Even the most honest of scientists are regularly misled by their cognitive biases. They often go to great lengths to find proof for whatever seems logical, while dismissing evidence to the contrary.
Yet this issue is rarely discussed — because it remains an embarrassing subject.
Boyd Norton / Public Domain
In 2015, a group of researchers tried to reproduce the results of 100 different psychology studies but succeeded in less than a third of all cases. In 2012, a team of Californian scientists reported that, after attempting to reproduce the results of 53 haematology and oncology studies, they only succeeded six times.
How is this still called ‘science’ and not ‘guesswork’?
Professor Robert MacCoun of Stanford Law School thinks that the time is ripe to change our scientific tools in the face of this issue. After all, the scientific community has already faced similar obstacles.
In the mid-20th century scientists discovered that both researchers and their test subjects subconsciously changed their behaviour to ensure their desired results. This has led to the emergence of double-blind studies. The current situation is no difficult — and can be overcome.
Statistically relevant studies (with a p-value of 0.05 or less) are much easier to publish. But this relevance is hard to measure correctly, given the amount of junk data in multi-dimensional datasets.
Keith Baggerly, a statistician from the University of Texas MD Anderson Cancer Center, claims that with datasets this large, the analysis can no longer safely rely on conventional mathematical methods, let alone the unassisted human brain. ??Andrew King of the Dartmouth College says that modern software has given researchers a way to easily analyse large amounts of data without really understanding the process, and arrive at the desired low p-value, even if it’s essentially meaningless.
The research has become gamified to its own detriment. The scientists are just chasing the best results.
One of the traps researchers fall into early on is the lack of regard for conflicting points of view. Jonathan Baron from the University of Pennsylvania in Philadelphia points out that those who want to prove their point phrase questions in a way most likely to guarantee an answer they like.
This principle can be observed at work in the judicial system. Brits might remember the 1999 case of Sally Clark — the woman who was found guilty of double infanticide. Both her kids died within weeks of being born. The persecution claimed that the probability of both these deaths being natural was 1 in 73 million.
Three years later, her conviction was overturned thanks to a different set of stats. The defence argued that double infanticides are 9 times rarer than families that have two of their children die of SIDS.
There’s an old American joke about an inexperienced sniper who shoots at a bare wall before drawing a target to make himself look good. A similar approach to science is, sadly, quite common among researchers.
Uri Simonsohn of Ramon Llul University calls it “P-hacking”: manipulating pre-existing data until its p-value drops below 0.05. ??A 2012 study of more than 2,000 American psychologists showed just how widespread p-hacking is. Half of all the test subjects dismissed the research that didn’t fit their narrative, and 35% presented surprising findings as if they were expected from day one.
Processing data can be just as dangerous as picking it. When doing analysis, we pay a lot of attention to inexplainable outliers and neglect less obvious error indicators, which can be more important overall.
A 2004 study on the data processing habits of three of the leading molecular biology labs presented an in-depth look into this issue. 88% of all the experiments that didn’t meet scientists’ expectations led to discussions of potential methodology errors. Their findings were not treated as logical, even if they, in fact, were.?
You can even go a step further and present your findings as valid despite not reaching the significance threshold. All that’s needed is an alternative ‘just-so’ explanation to justify the deviation of research data from the what was originally expected. Matthew Hankins of the King’s College in London collected more than 500 phrases commonly used to mislead the reader and re-assert the validity of studies’ unreliable findings.
Dan Meyers / Unsplash
The results with a p-factor of greater than 0.1 can be described as “flirting with conventional levels of significance”. Go a bit lower and you got yourself a “borderline significant trend” (p = 0.09), or results that are “not absolutely significant, but very probably so” (p > 0.05).
Each of the aforementioned ways of thinking is fairly convenient to both the researcher and the scientific community at large. These traps provide motivation, and speed up the research process. Forgoing them would lose us a lot of time. But in order to battle misinformation, we need to slow down and embrace the difficulty that comes with looking for truth.
First of all, scientists need to make an effort to remain open-minded. They need to seriously examine conflicting hypotheses, and conduct experiments to determine their viability before discarding them.
You can even ask your academic rivals for research assistance! Scientific environments that embrace diversity of opinions are conducive to early detection of logical and methodological errors.
Open-access journals also help. Everyone benefits from being able to access others’ research-related data: from raw experiment results to methodological insight and source code. A number of non-profit organizations, including the Virginia-based Centre for Open Science, are currently promoting this idea.
An even more radical step would be to peer review your research plan before conducting any research. If the international community approves of your methodology, the results will be published regardless of their p-value. More than 20 scientific journals are currently offering this opportunity.
Another way to eliminate cognitive bias is to employ so-called ‘blind analysis’. It is already widely used by physicists, but has yet to achieve significant adoption outside of this field. The idea is to deny researchers access to meaningful summaries until all the data is accounted for. Not knowing how ‘close’ they are to the desired result; researchers are less likely to fall into one of the aforementioned ‘mind traps’.
This can be achieved with special software that shuffles or corrupts data prior to its processing, and only reverts it back during the final stage of the analysis. Up until the very end, the researchers won’t know if the data they’re working with corroborates their initial hypothesis.
The scientists that support this technique call it bothersome, but necessary. It can give researchers an extra edge, and ensure the integrity of the undertaking. Some people call it ‘a method of intellectual hygiene’.
All in all, researchers aren’t that different from the rest of us. They, too, have their weaknesses, even if they’re embarrassed to admit it. The techniques that balance these tendencies out are effective — and are already being employed by a wide variety of people. Of course, we can’t live a life completely free of cognitive bias — but technology is making sure that we’re doing whatever we can to eradicate it where it matters.
Further reading:
Inside ITMO University:
Yet this issue is rarely discussed — because it remains an embarrassing subject.
Boyd Norton / Public Domain
In 2015, a group of researchers tried to reproduce the results of 100 different psychology studies but succeeded in less than a third of all cases. In 2012, a team of Californian scientists reported that, after attempting to reproduce the results of 53 haematology and oncology studies, they only succeeded six times.
How is this still called ‘science’ and not ‘guesswork’?
Professor Robert MacCoun of Stanford Law School thinks that the time is ripe to change our scientific tools in the face of this issue. After all, the scientific community has already faced similar obstacles.
In the mid-20th century scientists discovered that both researchers and their test subjects subconsciously changed their behaviour to ensure their desired results. This has led to the emergence of double-blind studies. The current situation is no difficult — and can be overcome.
The crux of the matter
Statistically relevant studies (with a p-value of 0.05 or less) are much easier to publish. But this relevance is hard to measure correctly, given the amount of junk data in multi-dimensional datasets.
Keith Baggerly, a statistician from the University of Texas MD Anderson Cancer Center, claims that with datasets this large, the analysis can no longer safely rely on conventional mathematical methods, let alone the unassisted human brain. ??Andrew King of the Dartmouth College says that modern software has given researchers a way to easily analyse large amounts of data without really understanding the process, and arrive at the desired low p-value, even if it’s essentially meaningless.
The research has become gamified to its own detriment. The scientists are just chasing the best results.
Not all hypotheses are created equal
One of the traps researchers fall into early on is the lack of regard for conflicting points of view. Jonathan Baron from the University of Pennsylvania in Philadelphia points out that those who want to prove their point phrase questions in a way most likely to guarantee an answer they like.
This principle can be observed at work in the judicial system. Brits might remember the 1999 case of Sally Clark — the woman who was found guilty of double infanticide. Both her kids died within weeks of being born. The persecution claimed that the probability of both these deaths being natural was 1 in 73 million.
Three years later, her conviction was overturned thanks to a different set of stats. The defence argued that double infanticides are 9 times rarer than families that have two of their children die of SIDS.
The sniper trap
There’s an old American joke about an inexperienced sniper who shoots at a bare wall before drawing a target to make himself look good. A similar approach to science is, sadly, quite common among researchers.
Uri Simonsohn of Ramon Llul University calls it “P-hacking”: manipulating pre-existing data until its p-value drops below 0.05. ??A 2012 study of more than 2,000 American psychologists showed just how widespread p-hacking is. Half of all the test subjects dismissed the research that didn’t fit their narrative, and 35% presented surprising findings as if they were expected from day one.
Paying attention to all the wrong signs
Processing data can be just as dangerous as picking it. When doing analysis, we pay a lot of attention to inexplainable outliers and neglect less obvious error indicators, which can be more important overall.
A 2004 study on the data processing habits of three of the leading molecular biology labs presented an in-depth look into this issue. 88% of all the experiments that didn’t meet scientists’ expectations led to discussions of potential methodology errors. Their findings were not treated as logical, even if they, in fact, were.?
Just-So Results
You can even go a step further and present your findings as valid despite not reaching the significance threshold. All that’s needed is an alternative ‘just-so’ explanation to justify the deviation of research data from the what was originally expected. Matthew Hankins of the King’s College in London collected more than 500 phrases commonly used to mislead the reader and re-assert the validity of studies’ unreliable findings.
Dan Meyers / Unsplash
The results with a p-factor of greater than 0.1 can be described as “flirting with conventional levels of significance”. Go a bit lower and you got yourself a “borderline significant trend” (p = 0.09), or results that are “not absolutely significant, but very probably so” (p > 0.05).
Solutions
Each of the aforementioned ways of thinking is fairly convenient to both the researcher and the scientific community at large. These traps provide motivation, and speed up the research process. Forgoing them would lose us a lot of time. But in order to battle misinformation, we need to slow down and embrace the difficulty that comes with looking for truth.
1. Working with conflicting hypotheses
First of all, scientists need to make an effort to remain open-minded. They need to seriously examine conflicting hypotheses, and conduct experiments to determine their viability before discarding them.
You can even ask your academic rivals for research assistance! Scientific environments that embrace diversity of opinions are conducive to early detection of logical and methodological errors.
2. Transparency
Open-access journals also help. Everyone benefits from being able to access others’ research-related data: from raw experiment results to methodological insight and source code. A number of non-profit organizations, including the Virginia-based Centre for Open Science, are currently promoting this idea.
An even more radical step would be to peer review your research plan before conducting any research. If the international community approves of your methodology, the results will be published regardless of their p-value. More than 20 scientific journals are currently offering this opportunity.
3. Blind analysis
Another way to eliminate cognitive bias is to employ so-called ‘blind analysis’. It is already widely used by physicists, but has yet to achieve significant adoption outside of this field. The idea is to deny researchers access to meaningful summaries until all the data is accounted for. Not knowing how ‘close’ they are to the desired result; researchers are less likely to fall into one of the aforementioned ‘mind traps’.
This can be achieved with special software that shuffles or corrupts data prior to its processing, and only reverts it back during the final stage of the analysis. Up until the very end, the researchers won’t know if the data they’re working with corroborates their initial hypothesis.
The scientists that support this technique call it bothersome, but necessary. It can give researchers an extra edge, and ensure the integrity of the undertaking. Some people call it ‘a method of intellectual hygiene’.
All in all, researchers aren’t that different from the rest of us. They, too, have their weaknesses, even if they’re embarrassed to admit it. The techniques that balance these tendencies out are effective — and are already being employed by a wide variety of people. Of course, we can’t live a life completely free of cognitive bias — but technology is making sure that we’re doing whatever we can to eradicate it where it matters.
Further reading:
- Juggling work and study at the Department of Photonics and Optical Information Technology
- Post-cyberpunk: what you need to know about the latest trends in speculative fiction
- Weekend Picks: light reading for STEM majors
Inside ITMO University:
- ITMO startups: machine vision edition
- Functional Materials and Devices of Optoelectronics
- A tour of the Museum of Optics
- The robotics lab
- The cyber-physical systems lab