Replace peer review with “peer replication” (2021)
blog.everydayscientist.com
blog.everydayscientist.com
This of course depends a lot on the specific field, but it can easily be months of effort to replicate a paper. You save some time compared to the original as you don't have to repeat the dead ends and you might receive some samples and can skip parts of the preparation that way. But properly replicating a paper will still be a lot of effort, especially when there are any issues and it doesn't work on the first try. Then you have to troubleshoot your experiments and make sure that no mistakes were made. That can add a lot of time to the process.
This is also all work that doesn't benefit the scientists replicating the paper. It only costs them money and time.
If someone cares enough about the work to build on it, they will replicate it anyway. And in that case they have a good incentive to spend the effort. If that works this will indirectly support the original paper even if the following papers don't specifically replicate the original results. Though this part is much more problematic if the following experiments fail, then this will likely remain entirely unpublished. But the solution here unfortunately isn't as simple as just publishing negative results, it take far more work to create a solid negative result than just trying the experiments and abandoning them if they're not promising.
It would mean disruption is no longer a useful tool for human development.
So entering into a paradigm where we test the known space - especially presently - would 1) help reduce cruft; 2) abate undersirable forward progress; 3) train the next generation(s) of scientists to be more diligent and better custodians of the domain.
Some experiments that study biological development or trained animals can take a year or more of fairly intense effort to start generating data.
This whole thread just shows how little the average HNer knows about the academic sciences.
And maybe smaller faculties at R2s pivot to replication hubs. And maybe this is easier for some sections of biology, chemistry and psychology than it is for particle physics. We could start where cost of replication is relatively low and work out the details.
It's completely doable in some cases. (It may never be doable in some areas either.)
Could it be simplified it even further to say x number of papers, but they only count if they’re replicated by others in the field?
I'd argue that since replication is somehow faster than original research, the requirement would count a replication somewhat lower than an original paper (say, at 0.75).
The non-easy part of this is once we start making changes to the criteria for tenure, this opens up people trying to stuff all the solutions for all of the problems that everyone knows already. (See Above.) Would some one try to stuff code-available for CS conference papers, for example? What does it mean for a poster session? At what point are papers released for pre-print? What does it mean for the tenure clock or the Ph.D clock? Does it mean that pre-tenure can't depend on studies that take time to replicate? What do we do with longitudinal studies?
I think you're looking at a 50 year transition where you would have to start simple and iterate.
First, people that want to be professors normally do so because they want to steer their research agenda, not repeat what other people are doing without contribution. Second, who works in their lab? Most of the people doing the leg work in a lab are PhD students, and, to graduate, they need to do something novel to write up in their dissertation. Thus, they can’t just replicate three experiments and get a doctorate. Third, you underestimate how specialized lab groups are — both in terms of the incredibly expensive equipment it is equipped with and the expertise within the lab. Even folks in the same subfield (or even in the same research group!) often don’t have much in common when it comes to interests, experience, and practical skills.
For every lab doing new work, you’d basically need a clone of that lab to replicate their work.
If we're talking about weird incentives and academia you hit on one of the worst ones right here, I think, since nothing there is very closely connected to helping students learn.
I know that's a dead horse, but it's VERY easy to find reasons that we shouldn't be too closely attached to the status quo.
> For every lab doing new work, you’d basically need a clone of that lab to replicate their work.
Hell, that's how startup funding works, or market economies in general. Top-down, non-redundant systems are way more fragile than distributed ecosystems. If you don't have the competition and the complete disconnection, you so much more easily fall into political games of "how do we get this published even if it ain't great" vs "how do we find shit that will survive the competition"
A Ph.D would be expected to perform some replication research as part of their package.
Finally, we would start with the fields that are easy to replicate and move up. We couldn't replicate CERN if we tried. But we could implement this in psychology tomorrow, for example.
If someone wins the Nobel Prize, do the people who replicated their work also win it? When the history books are written do the replicators get equal billing to the people who made the discovery?
When selecting candidates for prestigious positions, are they really going to consider a replicator equal to an original researcher?
This is, of course, a naive proposal without too much thought into it. But I was wondering what I would have missed here.
Though, it seems that it is possible to game the system, by creating positive/negative replication intentionally, to collude with/harm the author.
They also tend to over-estimate the effect of peer review (often equating peer review with validity).
> If someone cares enough about the work to build on it, they will replicate it anyway. And in that case they have a good incentive to spend the effort. If that works this will indirectly support the original paper even if the following papers don't specifically replicate the original results. Though this part is much more problematic if the following experiments fail, then this will likely remain entirely unpublished.
It can also remain unpublished if other things did not work out, even if the results could be replicated. A half-fictional example: a team is working on a revolutionary new material to solve complicated engineering problems. They found a material that was synthesised by someone in the 1980s, published once and never reproduced, which they think could have the specific property they are after. So they synthesise it, and it turns out that the material exists, with the expected structure but not with the property they hoped. They aren’t going to write it up and publish it; they’re just going to scrap it and move on to the next candidate. Different teams might be doing the same thing at the same time, and nobody coming after them will have a clue.
I have gone down the rabbit hole of engineering research before and 90% of the time I’ve managed to find an anecdote or subsequent research footnotes or actual subsequent research publications, that substantially invalidated the lofty claims of the engineers in the 70s or 80s (which is amazing still despite this, a genuine treasure trove of research unused and sometimes useful aerospace engineering research and development) and unfortunately outside the few proper publications, a lot of the invalidations are not properly reverse cited research material and I could have spent a week cross referencing before I spot the link and realise the unnamed work they are saying they are proving wrong is actually some footnotes containing the only published data (before their new paper) on some old work that has a bad scan copy on the NASA NTRS server under some obscure title and no related keywords to the topic the research is notionally about…
Academic research can genuinely suck sometimes… particularly when you want to actually apply it.
A research assistant would quickly be thrown out if he/she refused to record negative experimental results, yet we somehow decide that is fine when operating as a collective.
In my experience, scientists ate comfortably cynical about peer review- even those that serve as reviewers and editors- except maybe junior scientists that haven’t gotten burned yet.
Initially we spent probably a few hours on a paper for peer review because we were relatively unfamiliar with the field but eventually I spent maybe a couple of hours doing the review. Wouldn't say peer review is a joke but it's definitely overrated by the public.
Funny enough, I see exactly the opposite. I've seen this in both reviews I've done and reviews I've received. Just this week I reviewed and saw one of my fellow reviewers write in their justifications: I am not familiar with X, but I am skeptical that the method can scale to a more complex application. Their weaknesses section was extremely generic and it was very clear they didn't understand the work. They gave a weak reject. In fact, when I first started reviewing, I was explicitly told to _only_ accept if I was confident that the work was good. So in my experience, the bias goes the other way that you are proposing.
Btw, I've even seen undergrads acting as reviewers. I was asked to review in in my first year of grad school. I don't think I was qualified then, but I was always a junior reviewer rather than a full so idk.
It's kinda funny. A journal doesn't make the product it sells (the papers that it copywrites). It doesn't pay for the service it performs ("vetting" and editing). And both of these would be done regardless of their existence. I can understand distribution, but that hasn't been useful for over a decade now. What even do these things do anymore?
(btw, I've seen profs delegate to undergrads. And it is quite common for post-docs AND grad students to be reviewers. Trust me, I am one)
Networking, mostly, in the sense that an article in a high impact journal has a higher probability to be integrated in citations networks. The fact that there is some gate keeping means that it’s valuable to be in rather than out, and that’s something you can use to get a position. Also, better journals (which are not necessarily the highest-impact ones) tend to have more thorough peer review (such as 3 reviewers by default instead of 1 or 2, editors who are not afraid to ask for more reviews if the 3 are not conclusive, etc).
> (btw, I've seen profs delegate to undergrads. And it is quite common for post-docs AND grad students to be reviewers. Trust me, I am one)
I am lucky not to have been in that situation when I was a student, and I did not delegate any further when I got the occasional review from the prof when I was a post-doc. But I am unfortunately not surprised.
What field are you in where post-docs aren't getting calls to review directly from the conference? I'm in ML and it isn't my advisor assigning me reviews, it is the conference.
One simple example, (from memory) the Bard paper doesn't include results for experiments in which GPT-4 outperforms it. As a result, people come away from these works with an inflated understanding of their capabilities. This wouldn't pass peer review.
High Sales = A large number of people can attest to the engineered good or service being of high enough quality that they will exchange hard earned money for the ability to use it.
Peer Review = Some folks who derive self-worth from citations ask you to add a citation to their work and you do it because you'll probably need to ask them for the same some time later.
Market success isn't the same thing as validity.
Look at it like this: a Peer Reviewed (tm) article comes out saying "Foos cannot Bar, it is impossible". The same day, Apple releases "Bar for Foos, $8/month". Over the next year you see media outlets discussing how well Foo Barring works. Online reviewers talk about how they've incorporated Barring their Foos into everyday life and it has benefitted them in all these great ways. Your colleagues at work mention their fruitful Foo Barring adventure over the past weekend. Routine posts on HN come up where hackers describe how they've incorporated the Foo Barring API's into their own products in some novel way.
Your mother then calls you up to ask if she should get involved with this new Foo Barring thing. What do you say, can Foos Bar?
Grifters don't win in the long-term.
Plenty of religions suggest otherwise.
(Dear reader: I'm not referring to your religion - I'm referring to the other religions.)
> Butt-candling[1] must work, just look all these happy customers!
But history is replete with ineffective or downright harmful treatments being popular long after the evidence showed them to be ineffective or harmful. Homeopathy is a prime example of this, seeing as those concoctions contain either no active ingredients or (in cases of low dilutions still labeled "homeopathic") contain ingredients picked based on notions of sympathetic magic ("like cures like").
[1] A hopefully fictional example.
It's only in academia that they must instead work in a sterile environment that some person who has never stepped foot off a school campus thinks is "more authentic" for them to be seen as legitimate.
The point I make is that peer review can not be guaranteed to 'fix' science in any way we might like. The Sokal Affair [1] has now been replicated repeatedly, including in peer reviewed journals. The most recent one even got quite cheeky and published it under the names "Sage Owens, Kal Avers-Lynde III" - Sokal III. [2] It always preys on the same weakness - bias confirmation.
[1] - https://en.wikipedia.org/wiki/Sokal_affair
[2] - https://www.nationalreview.com/news/academic-journal-publish...
You'd have no idea if you were going down a well trodden path which would yield no success because you have no idea it was well trod. No one publishes negative results, etc.
Well, the trouble is that hasn't been the case in practice. A lot of the replication crisis was attempting for the first time to replicate a foundational paper that dozens of other papers took as true and built on top of, and then seeing said foundational paper fail to replicate. The incentives point toward doing new research instead of replication, and that needs to change.
In something like psychology that is likely harder, since the experiment you want to do might be related to but differ significantly from the prior work. I am no psychologist, but I’d like to think that they don’t take one study as ground truth for that reason but try to understand causal mechanisms with multiple studies as data points. If the hypothesis is correct, it will likely present in multiple ways.
The alternative is a bunch of stuff being published which people belief as "science" that doesn't hold up under scrutiny, which undermines the reliability of science itself. The current approach simply gives people reason to be skeptical.
the concern about skepticism is not irrelevant, but many of these skeptics also are skeptical of the earth being round, or older than a few thousand years, or not created by an omnipotent skylord, and I'm not sure it's actually a significant concern given the current number and expertise of those who are skeptical
so, we can hear their arguments for their skepticism, but that doesn't mean the arguments are valid to warrant the skepticism exhibited. And in the end, that's what matters: skepticism warranted by valid arguments, not just any Cletus McCletus's skepticism of heliocentrism, as if his opinion is equal to that of an astrophysicist (it isn't). And you know what? It isn't necessary to convince a ditch digger that the earth goes around the sun, if they feel like arguing about it.
I’m skeptical that these people truly exist outside of the internet wanting it to be true.
I'm not, just like I'm not skeptical that climate change denialists truly exist outside of the internet
there's simply no valid argument to warrant skepticism of either, given the ease of locating evidence that both do
I can walk outside and prove gravity.
I can find a million pictures of the earth taken from space, look at a globe and view trade routes that circumnavigate it. I can also look to the sky and see the sun and moon are clearly circular which makes a pretty good case for a pattern.
Climate change or the age of the earth are based on a whole lot more interconnected bits of science that even if you studied your entire life you could not truly say that you understand. You’re putting your trust in layers of science that add up to a certain conclusion (which is good). When people are given good reasons to believe that science and peer reviews aren’t always legitimate it undermines that process of trust building on trust.
Modern aviation is layer upon layer of science building on each other, but I can easily watch a plane takeoff to validate all of those processes.
That’s it in a nutshell. If you can easily replicate it, it’s easy to trust. If you can’t, it’s not…especially when it’s used to drive politics.
recall the tangent we deflected to was the mere existence of round earth deniers, climate change deniers (read: people), both of which do, in fact, exist as real people in the real world, as can be easily confirmed by anyone researching the topic in good faith, like I said
anyhow, back on topic: the proposed alternative just doesn't seem better than the status quo, no matter how you've sliced it so far, for the reasons given in my original post and ignored by you
I wasn’t proving either, just validating skepticism
anyhow, back on topic: the proposed alternative just doesn't seem better than the status quo, no matter how you've sliced it so far, for the reasons given in my original post and ignored by you
There’s not a lot left to say if you’re going to ignore that.
maybe, maybe not. Problem is, that's totally, absolutely, and completely irrelevant to your tangent, which was: you doubted the existence of round earth deniers in real life
(when they do, in fact, exist, just like climate change deniers, no matter how ridiculous either of them or their beliefs are)
see, you neglect that some people are also skeptical of things which CAN be verified quickly, for many varied reasons which we'll for expedience summarize here as "the dumb", and many of these round-earth-denying, perhaps climate-change-denying people are in the "we're losing faith in science!!!1" crowd
which brings us back on topic: the proposed alternative just doesn't seem better than the status quo, no matter how you've sliced it so far, for the reasons given in my original post and ignored by you
It's not just "can we replicate the analysis on sample X", but also "can we collect a sample similar to X and do we observe similar things in the vicinity" in many cases. That alone may require multiple seasons of rather expensive fieldwork.
Then you have tens to hundreds of thousands of dollars in instrument time to pay to run various analysis which are needed in parallel with the field observations.
It's rarely the simple data analysis that's flawed and far more frequently subtle issues with everything else.
In most cases, rather than try to replicate, it's best to test something slightly different to build confidence in a given hypothesis about what's going on overall. That merits a separate paper and also serves a similar purpose.
E.g. don't test "can we observe the same thing at the same place?", and instead test "can we observe something similar/analogous at a different place / under different conditions?". That's the basis of a lot of replication work in geosciences. It's not considered replication, as it's a completely independent body of work, but it serves a similar purpose (and unlike replication studies, it's actually publishable).
Either "peer reviewed" articles describe progress of promising results, or they don't. If they don't the research is effectively ignored (at least until someone finds it promising). So let's consider specifically output that described promising results.
After "peer review" any apparently promising results prompt other groups to build on them by utilizing it as a step or building block.
It can take many failed attempts by independent groups before anyone dares publish the absence of the proclaimed observations, since they may try it over multiple times thinking they must have botched it somewhere.
On paper it sounds more expensive to require independent replication, but only because the costs of replication attempts are hidden until its typically rather late.
Is it really more expensive if the replication attempts are in some sense mandatory?
Or is it perhaps more expensive to pretend science has found a one-shot "peer reviewed" method, resulting in uncoordinated independent reproduction attempts that may go unannounced before, or even after failed replications?
The pseudo-final word, end of line?
What about the "in some sense mandatory" replication? Perhaps roll provable dice for each article, and in-domain sortition to randomly assign replicators. So every scientist would be spending a certain fraction of their time replicating the research of others. The types of acceptable excuses to derelict these duties should be scrutinized and controlled. But some excuses should be very valid, for example conscientious objection. If you are tasked to reproduce some of Dr. Mengele's works, you can cop out on condition that you thoroughly motivate your ethical concerns and objections. This could also bring a lot of healthy criticism to a lot of practices, which is otherwise just ignored an glossed over for fear of future career opportunities.
If the cited work is unable to be replicated, and you try to replicate but get different results, then you would be doing original research, and then you can base further work on your initial original study that came to a different result.
On the flip side, if you are able to replicate it, then you are doing extra work initially, but after replicating the work you've cited, the work you've done is more likely to be reproducible by someone else.
The amount of citations needed to require replication could itself be a function of how easy it is to replicate work across an entire field.
A field where there's a high rate of success in replicating work could have a higher threshold for requiring replication compared to a field where it's difficult to replicate work.
Also your mentality is exactly part of the problem: you arrogantly believe that replication work is beneath you and that originality is all that matters.
High-impact work that people care about gets built on, and replicated. Low impact work does not.
My results can be confirmed by anyone who wishes to do so, because I always publish my code.
you arrogantly believe that replication work is beneath you
I’ve done plenty of replication work when it was needed for my research. Replicating other people’s work however is not my responsibility and it is not usually required to perform original research.
so you want to do science.... which requires there to be replication.... but somehow it's not your responsibility to replicate anything? last I checked the scientific method has to be followed through all the steps to be valid, not just the ones that you personally like
1. I don't trust the results
2. I want to understand their work in sufficient detail
Usually these reasons apply when I'm building directly on top of someone's work. This is not always the case. Often my research is based on ideas which could be tested by generating my own results. As long as I can replicate my own results I don't see a problem.
replicating your own results is not science though, it's anectodes
This could help your career by increasing the probability that the work you're citing is more likely accurate, and as a result, your work is also likely more accurate.
Does it really deserve to be called work if it doesn't include the a full, working set of instructions that if followed to a T allow it to be replicated? To me that's more like pollution, making it someone else's problem. I certainly don't see how "we did this, just trust us" can even be considered science, and that's not because I don't understand the scientific method, that's because I don't make a living with it, and have no incentive to not rock the boat.
I think it's fascinating that we can at the same time hold things like "one is none" to be true, or that you should write tests first, but with science we already got so used to a lack of discipline that we just declare it fine.
It's not hard to not climb a tower you can't get down from. It's the default, actually. You start with something small where you can describe everything that goes into replicating it. Then you replicate it yourself, based on your own instructions. Before that, you don't bother anyone else with it. Once that is done, and others can replicate as well, it "actually exists".
And if that means the majority of stuff has to be thrown out, I'd suggest doing that sooner rather than later, instead of just accumulating scientific debt.
Meanwhile, Alice has spent the last 5 years making the best possible use of the flawed theory, and published a lot of original research. Sure, many of her publications are rubbish, but a few contain interesting results. Contrary to Bob, Alice can show actual results and has publications.
Who do you believe will remain in academia? And, according to public perception, will seem more like an actual scientist?
To put a point on it, if public distrust of science becomes big enough, it all can go away before you can say "cultural revolution" or "fascist strongman". Then there'd be no more academia, and its shell would be inhabited by party members, so to speak. I'd gladly sacrifice the ability of Alice and others like her to live off producing "mostly rubbish" to at least have a chance to save science itself.
Academic science isn’t just the doing science part but the articulation and presentation of your work to the broader community. If Bob knows this space so well, he should be able to clearly communicate the issue and, ideally, present an easily understandable counter example to the existing theory.
Technical folks undervalue presentation when writing articles and presenting at conferences. The burden of proof is on the presenter, and, unless there’s some incredible demonstration at the end, most researchers won’t have the time or attention to slog through your mess of a paper to decipher it. There’s only so much time in the day and too many papers to read.
In my experience, the best researchers are also the best presenters. I’ve been to great talks out of my domain that I left feeling like I understood the importance of their work despite not understanding the details. I’ve also seen many talks in my field that I thought were awful because the presentation was convoluted or they didn’t motivate the importance of their problem / why their work addressed it
That is not to say that scientist should not document the process to their best ability so it can be reproduced in principle. I'm just arguing that it is impossible to easily reproduce other people's results. Again when chemical/manufacturing companies open another location they often spend months to years to make the process work in the new factory.
We're not talking about 1 of 10 reproduction attempts failing, we're talking about 100%. And no, companies don't time and time again try to reproduce something that has never been reproduced and fail, to then try again, endlessly. That's just not a thing.
> it is impossible to easily reproduce other people's results
We're also not talking about "easily" reproducing something, but at all. And in principle doesn't cut it, it needs to be reproduced in practice.
I agree that reproduction in scientific work is important, but it is also apparently impossible in the best possible circumstances. When dealing with physical materials, inexact measurements, margins of error, etc, I think we have to accept that there is no set of instructions that, if followed to a T, will ever ensure perfect replication.
But this is the other way around. Have you ever written a program that doesn't run anywhere except a single machine of yours? Would you release it and advertise it and encourage other people to use it as dependency in their software?
If it only runs on one machine of yours, you don't even know if your code is doing something, or something else in the machine/OS. Or in terms of science, whether the research says something about the world, or just about the research setup.
And yes, there are often overly excited press releases that accompany it - the "advertise it and encourage others to us it as a dependency" part of it analogy - but this is typically just noise in the context of scientific research. If that is your main problem with scientific publishing, you may want to be more critical of science journalism instead.
Fwiw, yes of course I've written code that only runs on my machine. I imagine everyone has, typically accidentally. You do it, you realise your mistake, you learn something from it. Which is exactly what we expect from scientific papers that can't be reproduced.
I disagree. It shows that when someone writes something in a text editor and publishes it, others can read the words they wrote. That's all it shows, by itself. Just like someone writing something on the web only tells us that a textarea accepts just about any input.
And even if it did show more than that, when someone "explores" it, is the result is more of that, something that might be true, might not be, but "is worth exploring"? Then at what point does falsifiability enter into it? Why not right away? To me it's just another variation of making it someone else's problem, kicking the can down the road.
> if it is exciting enough, and lots of eyes end up looking, then flaws are typically found fairly quickly.
If that was true, there wouldn't even be a replication issue, much less a replication crisis. It's like saying open source means a lot of people look at the code, if it's important enough. Time and time again that's proven wrong, e.g. https://www.zdnet.com/article/open-source-software-security-...
> yes of course I've written code that only runs on my machine. I imagine everyone has
I wouldn't even know how to go about doing that. Can you post something that only runs on one of your machines, and you don't know why? Note I didn't say your machine, I said one machine of yours. Would you publish something that runs on one machine of yours but not a single other one, other than to ask "can anyone tell me why this only runs on this machine"? I doubt it.
This thread discusses peer replication, this is not even an analogy.
The way you're describing a scientific publication is as if it were the end result of the scientific act. To use the software analogy, you're describing publication like a software release: all tests have been performed, all CI workflows have passed, QA have checked everything, and the result is about to be shipped to customers.
But talking to researchers, they see publishing more like making a new branch in a repository. There is no expectation that the code in that branch already be perfect (hence why it might only run on one machine, or not even run at all, because sometimes even something that doesn't work is still worth committing and exploring later).
And just like in software, where you might eventually merge those branches and create a release out of it, in the scientific world you have metastudies or other forms of analysis and literature reviews that attempt to glean a consensus out of what has been published so far. And typically in the scientific world, this is what happens. However, in journalism, this isn't usually what happens, and one person's experimental, "I've only tested this on my machine" research is often treated as equivalent to another person's "release branch" paper evaluating the state of a field and identifying which findings are likely to represent real, universal truths.
Which isn't to say that journalists are the only ones at fault here - universities that evaluate researchers primarily on getting papers into journals, and prestige systems that make it hard to go against conventional wisdom in the field both cause similar problems by conflating different levels of research or adding competing incentives to researchers' work. But I don't think that invalidates the basic idea of published research: to present a found result (or non-really), provide as much information as possible about how to replicate the result again, and then let other people use that information to inform their work. It just requires us to be mindful of how we let that research inform us.
Well some do, others don't. Like the one who wrote the article this is a discussion of.
https://en.wikipedia.org/wiki/Replication_crisis
> Replication is one of the central issues in any empirical science. To confirm results or hypotheses by a repetition procedure is at the basis of any scientific conception. A replication experiment to demonstrate that the same findings can be obtained in any other place by any other researcher is conceived as an operationalization of objectivity. It is the proof that the experiment reflects knowledge that can be separated from the specific circumstances (such as time, place, or persons) under which it was gained.
Or, in short, "one is none". One might turn into more than one, it might not. Until it does, it's not real.
more snippets from the above WP article:
> This experiment was part of a series of three studies that had been widely cited throughout the years, was regularly taught in university courses
> what the community found particularly upsetting was that many of the flawed procedures and statistical tools used in Bem’s studies were part of common research practice in psychology.
> alarmingly low replication rates (11-20%) of landmark findings in preclinical oncological research
> A 2019 study in Scientific Data estimated with 95% confidence that of 1,989 articles on water resources and management published in 2017, study results might be reproduced for only 0.6% to 6.8%, even if each of these articles were to provide sufficient information that allowed for replication
I'm not saying it couldn't be fine to just publish things because they "could be interesting". But the overall situation seems like quite the dumpster fire to me. As does software, FWIW.
I think it would be fine to half the productivity of these fields, if it means that you can reasonably expect papers to be accurate.
Peer replication is completely unfeasible in experimental fields of science. The current process of peer review is alright, people just need to learn that single papers standing by themselves don't mean too much. The "peer replication" happens over time anyway when others use the same tools, samples, techniques on related problems and find results in agreement with earlier papers.
Implementing the code for the simulation and analysis of the data? four months, at most. Running the simulation? almost three years until I had data with good enough resolution for publishing.
For a reductive example, the idea to solve P vs NP is a great one, but I’m not going to do that any time soon!
Even if someone meticulously documents their process, it could still take months to replicate the results.
I'm familiar with lithography/nanofabrication and I know that it is typically the case that a process developed in one clean-room can not be directly applied to a different clean room and instead one has to develop a new process based on what the other results.
Even in the same lab it can often happen that if you come back to a process after a longer time, that things don't work out anymore and quite a bit of troubleshooting ensues (maybe a supplier for some chemical changed and even though it should be the same formula it behaves slightly different).
I previously worked in agricultural research (in the private sector), and we spent YEARS trying to replicate some published research from overseas. And that was research that had previously been successfully replicated, and we even flew in the original scientists and borrowed a number of their PhD students for several months, year after year, to help us try to make it work.
We never did get it to fully replicate in our country. We ended up having to make some pretty extreme changes to the research to get similar (albeit less reliable) results here.
We never did figure out why it worked in one part of the world but not another, since we controlled for every other factor we could think of (including literally importing the original team's lab supplies at great expense, just in case there was some trace contaminant on locally sourced materials).
Doesn't that indicate further research is needed? It sounds fascinating to me. (I know it isn't interesting for the people who couldn't get it working.) It also might indicate that the original research was incomplete in the sense that it might be a fluke due to specific conditions in the original country which isn't universal.
You are almost stressing all the ways we are producing garbage rendered non-reproducible with deficient documentation of processes, changes in supply, and changes in the environment. All three can be minimized through peer replication.
Reproduction is hard, really really fucking hard. Just saying, that means we should replicate before trying to understand, means essentially cutting off the understanding.
And like others have said, if someone wants to build from it they'll depend on that information being correct, if no one can manage to ever build from it then the idea dies.
Also there's a huge difference between, replicate this study on infant response to stimulus, or spider colony behavior, and, replicate this incredibly intricate semiconductor that took years of configuration to correctly produce.
Currently, a significant proportion of research results in various fields cannot be reproduced. This essentially means that a lot of work turns out to be flawed, leading to wasted efforts (you can refer to the 'reproducibility crisis' for more context). Moreover, future research often builds upon this erroneous information, wasting even more resources. As a result, academic journals get cluttered with substandard work, making them increasingly difficult to monitor and comprehend. Additionally, the overall quality of written communication deteriorates as emphasis shifts from the accurate transfer and reproduction of knowledge to the inflated portrayal of novelty.
Now consider a scenario where 50% of all research is dedicated to reproduction. Although this may seem to decelerate progress in the short term, it ensures a more consistent and reliable advancement in the long term. The quality of writing would likely improve to facilitate replication. Furthermore, research methodology would be disseminated more quickly, enhancing overall research effectiveness.
Wouldn't this imply that worthwile results are already being replicated, so the primary cost has been paid and we just need to give some method to disseminate this work and then have it factor into the credibility of science? pre-print < peer reviewed < peer replicated, with the last step having internal rankings depending upon how much it has been replicated?
And it'll also show when someone is building on things but not replicating it, which I guess is an issue in some fields more than others.
This seems particularly problematic because it is already notoriously hard to get tenure and academia is already notoriously unrewarding to researchers who don't have tenure.
NSF grants distribute 8.5 billion dollars a year, which is less than Major League Baseball (and its Congressionally granted monopoly) makes. The US Congress has directed 75 billion dollars in aid to Ukraine to date.
There is plenty of science out there which financially, practically, or ethically simply by definition cannot be replicated. That doesn't mean their results should not be published. If peer review shows that their methods and analysis are sound, there is no reason to doubt the results.
> Where are you going to get the budget to build a second LHC solely for replication?
In cases like this you could simply have a second, independent team time-sharing the LHC and using it to replicate experiments run by the first team. (And vice- versa). It’s not a perfect replication but it’s probably still an improvement over the “just trust me bro” status quo.
> How are you going to replicate a long-term medical cohort study which has been running for thirty years?
Run two independent studies in parallel from the beginning.
> What about a paper describing a one-off astronomical event, like the "Wow!" signal?
There was a ton of effort invested into trying to replicate that observation! Since nobody else ever managed to do so, we can’t draw any conclusions from it.
> What if you research the long-term impact of high-dose radiation exposure during Chernobyl?
That doesn’t preclude replication unless, for some reason, you’re the only researcher researching the long-term impact of high-dose radiation exposure during Chernobyl.
That's not a true replication, and it isn't going to avoid issues like the OPERA experiment measuring neutrinos going faster than the speed of light due to a loose connector. It would not be any different from having the second team just run their own analysis on the data from the first team - at which point the second team can just as well simply validate the first team's analysis like peer review is currently doing.
> Run two independent studies in parallel from the beginning.
So all currently-running long-running research has to be thrown out? What if the two studies find very small differences, are you allowed to publish either of them? Are the two teams allowed to collaborate at all?
> There was a ton of effort invested into trying to replicate that observation! Since nobody else ever managed to do so, we can’t draw any conclusions from it.
You can't "replicate" an observation of a freak astronomical event because you can't trigger a freak astronomical event. At best you can do observations and hope it happens again. We indeed cannot draw any conclusions from it, but that doesn't mean you can't publish papers about it. If replication is mandatory, you would not be allowed to do anything with it at all.
> That doesn’t preclude replication unless, for some reason, you’re the only researcher researching the long-term impact of high-dose radiation exposure during Chernobyl.
It cannot be reproduced because it would be unethical to expose people to near-fatal levels of radiation simply for reproduction. Simply reusing data from the original test subjects isn't a reproduction, after all.
As you already quoted me as saying, it's not a perfect replication. Which is fine! I'm advocating a position of "replicate findings as much as reasonably possible", not a position of "we need to build redundant copies of every multi-billion-dollar research megaproject". My whole point is that this doesn't need to be an absolutist true-or-false sort of thing.
> and it isn't going to avoid issues like the OPERA experiment measuring neutrinos going faster than the speed of light due to a loose connector
Maybe not. I never claimed this would solve every problem in all of science forever.
> So all currently-running long-running research has to be thrown out?
No. I think it's reasonable to propose more rigorous standards for future studies without throwing out every in-progress study that didn't follow those same standards. After all, there are literally centuries of published science that didn't even follow the contemporary standards of peer review, and we haven't thrown any of that out.
> What if the two studies find very small differences, are you allowed to publish either of them?
That's an extremely broad question. You might as well ask, "what does it mean for a finding to be replicated?".
If you and I each independently go out to measure the length of the Golden Gate Bridge in millimeters, there are likely to be very small differences in the result you get and the result I get. There's an expected margin of error here where we can agree that our results are consistent with each other. Sometimes the differences are reasonable and can be explained, and sometimes they can't be explained.
Regardless, I also think it might even be valuable for the studies to be published in some form even if they don't replicate at all; just not necessarily with the imprimatur of some credible or prestigious journal.
> Are the two teams allowed to collaborate at all?
I would suggest keeping the two teams independent at least until they both have results. Afterwards, it might be valuable for the team to collaborate in terms of trying to reconcile their results.
> You can't "replicate" an observation of a freak astronomical event because you can't trigger a freak astronomical event. At best you can do observations and hope it happens again. We indeed cannot draw any conclusions from it, but that doesn't mean you can't publish papers about it. If replication is mandatory, you would not be allowed to do anything with it at all.
Yeah, I guess I'm fine publishing a paper that just says "here's this anomalous observation we had" if it's an especially interesting anomalous observation like that.
> It cannot be reproduced because it would be unethical to expose people to near-fatal levels of radiation simply for reproduction.
Obviously. I think you know that's not what I'm suggesting at all here.
> Simply reusing data from the original test subjects isn't a reproduction, after all.
At this point, I agree we're mostly stuck with whatever data we managed to get 37 years ago. But if a similar incident happened in the future, you could have independent teams collecting redundant sets of data.
Two. This drain of resources can't be done for free. Somebody will need to pay twice for half of the research [1], and faster. Peers will need to be hired and paid, maybe by the writer's grants. Researchers cant justify to give their own funds to other teams without a profound change in regulation and even in that case would be harming their own projects.
[1] as the valuable experts are now stuck validating things instead doing their own job
Would open also a door for foul play. Blocking competitors teams in molasses just trowing them secondary silly problems that they know that are a dead end, while the other team work in the real deal, and take the advantage to win the patent.
I don't see how the current system works really either. Fraud is rampant, and replication crisis is the most common state of most fields.
Basically the current system is failing at finding out what is true. Which is the entire point. That's pretty damn bad.
AFAICT meta-studies is the level where we as a society really can try to say something intelligent about how stuff works. If an important question is not included in a meta-study, we (i.e. universities and research labs) probably need to do more research on that topic before we really can say that much about it.
But publication is not a closed system. The "published, peer-reviewed paper" is frequently an artifact used to decide practical policy matters in many institutions both public and private. To the extent that Science (as an institution in its own right) wants to influence policy, that influence needs to be grounded in reproducible results.
Also, I would not be surprised if stronger emphasis on reproducibility improved the quality of conversation among scientists.
I never seen papers as "truth", but more as "possibilities". After many other "proofs" (products, papers, demos, etc.) you can assign some concepts/ideas the label "truth" but one/two papers from the same group is definitely not enough.
Implicit in this process is that the authors are acting in good faith. To treat the authors as hostile is both demoralizing for the reviewers (who wants to be that cynical about their field) and would require extensive verification of each statement well beyond what is required to return the review in a timely manner.
Unless your paper has mathematical theory (and mistakes do slip through), a publication should not be taken as proof of something on its own, but a data point. Over time and with enough data points, a field builds evidence to turn a hypothesis into a scientific theory.
This way papers are peer replicated in an emerging manner because the knowledge is passed from one group to another and they use parts of that knowledge to then apply it to their own research. You have to see this from a more holistic picture. Individual papers don't mean too much, it's their overlap that generates scientific consesus.
In contrast, requiring some random reviewer to instead replicate my full paper would be an impossible task. He/she would not have the required equipment (because there's only 2 lab setups in the whole world with the necessary equipment), he/she would probably not have the required knowledge (because mine and his research only partially overlap - e.g. we're researching the same materials but I use angle-resolved photoemission experiments and he's doing electronic transport) and he/she would need to spend weeks first adapting the growth recipee to the point where his sample quality is the same as mine.
One piece of evidence comes from software like GRIM and SPRITE. GRIM was run over psychology papers and found around 50% had impossible means in them (that could not be arrived at by any combination of allowed inputs) [1]. The authors generally did not cooperate to help uncover the sources of the problems.
Yet another comes from estimates by editors of well known journals. For example Richard Horton at the Lancet is no stranger to fraud, having published and promoted the Surgisphere paper. He estimates that maybe 50% of medical papers are making untrue claims, which is interesting in that this intuition matches the number obtained in a different field by a more rigorous method. The former editor of the New England Journal of Medicine stated that it was "no longer possible to believe much of the medical research that is published".
50%+ is a number that crops up frequently in medicine. The famous Ioannidis paper, "Why most published research findings are false" (2005) has been cited over 12,000 times.
Marc Andreessen has said in an interview that he talked to the head of a very large government grant agency, and asked him whether it could really be true that half of all biomedical research claims were fake? The guy laughed and said no it's not true, it's more like 90%. [2]
Elizabeth Bik uncovers a lot of fraud. Her work is behind the recent resignation of the head of Stanford University for example. Years ago she said, "Science has a huge problem: 100s (1000s?) of science papers with obvious photoshops that have been reported, but that are all swept under the proverbial rug, with no action or only an author-friendly correction … There are dozens of examples where journals rather accept a clean (better photoshopped?) figure redo than asking the authors for a thorough explanation." In reality there seem to be far more than mere thousands, as there are companies that specialize in professionally producing fake scientific papers, and whole markets where they are bought and sold.
So you have people who are running the scientific system saying, on the record, that they think science is overrun with fake results. And there is some quantitive data to support this. And it seems to happen quite often now that presidents of entire universities are being caught having engaged in or having signed off on rule breaking behavior, like image manipulation or plagiarism, implying that this behavior is at least rewarded or possibly just very common.
There are also whole fields in which the underlying premises are known to be false so arguably that's also pretty deceptive (e.g. "bot studies"). If you include those then it's quite likely indeed that most published research is simply untrue.
[1] https://peerj.com/preprints/2064v1/
[2] https://www.richardhanania.com/p/flying-x-wings-into-the-dea...
Organic Syntheses "A unique feature of the review process is that all of the data and experiments reported in an article must be successfully repeated in the laboratory of a member of the editorial board as a check for reproducibility prior to publication"
"If you can't reproduce a procedure in Org Syn, it's YOUR fault" - my PhD supervisor
99% of all papers mean nothing. They add nothing to the collective knowledge of humanity. In my field of robotics there are SOOO many papers that are basically taking three or four established algorithms/machine learning models, and applying them to off-the-shelf hardware. The kind of thing any person educated in the field could almost guess the results exactly. Hundreds of such iterations for any reasonably popular problems space (prosthetics, drones for wildfires, museum guide robot) etc every month. Far more than could possibly be useful to anyone.
There should probably be some sort of separate process for things that actually claim to make important discoveries. I don't know what or how that should work. In all honesty maybe there should just be less papers, however that could be achieved.
A lot of papers are done as a part of the process of getting a degree or keeping or getting job. The value is mostly the candidate showing they have the acumen to produce a paper of such quality that meets the publisher and peer review requirements. In some cases, it is to show a future employer some level of accomplishment or renown. The knowledge for humanity is mostly the authors ability to get published.
I understand that part of the reason for that is that people need to appear as though they are part of the "actually trying" crowd to get the desired job effects. But it is nonetheless a problem, and a large one very worth at least trying to solve.
There's another less obvious problem: we don't always know what is groundbreaking and what is not until something is published and is accepted at large as being a really big deal.
This is a direct result of the aggressive "publish or perish" system. I worked as an aide in an autonomous vehicles lab for a year and a half during my undergrad, and while the actual work we were doing was really cool cutting edge stuff, it was absolutely maddening the amount of time we wasted blatantly pulling bullshit nothing papers exactly like you describe out of our asses to satisfy the constant chewing out we got that "your lab has only published X papers this month".
> There should probably be some sort of separate process for things that actually claim to make important discoveries.
This used to be Springer Nature and the likes, but they've had so many retractions in the past years + they broke their integrity in the Schoen scandal, allowing lenience in the review process to secure a prestigious publication in their journal.
In reality, I mean you're probably my academic senior: How does true advancement get publicized these days? You post a YouTube video somewhere. See LK99. No peer review, no fancy stuff, a YouTube video was enough to get Argonne National lab on the case.
Because it presents an experimental result to other scientists that they may consider worth trying to replicate?
Lots of "replicated" "peer-reviewed" research have been found to be wrong. That's fine, it's part of the process of discovery.
A paper should be taken for what it is: a piece of scientific work, a part of a puzzle.
It would be even better if it was replicated of course.
Depending on what certainty you need you might have to wait for the result of one or several replications, but that is application dependent.
Maybe this is what needs to change. If we only reward discovery and success, then the incentive is to only produce discovery and success.
The US system, and others, even attack people who dare to try and make science more open. RIP Aaron Swartz, and long live Alexandra Elbakyan.
> All procedures and characterization data in OrgSyn are peer-reviewed and checked for reproducibility in the laboratory of a member of the Board of Editors
Never is a strong word.
> If someone cares enough about the work to build on it, they will replicate it anyway.
That's duplicative at the "oh maybe this will be useful to me" stage, with N different people trying to replicate. And with replication not a first-class part of the system, the effort of replication (e_R) is high. For appealing things, N is probably > 2. So N X e_R total effort.
If you move the burden at the "replicate to publish" stage, you can fix the number of replicas needed so N=2 (or whatever) and you incentive the orginal researchers to make e_R lower (which will improve the quality of their research even before the submit-for-publication stage).
I've been in the system, I spent a year or two chasing the tail of rewrites, submissions, etc, for something that was detectable as low-effect-size in the first place but I was told would still be publishable. I found out as part of that that it would only sometimes yield a good p-value! And everything in the system incentivized me to hide that for as long as possible, instead of incentivizing me to look for something else or make it easy for others to replicate and judge for themselves.
Hell, do something like "give undergrads the opportunity to earn Master's on top of their BSes, say, by replicating (or blowing holes in) other people's submissions." I would've eaten up an opportunity like that to go really really deep* in some specialized area in exchange for a masters degree in a less-structured way than "just take a bunch more courses."
In general I think undergraduate projects are a great space to attempt to replicate findings, but it heavily depends on the field. Fundamental physics experiments can be expensive and require equipment that's outside the reach of undergrads. But one thing I love about engineering as an academic field, by comparison, is that anything you research tends to be more achievable for others to replicate because as your end goal you are aiming for something that's practical in the field.
Then perhaps those papers shouldn't be published? Or held in any higher esteem than a blog post by the same authors?
A paper in a peer review journal is like posting a request for reproduction in a heavily moderated mailing list.
A paper in a predatory journal is like the "You are the best ___" price that you get if you pay to go to the "congress" invitation in spam.
Neither of them guaranty that the result is true. The publication in some peer review journals give a minimal guaranty that the paper is not horribly bad, but I've seen too much crap there too.
I know a few journals and author in my area that are serious and I can guess the result will hold, but I find very difficult to evaluate journals and authors in other areas.
Other research proves impossible to replicate because whatever experiment was not described in enough detail to actually replicate it, which should be grounds to immediately dismiss the research before publishing, but which can’t truly be caught if you don’t actually try to reproduce.
Finally these practical concerns don’t even touch on the biggest benefit of reproduction as standard which is that almost nobody wants to reproduce research as they are not rewarded for doing so. This would give somebody, namely those who want to publish something, a strong impetus to get that reproduction done which wouldn’t otherwise exist.
What are we supposed to do in a hundred years when the scientists of today are dead and we have a bunch of results with important implications that aren't documented well enough to replicate?
Only then could he even start building the experiment - total time to run it all seems to run across years.
If you build upon a result, you almost have to replicate it.
An acquaintance spent years building upon a result that turned out to be fraudulent/p-hacked.
In the distant past, publication was an informal process that mostly involved mailing around letters, or for a major result, self-publishing a book. Eventually publishers began to devise formal journals for this purpose, and some of those journals began to receive more submissions than it was feasible to publish or verify just by reputation. Some of the more popular journals hit upon the idea of applying basic editorial standards to reject badly-written papers and obvious spam. Since the journal editors weren’t experts in all fields of science, they asked for volunteers to help with this process. That’s what peer review is.
Eventually bureaucrats (inside and largely outside of the scientific community) demanded a technique for measuring the productivity of a scientist, so they could allocate budgets or promotions. They hit on the idea of using publications in a few prestigious journals as a metric, which turned a useful process (sharing results with other scientists) into [from an outsider perspective] a process of receiving “academic points”, where the publication of a result appears to be the end-goal and not just an intermediate point in the validation of a result.
Still other outsiders, who misunderstand the entire process, are upset that intermediate results are sometimes incorrect. This confuses them, and they’re angry that the process sometimes assigns “points” to people who they perceive as undeserving. So instead of simply accepting that sharing results widely to maximize the chance of verification is the whole point of the publication process, or coming up with a better set of promotion metrics, they want to gum up the essential sharing process to make it much less efficient and reduce the fan-out degree and rate of publication. This whole mess seems like it could be handled a lot more intelligently.
Today, publications do not serve the same purpose as they did before the internet. It is trivial today to write a convincing paper without research and getting that published(www.theatlantic.com/ideas/archive/2018/10/new-sokal-hoax/572212/&sa=U&ved=2ahUKEwjnp5mRtsiAAxVwF1kFHesBDC8QFnoECAkQAg&usg=AOvVaw0t_Bo31BrT5D9zHBdmNAqi).
I understand that it’s frustrating it didn’t happen instantly. And I also understand that it’s deeply frustrating that some undeserving person accumulated status points with non-scientists based on fraud, and that let them take a high-status position outside of their field. (I think maybe you should assign some blame to the Stanford Trustees for this, but that’s up to you.) None of this means we’d be better off making publication more difficult: it means the metrics are bad.
PS When a TFA raises something like “the replication crisis” and then entangles it with accusations of deliberate fraud (high profile but exceedingly rare) it’s like trying to have a serious conversation about automobile accidents, but spending half the conversation on a handful of rare incidents of intentional vehicular homicide. You’re not going to get useful solutions out of this conversation, because it’s (perhaps deliberately) misunderstanding the impact and causes of the problem.
In this case, we are not even aware of most events of fraud/"bad papers"/manipulation - the "crisis" is that we are losing faith in the science we are doing - results that were cornerstones of entire fields are found to be nonreproducible, making all the work built on top of it pointless.(psychology, cancer, economics, etc - I'm being very broad)
At this point, we don't know how deep the rot goes. We are at the point of recognizing that it's real, and looking for solutions. For car accidents, we're past that - we're just arguing about what are the best solutions. For the replication crisis, we're trying to find a way forward.
Like that scene in The Thing, where they test the blood? We're at the point where we don't know who to trust.
Ps: what's a tfa?
Fraud is so frequent and easy to find that there are volunteers who in their spare time manage to routinely uncover not just individual instances of fraud but entire companies whose sole purpose is to generate and sell fake papers on an industrial scale.
https://www.nature.com/articles/d41586-023-01780-w
Fraud is so easy and common that there are a steady stream of journals which publish entire editions consisting of nothing but AI generated articles!
https://www.nature.com/articles/d41586-021-03035-y
Despite being written as a joke over a decade ago, you can page through an endless stream of papers that were generated by SciGen - a Perl script - and yet they are getting published:
https://pubpeer.com/search?q=scigen
The problem is so prevalent that some people created the Problematic Paper Screener, a tool that automatically locates articles that contain text indicative of auto-generation.
https://dbrech.irit.fr/pls/apex/f?p=9999:1::::::
This is all pre-ChatGPT, and is just the researchers who can't be bothered writing a paper at all. The more serious problem is all the human written fraudulent papers with bad data and bad methodologies that are never detected, or only detected by randos with blogs or Twitter accounts that you never hear around.
None of this has any more bearing on fraud by professional scientists than, say, the existence of some garbage-filled Wikimedia server or a badly-edited Wikipedia page means that the Wikipedia editors themselves are fraudsters.
Auto-generated papers have been published in journals from the IEEE, Elsevier, Springer Nature and other well known publishing houses. These papers have supposedly passed peer review in western journals that have been around for decades, and have been signed off by professional academics. Invariably no satisfactory explanation for how this happens is provided, with "we got hacked" being a remarkably common claim. Quite how you publish an entire magazine full of fraudulent articles due to one person getting hacked is unclear; actual newspapers and magazines don't ever have this problem.
Here's an example. The Springer Nature journal "Personal and Ubiquitous Computing" was established in 1997 and has its own Wikipedia page:
https://en.wikipedia.org/wiki/Personal_and_Ubiquitous_Comput...
The Editor-In-Chief is a British academic, who also has his own Wikipedia page. So these aren't fly-by-night no-brand nobodies. Yet this journal somehow managed to publish dozens of obviously auto-generated papers, like this one:
https://static-content.springer.com/esm/art%3A10.1007%2Fs007...
"The conversion of traditional arts and crafts to modern art design under the background of 5G mobile communication network"
or
https://static-content.springer.com/esm/art%3A10.1007%2Fs007...
"The application of twin network target tracking and support tensor machine in the evaluation of orienteering teaching"
The papers are just template paragraphs from totally unrelated topics spliced together. Nobody noticed this had happened until months after publication, strongly implying that this journal has no readers at all (this is a common theme in all these stories, they never seem to notice themselves). The editor agreed the papers were nonsense (his words), but blamed peer reviewers. Yet this journal claims to have a large editorial board with over 40 people on it, mostly from universities in the Europe, USA and China.
What's amazing is that this exact same "attack" had happened before. The previous year Springer Nature had to retract over 400 papers which were auto-generated in the exact same way. They learned nothing and appear to treat the problem as a similar level of severity to filtering email spam.
And in the last six months alone we've seen major fraud scandals impacting Stanford (the President no less), Harvard and Yale. These are supposedly elite universities and researchers. Francesca Gino was earning over $1M a year. Yet their fraud is being uncovered by motivated volunteers, not any kind of systematic well funded science police.
So all the signs here point towards fraud being incredibly easy to get away with. Whole journals have literally no readers at all, academia relies on Scooby-Doo levels of policing, and supposedly prestigious brands are constantly having fraud uncovered by random tweeters, undergrads doing journalism as a hobby etc.
There are two parallel academias. There is the reputable high-trust one, where it's easy to get away with fraud, because people generally don't commit it. And there is the scammy one that exists to help people to game the metrics. While the two overlap a bit, they are mostly disjoint.
If you are an academic, you get a steady stream of spam from the scammy side of the academia. You get calls to submit papers to a conference with "Proceedings by Springer" (but the scope of the conference is barely mentioned), you get invited to become an "ΕԀitоrial ΜҽmƄҽr" of a journal, and so on. Those are like Nigerian letters. They make it very explicit that they are scams, in order to avoid wasting people's time.
You guessed that nobody reads the journal you mentioned, and that's trivially true. Of course nobody reads journals, because their scopes are too wide. No matter what you are working on, most articles in the journals you publish in are irrelevant to you. People read only articles that look interesting or relevant. If nobody cares about an article, it doesn't get read.
While the rest of the world is based on top-down hierarchies, that's not a good model for understanding research. In general, the higher up in the hierarchy you go, the less relevant things become. The article is more relevant than the journal, and the journal is more relevant than the publisher. A rank-and-file professor is more relevant than a department chair. A department chair is more relevant than a dean. And a dean is more relevant than a chancellor/president/whatever.
And recall that I said all this work pre-dates ChatGPT. Using LLMs to generate scientific papers works great, and you won't be able to find them using regexs.
The journals themselves admit there are serious fraud problems and that they don't know what to do about it. So it's very concerning. The world needs a trustworthy scientific literature.
The public perception of a publication in a prestigious journal as the established truth does not help, too.
it's not so much the public perception but what govs/media/tech and other institutions have pushed down so that the public doesn't question whatever resulting policy they're trying to put forth.
"Trust the science" means "Thou shalt not question us, simply obey".
Anyone with eyes who has worked in institutions knows that bureocracy, careerism and corruption are intrinsic to them.
I’ll pile on to say that you also have the variable of how the non-scientist public gleans information from the academics. Academia used to be a more insular cadre of people seeking knowledge for its own sake, so this was less relevant. What’s new here is that our society has fixated on the idea that matters of state and administration should be significantly guided by the results and opinions of academia. Our enthusiasm for science-guided policy is a triple whammy, because 1. Knowing that the results of your study have the potential to affect policy creates incentives that may change how the underlying science is performed 2. Knowing that results of academia have outside influence may change WHICH science is performed, and draw in less-than-impartial actors to perform it 3. The outsized potential impact invites the uninformed public to peer into the world of academia and draw half-baked conclusions from results that are still preliminary or unreplicated. Relatively narrow or specious studies can gain a lot of undue traction if their conclusions appear, to the untrained eye, to provide a good bat to hit your opponent with.
The reality is that science isn't about isolated findings; it's a cumulative effort. One paper might suggest a conclusion, but it's the collective weight of multiple studies that provides a more rounded understanding. Media's tendency to cherry-pick results often distorts this nuanced process.
It's also worth noting the trend of prioritizing certain studies, like large RCTs or systematic reviews, while overlooking smaller ones, especially pilot studies. Pilot studies are foundational—they often act as the preliminary research needed before larger studies can even be considered or funded. By sidelining or dismissing these smaller, exploratory studies, we risk undermining the very foundation that bigger, more definitive research efforts are built on. If we consistently ignore or undervalue pilot studies, the bigger and often more impactful studies may never even see the light of day.
> Still other outsiders, who misunderstand the entire process, are upset that intermediate results are sometimes incorrect. This confuses them, and they’re angry that the process sometimes assigns “points” to people who they perceive as undeserving. So instead of simply accepting that sharing results widely to maximize the chance of verification is the whole point of the publication process, or coming up with a better set of promotion metrics, they want to gum up the essential sharing process to make it much less efficient and reduce the fan-out degree and rate of publication.
Does not represent my experience in the academy at all. There is a ton of gamesmanship in publishing. That is ultimately the yardstick academics are measured against, whether we like it or not. No one misunderstands that IMO, the issue is that it's a poor incentive. I think creating a new class of publication, one that requires replication, could be workable in some fields (e.g. optics/photonics), but probably is totally impossible in others (e.g. experimental particle physics).
For purely intellectual fields like mathematics, theoretical physics, philosophy, you probably don't need this at all. Then there are 'in the middle fields' like machine learning which in theory would be easy to replicate, but also would be prohibitively expensive for, e.g. baseline training of LLMs.
Given that some experiments cost billions to conduct, it is impossible to implement "Peer Replication" for all papers.
What could be done is to add metadata about papers that were replicated.
At least that's my understanding
In theory editors (or rather copyeditors, the editors themselves have to handle too many papers to do this sort of thing) should help with things like style, grammar, and spelling. In practice, quality varies but it is often subpar.
I don't know how refereed conference proceedings work (we don't really use these). The only journals I know of that have professional editors (i.e., editors who are not active researchers themselves) are Nature and affiiliated journals, but someone more knowledgeble should correct me here.
Also, most work goes to conferences; journals typically publish longer versions of published works.
Yes, who do you think ask the reviewers to perform their reviews?
> peer review is more for checking if the methodology, scope, claim, direction, conclusion and relevances is sound&trustable.
No, the parent comment has it right. The only thing being reviewed is the paper, and the point is to make sure it communicates clearly, not that it’s “sound and trustable.”
If you can't replicate them it's like they didn't happen anyways
But yeah, in the grand scheme of things if it hasn’t been replicated, then it hasn’t been proven, but some works are credible on their own.
If someone wants to spend some time replicating something that’s only been described in a paper or two, that is valuable work for the community and should be encouraged. If the person is a PhD student using that as an opportunity to hone their skills, it’s even better. It’s not glamorous, it’s not something entirely new, but it is useful and important. And this work needs to go to normal journals, otherwise there’s just be journals dedicated to replication and their impact factor will be terrible and nobody will care.
It's basically a catch 22. We want replication in prestigious journals, but any Journal with replications becomes less novel and prestigious.
It all comes down to what people value about journals. If people valued replication more than novelty, replication journals would be the prestigious ones.
It all comes back to the fact that doing novel science is considered more prestigious than replication. Institutions can play all kinds of games to try to make it harder for readers to tell novelty apart from replication, but people will just find new ways to signal and determine the difference.
Let's say we pass a law that prestigious journals must published 50% replications. The Prestige from publishing in that journal will just shift to publishing in that journal with something like first demonstration in the title or publishing in that journal Plus having a high citation or impact value.
It is really difficult to come up with the system or institution level solution when novelty is still what individuals value.
As long as companies and universities value innovation, figure out ways to determine which scientists are innovative, and value them more
I have had a paper rejected twice in a row over the last year. Both times the comments include something like "paper was very well-wriiten; well-written enough that an undergrad could read it".
Peer review ensures the gates are kept.
There always seems to be a contingent of people that think that anything less than %100 solution is inadequate so nothing is done. Peer review has proven itself inadequate and people hang on to it tooth and nail. Some disciplines should require replication on everything - I won't name Psychology or Social Sciences in general but the failure to replicate rate for some is unacceptable.
At least in CS/ML there needs to be a “code or it didn’t happen”. Why? Papers are ambiguous. Even if they have mathematical formulas, not all components are defined.
Peer replication in these fields is an easy low hanging fruit that could set an example for other fields of science.
But your run of the mill niche topic will not have the dollars behind it to replicate everyones research.just because CS/AI research is very convenient to replicate does not mean this can be extended to all research being done.
That is exactly why peer review exists to weed out the implausible and low effort/relevance work. It is not fraud proof because it was not designed to be.
This should be standard now, in the age of GitHub, GitLab, et al. If a paper discusses an implementation, but doesn't provide code, it is probably BS.
1. Rebrand peer review as a "readability review" which is what reviewers tend to focus on today.
2. A "replicability statement", a separately published document where reviewers push authors to go into detail about the methodology and strategy used to perform the experiments, including specifics that someone outside of their specialty may not know. Credit NalNezumi ITT
In some fields, aside from specialized knowledge, good experimental work requires what we call "hands." For instance, handling air sensitive compounds, or anything in a condensed or crystalline state. In my thesis experiment, some of the equipment was hand made, by me.
Sometimes specialized facilities are needed. My doctoral thesis project used roughly 1/2 million dollars of gear, and some of the equipment that I used was obsolete and unavailable by the time I finished.
Wow I envy you. My doctoral thesis project spent like... USD2.5k directly for gears (half of it just to buy lego bricks to build our own instrument exactly because we can't afford to buy commercial one lol)
A place where you can spread out more is in dissertations. Mine contained an entire chapter on the experiment, another on the analysis, and appendices full of source code, schematics, etc. I happily sent out copies, at my expense. My setup was replicated roughly 3 times.
The former would be a back and forth between a reviewer that inquire and ask questions (based on the paper) with the goal to reproduce the result, but don't have to actually reproduce it. This is usually good to find out missing details in the paper that the writer just took for granted everyone in the field knows (I've met Bio PHD that have wasted Months of their life tracking up experimental details not mentioned in a paper)
The latter would be the result of the former. Instead of having pages long "appendix" section in the main paper, you produce another document with meticulous details of the experiment/methodology with every stone turned together with an peer reviewer. Stamp it with the peer reviewes name so they can't get away with hand wavy review.
I've read too many papers where important information to reproduce the result is omitted. (for ML/RL) If the code is included I've countless of times found implementation details that is not mentioned in the paper. In matter of fact, there's even results suggesting that those details are the make or break of certain algorithms. [1] I've also seen breaking details only mentioned in code comments...
Another atrocious thing I've witnessed is a paper claiming they evaluated their method on a benchmark and if you check the benchmark, the task they evaluated on doesn't exit! They forked the benchmark and made their own task without being clear about it! [2]
Shit like this make me lose faith in certain science directions. And I've seen a couple of junior researcher giving it all up because they concluded it's all just house of cards.
[1] https://arxiv.org/abs/2005.12729
[2] https://arxiv.org/abs/2202.02465
Edit: also if you think that's too tedious/costly, reminder that publishers rake in record profits so the resources are already there https://youtu.be/ukAkG6c_N4M
Same. Now, when I review manuscripts, I pay much more attention to whether there is enough information to replicate the experiment or simulation. We can put out a paper with wrong interpretations and that’s fine because other people will realise that when doing their own work. We cannot let papers get published if their results cannot be replicated.
> The latter would be the result of the former. Instead of having pages long "appendix" section in the main paper, you produce another document with meticulous details of the experiment/methodology with every stone turned together with an peer reviewer. Stamp it with the peer reviewes name so they can't get away with hand wavy review
Things that take too much space to go in the experimental section should go to a electronic supplementary information document. But then it would be nice if the ESI were appended to the article when we download a PDF because tracking them is a pain in the backside. Some fields are better than others about this, for example in materials characterisation studies it’s very common to have ESI with a whole bunch of data and details.
Large dataset should go to a repository or a dataset journal, that way the method is still peer reviewed and the dataset has a doi and is much easier to re-use. It’s also a nice way of doubling a student’s papers count by the end of their PhD.
> Another atrocious thing I've witnessed is a paper claiming they evaluated their method on a benchmark and if you check the benchmark, the task they evaluated on doesn't exit! They forked the benchmark and made their own task without being clear about it! [2]
That’s just evil!
This may be possible in some sciences, but not in epidemiology or biomed. Often the study is based on tissue samples owned by some entity, with permission granted only to some certain entity.
Datasets in epidemiology are often full of PII, and cannot be shared publicly for many reasons.
> What if all the experiments in the paper are too complicated to replicate? Then you can submit to [the Journal of Irreproducible Results].
Observational science is still a branch of science even if it's difficult or impossible to replicate.
Consider the first photographs of a live giant squid in its natural habitat, published in 2005 at https://royalsocietypublishing.org/doi/10.1098/rspb.2005.315... .
Who seriously thinks this shouldn't have been published until someone else had been able to replicate the result?
Who thinks the results of a drug trial can't be published until they are replicated?
How does one replicate "A stellar occultation by (486958) 2014 MU69: results from the 2017 July 17 portable telescope campaign" at https://ui.adsabs.harvard.edu/abs/2017DPS....4950403Z/abstra... which required the precise alignment of a star, the trans-Neptunian object 486958 Arrokoth, and a region in Argentina?
Or replicate the results of the flyby of Pluto, or flying a helicopter on Mars?
Here's a paper I learned about from "In The Pipeline"; "Insights from a laboratory fire" at https://www.nature.com/articles/s41557-023-01254-6 .
"""Fires are relatively common yet underreported occurrences in chemical laboratories, but their consequences can be devastating. Here we describe our first-hand experience of a savage laboratory fire, highlighting the detrimental effects that it had on the research group and the lessons learned."""
How would peer replication be relevant?
Would this require labs to improve their software environments and learn some new tools? Would this require labs to give up whatever used to be secret sauce? That's. The. Point.
I think when people talk about "replicate" they mean something more than that.
The dataset could contain coding errors, and the analysis could contain incorrect formulas and bad modeling. Reproducing a bad analysis, successfully, provide no corrective feedback.
I know for one paper I could replicate the paper's results using the paper's own analysis, but I couldn't replicate the paper's results using my analysis.
> Would this require labs to give up whatever used to be secret sauce? That's. The. Point.
That seems to be a very different Point.
Newton famously published results made from using his secret sauce - calculus - by recasting them using more traditional methods.
In the extreme cas, I could publish the factors for RSA-1024 without publishing my factorization method. "I prayed to God for the answer and He gave them to me." You can verify that result without the secret sauce.
I mean, people use all sorts of methods to predict a protein structure, including manual tweaking guided by intuition and insight gained during a reverie or day-dream (à la Kekulé) which is clearly not reproducible. Yet that final model may be publishable, because it may provide new insight and testable predictions.
But my point is this linked-to essay appears not only to exclude some areas of good science, but to suggest that any topics which cannot be replicated before publication is only worthy of publication in the Journal of Irreproducible Results.
I gave examples to highlight why I disagree with author's opinion.
Please do not interpret this to mean I do not think improvement is possible.
Yes in a perfect world we would also replicate the data collection but we do not live in a perfect world
Same is true for Drug Trials, there is always a battle over getting the raw data from drug trails as the companies claim that data is trade secret, so independent verification of drug trails is very expensive but if the FDA required not just the release of redacted conclusions and supporting redacted data but 100% of all data gathered it would be alot better IMO
For example the FDA says it will take decades to release the raw data from the COVID Vaccine trials.. Why... and that is after being forced to do so via a law suit.
Yes, but why must the first team wait until the second is finished before publishing?
What if you are the only person in the world with expertise in the fossil record of an obscure branch of snails? You spend 10 years developing a paper knowing that the next person with the right training to replicate the work might not even be born yet.
Other paleontologists might not be able to replicate the work, but still tell if it's publishable - that's what they do now, yes?
> but we do not live in a perfect world
Alternatively, we don't live in a perfect world which is why we have the current system instead of requiring replication first.
Since the same logic works for both cases, I don't think it's persuasive logic.
> the FDA says it will take decades
Well, that's a tangent. The FDA is charged with protecting and promoting public health, not improving the state of scholarly literature.
And the FDA is only one of many public health organizations which carried out COVID vaccine trials.
Nobody, obviously. You cannot reproduce a result that hasn’t been published, so no new phenomenon is replicated the moment it is first published. The problem is not the publication of new discoveries, it’s the lack of incentives to confirm them once they’ve been published.
In your example, new observations of giant squids are still massively valuable even if not that novel anymore. So new observations should be encouraged (as I am sure they are).
> Or replicate the results of the flyby of Pluto, or flying a helicopter on Mars?
Well, we should launch another probe anyway. And I am fairly confident we’ll have many instances of aircrafts in Mars’ atmosphere and more data than we’ll know what to do with it. We can also simulate the hell out of it. We’ll point spectrometers and a whole bunch of instruments towards Pluto. These are not really good examples of unreproducible observations.
Besides, in such cases robustness can be improved by different teams performing their own analyses separately, even if the data comes from the same experimental setup. It’s not all black or white. Observations are on a spectrum, some of them being much more reliable than others and replication is one aspect of it.
> How would peer replication be relevant?
How would you know which aspects of the observed phenomena come from particularities of this specific lab? You need more than one instance. You need some kind of statistical and factor analyses. Replication in this instance would not mean setting actual labs on fire on purpose.
It’s exactly like studying car crashes: nobody is going to kill people on purpose, but it is still important to study them so we regularly have new papers on the subject based on events that happened anyway, each one confirming or disproving previous observations.
Your comment concerns post-publication peer-replication, yes?
If so, it's a different topic. The linked-to essay specifically proposes:
""Instead of sending out a manuscript to anonymous referees to read and review, preprints should be sent to other labs to actually replicate the findings. Once the key findings are replicated, the manuscript would be accepted and published.""
That's pre-publication peer-replication, and my comment was only meant to be interpreted in that light.
Sorry I might have gone mixed up between threads.
Yeah, pre-publication replication is nice (I do it when I can and am suspicious of some simulation results), but is not practical at scale. Besides, the role of peer review is not to ensure results are right, that is just not sustainable for referees.
Due to the pressure of "publish or die" there is very little honesty in research. Fortunately there are some who are transparent with their work. But for the most part, science is drowning in a sea of research that lacks transparency and replication short falls.
CI/CD forces people to codify exactly how to build and deploy something in order for it to get into a production environment. Docker and VMs are ways around this by giving people a "my machine" that can be copied and shared easily.
I can't even imagine how hard it would be to write instructions for another lab to successfully replicate an experiment at the forefront of physics or chemistry, or biology. Not just the specialized equipment, but we're talking about the frontiers of Science with people doing cutting-edge research.
I get the impression that suggestions like these are written by non-scientists who do not have experience with the peer review process of any discipline. Things just don't work like that.
Not to mention that the cutting edge in many sciences are perhaps two-three research groups of 5-30 individuals each in varying research institutions around the world.
For the "hard" sciences, replication often isn't so difficult it seems. LK-99 being an interesting study in this, where people are apparently successfully replicating an experiment described in a rushed paper that is widely agreed to lack sufficient details. It's cutting edge science but replication still isn't a problem. Most science isn't the LHC.
The real problems with replication are found in the softer fields. There it's not just an issue of randomness or difficulty of doing the experiments. If that's all there was to it, no problem. In these fields it's common to find papers or entire fields where none of the work is replicable even in principle. As in, the people doing it don't think other people being able to replicate their work is even important at all, and they may go out of their way to stop people being able to replicate their work (most frequently by gathering data in non-replicable ways and then withholding it deliberately, but sometimes it's just due to the design of the study). The most obvious inference when you see this is that maybe they don't want replication attempts because they know their claims probably aren't true.
So even if peer reviewers or journals were just checking really basic things like, is this claim even replicable in principle, that would be a good start. You would still be left with a lot of papers that replicate fine but their conclusions are still wrong because their methodology is illogical, or papers that replicate because their findings are obvious. But there's so much low hanging fruit.
Is it really that hard for researchers to standardize around providing Dockerfiles? Environment replication is a solved problem.
Unfortunately, no. Dockerfiles aren't as portable as you think, and not architecture-independent. VMs are better, but even then, performance isn't portable either.
The last artifact I produced included builds of 3 web browsers from source--it was over 10GB. One doesn't just "build Chrome in a dockerfile".
On replication, it is a worthwhile goal but the career incentives need to be there. I think replicating studies should be a part of the curriculum in most programs - a step toward getting a PhD in lieu of one of the papers.
They have this idea that a single editor screens papers to decide if they are uninteresting or fundamentally flawed, then they want a bunch of professors to do grunt work litigating the correctness of the experiments.
In modern (post industrial revolution) branches of science, the work of determining what is worthy of publication is distributed amongst a program committee, which is comprised of reviewers. The editor / conference organizers pick the program committee. There are typically dozens of program committee members, and authors and reviewers both disclose conflicts. Also, papers are anonymized, so the people that see the author list are not involved in accept/reject decisions.
This mostly eliminates the problem where work is suppressed for political reasons, etc.
It is increasingly common for paper PDFs to be annotated with badges showing the level of reproducibility of the work, and papers can win awards for being highly reproducible. The people that check reproducibility simply execute directions from a separate reproducibility submission that is produced after the paper is accepted.
I argue the above approach is about 100 years ahead of what the blog post is suggesting.
Ideally, we would tie federal funding to double blind review and venues with program committees, and papers selected by editors would not count toward tenure at universities that receive public funding.
From my point of view, the biggest issue is accepting/rejecting papers based on first impressions. Because there is often only one round of reviews, you can't ask the authors for clarifications, and they can't try to fix the issues you have identified. Conferences tend to follow fashionable topics, and they are often narrower in scope than what they claim to be, because it's easier to evaluate papers on topics the program committee is familiar with.
The work done by the program committee was not even supposed to be proper peer review but only the first filter. Old conference papers often call themselves extended abstracts, and they don't contain all the details you would expect in the full paper. For example, a theoretical paper may omit key proofs. Once the program committee has determined that the results look interesting and plausible and the authors have presented them in a conference, the authors are supposed to write the full paper and submit it to a journal for peer review. Of course, this doesn't always happen, for a number of reasons.
That said, it infuriates me to no end when I read a Phys. Rev. paper that consists of a computational study of a particular physical system, and the only replicability information provided is the governing equation and a vague description of the numerical technique. No discretized example, no algorithm, and sure as hell no code repository. I'm sure other fields have this too. The only motivation I see for this behavior is the desire for a monopoly on the research topic on the part of authors, or embarrassment by poor code quality (real or perceived).
https://scitechdaily.com/gravitational-waves-detected-using-...
Hydrodynamic quantum analogs can be uses to study quantum particles at macro-scale:
https://en.wikipedia.org/wiki/Hydrodynamic_quantum_analogs
ESA Euclid near-infrared telescope launched few weeks ago:
https://www.esa.int/Science_Exploration/Space_Science/Euclid...
But the information gets around. In my former field, everyone knew which were the dodgy papers, with results no-one could replicate.
UPDATABLE COVER PAGE:
Title Authors
Abstract
Blah, blah, ...
Not reproduced.
Successful reproductions: ...citations...
Reproduction attempts: ...citations...
Countering reproductions: ...citations...
Reproduction resources:
Data, algorithms, processes, materials, ...
Cost, time, one-off events, ...
Having a standard section for this, with an initial state of "Not reproduced" provides more incentive for original workers to provide better reproduction info.
For algorithm and math work the reproduction could be served best with downloadable executable bundle.
Seems like article is not about software code.
If on, the other hand, they just want the raw data, and let others go to town on it in their own way, that's fine, probably. Results that don't depend on very particular details of the processing pipeline are probably more robust anyway.
Human subjects research, for one e.g. very often they involve clinical populations that are very hard to recruit. You can spend tens of thousands in advertising, and multiples more in labor, to get a hundred participants in, over the course of an entire year of effort, and that's not even counting the money spent on a clinician doing a diagnosis. And then, when you do, you may, say, pay $1000 for the MRI per subject, plus the $100 bucks you pay directly to the participant themselves.
Or even worse: Some passing by comet, or planet/moon/whatever in the solar system. And just to make things EVEN worse, you need to analyze the data in some destructive way.
Certainly very plausible scenarios, but also some which could prohibitively expensive to do multiple times.
Science is a process. Peer review isn't perfect. Replication is important. But it doesn't seem like the author understands what it would take to simply replace peer review with replication.
All PhD programs have requirement for a minimum number of novel publications. We could add to the requirements a minimum number of replications.
But truth to be told, a PhD in science/ engineering will probably spend their first two years trying to replicate the SOTA anyway. It’s just that today you cannot publish this effort, nobody cares, except yourself and your advisor.
Publication is a starting point, not a conclusion
Publication is submitting your code. It still needs to be tested, rolled out, evaluated, and time-tested.
Having running demos is another step in the right direction (see https://blog.arxiv.org/2022/11/17/discover-state-of-the-art-...).
But outside of computer science replication is even more difficult. Maybe if people would use standardized laboratories and robots, one could replicate findings by rerunning the robots code on another standard robot lab ( Basically the idea here is to virtualize laboratory work).
But even then for the biggest most complex experiments this will not work: Replicate CERN anyone?
It may be more accurate to suggest that repeatability is part of the scientific method. But even that is not strictly true.
Consider, the single longest running scientific work was not repeatable, and was not shared with anyone outside the cadre of people doing it. Around 3000 years ago, a secretive caste of astrologers/scribes watched the heavens, and recorded their observations for several centuries. They did not publish their findings, thus making them anecdotal (yes, that's what anecdotal means, just that it wasn't published). The exact circumstances and variables were never repeatable, due to the movements of the celestial bodies, precession, etc.
Similarly, the UQ pitch drop experiment, having not yet completed, has not been repeated. But it's still an entirely valid scientific experiment.
At that point I thought about making a TeX interpreter so one could easily "run a paper" on their own data to see if the papers claims hold. As it turned out, people often write the same formula in multiple ways and to make a TeX interpreter you'd have to specify a "runnable" subset and convince anyone to use that subset instead of what they got used to. So the idea stalled.
In a few years, publishing a GitHub link along the paper became the norm, and the problem disappeared. At least in applied geometry, people do replicate each other results all the time.
Tools like nextflow or snakemake help with respect to having a one liner to generate all data in a paper potentially, handle dependencies, list resource expectations, use your own profile to handle your environment specific job scheduling commands and parameters. However, this still doesn’t do anything for whether you have access to the resources needed.
Peer review might (or might not) weed out a few papers before they ever get to being reproduced - and that a paper "passed" peer review often means very little. (In some journals more, in some less).
You can't replace peer review with peer replication. Reviewers often do volunteer work - supporting their field and the journal by checking submissions just for any grave errors/mistakes. They often spend just 10 to 15 minutes per submission - for hundreds of submissions. It's not realistic to ask those reviewers to do a full replication attempt for hundreds of submissions.
So any attempt to "replace" review with replication, would end up basically removing review altogether, without increasing the amount of replication attempts being made.
The review score of the abstract was only used to decide on the best topics to invite for a presentation or talk - and the review score of the paper was used to hand out awards, decide "highlighted" papers, and it also influenced how high up in the search results a given paper might appear.
Let people review what they want, where they want, how they want. Let people replicate when they find interesting and motivating to work on.
It's a ton of unpaid, volunteer work, if I want to be a high quality reviewer then it's at least a day (at least 3 thorough reads, taking notes, writing the review, reviewer discussions, post rebuttal, back-and-forth for journals). I am lucky and privileged that my employer counts this towards work time. Only 20% papers get accepted in my domain.
Now if I had to spend a week on replicating a paper - and this is CS/graphics, where it's easy and "free" - I'd never volunteer to being a reviewer.
You'd need professional "replicators", but who will pay for them? And who will be them - you need experts, and if you are an expert, you don't want to merely replicate others people work full time, instead of working on your own innovation.
The main point is that the paper seriously underestimates the difficulty and time it requires to replicate experiments in many experimental fields. Who will decide which work needs to be replicated? Should capable labs somehow become bogged down with just doing replication work? Even if they don't find the results not interesting?
In reality if labs find results interesting enough to replicate they will try to do so. The current LK-99 hurrah is a perfect example of that, but it happens on a much smaller scale all the time. Researchers do replicate and build on other work all the time, they just use that replication to create new results (and acknowledge the previous work) instead of publishing a "we replicated paper".
Where things usually fail is in publication of "failed replication" studies, and those are tricky. It is not always clear if the original research was flawed or the people trying to reproduce made an error (again just have a look at what's happening with LK-99 at the moment). Moreover, it can be politically difficult to try to publish a "fail to reproduce" result if you are small unknown lab, if the original result came from a big known group. Most people will believe that you are the one who made the error (and unfortunately big egos might get in the way, and the small lab will have a hard time).
More generally, in my opinion the lack of replication of results is just one symptom of a bigger problem in science today. We (as in society) have essentially turned the scientific environment increasingly competitive, under the guise of "value for tax payer money". Academic scientists now have to constantly compete for grant funding, publish to keep the funding going. It's incredibly competitive to even get in ... At the same time they are supposed to constantly provide big headlines for university press releases, communicate their results to the general public and investigate (and patent) the potential for commercial exploitation. No wonder we see less cooperation.
But - a quick counterexample - as far as replication goes: What if the experiments were run on custom made or exceedingly expensive equipment? How are the replicators supposed to access that equipment? Even in fields which are "easy" to replicate - like machine learning - we are seeing barriers of entry due to expensive computing power. Or data collection. Or both.
But then you move over to physics, and suddenly you're also dealing with these one-off custom setups, doing experiments which could be close to impossible to replicate (say you want to conduct experiments on some physical event that only occurs every xxxx years or whatever)
The junior faculty will clear the rotten apples at the top by finding flaws in their research and then will win the tenure that was lost in return
This will create a nice political atmosphere and improve science
Replicator: Do you know how much data I'll need to collect? 11,000 particpants followed across multiple timepoints of MRI scanning. Show me the money.
"Science needs accounting" is a search I had saved for months which really resonates with the idea of "peer replication."
In accounting, you always have checks and balances, you never are counting money alone. In many cases, accountants duplicate their work to make sure that it is accurate.
Auditors are the corollary to the peer review process. They're not there to redo your work, but to verify that your methods and processes are sound.
The pay difference between research and industry in many areas is not even funny.
Then I remembered that my main issue with modern academia is that everyone is incentivized to publish a huge amount of research that nobody cares about, and how I wish we would put much more work into each of much fewer research directions.
Would code replication result in fewer use after free, or off by one than code review? Or would it mostly be a waste of resources including time?
[1]: if you have delivered telco code to Softbank you may have heard this sentence
1 mg of anti-rabbit antibody (a common thing to use in a lot of biology experiments) is $225 [1]. Outside of things like standard buffers and growth medium for prokaryotes, this is going to be the cheapest thing you use in an experiment.
1/10th of that amount for anti-flagellin antibody is $372. [2]
A kit to prep a cell for RNA sequencing is $6-10 per use. That's JUST isolation of the RNA. Not including reverse transcribing it to cDNA for sequencing, or the sequencing itself. [3]
Let's not even reach things like materials science where you may be working on an epitaxial growth paper, and there are only a handful of labs where they could even feasibly repeat the experiment.
Or say something with a BSL-3 lab where there are literally only 15 labs in the US that could feasibly do the work, assuming they aren't working on their own stuff. [4]
[1] - https://www.thermofisher.com/antibody/product/Goat-anti-Rabb... [2] https://www.invivogen.com/anti-flagellin [3] https://www.thermofisher.com/order/catalog/product/12183018A [4] https://www.niaid.nih.gov/research/tufts-regional-biocontain...
I guess we should not talk about the Higgs before someone else builds a second one and replicates the papers.
And replication can be done by the same party (although an independent party is better), and that may mean many trials.
And do we even set policy based on existence or non-existent of higgs bosons?
I am particularly unhappy with soft sciences in terms of replication.
Probably not gonna catch on.
High impact journals [6] tend to prefer exciting, novel, and positive results (we tried new thing and it worked so well!) vs negative results (we mixed up a bunch of crystals and absolutely none of them are room-temp superconductors! we're sure of it!).
The result is that cherry picking data pays, leaning into confirmation bias pays, publishing replication studies and rigorous but negative results is not a good use of your academic inertia.
I think that creating a new category of rigor (i.e. journals that only publish independently replicated results) is not a bad idea, but: who's gonna pay for that? If the incentive is you get your name on the paper, doesn't that incentivize coming up with a positive result? How do you incentivize negative replications? What if there is only one gigantic machine anywhere that can find those results (LHC, icecube, etc, a very expensive spaceship)?
There might be easier and cheaper pathways to reducing bad papers - incentivizing the publishing of negative results and replication studies separately, paying reviewers for their time, coming up with new metrics for researchers that prioritize different kinds of activity (currently "how much you're cited" and "number of papers*journal impact" things are common, maybe a "how many results got replicated" score would be cool to roll into "do you get tenure"? See [3] for more details). PLoS publish.
I really like OP's other article about a hypothetical "Journal of One Try" (JOOT) [2] to enable publishing of not-very-rigorous-but-maybe-useful-to-somebody results. If you go back and read OLD OLD editions of Philosophical Transactions (which goes back to the 1600's!! great time, highly recommend [4], in many ways the archetype for all academic journals), there are a ton of wacky submissions that are just little observations, small experiments, and I think something like that (JOOT let's say) tuned up for the modern era would, if nothing else, make science more fun. Here's a great one about reports of "Shining Beef" (literally beef that is glowing I guess?) enjoy [5]
[1] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6668985/ [2] https://web.archive.org/web/20220924222624/https://blog.ever... [3] https://www.altmetric.com/ [4] https://www.jstor.org/journal/philtran1665167 [5] https://www.jstor.org/stable/101710 [6] https://en.wikipedia.org/wiki/Impact_factor, see also https://clarivate.com/