What's wrong with social science and how to fix it (2020)

"It’s analogous to money laundering. Here’s how it works: First, various academics have strong moral impulses about something. For example, they perceive negative attitudes about obesity in society, and they want to stop people from making the obese feel bad about their condition. In other words, they convince themselves that the clinical concept of obesity (a medical term) is merely a story we tell ourselves about fat (a descriptive term); it’s not true or false—in this particular case, it’s a story that exists within a social power dynamic that unjustly ascribes authority to medical knowledge.

Second, academics who share these sentiments start a peer-reviewed periodical such as Fat Studies—an actual academic journal. They organize Fat Studies like every other academic journal, with a board of directors, a codified submission process, special editions with guest editors, a pool of credentialed “experts” to vet submissions, and so on. The journal’s founders, allies and collaborators then publish articles in Fat Studies and “grow” their journal. Soon, other academics with similar beliefs submit papers, which are accepted or rejected. Ideas and moral impulses go in, knowledge comes out. Voilà!

Eventually, after activist scholars petition university libraries to carry the journal, making it financially viable for a large publisher like Taylor & Francis, Fat Studies becomes established. Before long, there’s an extensive canon of academic work—ideas, prejudice, opinion and moral impulses—that has been laundered into “knowledge.” (source: https://www.wsj.com/articles/idea-laundering-in-academia-115...)

I was one of the extreme "trust the science" people, until I joined a startup that worked with academia. The amount of pettiness, vindictiveness, cutthroat power games I have seen surpassed even the most hardcore of startups.

1. Publication venues aren’t created equal. People outside of academia don’t understand that anyone can and will start a journal/conference. If I want to launch the Proceedings of the of Confabulatory Results in Computer Science Conference, all I need is a cheap hosting account and I’m ready to go. In countries like the US this is explicitly protected speech, and the big for-profit sites run by Elsevier et al. will often go ahead and add my publications to their database as long as I make enough of an effort to make things look legitimate.

In any given field there are probably a handful of “top” venues that everyone in the field knows, and a vast long tail ranging from mid-tier to absolute fantasist for-profit crap. If you focus on the known conferences you might see bad results, but if you focus on the long tail you’re doing the equivalent of searching the 50th+ page of Google results for some term: that is, you’re deliberately selecting for SEO crap. And given the relatively high cost of peer-reviewing the good vs not peer-reviewing the bad, unfiltered searches will always be dominated by crap (surely this is a named “law”?) Within TFA I cannot tell if the author is filtering properly for decent venues (as any self-respecting expert would do) or if they’re just complaining about spam. Some mention is made of a DARPA project, so I’m hopeful it’s not too bad. However even small filtering errors will instantly make your results meaningless.

2. Citations aren’t intended as a goddamn endorsement (or metric). In science we cite papers liberally. We do not do this because we want people to get a cookie. We don’t do it because we’re endorsing every result. We do it because we’ve learned (or been told) about a possibly relevant result and we want the reader to know about it too. When it comes to citations, more is usually better. Just as it is much better to let many innocent people go free rather than imprison one guilty one, it is vastly better to cite a hundred mediocre or incorrect papers than to miss one important citation. Readers should not see a citation and infer correctness or importance unless the author specifically states this in the text at which point, sure, that’s an error. But most citation-counting metrics don’t actually read the citing papers, they just do text matching. Since most bulk citations are just reading lists, this also filters for citations that don’t mean much about the quality of a work.

The idea of using citation counts as a metric for research quality is a bad one that administrators (and the type of researchers who solve administrative problems) came up with. It is one of those “worst ideas except for any other” solutions: probably better than throwing darts at a board and certainly more scalable than reading every paper for quality. But the idea is artificial at best, and complaints like “why are people citing incorrect results” fundamentally ask citations to be something they're not.

Overall there are many legitimate complaints about academia and replicability to be had out there. But salting them with potential nonsense does nobody any favors, and just makes the process of fixing these issues much less likely to succeed.

I'm a physical scientist in industry, so I don't publish, or use p-values. But if the familiar "square root of n" rule holds, than a 10x improvement in any kind of "noise" measure requires a 100x increase in sample size. This is probably impossible for many kinds of experiments. Even in my work, desiring such an improvement usually requires me to scrap an experiment and come up with a better one, rather than slogging away to collect more data.

Worse, we may suffer from the fact that for some kinds of studies, sample size is limited by the number of potential subjects per researcher. You can only ask undergraduates to do so much. More researchers in the field means weaker statistics.

A possible minor improvement is to eliminate the requirement of publishing a masters thesis project. My friends who were psych grad students all had to do a master's en route to the PhD, and it had to be published to fulfill their requirement. This might have seemed like a quality check, but instead created an incentive to load the literature with low quality research. And many of the students were not looking forward to research careers anyway. Their term for finishing the degree was "published and outa here."

Even if all of the published papers were peer reviewed and replicated, there's a lot of science that never sees the light of day.

Even non-results are important - if not glamourous, and (even worse) publishing is disincentivised if the results go against what the funders wanted or expected.

A good example of a field with the same problems is epidemiology. I wrote a brief article with many examples a couple of years ago [1]. Unfalsifiable claims, objectively false claims, non-replicable work, citations that were invalid in various ways (e.g. citing a paper in support of a claim which didn't actually support the claim), and most surprisingly, a lot of circular logic. This field seems to routinely lose track of what comes from assumptions vs simulations vs reality. Papers would make claims that sounded like factual statements about observed data, but when the citation was checked the number would turn out to be a simulation output or even an arbitrary assumption used as simulation input. Validation had gone AWOL long ago with the substitutes being unscientific, e.g. claiming a model was validated because it roughly matched the output of other models.

You could argue that epidemiology is actually a social science rather than a hard science. Maybe the distinction isn't interesting though.

[1] https://plan99.net/~mike/epidemiology.pdf

100 comments at the time: https://news.ycombinator.com/item?id=24447724

The academic influence of social science papers is uncorrelated with their scientific quality.

e.g. https://paperswithcode.com/ and the recent phenomenon of people hosting their system demonstrations on huggingface.

Especially as this comes from "a part of DARPA's SCORE program, whose goal is to evaluate the reliability of social science research"

Everyone else in academia is working towards this end result or drops out somewhere along the path and works in industry. The exception being those specialist degrees that are quasi-academic such as law and medicine, but which have more applied (and therefore measurable) results.

And yet, we don't hold these people to the standards they set for themselves as a cohort? I wouldn't expect every paper to be replicable, or every researcher to always be right. But I see things like the https://en.wikipedia.org/wiki/EmDrive and wonder how many billions of dollar were put into this thing. It's essentially a state backed way of defrauding the public out of tax dollars to pay people with doctorates that made friends with other people with doctorates. What if the US spent the money that was used on the EmDrive on housing or social services for instance? Or building libraries or bridges?

All it would take would be for a few academic journals to demand that the statistical replication of papers be required, but it would mean that the editors would have to be guided by principles other than maximizing revenue.

How would you implement changes?