Individual studies shouldn't be the primary focus. It shouldn't fall to every individual researcher to be fully accurate and honest. This is just not possible because people are imperfect, and have all the wrong incentives in many cases.

What we need is a system designed to deal with this better. My suggestion would be greater incentives to the publication of negative results, and incentives to reproduce research. Create a system that encourages reproduction as a second step to peer review. This should also increase quality. If your research can't be easily reproduced it's probably not doing a whole lot of good even if ultimately you got the right answers.

Test driven research, if you will.

Peronsally, IMO: you should submit methodology--including the data analysis and statistical tests you'll perform, to a journal before the study has been conducted. Then, they say yes or no and if yes you perform the study and as long as it is performed to the approved specifications, you get published.

It sounds nice, but this will make for very boring journals with 90% negative results due to problems with the methodology.

Here an example of a typical project, where researchers are interested in the role of gene X in blood vessel formation in cancer, which could make it a good target for a cancer treatment. They want to knock the gene out and measure the blood vessel formation in an in vitro model.

Publication 1.

- We attempted to knock this gene out in this cell line, but due to an infection the cell line died. The installation was sterilized, the stored master cell bank was unfrozen, and we updated our sterilization protocols.

Publication 2.

- We attempted to knock this gene out in this cell line, but sequencing showed that there were off-target effects introduced. Reagents should be changed to be more specific.

Publication 3.

- The gene was succesfully knocked, but we found that a homologous gene appeared to activate the same pathway. The homologous gene should be knocked out as well to study the function of the gene.

Publication 4.

- The gene was successfully knocked out by disabling the gene and its homologue. The assay we used to study blood vessel formation was flawed. Due to insufficient oxygen supply, the positive and negative controls for blood vessel formation did not grow. New experiments should increase oxygen flow.

Publication 5.

- We successfully studied the function of the gene and found that it does not impact blood vessel formation in cell line x.

This is how the majority of research projects end up.

However, other researchers will be looking for a biological target to treat cancer X, and the information obtained by this series of experiments is not interesting. This is just one out of 99% of 30,000 genes that are not suitable treatment targets.

The conclusions are only interesting for researchers focused on gene x (and maybe its homologue), and potentially for researchers developing methodology. Questions answered were different than the original ones posed.

Wouldn't that just put the journals in control of what studies get performed (as well as what gets published)? Seems weird to give all that power to the journals.

Well, they are already in control of what studies get published, what isn't much different.

I do think they shouldn't control that, but it's a deeper change, for a different gain.

It is a lot different because if the study isn't published in one journal, it could always be published in another. In this situation, the study wouldn't be done at all until a journal agreed to publish it.

You could submit your proposal to other journals as well though

If I was a journal, and I was promising to publish a study once it is completed, I'd ask for exclusivity rights.

Which goes to show the whole issue with this idea.

Again that didn't change anything when compared to the current situation.

Journals do ask for exclusivity when they publish things. The don't ask for it just for submission. And the same reasons for why they don't now (nobody would submit anything to them) apply for the OP's idea.

In fact, all the changes only make the life of the researcher easier, because he would only have to present a proposal for the journals to deny. Now he has to complete his research before he gets the denial. (It also makes the life of the journals much harder, for the same reason.)

> Again that didn't change anything when compared to the current situation. Journals do ask for exclusivity when they publish things. The don't ask for it just for submission.

You contradict yourself. It absolutely changes things.

> Now he has to complete his research before he gets the denial.

It could degrade the incentive for the researcher to do great work, if they knew that they would be published.

There are already pre-print servers where you can pre-register your studies. It is not mandatory, but it is a think.

Despite his dubious past, the billionaire John Arnold has focused time and money to amend scientific research. I've read maybe 2 years ago about that endeavor and followed him on Twitter ; I found it quite interesting and original. For example, he offers researchers to save their initial data (before any screening) so that other researchers can check which cases were thrown out of the study. He also funded reproduction tests. I haven't followed since, so I hope he kept on doing good work.

This is my first time learning about him and haven't been exposed to much more than the wikipedia article - what about his past is dubious?

> It shouldn't fall to every individual researcher to be fully... honest.

Really? I don't know, "Don't be a fraud" might be a good Rule No. 1

Fraud is only a small part of the problem. No open data, poorly described methodology so reproductions sometimes aren't even possible, too little funding for reproductions, p-hacking, no incentives to publish negative results, journals incentivized to publish sensational results to increase their impact factor even though sensational results are likely false thus perpetuating misinformation. These are the main drivers.

It’s more that a system should not allow a single fraud to shape its understanding and its policy. Assume fraud exists.

Or even requirements for replication amongst groups with separate funding sources.

If everyone needs replication for publishing, the incentive to replicate other’s work increases.

In addition there needs to be tiers of publication that are purely objective. Maybe something along the lines of this paper was replicated by 3 different teams and the results hold up. Like a replication score or count- this would be desirable since then students could apply for funding to replicate papers.

The problem is that for any sufficiently complex or advanced research there are enough variables to contend with that there are 1000s of ways to get "no repeatable result" and only a few to actually replicate. Knowing which variables need to be controlled is part of the expertise, training, and lore. Knowing when those variables are getting out of control and how to fix them early (knowing which knob or often combination of knobs are required) is yet more expertise and experience. This is true in physics, chemistry, biology, and electronic design. Getting a null result really is often simply a failure to properly run the experiment... and as a scientist it would be very easy to follow any documented "recipe" and fail to get the proper result in hundreds of different ways. That does not make one a great scientist. You can't even tell me every way it could be screwed up. I can always come up with a new one!

Programming is such a wonderfully controlled environment, it's digital, replicable, and _almost_ math-like. You can know the variables. You can trivially replicate a program, or dig down into the OS... but every physical experiment out there has much more common and worse problems than Spectre or Rowhammer waiting to destroy it. It's the cosmic rays.

Your equipment/materials/specimens gets contaminated or reacts or ages. Your ultrapure water isn't. There's a vibration that changes everything. Your calibration sample changes. Cleaning staff used a new cleaner. The AC turned on. Build your experiment to be impervious to all of these and everything else I can't think of... or use the intelligence of an expert. Get there faster, cheaper, and find better ways of doing it along the way.

You'ld be right to say that it should all be documented, but the problem is that you basically need to be an expert to even understand the variables and to those experts many variables are so obvious to control they wouldn't think of doing otherwise. To give concrete and extremely valuable examples in manufacturing: EUV systems for lithography, OLED deposition and patterning, Lithium battery electrode forming. You can have the very best and exquisitely designed equipment for manufacture and yet without experienced experts running the process (replicating the experiment day in and day out) it will go out of control and you will get garbage in a million ways. It's like amateurs playing chess against a master... where the master is the natural variation of the environment.

High value semi manufacturing is a great example of how hard it is. The processes are designed to be as simple as possible to replicate. You replicate every day as many times as possible. You have huge amounts of money and resources to succeed and sometimes you still fail. Getting to 10% yield/replication is a big step... getting to 50% maybe harder... then 90%... and higher. Novel science is often harder and there's much less money and the time to replicate could be months/result rather than 1000s of testable results/min.

If you can't replicate that means the original research didn't document the conditions adequately.

The whole point is that some people CAN replicate, but not everyone can, because the documentation for real (interesting) experiments would be so complex for a layman that they would likely become confused and never finish. If I give you and your 20 best friends a $300M EUV machine with all the documentation in the world, the best thing you could do is sell it for $10M in parts. Give it to TSMC or Intel's experts and they will make $Bs in chips at 50% margins. YOU can't replicate the experiment. Most people can't replicate the experiment, and being unable to do so does NOT make you special. However, some people can replicate it many millions of times a year.

Who's results would you publish?

Undermining the whole concept of "science".

Perhaps with some justification, perhaps not. But if replication is impossible because of "irreducible complexity" (seems to be the argument?) then why bother with science at all?

Things are not that bad. But in some fields other approaches may be more productive.

For example clinical medicine may have more to contribute than scientific medicine. Psychological studies in particular are notorious for being impossible to replicate. Does not mean we should give up on psychology. It does not have to be science to be useful

Prediction markets are an interesting alternative solution. If people can make money by correctly predicting the outcomes of experiments, and you have a way to prevent profit-driven bias impacting outcomes (e.g. independent experiment designers and performers, and the performers can't participate in the market but maybe get a cut of the transaction fees so that more "interesting" experiments get performed), then you get a pretty different research dynamic.

(I am paraphrasing Eliezer Yuskowsky here, specifically a small part of his NSFW and frankly weird collaborative fiction Mad Investor Chaos.)

If prediction markets could help, we wouldn't need science.

Then how would you resolve the markets? The predictions would be about the outcomes of experiments...

Personally this is one of the reasons I chose to go into software instead of statistics when I was a young math major. The idea that you could make a small mistake that invalidates all your work and you might not even know it until much later or even never always made me uncomfortable. With software I can write tests till my neuroticism has been placated

If all the programmers who write tests go into industry and none of them go into academia, then the state of software in academia will be... oh, I see, the status quo.

No, they do write tests, just

1. Include the ones that passed and discard those that failed. 2. Run the tests several times and report only the best result. 3. Keep some of the details secret (data, some weird parameters, or maybe the whole repository), tweak something, run again, get some nice figures and publish a new paper.

Caltech has Millikan's (oil drop experiment demonstrating quantization of charge) notebooks. He repeated the experiment until the data fit his theory.

Fortunately, he turned out to be right.

"Poorly aligned incentives" in the nutritional sciences is, as far as I can tell, mostly various food lobbies getting their product to look good.

The reason there is "confusion" about whether many animal products are good or bad is because the corporations making billions of dollars off the product fund biased research.

Consider looking at the research digests from a non-profit (that sells no products, has no ads, etc) focusing on improving your health: https://nutritionfacts.org/

How a non-profit pays for hosting website cost, writing report, and so on and so on… being a non-profit doesn’t mean you are tied to corporations and free of any conflict of interest. If you start looking carefully you will see that most non-profit org aren’t unbiased and honest. And since long time…

For instance this whole vegan trend is based on some radical religious opinion to control fertility and lower class reproduction by promoting a hormonal dysfonctionnal diet. All those from “official scientific paper” written by those religious fanatic few decades ago.

> vegan trend is based on some radical religious opinion

What a bold empirical claim. Could you share any evidence of this?

It has probably been a mistake to have so many articles that are effectively "We measured A and its value suggests the following theory" rather than "To test theory B we measured A and found it did(n't) agree"

I reckon that lack of a good theory is the by far best predictor for non-replicable results.

I think there are several issues with this. In these fields, theories (not so fundamental ones) may not be very useful, in contrast with fields like Mathematics or Physics where a hypothesis may lead to a lot of other theoretical results. For example, if material A has property X, there will be a lot of applications for it, but this hypothesis cannot be used as a foundation for other theoretical research. As a result, it is not useful to just state a hypothesis: you have to verify it as well, and negative results are not really valued because it means nothing to others if you rejected something they don't care about.

Not sure if properties of materials is the best example, those tend to have some theory behind them that, if not explains, at least defines the property and its consequences. If one research team measures the stiffness of a material it's easy enough to make a test that replicates their result.

Medical research in particular seems to suffer a lot from papers where something has been measured with no supporting theory whatsoever and consequently no real information other than "If you do what we did you probably get the same result". Rather unsurprisingly this is hard to replicate because there's no theory telling you what is and isn't relevant to replicate.

I'm not saying this is easy to fix, but maybe we should focus more effort on papers that expand the theory, rather than papers that at best show that something might be the case but we don't know why.

It makes sense to me to first collect small amounts of data, only then formulate theories and only then verify those with larger experiments. Especially when experiments are expensive.

It does not make sense to act like small studies are proof of anything the way people like to use those studies on the internet. It is wrong that researchers are effectively punished for reproduction, verification etc. But the existence of small data collection studies on itself is not wrong.

I think this problem even has a name: HARKing - Hypothesis After Results Known.

One way to mitigate it is to pre-register studies as it's done in some areas like clinical studies.

> what needs to change in the education process so studies are better designed

Remove politics, free kidnapped and gagged science.

Politics and education are certainly factors, but I would argue that they are minor. The bigger issue is the natural incentive structures that academia has created. If you want to have a career, you get both attention and funding in return for headline-grabbing results. This creates two huge second-order effects: first, everyone's incentivized to look for big impressive new effects. Second, focusing on confirmatory work is a great way to tank your career. And those create a third order effect: people who cut corners thrive, and there's not much to keep that in check because people who cautiously vet their work and others' get stuck in adjunct professor hell until they finally give up and exit the field.

Politics just adds some spin onto that. Education, I just don't think that's even a possible solution. The math involved in some of these fancier statistical techniques can be devilishly hard to get right, and it's often simply asking too much to expect someone to be an expert at both that and molecular biology. You could argue that that means people should look for competent co-authors to cover that side of things. But, see above - current incentive structures don't really favor the researcher who says, "I know! Let's bring on some nerd whose only job is to find reasons why we shouldn't publish!"

If we assign grant points to people that flood the precious journal space with fake articles claiming over-hyped promises, maybe we should give grant points also to the people able to spot the lies on those kind of articles.

After all they are serving to the advance of science also (and saving thousands of expensive hours of research to other). If you want to promote it, just create incentives.

Otherwise, If the cheater wins in a 95% of the times and the verifier does not have anything to gain (and will damage their own research by the time invested on this) after a while only cheaters remain [1]. At middle term, all research tend towards a sort of expensive shamanism in well illuminated white rooms.

[1] (And of course the non cheaters will be actively blocked from this point).

My proposal would be: If there was a clear, proven, intention of cheating and not just an error (we all commit stupid errors sometimes), the letter or communication of the people disclosing it will receive the same citation rank that the fake article had accumulated until that date. Good for the journals, good for the research.

Like a bug bounty for academic publications?

Yes, something like that

> Politics and education are certainly factors, but I would argue that they are minor

Are they? If your proposed project does not align with current politics, no funding for you.

From reading papers of various different types in the past few years it really depends a lot on the exact field in question.

Nutrition? Would be surprising if there's much political distortion there but there's lots of really bad methodologies, partly because it's hard to get good data (people are bad at reliably writing down what they eat).

Social bot/misinformation research? Basically just politics masquerading as science. Methodologies are invariably bogus and conclusions are clearly determined before any 'research' is done.

Epidemiology? A mix of bad incentives and ideological distortion. The focus on policy advice isn't really scientific, and their policies somehow always end up requiring massive state intervention in people's private lives. This isn't something that falls naturally out of the science (which is unreliable anyway) but more the fact that they don't even consider any other possibilities.

The above three are pretty bad fields. There are plenty of better ones. Computer science is mostly OK. AI and cryptography are slightly dubious. Stuff like databases or computer graphics are fine. So IMO it doesn't make sense to talk about research or academia in general, or it does but only when discussing the universal incentive problems. Some fields just seem naturally resistant to political game playing because they don't have much/any impact on social policy to begin with. Obviously I mean politics as in democratic politics here, not the more mundane everyday departmental politics.

Academia is pretty corrupt or corruptible (well, it's humans after all). If CS research isn't corrupted at the moment, it will be if it being corrupted becomes useful.

I don't know about hard CS, and I don't know about that word "corrupt" because it implies malice and I think that Hanlon's Razor is probably a useful heuristic here. But, in machine learning, I've see plenty of the kind of thing I'm talking about. Papers often introduce new algorithms or techniques whose improvement over the competition can plausibly be attributed to excessively tuned hyperparameters that overfit the data on one or more benchmarks. It's really not all that far removed from p-hacking in my opinion.

as has been the case in many, if not most, times, places, and contexts.

Politics is an emergent property of human interaction. It cannot be removed.

I think that's a separate issue that would not have affected this particular case.

That's just hard to do with a topic like this. There are committed people on both sides of the issue who know what they want to hear. That's not a good environment for science to find a neutral answer.

> Remove politics

That needs the remotion of taxation. I agree.

You're framing this as if they're mutually exclusive options, which they're not. My impression is that it's "all of the above": the incentive structure in academics is completely broken, and this incentivizes both unethical practices, as well as being sloppy "to get things done".

I can think of numerous examples of both from my career, from working with various colleagues and in stories I've heard. All of this is the tip of the iceberg I think: what you hear about usually are cases of blatant fraud, with high-profile or high-importance research. The bulk of greyer malpractice, with softer squishier levels of ethical dubiousness, and more indirect, cumulative impacts on peoples' careers and literature trajectories, are more hidden.

It's gotten to the point where I no longer put much stock in the reputation of individual researchers. Too many times I know in the best case there's an institution, or department, or subcommunity, or doctoral students or postdocs behind it, with who knows what sort of manipulation or luck. In the worst case there's years of fraud or other forms of corruption. Call me inappropriately cynical but I didn't start out this way.

> I do wonder how much of this is due to poorly aligned incentives causing researchers to error on the side of what will get the most attention versus researchers in general simply being sloppy.

I think we should also consider the possibility that some of it might be ideologically motivated.

For example someone who thinks people shouldn’t eat red meat due to climate change, animal rights, etc. (and who possibly belongs to a social circle where such beliefs are popular) might be motivated to fabricate results that could convince some people to eat less red meat (hypothetically, not saying that is the case here).

In this case I'm pretty sure that the research in question was about the same quality as the work produced by Tobacco Industry Research Committee or the Sugar Research Foundation.

I think it is not enough of a sense of quality and too much pressure or gamification of getting things published. And fundamentally lots of researchers need to be more interested in finding the truth and considerably better at statistics.