAdHuowD30n

Also, Feynman was very clear that the field of physics is not immune to the underlying dynamics, citing both the historical reluctance to correct the Millikan oil drop experiment and sloppy methods being used in contemporary particle accelerator research.

He might be surprised to learn how many papers published in STEM are cargo cult science.

I'd say conservatively ;-) about 75%

And yet, this is a kind of figure that possibly could be proven, if you either assign a lot of people to read / analyze the papers and mark them "cargo cult" and "not cargo cult", or automate it with modern day text analysis tooling. Getting all the papers in one place will be a logistic challenge though.

Even just 100 or 200 random papers would be okay for a rough indication. There must be someone who has done this?

That some papers are bollocks is not a controversial notion, but I genuinely have no idea if it's 10%, 50%, or 90% that's bollocks. Even just a rough indication of "about 20%" or "about 80%" would be hugely helpful. It's also unclear that non-STEM is more bollocky than STEM – Oxman's thesis is STEM, and also bollocks. STEM word salad papers don't make the news because the conclusions are rarely interesting, whereas sociology and psychology papers do, because the conclusions tend to be more interesting.

Leaving aside the vast semantic Gulf between what Feynman was talking about and things being "bollocks", how do you evaluate the truthiness of novel arguments? The typical solution here is domain experts, aka peer review. The author's point is that they don't trust peer review, so you need some other method that's both practical and universal. I can't imagine what that could possibly look like.

I've probably read 100 ML papers. They're pretty bad. Even the most cited ones like Adam, Attention, and Latent Diffusion are messy, unclear, or just straight up wrong at times.

Adam - Their algorithm is inefficient. Their paper can be summarized as "use the signal-to-noise ratio," but the key idea is hidden in a page-long paragraph halfway through the paper.

Attention - I probably read "keys, query, values" a dozen times before I realized the key and query matrices multiply with the same vector (where the "self" comes in).

Latent Diffusion - Their theoretical grounding was just completely wrong. It's not some stochastic diffeq, or Markov process, or w/e terms they threw around. It's just finding the gradient of log-likelihood using finite differences (the error they add).

You could also do a random sampling.

SciHub provides an index of their DOIs [1] as well as torrents for the actual content. I presume one could cross reference citation counts and pull the top few hundred papers in different fields to get a pretty good sample over time and research fields.

[1] https://sci-hub.se/database#download

The fact that so many papers are published rather works against your argument. The more papers are published in a given field, the more likely they are to be bullshit.

Of course, that applies for STEM fields as well as the social sciences. However, if we do comparisons across fields, such as looking at how many citations a typical paper in some field will receive, we find that STEM tends to do a lot better than the social sciences[1].

>Pick a field and take it down if you can. But trying to build a case against all of social science using the example of someone notoriously elevated despite a lackluster career as a publishing academic doesn't get you anywhere!

The longer such fraud goes uncorrected, the more rewarded a person is for their fraud, and indeed the more lengths other people go to cover up such fraud (as was argued by the OP) the more likely such fraud is normalized in a particular field[2]. The Claudine Gay story suggests that, if anything, XKCD was wildly optimistic about the social sciences[3].

[1] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5446096/ [2] https://danluu.com/wat/ [3] https://xkcd.com/451/

How does that follow? Is your premise that a static number of people are producing ever-increasing numbers of papers? That's not what's happening at all. There are lots of papers because there are lots of people doing research. As the number of vulnerability researchers increased, drastically in the 2000s, the number of vulnerabilities went up --- there was a point where people we tracking new vulnerabilities, across all platforms, by hand! But the median quality of a vulnerability research result didn't fall; in fact, while the variance increased, the peak quality of vulnerabilities improved.

I do not assume a static number of researchers, nor do I see why you would infer that from what I said. Of course there are more researchers going into these fields. The more people enter into a field, the more likely the typical person entering that field is going to be a mediocre researcher, and the more likely that their work is going to be crap.

Sure, in the case of a very small and new field with promise of intellectually stimulating work and structural advantages for eliminating bullshit, adding more people and papers is likely going to be a net benefit. But realistically many established disciplines can only add people by lowering standards.

As an aside, why is this the thing you chose to focus on in responding to me? While I do think it is true and I am willing to defend it, it's not really the central issue.

What I understand as the central point of your original comment was that we cannot condemn all of the social sciences (or 90% of the work in them) on the basis of the Claudine Gay story. I refute that with this sentence:

>The longer such fraud goes uncorrected, the more rewarded a person is for their fraud, and indeed the more lengths other people go to cover up such fraud (as was argued by the OP) the more likely such fraud is normalized in a particular field. The Claudine Gay story suggests that, if anything, XKCD was wildly optimistic about the social sciences.

Put another way, in the spirit of Scott Alexander's recent essay "Against Learning From Dramatic Events", the Claudine Gay story represents a significant deviation from what would be predicted by a model that assumed the social sciences had a more robust level of intellectual integrity.

As a STEM prof, all my papers are cargo cult science.

Reply on news.ycombinator.com