More Fakery
science.org
science.org
The main difference between the excellent PI and the two fraudsters was that the former insisted on everyone in her lab keeping highly accurate and detailed daily lab notebooks, while the other two had incredibly poor lab notebook discipline (and often didn't even keep records!). She actually caught one of her grad students fudging data via this method, before it went to publication. Another requirement was that samples had to be blindly randomized before we analyzed them, so that nobody could manipulate the analytical process to get their desired result.
If you're thinking about going into academia, that's the kind of thing to look out for when visiting prospective PIs. Shoddy record keeping is a huge red flag. Inability to replicate results, and in particular no desire to replicate results, is another warning sign. And yes, a fair number of PIs have made careers out of publishing fraudulent results and never get caught, and they infest the academic system.
In my hard-tech experiment-heavy start-up there's no way we could have made any actual technical progress without setting up a solid data preservation and analysis framework first. For every experimental run, all the original sensor data are collected and immediately uploaded along with any photos, videos, and operator comments to a uniquely-tagged confluence page. Results and data from any further data or product analysis are linked to this original page.
As an anecdotal example, we recently caught swapped dataset labels in results from analysis performed on our physical samples by a third-party lab. We were able to do this easily just because we could refer back to every other piece of information regarding these samples, including the conditions in which they were generated months prior to this analysis. As soon all the data were on display at once, the discrepancies were obvious.
My colleagues and I are trying to address this by creating a platform to discuss and publish such "bad" or "negative" results. More info here:
Obviously, there's a few fields where this might not work (you can't just create a second Large Hadron Collider for validation), but in areas from sociology to organic chemistry to environmental science, I think there's a lot of promise in that method for helping to re-align incentives around producing solid, replicable research.
[1] https://www.worksinprogress.co/issue/escaping-sciences-parad...
"You all got it wrong," he said, "the news is not that Amy Wagers could not make things work with mouse stem cells the way this prior paper said this one time. The news is that Wagers-ize is becoming a verb which means 'to disprove an amazing result after attempting to replicate it'. The lab has Wageres-ed another pluripotent stem cell result. The news is about how often this happens and what it means for this kind of science."
This class was in 2006 and this later profile in 2008 seemed to bear things out the way he said: https://news.harvard.edu/gazette/story/2008/07/amy-wagers-fo...
The former is pretty easy, but only catches errors in the analysis phase (i.e., the data itself could be flawed). The latter is very comprehensive, but you essentially have to double up the effort on re-doing the study---which may not always be possible if you're studying a moving target (e.g., how the original SARS-CoV-2 variant spread through the initial set of hosts).
- Require all research to publish their source code.
- Require all research to publish their raw data minus "PII".
* Note: I use "PII" here with the intention of it taking the most liberal meaning possible, where privacy trumps transparency absolutely and where de-anonymization is impossible. This would rule out a lot of data, and personally I think we could take a more balanced approach, but even this minimalist approach would be a vast improvement on the current situation.
That isn't a replication in any meaningful sense. But a replication can certainly take many forms. An exact replication is one, another could be to do a conceptual replication, so studying the same effect but with a different design, or combining these with a new analysis pooling the data from both the study and the new study with (possibly) improved statistical analysis.
How can replicable research go wrong? Here's just a fraction of the things I've seen reading papers:
1. Logic errors. So many logic errors. Replicating something that doesn't make sense leaves you with two things that don't make sense: a waste of time and money.
2. Tiny effect sizes. Often an effect will "replicate" but with a smaller effect than the one claimed; is this a successful replication or not?
3. Intellectual fraud. Often this works by taking a normal English term and then at the start of your paper giving it an incorrect definition. Again this will replicate just fine but the result is still misinformation.
4. Incoherent concepts. What exactly does R0 mean in epidemiology and precisely how is it determined? You can replicate the calculations that are used but you won't be calculating what you think you are.
5. A lot of research isn't experimental, it's purely observational. You can't go back and re-observe the things being studied, only re-analyze the data they originally collected. Does this count?
6. Incredibly obvious findings. Wealthy parents have more successful children, etc. It'll replicate all right but so what? Why are taxpayers being made to fund this stuff?
7. Fraudulent practices that are nonetheless normalized within a field. The article complains about scientists Photoshopping western blots (a type of artifact produced in biology experiments). That's because editing your data in ways that make it fit your theory is universally understood to be fraud ... except in climatology, where scientists have developed a habit of constantly rewriting the databases that contain historical temperature records. And by "historical" we mean "last year" here, not 1000 years ago. These edits always make global warming more pronounced, and sometimes actually create warming trends where previously there were none (e.g. [1]). Needless to say climatologists don't consider this fraud. It means if you're trying to replicate a claim from climatology, even an apparently factual claim about a certain fixed year, you may run into the problem that it was "true" at the time it was made and may even have been replicated, but is now "false" because the data has been edited since.
Epidemiology has a somewhat similar problem - they don't consider deterministic models to be important, i.e. it may be impossible to get the same numbers out of a model as a paper presents, even if you give it identical inputs, due to race conditions/memory corruption bugs in the code. They do not consider this a problem and will claim it doesn't matter because the model uses a PRNG somewhere, or that they "replicated" the model outputs because they got numbers only 25% different.
What does it even mean to say a claim does or does not replicate, in fields like these?
All this takes place in an environment of near total institutional indifference. Paper replicates? Great. Nobody cares, because they all assumed it would. Paper doesn't replicate, or has methodological errors making replication pointless? Nobody cares about that either.
Your proposal suggests blocking publication until replication is done by independent labs. That won't work, because even if you found some way to actually enforce that (not all grants come from the government!), you'll just end up with lots of papers that can be replicated but are still nonsensical for other reasons.
Even ignoring how, frankly, paternalistic and condescending this sentence comes off as, let's take the rest of your comment on face value.
First, I can agree with you that the parent comment's idea about not allowing publication without replication is a non-starter; it would basically be impossible to implement (in the U.S., you'd immediately run into 1st Amendment issues), and on a practical level, that doesn't match with almost any realistic scenario of how research both is or could be conducted. Anyone that can get published should be welcome to do so.
However, I think that the idea of funding a subset of replication studies would have great value, and your list of "issues" are largely knocking down strawmen. Let's look at this from the point of view of a funding agency, which annually pour billions of dollars into research.
1. Logic errors would be great! A replication report stating "the conclusions of this research are deeply invalid due to x,y, and z logic errors here, here, and here" would make the agency less likely to fund authors that make those errors. If the errors don't invalidate the conclusions, they're still worth noting and can be taken into account.
2. Effect sizes varying in replicated studies is still useful information. Not all "does/doesn't replicate" questions need a binary yes/no answer.
3. Similarly to #1, catching and publishing instances where authors are making invalid claims (even if not doing invalid research) seems like a good thing. "Hey, these researchers have a tendency to claim X when really they're talking about Y. What the heck?" is great information to know for a funding agency to know.
4. See #1 and #3. Generally, the authors of the replication studies are also going to be intelligent; if they're being paid to check for logical fallacies and incoherent concepts (one can imagine a first-year required grad course that lays out these ideas and case studies of how to catch them).
5. I've been thinking about a situations where this still wouldn't produce useful meta information, and I'm having a hard time coming up with them. Imagine a fish survey, where divers are doing fish counts along transects. You can't go re-observe those fish, but a) research groups might want to change their methods to include using cameras instead of marking dive slates, or b) if you are able to re-run their transects due to good record-keeping (yay, replication!) but don't see the same fish, the funding agency now knows that a) the original research group has solid experimental methods and b) maybe they should keep funding groups to run that transect because of the variation, or not fund groups using a single transect data as the basis for forming conclusions. Happy to consider non-outlier (i.e. building a second LHC) examples where this breaks down.
6. Yes, maybe don't fund groups to try replicating these findings. Or maybe do occasionally anyway--if the research groups are doing good work, it should be cheaper to conduct replication-based research than doing a novel one anyway.
7. This is a good argument for replicable research. A huge number of research proposals go unfunded, which is to say that many researchers are somewhat competing for funding. In this paradigm, saying "hey, my group found and thoroughly documented a huge, glaring issue in the methods used in these studies due to a bad-faith corruption of historical data," in a situation where agencies would be looking to pay groups for valid claims. Increasing the number of replication studies adds a greater adversarial capacity to a research system which helps to catch/prevent this kind of fraud in the first place.
This is already a long comment, but I can make similar points about almost all of your claims. In summary: you seem to be making the assumption that replication studies are basically done by robots blindly following an instruction manual of someone else's research. Especially given the funding incentives mentioned above, I don't think that would be true; rather, many errors would be caught (or prevented ahead of time) and lead to better, more solid research being done in the first place.
> (1) (3) (4) you seem to be making the assumption that replication studies are basically done by robots blindly following an instruction manual of someone else's research.
I guess we're using the word replication differently. You seem to be using it to mean a general re-review and re-analysis of everything - basically a funded more aggressive peer review followed by an actual re-running of the study sometimes, if it makes sense, whereas I'm using it to just mean re-running the study exactly as originally described to check the results are the same.
I think in science the term replication normally just means re-running the study exactly as described. It doesn't mean arguing with the authors over their definitions or the logical basis of the study itself. I've seen a bunch of cases where a replication fails and the original study team basically rejects the whole exercise by saying "They didn't follow our instructions so of course they got different results". And I mean, that's kind of fair, right? If you say "I did X and saw Y", and then someone else says it's not true but they didn't actually do X, surely the scientists have every right to be annoyed and reject the exercise? So good replications are exactly what they sound like - a replication of the original process. They aren't generalized adversarial funded peer reviews.
And that's why it's important to highlight that replication is only one aspect. If Congress or whomever goes and earmarks money for replication, then someone with money will eventually ask you to replicate a study with a flawed design. What do you do? You could:
a. Say no: the study is flawed. Replication would be useless because you'd just be repeating a nonsensical procedure. You get no money.
b. Take the money and re-run it. The study conclusions are still wrong but now you got a paper published, and the original is successfully replicated so journalists/professors will use that as a stamp of approval. You may even want the study to replicate because it's professionally or ideologically useful.
c. Try to 'fix' the original design and do an improved study. This just starts the process over from scratch, it doesn't yield evidence.
It'll be (b) or (c) of course, every time. That's just how the incentives are set up and besides, doing another study can be done without involving the original authors. The moment you go down route (a) you're going to get pushback. They won't agree their definitions are illogical. So, it'll devolve into an exchange of letters that nobody ever sees or cares about, and granting agencies won't know how to value it.
> 2. Effect sizes varying in replicated studies is still useful information. Not all "does/doesn't replicate" questions need a binary yes/no answer.
But the bureaucracy needs actionable outcomes, otherwise there's no point. If someone can point at a replication and say "ah yes, we claimed our educational intervention would boost grades by 2x in 15 year olds and governments spent money on that basis, but a replicated 0.1% improvement nonetheless proves us right" then nobody outside the system will consider this a valid replication. There must be actual outcomes from a failure to replicate meaning you'd have to draw the line at how much delta is allowable from the original numbers, but nobody is even having that discussion let alone doing anything about it.
For (5), it was more of a question than anything else. Replication using originally collected data might be useful sometimes but the question is whether the general public would consider this to be a genuine cross-check. My guess is no. Science has no formal mechanisms to detect made up or fraudulent data sets, which is a real problem, so if you allow re-analysis of originally collected data you'll get a steady stream of situations where a study is announced, some skeptics say "uhhh that sounds wrong", the media/academic institutions beat up on them claiming it's peer reviewed and replicated which is gold standard so you've got to believe, and then it turns out the whole original study + replications were all based on fraud. What could be more damaging than that? A replication is a type of audit. These have to take into account the possibility of people playing games for profit.
Yes, I think that this is the crux of the matter here, and I agree that merely funding direct, vanilla replication is not the optimal solution here and would not solve many of the problems we're facing. I'm really talking about the idea from the thread I linked in my original comment--the idea of "red teams" for science. There's probably a number of ways to accomplish this, and direct replication could play a role, but I imagine that a funding agency tasked with funding red teams for other government-funded research would lean towards a more diverse and holistic toolkit. You'd certainly need to have some mechanisms for "watching the watchers," but again, very few researchers I've met are stupid, many of them actually care pretty deeply about ethics and scientific integrity, and would probably jump at the chance for a steady paycheck while helping to advance that.
> Science has no formal mechanisms to detect made up or fraudulent data sets, which is a real problem [...]
This isn't strictly true (I'd like encourage people not to use "Science" as a proper noun; I think it's actually a bit unhelpful). I think a common approach to performing research using the scientific method often (almost always?) involves using observation across multiple instances, then using statistics to draw a generalized conclusions based on the group of observations. And statistics does have formal methods for identifying potentially fraudulent datasets, some of which a Dr. Hilgard does a great job of explaining in an actual case of research fraud he found in this article[1].
[1] https://crystalprisonzone.blogspot.com/2021/01/i-tried-to-re...
Full blown red teams would be great, but seem very hard to set up. I've actually talked to British MPs about this and made some concrete proposals along these lines, but there are several blockers:
1. General apathy / lack of political will. Some of the more savvy MPs know their scientific advisors weren't reliable in COVID and understand the underlying issues, but the vast majority do not and would rather not think about it.
2. The moment you say let's set up a red team, you face the question of how to find high integrity well trained people that have exactly the same idea of what good science is as you do. Particularly problematic: do you have field-specific red teams? If so what stops Ferguson-style problems where the entire field agrees that it's unreasonable to even expect mathematical models to be replicable? You really need outsiders with the highest standards, who won't accept justifications like "We rewrote our dataset with a model because that's just how we do things in our field and who are you to argue with us?".
I think there's a meme about how physicists are always telling other fields they're doing it wrong. Maybe red teams need to be made of physicists :)
3. What exactly is Science, proper noun intended? To red team you need to have a very clear definition of what the scientific method actually is. Who should come up with this? I ended up suggesting a committee of MPs should do it even though that sounds wrong, because ultimately they're the ones authorizing the flow of money, and they're the ultimate outsiders. The moment you start picking insiders it's going to turn into a huge fight (who are these physics nerds to tell us we can't do experiments with 30 undergrads? etc)
"That's academic" is already a mild insult meaning useless or irrelevant, but that perception never had any impact on academia so far. The risk for the academy is that negative feelings grow, and then people start wondering why they're paying for so many studies where they're either shoddy or obvious. The justification for public funding is really only studies that:
a. Yield non obvious conclusions.
b. Correctly.
c. And which wouldn't have been funded by industry.
It's possible that the set of such studies is small, and in some fields there are probably zero such studies (e.g. my bête noire, twitter bot research).
Today the risk to academia is low because the political elites in western countries have all self-selected through credentialism and university based social networking, more or less. But all it takes is a populist candidate to get a big enough base and universities may find themselves in the firing lines, with little in the way of defense. It'd be better to prune the obvious studies now.
I think it depends on what you're investigating, and how much is at stake. I doubt it would be much fun to be put on a corporate hit list.
>The court was told that James Fries, professor of medicine at Stanford University, wrote to the then Merck head Ray Gilmartin in October 2000 to complain about the treatment of some of his researchers who had criticised the drug.
>"Even worse were allegations of Merck damage control by intimidation," he wrote, ... "This has happened to at least eight (clinical) investigators ... I suppose I was mildly threatened myself but I never have spoken or written on these issues."
https://www.cbsnews.com/news/merck-created-hit-list-to-destr...
1. You download a paper and read it. It's got major, obvious problems that look suspiciously like they might be deliberate.
2. You report the problems to the authors. They never reply.
3. You report the problems to the journals. They never reply.
4. You report the problems to the university where those people work. They never reply.
5. Months have passed, you're tired of this and besides by now the same team has published 3 more papers all of which are also flawed. So you start hunting around for people who will reply, and eventually you find some people who run websites where bad science is discussed. They do reply and even publish an article you wrote about what is going wrong in science, but it's the wrong sort of site so nobody who can do anything about the problem is reading it.
6. In addition if you red-teamed the wrong field, you get booted off Twitter for "spreading misinformation" and the press describe you as a right wing science denier. Nobody has ever asked you what your politics are and you're not denying science, you're denying pseudo-science in an effort to make actual science better, but none of that matters.
7. You realize that this is a pointless endeavour. The people you hoped would welcome your "red teaming" are actually committed to defending the institutions regardless of merit, and the people who actually do welcome it are all ideologically persona non grata in the academic world - even inviting them to give a talk risks your cancellation. The End.
An essay that explores this problem from the perspective of psychology reform can be found here:
https://www.psychologytoday.com/us/blog/how-do-you-know/2021...
where it would get complicated is also the same as everything. when the verification effort neither supports nor refutes the original one. many would argue that it means it wasn't done right, but lots of things aren't done right in life.
then there can be the triple replication revolution! so it goes...
I'd also argue that your reduction of this problem sort of misses the point. One of the big problems with the way that studies are done is not that replication efforts aren't conclusive (it's very difficult to prove something doesn't exist), it's that a) non-replicable studies are generally considered as valuable as replicable ones, and b) as a result, it's extremely difficult to replicate many studies to begin with, because there's no incentive to take the time to make it possible. Even if the end result of a replication paper is "we couldn't produce the same results", the people working on it can say "this author's experiments were exceedingly difficult to even try to reproduce," or conversely "we didn't get the same results, but their data collection methods and analysis code were well-documented and accessible." That has a lot of value!
If you tried doing triple replication for every paper, I agree that maybe wouldn't be the best use of resources. But the current state of affairs is so bad that a well-organized drive to create single-attempt replication on a fraction of publicly-funded projects has the potential to be a significant driver of change.
Note that this is optimistic: I'd expect the percentage of publications where a reproducibility study makes sense to be above 50%.
This means that the vast, vary majority of works will never be considered for replication - even with a dedicated replication institute. So for most applicants, the amount of replicated results will be 0.
1) for many people, the idea that science has widespread fraud is just hard to accept; in this respect it is similar to the difficulties that many religious communities have in accepting that their clergy could have a corruption problem
2) the solutions require thinking about problems like p-hacking, incentives, selection effects, and other non-trivial concepts that are tough for the average person to wrap their heads around.
Also: perception is ~effectively reality.
[1] https://arstechnica.com/science/2015/06/study-claims-1-in-4-...
[2] https://www.biorxiv.org/content/biorxiv/early/2016/04/20/049...
Most people don't understand astronomy and physics well enough to prove to themselves that the earth orbits the sun and not vice-versa. Yet they believe it does, with certainty, because they have been taught that it is true.
Most CS work is 90% maths, I don't see how you can have reproducibility issues?
One of those papers is about counting the scale and scope of online political advertising during the 2020 election. How does one reproduce that study? The 2020 election is long past, and that data isn't archived anywhere other than what the researchers have already collected. This is a pretty simple empiracal data collection tastk, but you can't just re-measure that today because that study is about a moving target.
In general, performance comparisons are hard to reproduce. For instance, when benchmarking network protocols, often a tiny change in configuration makes a big change in a results. You might change the size of a buffer from 150 packets to 151 packets and see performance double (or halve.)
Instead of making measurements with some arbitrary choices for parameters, you can take lots of measurements with parameters randomly varied to show a distribution of measurements. It's hard work to track down all the possible parameters and decide on a reasonable range for them, so it's rarely done. I found many 10x variances in network protocol performance (like how fairly competing TCP streams can sharing bandwidth).
The big idea was to show that by randomizing some decisions in the protocol (like discarding packets with some probability as the buffer gets full) you can make the performance less sensitive to small changes. ie, more reproducible. Less sensitivity is especially good when you care about the worst-case performance rather than average. It can also make tuning a protocol much easier, since you aren't constantly being fooled by unstable performance.
Performance sensitivity analysis is hard work, so most papers are just like "we ran our new thing 3 times and got similar numbers so there you go."
For example, if some on the right suggest some difficult but needed reforms, it tends to be spun as an attack on science. Or complaints that trivial projects are being overfunded get misinterpreted by the right and they try to make an example of the wrong studies for the wrong reasons.
The pandemic was a good example of this in my mind, in that I think there were some serious systematic problems in academics and healthcare that were laid brutally bare, and many people suffered or died as a result. But then the whole thing got misidentified and sucked into the political vortex and all you end up with are hearings about how to rehabilitate the CDC, as if that is the problem and not a symptom of even bigger problems.
I still think there are ways for things to change, but the most likely of them involve unnecessary suffering and chaos.
But then how do you do promotions? How do you judge output? Worse still, how does US Congress justify spending taxpayer dollars. Rather than acknowledging that any short term measurement of the quality of science is a fool's errand, we have doubled down on meaningless metrics like impact factors and h-indices. And this is what we have as a result.
When I was working on my PhD, different lab were known to have different rigor. If a new discovery was made, skepticism was the norm. Until reproduced and proven, a new discovery was suspect.
I mean it's clear that there is benefit to stamping out fraud just because it's inefficient, but in the end, the scientific method filters it out.
One solution is to couple this quantitative criteria with independent committees that assess people beyond the metrics, but that requires a lot of human effort and not scalable.
Assessing people in ways that don’t scale seem to be the way to avoid this gaming trap in academia.
I agree. The egregious fraud is just the high-sigma wing. It's a symptom, but the real problem is how it affects the majority.
Google periodically changes the evaluation, in theory to reward good content and penalize bad, but people still try to game the system, rather than improving content.
I doubt there's an easy answer.
Trying to better align the short term objectives with longer term ones could help, but that just makes it harder to game, doesn't eliminate it.
For much of human history, science was something you did in your spare time - or, if you were exceptional, you might have a patron. Then nation states discovered the value of technology and science, and wanted more, and so have created science factories. But, perhaps unsurprisingly, the rate of science production cannot really be improved in this way, and yet the economics of science demand that is does. This disconnect between reality and expectation is the root of this problem, and many others.
Unfortunately some of the effect sizes are so small that it's hard to tell what's working or not. The results of papers on body building, for instance, are definitely put into practice by some people. If the claim of the paper is that eating pumpkin [EDIT] decreases muscle recovery time by 5%, how is an individual who starts eating pumpkin supposed to notice that he's not getting any particular benefit from following its advice? Particularly if he's also following random bits of advice from a dozen other papers, half of which are valid and half of which are not?
To some extent, this is the curse of being the creator of the global reserve currency. The US government can, in theory, print as much money as it wants and pay off whoever it wants to do whatever it wants. This also extends to the academic and financial (VC) sectors, because a lot of that liquidity comes directly from the Government/Fed.
Unfortunately this leads to a culture of corruption (who gets the grants/contracts/funding?) and widespread fraud. This causes the ROI of money printing to go down, the money printer accelerates and we get inflation too.
https://www.youtube.com/playlist?list=PLAB-wWbHL7Vsfl4PoQpNs...
I was a chemistry researcher working on renewables, and during my master's thesis 9 months were spent validating fake results (from a publication of a scientist who worked in our group moreover).