I think reproducibility is one of the biggest challenges to face science in our time. Luckily we have the tools we need to solve the problem already. Of corse paper journals don't want to waste space and reproducing results, but no such constraints are on online publications. And while moving publications online we can solve the issues of journal publishers owning the copyright on papers written with public money. The big question that's left is trust. A lot of the journals that have gone online are essentially scams publishing whatever you want for a fee. We still need the peer review step the close the gap.
The big problem isn't that journals don't want to "waste" space publishing reproduced results, but that scientists don't want to "waste" their time reproducing other peoples results. Unless the result is a once a decade game changing result for your field, there is no incentives in reproduction. You cannot change peoples behaviour without changing the underlying incentive structure. Do that and everything else will trivially fall into place.
Nah i think the bigger problem is the MBAs walking around, trying to apply widget factory metrics to academia, healthcare, and other hard to measure environments.
This in turn has lead to the whole "publish or perish" environment, as the MBAs use published articles as a replacement for widgets made and citations as sales.
> scientists don't want to "waste" their time reproducing other peoples results
Completely agree, and it is actually even worse than you imply: if you think it is hard to get scientists to reproduce others work, imagine trying to pitch NSF/DOE, etc. to fund the study!
Besides reproducibility, you also need to make testable predictions for it to be science. I wonder what would the testable predictions be for most evidence that all (or almost all?) evolution proceeds entirely by random mutation and natural selection. If this isn't true then a lot of the stuff in evolutionary psychology and other "just so" explanations may be completely unscientific. Exemplified by crap like this: https://www.theguardian.com/science/2007/aug/25/genderissues
If you are looking for solid evidence that current living things originated by common descent and speciation through evolution by natural selection, that evidence is abundant.[1] Some of the evidence backing up what is called "evolutionary psychology"[2] is much more debated, and much more problematic, and other psychologists are quick to criticize many evolutionary psychology publications.[3]
Those are two different things: the theory of common descent has a lot of evidence in its favor. On the other hand, the claim that all features observed today have been the result solely of random mutation and natural selection, is rather untested. It doesn't make many testable predictions. The mathematical issues with the theory (wherein, on its own terms, the chances are infinitesimal of it being true) are rarely addressed in mathematical terms, instead most responses are of the type "well X1 feature might have helped Y animal survive and reproduce better until it led to X2 feature". If you look at tons of published literature that mentions evolutionary explanations, it is all "just-so" stories which already assume the theory is true.
I am talking about the following criterion in order to qualify as a real scientifically tested theory:
According to this criterion, the theory that random mutation and natural selection alone produced all the genotypes we see is like Marxism and Freud's psychological theories. If we assume they are true, we can explain everything with just-so explanations. But that's not falsifiable, it doesn't make testable predictions. NOTE: this is different than saying all animals had common ancestors. It says we know how the changes happened that led to the current features. The fact is, we don't know. We barely have scratched the surface. We are still in stages comparable to the luminiferous ether theory in ohysics or humors in biology. I think in 100 years people will look at our theories of random mutation and natural selection (whether classical, punctuated equilibria etc) and consider us naive. And yet we have proponents like Dawkins yelling from the rooftops that "evolution is as proven as gravity". That's equivocating the word "evolution" with common descent. There is also a lot of pressure and political interest from demand for a naturalistic explanation, similarly to how there is demand from the religious camp for "the institute for creation research". It affects who gets research and who gets published. Science is a human endeavor and it is affected by political and organizational pressures just like everything else. But when something is testable and is being tested, it's obvious. Here nothing is obvious - even the math is dubious. These theories are just riding on the coattails of the theory of common descent because we don't have better naturalistic explanations at the moment. But just because a theory is the best we have doesn't mean it's true. Overstating it is, is more advocacy than science.
That hypothesis isn't falsifiable. We know "X effects Y" from experimental result. You're asking for proof that "X and only X effects Y", which is the same as asking for "There does not exist some X' that effects Y" where the set of X' is unbounded. No, that's not testable and doesn't need to be.
Just like we didn't need to know about the X' of Einstein's equations to know that the X gravity was the primary force holding you to the ground.
There very well may be an X' that effects observable biological features other than evolution, but that X' needs to be tested on an individual basis. So while it's productive to hypothesize an X' and test it, there's not much we can do to rule out all X' as a class.
So if you can't rule them out, your theory shouldn't postulate that they don't exist. It's one thing to do 100,000 experiments and have the results COMPLETELY predicted by Einstein's equations beforehand. It's quite another to invent just-so stories in terms of the theory itself after each observation!
The delicious irony is that article which is published in The Economist makes no mention of rampant irreproducibility in the field of economics. Honi soit...
Is that because those studies are inherently wishy-washy, or as with many other scientific studies, the processes aren't actually detailed enough, or the code itself is missing?
At my day job I work a lot with data, from a multitude of sources. It amazes me how hard it is to build data products (you could think of it as an analog to scientific experiments) and know exactly what parts of the processing pipeline influenced what parts in the result.
On the other hand: When I hear that many scientists use Excel for complicated processing chains and are sometimes themselves hardly able to reproduce anything a few month after the paper was published, I believe we hit some new low point in scientific activity. Combined with an infrastructure focused on research impact and number of papers published, this is really a sad situation.
Real science is random, see No Method. When you make it a career that has requirements and expectations you set this sort of process to take place. Still even through this process, if enough people are doing it, there is bound to be some breakthrough science. But most of the people are wasting their time cooking data to make their careers go further, rather than make progress in science.
I have not worked directly in research, but that has been my experience when trying to reproduce results from papers: it hardly works even after following each step the researcher did (confirmed via email) and a lot of conditions that are not true in the real world have to be met.
I assume you're referring to "Against Method" by Feyerabend?
His argument is absolutely not that science is "random". It's that the methods of science cannot be predicted, prescribed, or universal, even over very short time scales and between similar fields.
Even if Feyerabend is right [0], it's still perfectly reasonable to expect studies to be repeatable.
[0] I'm sure he is but I think that the degree and practical significance of his observation is questionable.
Yeah, I have seen this in person. What is more sad, is that the scientists using Excel are aware of the limitation but they just don't want to spend time learning to work with R or Python and csv files and version control systems, after all they rather spend this time for grant seeking. For them convenience trumps reproducibility, they just redo the analysis until one one could get the "right" result. This indeed is very bad for academia and drives smart people away from postgraduate school. I also encountered use of proprietary software that scientist written themselves just to prevent reproducibility and make it harder for competing research groups. How this people can get published in high impact papers is beyond my understanding.
The Planet Money podcast #677 (The Experiment Experiment) discussed an effort to reproduce some experiments.
One of the methods they discussed to both increase reproducibility and reduce experimenter bias was to register the experiment procedure and hypothesis with the journal before performing the experiment. It's been a while since I listened but I think one or more journals is supporting this workflow.
I'm glad we are working towards a better scientific process. These days sensationalism scores more grant money and Scientific American articles. We need incentives to improve our body of knowledge not just make headlines.
The article uses both the terms reproduce and replicate, but doesn't go into the difference, and seems to use the them with just the opposite meanings they've had in academic use - roughly speaking:
Reproduce - to verify a study's results by re-analyzing the same data.
Replicate - to re-do the whole experiment.
Though the terminology is not quite intuitive, and there are other relevant distinctions to make. See http://languagelog.ldc.upenn.edu/nll/?p=21956 for some discussion of the history.
I had never heard of this distinction, and I also dislike it. This definition of "reproduce" seems to fly in the face of the word's more general meaning. "Verify" and "re-analyze" seem like a more natural fit.
Yes, we need a way of describing those two different things. But using "reproduce" and "replicate" (which, as the article showed by its actual usage, tend to be seen as synonymous) does not seem like a good way to go.
> which, as the article showed by its actual usage, tend to be seen as synonymous
In the sciences, language often takes on an entirely different meaning than in popular lexicon. This is one of the reasons I cringe every time I hear someone state, 'It is just a theory'.
> "Verify" and "re-analyze" seem like a more natural fit.
At least in my field (computational science), verification already has a very specific meaning. Overloading terms is avoided, although that also happens.
In terms of successful communication between scientists, it is far more important for everyone to agree on the meaning of a term than what particular word is used itself. But perhaps this is also accounts for difficulties communicating important results to non-specialists...
Possible compromise: let the scientists have their terminology, and let general audience publications like the Economist have theirs.
Mostly I just don't like the idea that the Economist article's usage is "wrong" just because some scientists use different terminology. Different spheres of interest will always be tripping over each other's terms when they come into contact; too many pixels have already been spilled trying to remedy this inevitable occurrence (mine included!)
> An analysis of 98 psychology papers, published in 2015 by 90 teams of researchers co-ordinated by Brian Nosek of the University of Virginia, managed to replicate satisfactorily the results of only 39% of the studies investigated.
The Economist is overstating the results a bit. From the coverage at the time:
Strictly on the basis of significance — a statistical measure of how likely it is that a result did not occur by chance — 35 of the studies held up, and 62 did not. (Three were excluded because their significance was not clear.) The overall “effect size,” a measure of the strength of a finding, dropped by about half across all of the studies. Yet very few of the redone studies contradicted the original ones; their results were simply weaker.
Reproduction, replication, are words not accurate enough. The point is that access to the data, the code, procedures, and so on is the rigorous way towards independent validation.
The best part of the scientific method is that any scientific contribution has to pass the filter of independent validation. Now it became technically possible for the authors of the research to give all they have in order to facilitate independent validation.
This is very good news for the researchers, because it is much more feasible for them to make open their results than to wait for the publishing system to recognize that we are in the 21st century.
31 comments
[ 4.7 ms ] story [ 83.8 ms ] threadThis in turn has lead to the whole "publish or perish" environment, as the MBAs use published articles as a replacement for widgets made and citations as sales.
Completely agree, and it is actually even worse than you imply: if you think it is hard to get scientists to reproduce others work, imagine trying to pitch NSF/DOE, etc. to fund the study!
[1] http://www.talkorigins.org/faqs/comdesc/
[2] http://www.cep.ucsb.edu/primer.html
[3] http://www.larspenke.eu/en/publications.html
I am talking about the following criterion in order to qualify as a real scientifically tested theory:
http://www.stephenjaygould.org/ctrl/popper_falsification.htm...
According to this criterion, the theory that random mutation and natural selection alone produced all the genotypes we see is like Marxism and Freud's psychological theories. If we assume they are true, we can explain everything with just-so explanations. But that's not falsifiable, it doesn't make testable predictions. NOTE: this is different than saying all animals had common ancestors. It says we know how the changes happened that led to the current features. The fact is, we don't know. We barely have scratched the surface. We are still in stages comparable to the luminiferous ether theory in ohysics or humors in biology. I think in 100 years people will look at our theories of random mutation and natural selection (whether classical, punctuated equilibria etc) and consider us naive. And yet we have proponents like Dawkins yelling from the rooftops that "evolution is as proven as gravity". That's equivocating the word "evolution" with common descent. There is also a lot of pressure and political interest from demand for a naturalistic explanation, similarly to how there is demand from the religious camp for "the institute for creation research". It affects who gets research and who gets published. Science is a human endeavor and it is affected by political and organizational pressures just like everything else. But when something is testable and is being tested, it's obvious. Here nothing is obvious - even the math is dubious. These theories are just riding on the coattails of the theory of common descent because we don't have better naturalistic explanations at the moment. But just because a theory is the best we have doesn't mean it's true. Overstating it is, is more advocacy than science.
Just like we didn't need to know about the X' of Einstein's equations to know that the X gravity was the primary force holding you to the ground.
There very well may be an X' that effects observable biological features other than evolution, but that X' needs to be tested on an individual basis. So while it's productive to hypothesize an X' and test it, there's not much we can do to rule out all X' as a class.
On the other hand: When I hear that many scientists use Excel for complicated processing chains and are sometimes themselves hardly able to reproduce anything a few month after the paper was published, I believe we hit some new low point in scientific activity. Combined with an infrastructure focused on research impact and number of papers published, this is really a sad situation.
His argument is absolutely not that science is "random". It's that the methods of science cannot be predicted, prescribed, or universal, even over very short time scales and between similar fields.
Even if Feyerabend is right [0], it's still perfectly reasonable to expect studies to be repeatable.
[0] I'm sure he is but I think that the degree and practical significance of his observation is questionable.
One of the methods they discussed to both increase reproducibility and reduce experimenter bias was to register the experiment procedure and hypothesis with the journal before performing the experiment. It's been a while since I listened but I think one or more journals is supporting this workflow.
I'm glad we are working towards a better scientific process. These days sensationalism scores more grant money and Scientific American articles. We need incentives to improve our body of knowledge not just make headlines.
Also alot of biological labs do try to reproduce experiments to further their own research. They just dont publish negative results.
Reproduce - to verify a study's results by re-analyzing the same data.
Replicate - to re-do the whole experiment.
Though the terminology is not quite intuitive, and there are other relevant distinctions to make. See http://languagelog.ldc.upenn.edu/nll/?p=21956 for some discussion of the history.
Yes, we need a way of describing those two different things. But using "reproduce" and "replicate" (which, as the article showed by its actual usage, tend to be seen as synonymous) does not seem like a good way to go.
In the sciences, language often takes on an entirely different meaning than in popular lexicon. This is one of the reasons I cringe every time I hear someone state, 'It is just a theory'.
> "Verify" and "re-analyze" seem like a more natural fit.
At least in my field (computational science), verification already has a very specific meaning. Overloading terms is avoided, although that also happens.
In terms of successful communication between scientists, it is far more important for everyone to agree on the meaning of a term than what particular word is used itself. But perhaps this is also accounts for difficulties communicating important results to non-specialists...
Mostly I just don't like the idea that the Economist article's usage is "wrong" just because some scientists use different terminology. Different spheres of interest will always be tripping over each other's terms when they come into contact; too many pixels have already been spilled trying to remedy this inevitable occurrence (mine included!)
The Economist is overstating the results a bit. From the coverage at the time:
Strictly on the basis of significance — a statistical measure of how likely it is that a result did not occur by chance — 35 of the studies held up, and 62 did not. (Three were excluded because their significance was not clear.) The overall “effect size,” a measure of the strength of a finding, dropped by about half across all of the studies. Yet very few of the redone studies contradicted the original ones; their results were simply weaker.
More here:
https://news.ycombinator.com/item?id=10132993
The best part of the scientific method is that any scientific contribution has to pass the filter of independent validation. Now it became technically possible for the authors of the research to give all they have in order to facilitate independent validation.
This is very good news for the researchers, because it is much more feasible for them to make open their results than to wait for the publishing system to recognize that we are in the 21st century.