The mathematics of science’s broken reward system
nature.com
nature.com
The reason why scientists have a natural "disdain" for sociological studies, is because they don't have these qualities. Many of the arguments aren't convincing from a critical angle, and only convincing to people who have a pre-existing bias towards certain conclusions. Yet the scientific culture of today is falling into that trap itself.
The article criticises metric-driven incentives, but it is not metrics (the general concept) that are at fault. It is the choice of metric, and the meta-analysis of this that is lacking. These choices are themselves often backed up by non-scientific vague arguments, of the similar sort that scientists often criticise other fields as depending upon.
We must certainly not conclude from these studies that quantitative analysis is itself what is at fault. I know the article doesn't explicitly say this, but it hints at it - using suggestive phrasing like "disdained sociological studies" and referring to all "metric incentives" as a single group - and it is a point I have seen made by many non-scientists. That is, using these flaws as a straw man to attack the very qualities of what has made science so successful and useful.
To improve the situation, we must reject these straw-man arguments against science, and develop better methods that are more quantitive, over a broader spectrum of what is being analysed, and that are more self-critical.
There are plenty of brilliant scientists who are held back by their excessive rationality, or whatever you want to call it. Another great book on this topic as it relates to econometrics (as opposed to physics) is "The Romantic Economist" by Nicolson.
Sociology is very important to study. The tools we have for it are not great, in various ways. For example, one can study people and societies by reading Balzac's writing, but the number of people who can produce that sort of thing is fairly limited. We can try to do controlled experiments, but the way we do it in practice is not great. We really really need better tools here...
So yeah, personally, I have utmost respect for the complexities involved in sociology - while at the same time I absolutely hate all the bullshit that's being done because doing actual research feels too hard.
There is no stereotypical "scientists' disdain" for those topics. It is vital to understand history, social sciences, economics, politics, and the arts. What is disdained is the way these topics are, in the present world, not pursued along more rigorous lines.
Too many people like to look good by producing grand-sounding theories about these topics, without putting their ideas through a more rigorous process to check whether they are actually true or not. Then we get a cohort of followers who believe these theories as if they were "as true" as mathematical theorems, but they are not. This wastes everyone's time, and worse.
One can certainly picture an overly-specialised scientist who doesn't know anything other than the specific field they chose - I guess this is what you mean by "excessive rationality" - but I don't see many cases of this actually existing. What I see in far greater numbers is the problem I mentioned in the previous few paragraphs, as well as people using the "excessively rational" image as a straw-man argument to attack scientists, to distract the world from their own flaws, of "not enough rationality".
However, it's too easy for bullshit-artists to hide behind this, as a way to deflect criticism of the theories that they're proposing, or the general area of work that they've chosen to pursue.
We're all human so it's perfectly reasonable to believe in something that's unproven - but then be honest about it, and explore the topic from a critical viewpoint without getting personally defensive. Unfortunately I see many more cases of this, than instances of the fallacy-fallacy.
Edit: maybe he was a nihilist, I'd tolerate nil as the only non quantitative number. Sorry for the rant, but that's what you get for anecdotal evidence.
Feynman certainly did.
It's not obvious to me that measuring scientists or research quality has anything to do with the "spirit of science." Just because science involves very carefully measuring the objects of study, it doesn't follow that science is best served by trying to meta-measure that process. And there is evidence that it can be harmful.
And pandemonium models [1] of computation show that the right rules of cummunication can produce astonishing order from chatic self interest. If scientst were forced to pre publish their methods and share all data -- science would probably work better. This is true even if this change had no effect on their behavior, or even if they made every effort to game the system.
This imrpovement to science doesn't ask a committee to define or measure anything ineffable, and it doesn't expect individuals to change thier behavior. It just changes the rule of interaction in a way that better favors the systems epistomological progress.
Of course the more noise you have, the less efficient the system is, so it's good to incentivize people to do honest, objective work - thus pre-publishing / pre-registering, sharing data and algorithms, etc. are all good and important goals. But so should be changing the metric affecting the aggregate - like making sure scientists are actually incentivized to replicate previous work.
People just don't like this real solution because it means many (really, most) of the wild speculations they hold as dear dogma (and have been calling science) will need to be reassessed using the correct metric. There will undoubtedly be a lot of egg on many faces.
Our need for "objectivity" here is to a large extent driven by the centralization of the funding allocation that http://mcadams.posc.mu.edu/ike.htm warned about and has since come to pass. If funding were more decentralized, lack of objectivity in any particular funding source would be _much_ less of a problem.
There's lots more in there that usually doesn't get quoted (unlike the bit about the military-industrial complex) and is spot-on prescient.
We follow these principles out of observing history. But part of it is an act of faith, I'll give you that. What sort of "evidence" are you referring to, that it would be harmful?
A problem is that any such metric is only ever going to be a proxy for some assumed real value, and such a proxy is always going to be controversial, particularly if it's used to control access to funding.
Which I think at the end of the day is the big issue here. All this soul-searching is just a big attempt to decide what we spend money on. Who gets to be a scientist, and who gets to become a person with a lot of useless knowledge and an incurable sense of failure.
And who goes into finance, software engineering, or some other tangentially related field where the skillset is very useful.
Certainly we need to look at our social reward/incentive systems from an adversarial point of view, in science, economics, politics and elsewhere. Game theory is useful for that, and it can help us develop systems that are less-easily susceptible to the flaws that the article mentions.
Academic scientific activity is still mostly a social activity because it is done through the communities and the social interaction between scientists and institutions. Specific problems may be solved by specific brains or groups of brains, but the output and the collaboration is all social. The rewards and incentives are all social systems. And at the end of the day, any time we write a blog or post a comment, it's all social. This is social.
But the disdain for social sciences by non-social scientists is discrimination only warranted to the extent they can avoid social and professional interactions with them, and feel immune from any criticism for their bigotry. The moment a scientist wishes to study the social aspects of science itself, they have no choice but to accept social science as a science.
But this is okay because reputation has no consequence in problem solving, because reputation is a social device, and not a solvent. The track record of a field, a department of an institution, the history of publications, or the publications of an individual social scientist are all largely irrelevant when faced with your own research targets. The only thing relevant is the available data and the premises chosen for any model. There is no need to criticize the scientific integrity of the work of others if you can do better. Do it yourself if you have to.
Actually I thought yesterday about how to misconstruct a human detector for photos, one possibility would be to have face detection and then measure the average pixel color of the face. That would get me probably quite nice detection rates if I test on a dataset that has a racial markup as the average CS lecture. The thing here is, that I push the false negatives purposely to people of color. It is not hard then to invent a story about how 'mathematics proves' that black people are more similar to apes than white people.
We currently just don't have anything to ground models of human interaction in,^1 and what is worse most people treat models just as previous generations treated prophecies. They don't understand math and believe in it because they don't understand it.
^1 I am actually not saying that models are worthless, I am saying that most complex models only show what the author wants them to show and they can be as easily manipulated as an essay.
Someone goes and builds a bridge. The bridge collapses because they applied mathematics in a way to maximise profit, with a safety margin just beyond what they can get away with by existing regulations. 5 million other bridge engineers also do the same, so that eventually bridge engineers get a bad rep, and people think bridges are awful awful things. This doesn't mean mathematics or quantitative analysis is at fault - and in fact these tools can be used to examine the incentives and other social dynamics that led to these situations, to be able to fix them more effectively in the future.
no, not of gravity or dark matter.
As for dark matter, the reason we don't have a good idea what dark matter is, is that is has almost no effect on normal matter. So the reason why we can ignore dark matter everywhere except at the frontiers of physics is, that dark matter only matters when one spends awesome resources to try to detect it.
No, you can't. What an outrageous claim to make. Prove it constructively, preferably.
I disagree, because of https://en.wikipedia.org/wiki/Goodhart%27s_law
The problem is that we want "quality science" (whatever that means!), but we don't know how to quantify that, or indeed really define it. So a quantitative metric will necessarily be measuring some sort of proxy or set of proxies for "quality", and then you will get people optimizing those proxies, not "quality". To the extent that those proxies miss something important, it will be underinvested in.
Unfortunately, I don't have a better proposal, perhaps short of taking the warnings in http://mcadams.posc.mu.edu/ike.htm to heart (the ones that are NOT about the military-industrial complex). Doing that might change the general funding climate sufficiently that the need for deciding "quality" like we do now may simply become less critical.
I've run into that myself, as described here: https://news.ycombinator.com/item?id=12271097
Mechanism design (reverse game theory) is subfield of game theory deals with these issues.
Edit. Prof. Google to the rescue [1].
They need something like a "Journal of Meta-Research", which would research the best ways to perform research. You'd think something like that would get a lot of funding...
You're probably being too kind here. The moment of actual falling was decades ago, or more.
The situation is only made worse by the fact that universities reward this; because they too are rewarded more for high-profile research than they are for scientific validity.
My guess is that most studies in softer sciences are simply erroneous. Many social sciences have already been shown to be unreliable by the replication crisis.
My personal experience comes from reviewing hundreds of published studies that evaluated the applied effectiveness of machine learning models. Half had significant statistical errors.
In the social sciences? Just curious.
> Half had significant statistical errors.
Such as? Have you enumerated the typical errors somewhere, so that other can learn to avoid these common pitfalls?
As for a place that has common stats errors and how to fix them? Well, just about every single intro to stats book, the entire R language, just about any library for just about any language, google, etc. The issue is not that the folks are making honest mistakes, that happens, it's that the system is perverse and incentivizes them.
You mean this? https://cran.r-project.org/web/packages/statcheck/index.html
It's obvious that citation counts are open to manipulation (e.g. through self-citation and encouraging citations from others via various mechanisms) but there's also a very large factor that can't be easily quantified: an enormous amount of good will has been lost in the system.
Many scientists who are interested mainly in satisfying their curiosity and contributing to society resent the top-down mismanagement of science in the UK and US. They see a system where those who play political games do well, but the smartest, most dedicated, and passionate researchers are often sidelined or ignored because they do not spend the requisite time playing the political system to artificially boost their reputations. Many times I've seen very talented academics with huge potential leaving academia as they were passed up for promotion, left on a temporary contract indefinitely, or neglected in other ways.
Strangely enough, I think that if you don't have pressure to climb the ladder (e.g. a family), then the current system offers great opportunities to do outstanding science. If you choose to stay low on the ladder and spend the vast majority of your time actually focusing on research problems, you can get far more actual research done than those chasing promotion or esteem. And because so many people are focused on their citation count, or some political game, the competition isn't as strong as it should be - you can make real advances if you quietly focus and leave the politicians to fight each other.
Never really got past 'Scientology', and that probably won't stick.
What's a better term than the vapid 'Metascience', or the clunky 'scientometrics'.
How about 'Superscience'? That would be an awesome Doctorate to hang on the wall.
There's no reason why not to mix theoretical and practical (as in data-driven) tools.
“Philosophy of science,” already mentioned, is probably a better fit. Once you go meta on science, after all, you are (arguably) not doing science anymore, so there is no harm in naming it a kind of philosophy.
Things have only been getting worse and worse as the "old guard" in each field retires/dies leaving behind only people trained to think rejecting a strawman hypothesis according to an arbitrary metric, then concluding something about your hypothesis, counts as science.
ie Hate the game, not the player.
Did anybody construct it, or has it just emerged out of the surrounding world?
Is it the same across the planet?
The current system was deliberately constructed by politicians after WWII when we had more money (and demand for science) than scientists. They set out to create more scientists by designing a system where each scientist trained 10 to 100s of additional scientists (PhDs). This worked great until around the mid 1970s when the supply of scientists finally caught up with the money. Since then the problem has been getting worse.
Here are some I can list: P-values, the h-index (article discusses), the funding crunches and (US specific) the cyclic nature of NIH funding, the massively skewed incentives of publish or perish, the entire idea of trying to assign value to discoveries (article discusses), the 2-body problem of family life and science (at least the Germans expect you to have no life and are clear on that), the work-life non-balance, the relatively very poor pay, the hyper-competition and sabotage (in some fields, labs, or universities), the total ignorance of doctors/politicians that take your work and apply it in appalling ways, the entire racket of scientific publishing, etc.
As always, we should point out to newcomers to this discussion of one of the most read scientific paper EVER by John P. Ioannidis "Why Most Published Research Findings Are False" [0]. If you don't have access and don't want to use Sci-Hub or ICanHazPDF, here is a youtube of John discussing his work [1]. You can use the google to really jump into a rabbit hole here. Essentially, flip a coin, heads the paper is right, tails the paper is wrong, it really is that bad these days.
Ok, so lets just scrap the whole thing then, right?
Here is why not:
https://www.youtube.com/watch?v=_Px5sZyxFYc
https://www.youtube.com/watch?v=XcPuRaSEq1I
https://www.youtube.com/watch?v=5EL_bLOK8_A
https://www.youtube.com/watch?v=qXWYSdijCGU
https://www.youtube.com/watch?v=uBh2LxTW0s0 (a good showing of why science is right and helps folks)
https://www.youtube.com/watch?v=gs0JQRT6TpY
So, yeah, science is a mess right now. But if we scientists don't step up and make it not a mess, those scummy money-grubbing scammer dirtbags and their ilk are more than happy to take up the slack. These fucking asshats are going to kill people like you poor uninformed cousins and their kids and take all their money because they think homeopathy is right and the moon landing was faked.
So, you young scientist that looks at this pile of garbage that is modern academia, do not despair! Fight the good fight! Yes, you may end up shirtless and ridiculed by the older scientists. But you have to do what you think is right because the rest of this world is depending upon you! Even for that tiny little bullcrap paper you are getting out just to graduate, that matters too. Be in the mud, be in the arena, fight for truth!
[0]http://journals.plos.org/plosmedicine/article?id=10.1371/jou...
Where would you not have access to a PLoS paper? The entire company is dedicated to open-access publishing.
First link in Google pops up the PDF, FWIW.
http://robotics.cs.tamu.edu/RSS2015NegativeResults/pmed.0020...
(It is a wee bit disturbing that first link isn't to PLoS)
Post all your papers as preprints. Don't bottom for some grumpy old PI. Academia is not scholarship -- the latter is frowned upon within the former, as it takes longer than simply churning out some useless publoid garbage result.
Be like Faraday, Newton, Varmus and Eisen... not like Bem.
They have already taken over healthcare and passed laws to force you to pay for their "help". To get you started: people are quitting cancer reproducibility projects out of disgust for the low quality, and just trying to figure out wtf was done to generate the data is draining all the funds before they can even attempt replication:
"Early on, Begley, who had raised some of the initial objections about irreproducible papers, became disenchanted. He says some of the papers chosen have such serious flaws, such as a lack of appropriate controls, that attempting to replicate them is “a complete waste of time.” He stepped down from the project's advisory board last year.
Amassing all the information needed to replicate an experiment and even figure out how many animals to use proved “more complex and time-consuming than we ever imagined,” Iorns says. Principal investigators had to dig up notebooks and raw data files and track down long-gone postdocs and graduate students, and the project became mired in working out material transfer agreements with universities to share plasmids, cell lines, and mice.
[...]
ALTHOUGH ERRINGTON SAYS many labs have been “excited” and happy to participate, that is not what Science learned in interviews with about one-fourth of the principal investigators on the 50 papers. Many say the project has been a significant intrusion on their lab's time—typically 20, 30, or more emails over many months and the equivalent of up to 2 weeks of full-time work by a graduate student to fill in protocol details and get information from collaborators. Errington concedes that a few groups have balked and stopped communicating, at least temporarily."
http://www.sciencemag.org/content/348/6242/1411
Also, it seems 4/5 doctors will tell you there is 95% probability you have a disease when it is actually 98% chance you do not have it, 40 years of education reform has not affected this at all (admittedly, this needs to be repeated on a larger scale, but from personal experience I have no doubt it will hold):
"Nearly 40 years ago the New England Journal of Medicine published a short survey of doctors’ understanding of the results of diagnostic tests.1 The participants, all doctors or medical students at Harvard teaching hospitals, were asked, “If a test to detect a disease whose prevalence is 1/1000 has a false positive rate of 5%, what is the chance that a person found to have a positive result actually has the disease, assuming that you know nothing else about the person’s symptoms or signs?” This wasn’t a very difficult question, which made the results all the more shocking. Fewer than a fifth of participants gave the correct answer, and most thought that the hypothetical patient had a 95% chance of having the disease.
Of course, this was a long time ago, and medical curriculums now contain much more in the way of statistics and probabilistic reasoning. You might expect that if the exercise were repeated today almost everyone would give the right answer. But you’d be wrong. Earlier this year a similar study was carried out, also in hospitals in the Boston area of Massachusetts, and the results were no better.2 Most doctors who were asked exactly the same question thought that the patient had a 95% chance of having the disease." http://www.bmj.com/content/349/bmj.g5619
Medical errors were "unintentionally" left out as a possible official cause of death, then when people actually estimate this, it is a leading cause:
"In 1949, Makary says, the U.S. adopted an international form that used International Classification of Diseases billing codes to tally causes of death...medical errors were unintentionally excluded from national health statistics...based on a total of 35,416,020 hospitalizations, 251,454 deaths stemmed from a medical error, which the researchers say now translates to 9.5 percent of all deaths each year in the U.S...According to the CDC, in 2013, 611,105 people died of heart disease, 584,881 died of cancer, and 149,205 died of chronic respiratory disease—the top three causes of death in the U.S. The newly calculated figure for medical errors puts this cause of death behind cancer but ahead of respiratory disease."
https://hub.jhu.edu/2016/05/03/medical-errors-third-leading-...
I could go on if you are interested.
http://www.nature.com/nature/journal/v497/n7450/full/497433a...
This is the guy that published the irreproducible results paper whose results were (of course) irreproducible (since it did not name any). I.e., "trust me on this", the opposite of science.
Note above that you can pay for more of his sage wisdom, thanks to the generous NatureMacSpringer megaconglomco.
Post everything as a preprint. Submit your trial results to OpenTrials or OpenFDA. Quit killing patients. Maybe work in pediatrics so that, right or wrong, you still won't make any money, so you might as well not kill any excess kids.
I'm not bitter or anything, after participating in clinical trials and analysis for over a decade...
I'd love to hear your story, up to you.
Take away the profit motive and it's astounding how much more reliable the trials get. Of course reviewers whine about how they're often equivocal or negative, because the thought of the literature being horribly biased either
a) has never crossed their mind (too busy with protocols),
or
b) is precisely what they want, for "showing progress".
So in order to keep up the appearance of positive results and keep the gravy train flowing, the trials that get published tend to be the ones that "show" "progress".
Mind you, these are clinical trials with preregistered endpoints and protocols. Pretty much the entirety of experimental science outside of physics is much worse. So what I'm telling you is that this is the BEST-CASE SCENARIO for an awful lot of science.
I'm not going to write up "my story" beyond this, because it's the same as most everyone else's story in the field. Killing humans for profit is not my idea of a good time, even if adult trials tend to pay better than the Army.
YMMV. I may be jaded, but I still believe there are populations that, through no fault of their own, end up with few or no options. There is room in this world for integrity -- but get (real) money involved and the bad money drives out the good. The linked paper is dead on.
The most disturbing thing appears to be that many folks aren't aware of Foucault's work.