This post was sparked by an interesting e-mail exchange with Jeremy Fox, over at Dynamic Ecology. We’d both come across the same announcement of a (very likely) case of research fraud, and had some similar reactions to it. We both knew there was a blog post in it! We agreed to post at the same time, but not to share draft posts. My prediction: we agree on some parts, not on others; but Jeremy’s post is better.
Behavioural economics got a bit of a black eye last week with the revelation that a major study by some very prominent authors is, virtually certainly, based on fraudulent data. What’s really astonishing, if you read that post (and you should) is that the fraud was so stunningly obvious with even a rather shallow dive into the data. Just to pick one thing, a treatment effect in the paper seems to have been generated by taking one variable, and adding to it a random number pulled from a uniform distribution bounded by 0 and 50,000. (Seriously, read the post.) This is such an implausible distribution for a real experimental effect that, once it’s been noticed, it’s about the most flagrant red flag you could imagine.
It’s not just this paper, though. Think about recent fraud cases in ecology and evolution, at least some of which were detected because investigators apparently copied-and-pasted large blocks of numbers in spreadsheets – spreadsheets that were posted publicly under a journal’s data archiving policy.
But why? Not why fraud (that’s an interesting question that Jeremy has dived into before, for instance here, but it’s not my question today). Instead, why clumsy, obvious fraud? There’s an interesting contrast here, I think, with forgeries in art. The “add a uniform random number from 0 to 50,000) strategy seems like taking a can of house paint and a ruler and dashing off a ten-minute “Mondrian”. Art curators, collectors, or auction houses wouldn’t be fooled for a moment. Instead, modern art forgery is an extremely sophisticated business – forgers find authentic period furniture, for example, that can be disassembled to yield wood panels of the right age to paint on. Why the contrast? Why do scientists seem so bad at fraud?*
Well, I don’t have an answer for you. (Perhaps fortunately, I’m not a scientific fraud expert.) But here are a few (interrelated) ideas, and you can suggest others in the Replies.
- Science is fundamentally built on trust. As modern science became professionalized, in the late 19th and early 20th centuries, we changed our minds about how a piece of scientific writing gained authority. In science’s early days, investigators were concerned with having witnesses to their work who could vouch for it. With professionalization, we began to assign authority to research results in part because their author was a member of the profession, with credentials from, and appointment at, reputable institutions. That doesn’t work completely, of course, but the fact that fraud rates (or at least, detected fraud rates) are very low suggests it works to a considerable extent. And since the system operates by trust, it’s vulnerable to even crude attempts at deceit. (Jeremy has explored that idea here.) You can think of this as good news: fraud isn’t testing the system enough to require attempts at fraud to be sophisticated. In art, this isn’t true; art forgery has been a lucrative enterprise for long enough that the fraud-detection system has had to be well practiced.
- Building on that last point: we’re mostly, probably, amateurs at fraud. With perhaps a very few exceptions, there aren’t professional science fraudsters (serial fraud appears even rarer than fraud itself). Art forgers have expertise and practice, and sometimes long careers producing dozens or hundreds of forgeries. Perhaps most scientific fraud is committed by someone who’s just decided to give it a whirl, under time pressure as the job market, a tenure deadline, or some other motivator looms. If fraud happens that way, perhaps it’s no surprise it’s slapdash. But to undercut my own suggestion: are we really that amateurish? We’re pretty good at experimental design, data analysis, and all the rest. Do our skills really not extrapolate to fraud?
- What if our skills do in principle extrapolate to fraud, but in practice, we don’t apply those skills? Here’s the thing: it’s likely that sophisticated fraud is difficult. Sure, it would have been trivially easy, for that economics paper, to draw “treatment effect” values from a normal distribution instead of a uniform one; but that’s only one of the many red flags in that paper. Nature does really weird and complicated things to our real data. Simulating that convincingly has always struck me as very difficult. I mean, I don’t actually need reasons not to fake my data; but if I did, my complete confidence that I’d suck at it would provide one! To take this one step further: it seems likely to me that conducting a really convincing, sophisticated scientific fraud would be more work than just doing the experiment for real.
- Finally, the most troubling possibility (and it’s so obvious that I’m sure you’ve gotten there before me): maybe scientists do conduct sophisticated fraud, but it goes undetected. Under this explanation, the fraud cases in the news are the crude ones simply because those are the ones we notice easily; but if we put the effort in to thoroughly review our literature, we’d unveil a wide distribution of fraud “quality”. Imagine a normal (not uniform, for Pete’s sake!) distribution of fraud quality, with the veil line due to limited inspection currently hiding all but the rare crude cases in the left (low-quality) tail. As we poured in more resources, we’d see more and more sophisticated cases, and likely more of them. (Because it would be, frankly, embarrassing if the crudeness of the uniform-0-to-50,000 case was the mode!). This idea will appeal to those who think that everything in science is “broken”, and I have to admit it isn’t impossible – but it’s very difficult to test. In art, substantial resources go into the investigation of forgery, because the costs of forgery are very high: a work could sell for $100,000,000 if judged genuine, but be near-worthless if that judgment is wrong. Do we have these same incentives in science? Arguably not – I’d take the perhaps controversial position that it’s unusual for a fraudulent paper to have very high costs. That’s because we rarely base our understanding of nature on a single paper – especially when we intend to build major investments (science, policy, or otherwise) on that understanding. I think it’s just hard for a single paper (fraudulent or not) to move the needle. So it’s rational for us not to invest more in fraud detection – even if it’s disturbing to think about what we might find.
So, four ideas (and please offer others in the Replies). They each seem plausible to me – but none stop me from being utterly gobsmacked each time I see another example of fraud that’s so crude it’s almost clownish. It would be amusing, if it wasn’t – no, wait, I have to admit: it IS amusing.
© Stephen Heard August 25, 2021
Image: How an ecologist might forge art. Public domain, via Maxpixel.net
*^Jeremy alerted me to this Retraction Watch post (from yesterday), which describes a faker’s “elaborate steps to cover her deception”. However, the faker in question has had multiple papers retracted, grant funding yanked, and her medical license (temporarily) revoked, all without the apparent need for Sherlock Holmesian deduction. So I’m not sure she’s really that good an example of artful scientific fraud.
Pingback: Scientific fraud vs. art forgery (or, why are so many scientific frauds so easy to detect?) | Dynamic Ecology
Your post has a better header image than mine.
I will take this very small triumph.
Good on you for bringing up the argument that, if the frauds we’re detecting are all obvious, that frauds must actually be numerous, because obvious frauds are surely a tiny fraction of all frauds. I couldn’t find a way to work this into my post without breaking up the flow, or else saddling it with even more p.s.’s than I did.
I’m not sure this argument is right. I mean, yes, it’s true that we’d expect the average *detected* fraud to be more obvious (i.e. detectable) than the average fraud. But I don’t see how you can use the obviousness of the average detected fraud to infer the frequency of frauds among all scientific papers, or the frequency of fraudsters among all scientists. To make those inferences, I think you have to make very strong assumptions about the shape of the distribution of fraud detectability. And I just don’t buy those assumptions. I’m not willing to assume that, for every laughably obvious fraud we detect, there must be thousands of well-disguised frauds out there. It seems more plausible to me that, for every laughably obvious fraud we detect, there must be (say) several more laughably obvious frauds out there, plus a very few well-disguised ones. The old posts of mine that you were kind enough to link to discuss surveys and other lines of evidence indicating that frauds and fraudsters are rare, not common-but-undetected.
So, do you think we’re at the start of an evolutionary arms race in science between fraud and fraud detection? As those old posts you linked to noted, the rate of retractions in science increased a lot in the oughts, as many individuals and journals started taking fraud seriously and putting measures into place to detect it and address it (retraction policies, data sharing policies, use of automated software to detect manipulated images, PubPeer, etc.) In response, some forms of fraud may be dropping in frequency. IIRC, rates of image manipulation have been going down at journals that have started using software to check for it. So will scientific fraudsters start to raise their games, and if so, how? I wonder if we’ll start seeing cross-paper duplications and relabelings of real data. That is, data that are only obviously fake if you compare data across papers. We’ve already seen a couple of instances of that in EEB (that we’ve detected).
Or maybe fraudsters won’t raise their games, at least not much. As my post today notes, my biostats undergrads are well aware that I look for plagiarism, and can easily guess how I do it (text-matching software). Once word started getting around that I was catching a lot of plagiarists, plagiarism became rarer and the remaining plagiarists raised their games (started doing more paraphrasing). But they only raised their games a bit, and not enough to avoid detection as best I can tell.
LikeLiked by 1 person
It’s a hypothetical, and I refer to it as “not impossible”. You’re right, I’m making an assuption about the shape of the distribution of fraud quality – that it’s roughly normal, or at least, not super-strongly skewed. That is, we’re observing the leftmost (lowest quality) “bit” of the fraud quality curve, and I’m assuming (in a thought experiment kind of way) that that leftmost bit isn’t also the mode. Why assume that? Well, how often in real life is the most extreme example of something also the most common kind of that something? (I know, you were expecting some kind of sophisticated reasoning; but nope, that’s all you get.)
LikeLiked by 1 person
Hi Jeremy, I wonder if you’re missing a dimension in your evaluation of the detectability of obvious vs. non-obvious fraud in that it is not just about how convincing the fraud is, but also about the hubris of the fraudster.
I am almost certain that a dedicated fraudster will never be caught as long as he (they seem to be mostly men) is slightly less ambitious in the journals he submits to. For example, would anyone have bothered sifting through Jonathan Pruitt’s data had he published his work in obscure journals? Even if someone spotted an anomaly in paper from a small journal, they would most likely shrug it off as a indicative of poor standards without giving it more thought.
I suspect that the amount of research fraud is way more than we are aware of. I agree with you that it’ll be hard to fool editors at top journals, or colleagues are high-ranking universities for long, but these situations only represent a small minority of global science. In many other universities across the world, getting published anywhere already has career benefits, so a pragmatic fraudster would only have to aim high enough to reap the career benefits, but not so high as to attract scrutiny. This seems to be the strategy for paper mills, for example.
If you haven’t already seen it, you might be interested in this fictionalised piece about paper mills. The Mills have Ayes: https://osf.io/ds6hk/
LikeLiked by 1 person
You’re absolutely right that fraud is more prevalent in low-impact journals, and among scientists based in certain less-wealthy countries. The estimates of the overall prevalence of fraud that I’m aware of do account for that, at least to some extent.
I agree with you that trying to fake your entire career as a famous scientist at a high-ranking university by faking a bunch of high-profile papers is a very risky thing to do. And I think it’s recently become more risky thanks to mandatory data sharing at many leading journals. Hopefully those policies will become even more widespread, and better enforced.
I also agree that it seems like the optimal risk: reward ratio for scientific fraud is to do just enough fakery to appreciably improve your odds of getting a decent, secure job in science, while attracting as little attention as possible. As you note, in some countries that might only require one or a few publications in very low-impact journals. I think this pragmatic level of fraud is harder to pull off (though surely not impossible) in, say, the US or Canada.
LikeLiked by 1 person
I’m surprised just how similarly the opening bits of our posts are structured. The similarity goes beyond what I’d have expected, given what we discussed in our emails. I’m not sure if we’re great minds, but apparently we think alike.
It is striking, isn’t it? And for others reading this: honestly, we didn’t exchange drafts or thoughts AT ALL beyond the broad topic of “why aren’t science frauds like art forgeries”!
Now I’m thinking about other ways I could’ve structured my post. The only one that occurs to me would be to “bury the lede”. Do a long intro about some art forgery. Then say “Scientific frauds work the same way–but it might not seem like it at first glance.” Then talk about an example of scientific fraud. Then compare it to the art forgery and spell out the take-home message.
I think that’d have been a worse post, though I guess we’ll never know.
LikeLiked by 1 person
One contrast between scientific frauds and at least some art forgers: motivation. In reading up a bit on art forgery, I encountered two motivations that some art forgers have, that I don’t think any scientific fraudsters have.
One is a desire to prove that the art matters more than the artist. Prove that people are silly for caring who the artist was, that all that matters is the art itself. Some writers have the same motivation–writing a fake memoir purportedly by someone of a different sex/race/class/etc., to undermine the notion that “authenticity” matters in writing. Science fraudsters don’t (and can’t) have that motivation.
Another motivation of some art forgers is to get revenge on critics who doubted their talent as artists. If someone thinks you’re a bad painter, one way to prove them wrong is by painting paintings indistinguishable from those of a great painter. Again, scientific fraudsters don’t have this motivation.
There may be other motivations that are shared between science fraudsters and art forgers, though. I’m not sure.
That’s a good point. Science fraudsters definitely don’t have motivation #1, and probably not #2 (it’s easier just to produce good real science). But maybe the Sokal hoax would fit in there somewhere? Although that’s more to do with showing the (claimed) low quality of critics’ criticism.
Yes, the Sokal hoax was intended to show that certain science critics have no idea what they’re talking about. Very similar to some art forgers. The forgers think critics don’t actually know good art when they see it, and are instead judging the art merely by asking who the artist was.
I have wondered, half-jokingly (but only half), if one day some serial scientific fraudster is going to respond to getting caught with a “Sokal hoax defense.” Respond to getting caught by saying “You finally got me! I’ve been faking it all these years to expose how bad the peer review system is at distinguishing real science from fake science!” Of course, this hypothetical defense would be much more convincing if the fraudster did as Sokal did–announced the fakery *before* it was discovered by others, rather than after.
I am not a scientist, but I guess that the motivations for scientific fraud are to gain respect. kudos, and grant funding. I would then argue that fraud is ultimately self-defeating because the more respect and kudos an author gets, the more their papers will be read, and the more likely it is for them to be found out.
“In art, substantial resources go into the investigation of forgery, because the costs of forgery are very high: a work could sell for $100,000,000 if judged genuine, but be near-worthless if that judgment is wrong. Do we have these same incentives in science? Arguably not – I’d take the perhaps controversial position that it’s unusual for a fraudulent paper to have very high costs.”
Heh. I said pretty much exactly the same thing in a draft of my post, but cut it because I felt like it interrupted the flow. So, as strikingly similar as our posts were, they were almost even *more* similar!
LikeLiked by 1 person
In my opinion, “fraud” is probably rather common and most of it is likely not obvious because it’s based on non-reporting. You just don’t report data you’ve gathered or methods you’ve used that are inconsistent with the conclusions you want to draw. Since they’re not reported, they can’t be detected. Or you draw conclusions that can’t be supported by the results to the exclusion of other possible explanations and neglect to mention that the other possible explanations exist (if they even occur to you at all). As long as your results and interpretation fit a prevailing narrative, confirmation bias on the part of the editors/reviewers/readers means that other researchers reading the paper are unlikely to examine it critically enough to detect these problems (if it is possible at all). “What you see is all there is”, as they say.
And I think many scientists would not feel guilty about committing this kind of “fraud” (and may not even consider it to *be* fraud), because they can rationalize by arguing backwards from “desired results” to “acceptable methods”. I.e., if your results and explanation fit a prevailing narrative you believe, that means that you must have used the correct methods, excluded the correct outliers, etc. You can argue against this point by saying that scientists are incentivized to publish high-impact results that overturn the status quo, but we all know that any single paper that we publish is unlikely to rise to the level of overturning an entire narrative. And we may be entirely invested in and fully believe that narrative ourselves.
It’s also not difficult to change just a few numbers in your results to get the outcome you want. How would anyone ever know? Your experimental results might not replicate by another lab group, but you can explain that away pretty easily. Especially in a field like ecology where exact replication is often not possible.
And given that a lot of fraud could occur in the direction of forcing one’s results to fit a prevailing narrative (because the researcher thinks that result is more likely to be true and less likely to be scrutinized), I think this means that this kind of fraud will *not* come out in the wash when a systematic review is conducted because the literature inputted to the review has already been homogenized. Thus further strengthening belief in the narrative.
LikeLiked by 1 person
I agree with a lot of this…. but not “It’s also not difficult to change just a few numbers in your results to get the outcome you want. How would anyone ever know? ” The examples we’ve been seeing over the last while suggest that, as I say in my 3rd point, it ISN’T particularly easy to alter or invent data that give you the answer you want, while also being unobvious. Things like how bad humnas are at picking “random” numbers get in the way!
LikeLiked by 1 person
But that’s only true for larger sample sizes, isn’t it? For experiments that tends to have small sample sizes, I think you could probably change those few values to whatever you like and there would be no way of detecting the fraud. Like if you have a 2*2 factorial design with three replicates per cell, I think you only need to edit a maximum of nine values to make the results say anything you want? I’m not sure that could be detected.
LikeLiked by 1 person
“Or you draw conclusions that can’t be supported by the results to the exclusion of other possible explanations and neglect to mention that the other possible explanations exist (if they even occur to you at all)…And I think many scientists would not feel guilty about committing this kind of “fraud” (and may not even consider it to *be* fraud), because they can rationalize”
This illustrates the other big reason why I don’t leap from “the frauds we detect are obvious to anyone who looks” to “there must be tons of careful frauds that we’re not detecting”. Because I think the stuff we’re not detecting isn’t fraud, it’s merely suboptimal research practice.
Further, I think scientists are *right* not to consider suboptimal research practices to be fraud.
Yes, there are shades of grey here. There’s a fuzzy boundary between “fraud” and “suboptimal practices that are not fraud.” But I do think there *is* a boundary; it’s not *all* an undifferentiated schmear of grey. I’m not a fan of attempts to expand the definition of “fraud” to include any and all suboptimal research practices, or of claims that various suboptimal research practices are tantamount to fraud. Is it really “fraud” (or tantamount to fraud) if you don’t preregister your study? If you don’t correct for multiple comparisons? If you spot an apparent pattern in your data and do a null hypothesis test to check whether it’s real? If you test a silly clickbait hypothesis with no good theory behind it? Etc.
One absolutely can and should argue for improvements to scientific research practices, without claiming that every suboptimal practice is fraud, or tantamount to fraud. After all, many people have done so, quite successfully! For instance, the movement towards data sharing and preregistration are not succeeding because of the argument that many/all scientists are fraudsters (or engaging in practices that are tantamount to fraud) who would be stopped by data sharing and preregistration.
Indeed, I worry that trying to reform science by claiming that “everything bad is fraud” is counterproductive. If you claim that most/all scientists are all fraudsters (or are tantamount to fraudsters), because they’re all operating within a corrupt system that can’t be reformed piecemeal, well, it’s only a short step from there to *justifying* fraud. You’re not sounding the alarm, you’re giving people a good reason to ignore the alarm. There’s not much daylight between somebody who says “academic science is all just a corrupt, pointless, careerist game, therefore it needs radical change” and someone who says “academic science is all just a corrupt, pointless, careerist game, therefore why bother following the rules, I might as well cheat to win”. If you keep saying that everyone and everything is broken and unfixable, well, you might not like what happens when people start to believe it. (As an aside, I think this point generalizes. It’s the problem with “doom and gloom” messaging on climate change, for example.)
I think another reason why it goes undetected is because “primary” research in science is so ephemeral. Rare is the beast that will actually read an issue of any particular journal cover to cover, articles get dipped into to cite or during literature research but what makes me laugh about the whole premise of publishing or perishing culture is actual readers of published articles is almost lost from the whole thing and depending on the field, the literature impenetrable unless you’re in that field.
It also speaks to that rich fallacy at the heart of scientific endeavour too (that we all learn at high school) that science should be repeatable. The silly structures and scales of science today means that, aside from some medical fields perhaps, very few are replicating or trying to replicate studies.
So perhaps it’s in part a function of a) Publishing has never been more important to sustaining a career in science b) In most fields papers are of limited interest beyond a tiny tiny group c) Nobody is really going to come calling to follow up you study anyway. d) There’re a lot of papers out there. So, it’s extremely easy to get away with fraud and competitively, a winning strategy. To go back to the art fraud comparator, there’s not so much of a ‘top art historian’ status as an art dealer or handler, there’s a broad network of experts with a vested interest and determining and redetermining original artists of works a core and active field of art history and the field far, far, smaller (basically tens of journals of note, more established boards of experts and foundations associated with many artist’s estates).
There is more attention these days being drawn to fraud and the astonishing ways in which it is done, data manipulation, basic (really basic) image manipulation and likely much more, nigh impossible to detect short of a fundamental change in practice and habitual duplication of studies (this is embedded in some countries and some fields).
LikeLiked by 1 person
How much fakery is required for a study to be classified as fraudulent? I didn’t see it discussed above, but I wonder about whether there’s a quality dimension to the problem (thoroughly discussed in both posts) and a quantity dimension. In addition to the possibility that for some research designs and studies messing with only a few observations may be sufficient to manipulate a result (RMs comment vis. small number of factor with few observations), what about a tweak here or there to an image, a deliberate choice to leave out obvious confounders and then argue them away if they come up in review, or dropping pesky non-conforming observations in small n studies? Maybe these are all just the “bad scientific practices” Jeremy was talking about. But, on the other hand, maybe most “fraud” doesn’t involve a collage cranium, duplicated images across papers, or a completely phony dataset. Instead, most fraud may be small. Essentially, fraud may have multiple dimensions and it’s the join-density that matters for ultimately determining the amount of fraud in science more generally. The selective forces at work (to keep fraud out and to make some fraud tempting) could be acting on both dimensions.
I kind of think that it is just too much effort relative to the reward – As an example, after a while trying to make reasonable synthetic datasets for the stats class, I found that there are actually a lot of fantastic real life datasets available for free.
Pingback: Farewell (sort of) to Dynamic Ecology | Scientist Sees Squirrel