Slashdot Mirror
A New Record For Scientific Retractions?
sciencehabit writes "An investigating committee in Japan has concluded that a Japanese anesthesiologist, Yoshitaka Fujii, fabricated a whopping 172 papers over the past 19 years. Among other problems, the panel, set up by the Japanese Society of Anesthesiologists, could find no records of patients and no evidence medication was ever administered. 'It is as if someone sat at a desk and wrote a novel about a research idea,' the committee wrote in a 29 June summary report."
news at 11pm
During World War II, Americans were very keen and excited to get their hands on scientific data from the Japanese after nuking them, especially all the data from human experiments which were not feasible in US. When they got the data, they realized most of it was non-sense. They had been randomly doing experiments on humans without any clear hypothesis or theory and most of the data did not make much sense.
And this, ladies and gentlement, is why real science is done by not only performing the experiement and recording the results, but by writing up your method with sufficient clarity that your results can be replicated by independent researchers.
Once that has been done sufficient times, if your method itself is sound, then the results are valid.
From TFA
German anesthesiologist Joachim Boldt is believed to hold the dubious distinction of having the most retractions—about 90. Boldt's scientific record also came under fire several years ago by some of the same journal editors questioning Fujii's work.
Is this coincidence or a pattern? I have no idea how the journal publishing is supposed to work, but being the "victim" of the two most prolific forgers leaves me a little suspicious of the quality of the publishing in general.
I am Slashdot. Are you Slashdot as well?
Peer review is not designed to catch fraud, although it can, as to work on the assumption that the work may be fraudulent would cost too much and would not even be effective against cleverer frauds. The only way to catch a clever fraud is to try and replicate their work, and this can only happen after publication, usually when another researcher tries to build on the original work. If you do this all fraud will eventually be caught, the best you can hope for in the long run, as a scientist committing fraud, is to be thought of as critically incompetent. For this reason fraud is rare among academic scientists, but is unfortunately more common among their commercial counterparts.
Except Americans. Unless it matches their beliefs in religon, politics, nature and economics.
So this guy was writing, what, approximately nine or ten papers a year on average? Was anyone paying attention? Didn't anyone notice something strange about his "discoveries?"
What does that say about the field of academic medical research?
Nearly fifty percent of all graduates come from the bottom half of the class!
he has 50 years of education, anything he writes is fact
20 years from now people will be saying the same thing about this supposed global warming. in the northeast it has actually been cooler than 30 years ago when i was a kid. almost every ridiculous theory about super hurricanes destroying NYC by 2010 have not happened.
To summarize: straw man, nonsense, nonsensical anecdote that doesn't matter even if true, straw man.
"I zero-index my hamsters" - Willtor (147206)
Not really, I've heard that he is so boring that most people who met him ended up sleeping.
Unfortunately, it is upon publication when a study is picked up by the media and exposed to the general public. By the time other scientists try to replicate the experiments and find they're bullshit, it's "too late" in a sense.
Some sort of independent verification needs to be worked into the process before a new study is put out there for general consumption. This either means before "publication" or the media needs to learn (hah) not to cite studies that haven't been independently verified, no matter how sensationalistically important they sound.
Liberty in your lifetime
It seems they were, but catching fabricated results like these isn't exactly easy and it won't happen in the review process.
In order to catch fabricated results you'd either have to repeat the experiment, which nobody wanted to do since the research was low impact, or catch discrepancies in the data, which was how he was caught out.
Your comment has already been made, and addressed, in this discussion.
Yo dawg, I heard you like the Ackermann function, so OH GOD OH GOD OH GOD
So much for frauds being caught by peer-review, huh?
That's an odd comment for an article about the peer-review system catching a fraudster. The system will find it eventually, though the time scale can be quite large, especially if you don't publish in a 'hot' field. You can't expect the peer review system to catch every bit of fraud as it comes in, it doesn't appear like a glowing fireball in the sky. This is likely a small amount of fabricated data about medical proceedure that didn't happen. Moreover, the author is probably quite bright (just lazy) and had a good idea of what would be expected in those experiment, so it is easy to make up reasonable data.
"Some sort of independent verification needs to be worked into the process before a new study is put out there for general consumption."
The media, and the public, need to learn. Publishing and dissemination are a critical part of science and shouldn't be compromised to make some reporters' jobs easier. Fraud isn't even the big problem with jumping to conclusions based on unverified studies - FAR more studies will be incorrect simply due to honest false positives than to fraud.
I'm not so sure that fraud is more or less prevalant among academic scientists than commercial scientist. As you alluded to, peer review is not designed to catch fraud. If a researcher published in an obscure backwater field that no-one would likely try to replicate, a researcher could go on for a long while w/o being caught or even being considered incompetent. The same is true for a commercial scientist (witness the cosmetic and nutritional suppliment industries).
On the other hand, scientist in "hot" fields will have hoards of researcher in that field and if you do research in an uninteresting niche in that field, you might also escape detection for a while (or not, as bayer and pfizer have recently leaked out, most research that they have internally attempted to replicate to try to find new avenues for drugs had non-replicable results).
People are people no matter in academia or industry...
A professor at my local university periodically gets undergraduate and starting graduate students to try reproducing the work of interesting research papers.
One engineering professor at my local institution figures about one-third of the papers can be reproduced to demonstrate the effect in question.
At most schools, graduate students are required to published paper on a "new" idea in an academic journal in order to receive their degree. As such, journals must exist to collect all the ideas students generate, and this is the driver behind the modern academic journal system. Huge pressure is put upon the students to describe "new" ideas, and as such, the paper must sell itself as being "new".
Complicating this effort, it the reality that most students are working on student projects. These projects don't have the necessary resources (time and money) to be developed into fully effective and reproducible ideas. As such, the results from these projects are fundamentally suspect. Also, the students working on the project, may not fully understand all the relevant effects on their research (because they are students). In particular, many students do not understand statistics. As such, students deliberately or inadvertently conduct biased experiments to show the desired effect in question, because the academic requirement is a "new" idea and not a "new and reproducible" idea.
The result is a collection of papers that all describe themselves as having "new" and "brilliant" ideas on topics that cannot be easily reproduced. When the ideas are reproduced, practicing engineers quickly discover they are reproducing a marginal student project. It is actually really tough to find reproducible, inventive and commercializable research ideas in academic journals, because of all the noise.
I disagree about the rarity, on the basis of empirical evidence, for example the recent paper in Science (IIRC, sorry I don't have the reference handy) in which a cancer researcher failed to replicate 46 out of 53 papers published -- all of them with peer review -- prior to embarking on new research in the field. Similar meta-studies have turned up astounding rates of non-reproducible results in other fields (some more than others -- sociology and IIRC social psychology topping the non-medical list).
One problem is that we have constructed a system that rewards the publication of positive results and punishes negative results published or unpublished. Punishes as in makes or breaks the entire career of young researchers, if the negative result occurs when they are up for tenure. Rewards as in ensures research funding and professional advancement as long as positive results keep flowing out.
Another fundamental problem that peer review has a terrible time with is confirmation bias. Science in general has a serious problem with confirmation bias. If one ever embarks on a study where one seeks evidence for some causal linkage associated with some phenomenon in a general population where the phenomenon occurs, one can always find exemplars that support your hypothesis. Lacking actual work to replicate your results using sound methodology (e.g. double blinded and/or conducted using competent statistical analysis, something still as rare as hen's teeth in science in general because to it is difficult to do statistics correctly in a complex problem, not easy, and certainly not easy as in covered in one or two undergrad stats courses which is all that it is probable that the researcher has ever taken) confirmation bias can not only worm its way into the literature, it can come to dominate entire fields as a significant fraction of scientists who do the reviewing for both publication and grants are "descended" from one or two original researchers and their papers. It can take decades for this to be discovered and work out in the wash.
Peer review works better in some disciplines than others. Math it works well, because there is literally nothing up a publisher's sleeve -- fraudulent publication is indeed impossible and even mistaken publication is relatively rare and conditional on involving math so difficult even the reviewers have a hard time following it. Physics and the very hard sciences are also fortunate in that it works decently (although less perfectly), at least where there is competition and the proper critical/skeptical eye applied to results new and old. At least there a mix of laboratory replication and strong requirements of consistency usually keep one out of the worst trouble.
A simple rule of thumb is: The more a result relies on population studies, especially ones conducted with any kind of selection process or worse selection process plus the actual modification of the data according to some heuristic or correction process, where the study itself is conducted from the beginning to confirm some given hypothesis, the more likely it is that the result (when published) is bullshit that will eventually, possibly decades later, turn out to be completely wrong. If you have enough places for a thumb to be subtly placed on the scales and the owner of the thumb has any sort of vested or open interest in the outcome, it is even odds or better that a teensy bit of pressure will be applied, quite possibly without even the intention of the researcher. Confirmation bias is not necessarily "fraud" -- it is just bad science, science poorly done.
There is a move afoot to do something about this. We know that it happens. We know why it happens. We know a number of things that we can do to reduce the probability of it happening -- for example requiring the open publication of all data and methods contemporary with any paper produced from them, permitting absolutely anybody to look at them and see if t
Even when the experts all agree, they may well be mistaken. --- Bertrand Russell.
Faked studies are only detected if someone attempts to reproduce them. People will only try to reproduce them if journals adopt a policy of publishing papers that are either confirmations or refutations of prior studies. On the whole this isn't the case.
I'd like to know how many of the studies could have been detected as fake through thorough enough statistical analysis of results - humans are notorious bad at faking data, even when they're trying their hardest to make it believable (as they then make it too believable).
Also FatPhil on SoylentNews, id 863
The whole point of the scientific method is that putting work out for general consumption is the best avenue for independent verification (to adapt a phrase familiar to this audience, one might think of it as "with many eyes, all non-reproducible results are shallow".)
The fact that reporters covering science in the popular media lack a basic understanding of the scientific method is a reason to change something, but the thing that needs change isn't scientific publishing.
Great summary here. As a young scientist I see this a serious problem for the credibility of science in general. The gross fraud cases seem to be mostly limited to a few fields - anaesthesiology in particular has had a few major retractions bouts recently.
As you point out, medicine and other fields involving population studies are much more prone to confirmation bias. In a similar vein, any field where the cutting edge involves extremely expensive experiments is open to direct abuse or failure of scrutiny to discover mistakes because it's prohibitively expensive to replicate experiments. Open data is one part of the solution to that problem, but to understand the data you need a good precise methodology published along with it, and often methods are lacking in detail to the point where they could never be accurately replicated. I think openness with data and methods need to go hand in hand.
The major lesson I've taken from all this is not to allow myself space for confirmation bias. In my field that means always performing the complete set of experiments to confirm the causative link you are exploring, not just getting a fat load of correlations. That needs to go hand in hand with a thorough understanding of the relevant statistics, not just blindly working with standard confidence intervals.
not according to TFA
insensitive clod overlords obligatory xkcd car analogy russian reversals whoosh pedant fanbois ftfy in 3...2...1..PROFIT
If you're going to fabricate 1 paper, you might as well fabricate 172.
Sheesh, evil *and* a jerk. -- Jade
I saw a study done recently where the author found that the results of many studies are quite difficult to reproduce, and he found that the more you tried to reproduce them and the more you talked publicly about your results, the more difficult it became to reproduce the results.
The problem is that researchers usually aren't approaching a study as "Lets do xxx and see what happens, then write about that". They've been funded by someone who has a particular result or proof point they'd like to see, or the study operator has a vested interest in the study outcome. At least an expectation of what they think they'll find.
Our happy little brains then lead us to that conclusion or desired outcome, and we'll gleefully ignore the things that detract from the results.
And yes, the guy who did this study of study results also found that his ability to reproduce his own results became more difficult as time went by.
For a couple of good examples of how this works, see the studies on salt and saturated fats in our diet. The intersalt study folks threw out 40% of the data that said that salt had no effect on health, suggesting that since its well known that salt affects your health that the people who weren't affected must be lying about their salt consumption. So almost half the data suggested no result, but it was discarded because it didn't fit with the desired determination. Same thing happened with saturated fats. The original researcher took 21 countries worth of data but only 5 of the 21 showed health issues that were allegedly correlated to the consumption of saturated fats. The other 16 showed no correlation at all. In fact, the real correlation was to high caloric, high sugar/carb, highly processed foods and health issues, not anything to do with saturated fats. There are cultures that eat 50-70% of their food intake as fat and they have little to no cancer, obesity or diabetes. Take one of those people and move them to the US or England and put them on our diet? They get fat and sick.
Of course, even when the study obviously sucks, the press can be counted on to come to conclusions that the study didn't even address.
Precisely. Indeed, a major part of the solution is to make scientists their own greatest skeptic, to mistrust our own pet ideas, to hesitate to claim "proof" to a fault even if evidence or model computations seem to support it.
The latter are an entire category in and of themselves. I "do" predictive modelling and moderately advanced statistics on a professional basis, and even have a patent pending in the field. I've done Monte Carlo computations in physics for well over a decade, and know a lot about randomness and hypothesis testing compared to your average scientist in the street, so to speak. I am all too aware that model computations are among the least trustworthy kinds of evidence and usually have far less predictive power than that which is claimed by the modeler. The problem there is subtle and related to complexity and nonlinearity. A highly multivariate, semi-empirical, nonlinear theory implemented as a model is often implemented by "fitting" some or all of its parameters to some (sub)set of data. This in turn is often equivalent in modelspeak to using hill-climbing (gradient search) to find an optimum fit to the data relative to some selected parametric starting point (this is sometimes referred to as making a "Bayesian" choice of the parameters based on some set of data used as priors").
There are many problems with this. One is the problem of omitted variables. In many of the problems where this is done, the choice of parameters (dimensions in the parametric space) is highly model dependent. Heuristics are often used to limit the size of the parametric space simply because doing anything in a really high dimensional space is a lot of work and introduces a substantially higher (but honest) estimate for errors in the final result. Heuristics, of course, is code for "I don't think these variables will significantly contribute", an open opportunity to omit variables that you don't want to be significant because they confound your hypothesis. A second problem is that many models are de facto parametric nonlinear function approximators. This means that -- especially if the data being fit to the parameters is "simple", e.g. monotonic or otherwise simply nonlinear over the range of the fit -- it is often perfectly easy to fit the data with a set of parameters, have the fit be "optimal", have the fit produce a perfectly reasonable chisq, and have the parametric fit be perfectly meaningless. This is all elementary modeling theory 101, but somehow "hiding" the basis by turning it into the solution of a set of coupled ordinary differential equations with nearly e.g. sinusoidal or nearly e.g. polynomial or exponential behavior makes the problem somehow disappear in the minds of the modellers. A third problem is that of complexity -- in many cases (especially for highly multivariate nonlinear models) there may be many local optima and hill-climbing from a selected starting point can easily be yet another form of inadvertent confirmation bias. A global search might find a better optimum, or might reveal that there are several constellations of parameters (especially when omitted variables are included) that can fit the empirical data within its precision, (properly) reducing confidence in the final prediction. A fourth is that even a model built with the variables that were important in the past (where the data being fit resides), that is robust in any parametric/Bayesian search, that uses any of several methods to "validate" the model (using past data) can easily fail in the future because the model simply does not extrapolate. The real space of variables and data is much larger, the model is always being built on some sort of optimistic projection onto a manageable subspace, and ignored stuff eventually becomes important and causes a complete deviation from the model. Chaotic models, stiff differential models -- there is no lack of examples, but somehow this sort of thing doesn't get factor
Even when the experts all agree, they may well be mistaken. --- Bertrand Russell.
Technically atheism is a "belief", since the absence of certain supernatural forces, and parallel universes purported to be accessible upon death isn't completely proven.
When people are prevented from attempting to carry out (nuclear tests are banned by international treaty), cannot (because they lack the means or large equipment like the LHC) or simply do not carry out experiments themselves (out of sheer laziness, or dropping out of school), then they must take the ones who actusally DO carry out scientific experiments at their word.
Scientists, then, take on the role of holy men, do they not? Isn't this where the fundamental conflict between science and religion emerges? Who are our social leaders, our bastions of sage advice? As a social problem, it's essentially the same conflict as with capitalism vs. communism. Who get to be "The Leaders"? The Government and/or Monarchs, or wealthy Corporate Executives who are "free"? With science vs. religion, it instead becomes a choice between The Scientists or The Elder Shamans.
A simple rule of thumb is: The more a result relies on population studies, especially ones conducted with any kind of selection process or worse selection process plus the actual modification of the data according to some heuristic or correction process, where the study itself is conducted from the beginning to confirm some given hypothesis, the more likely it is that the result (when published) is bullshit that will eventually, possibly decades later, turn out to be completely wrong. If you have enough places for a thumb to be subtly placed on the scales and the owner of the thumb has any sort of vested or open interest in the outcome, it is even odds or better that a teensy bit of pressure will be applied, quite possibly without even the intention of the researcher. Confirmation bias is not necessarily "fraud" -- it is just bad science, science poorly done.
The more interesting aspect of this is how economic facts are so rooted, mainstream and rehashed, using this very same process, that they become political ideology... Sad world...
I expect fraud will be most common where there is the most motivation and the most opportunity. Motivation can be a direct profit motive - so drug tests need special scrutiny. Motivation can also be for career / academic success - so universities might want to carefully consider policies that promote scientists based heavily on their number of publications.
Fields where small groups of researchers work on expensive to reproduce projects are more suspect. Medical tests requiring large numbers of subjects are an example. Research that requires unique equipment might also be suspect (say experiments at LHC), but fortunately most of those collaborations involve large groups of scientists, so fraud would me much more difficult to hide.
A particular problem with reports of this type of fraud is that it may encourage other researchers to do the same. If they see someone who had a successful career for 20 years, it might be tempting. I'm NOT suggesting that we cover up fraud, just that reporting may have an unfortunate side effect.