Pitfalls of Consciousness: How researchers cheat themselves

People are surprisingly good at deceiving themselves, so researchers often fail to reproduce the results of their experiments. It is not customary to talk about this rather large problem in science.

Even the most honest person is a master of self-deception. We are able to quickly isolate anomalous results, however, we often take for granted all, as we see it, “logical” conclusions. Thus, we unconsciously move away from reality.
')
In 2015, an attempt was made to repeat the results of a hundred psychological studies, but this was only possible in three cases out of ten.

In 2012, scientists from Amgen, a biotechnology firm in Thousand Oaks, California, reported that they were able to reproduce the results of only six studies in oncology and hematology out of fifty-three.

“The emergence of similar crisis situations is our chance to improve the scientific toolkit,” says Robert McCone, a sociologist from Stanford. This has happened before.

Back in the mid-twentieth century, scientists found that experimenters and test subjects often unconsciously changed their behavior in order to “tailor” the results of the study to their expectations. This discovery was the cause of the double-blind research method.

For this reason, researchers come up with a variety of ways to eliminate errors in data analysis: strategies that include working with rivals (opponents of the theory, for example) and analyzing “fake” data.

Problem

There is an opinion that only studies with statistically significant results , that is, results whose p-value is 0.05 or less, should be admitted to the publication. Difficulties arise primarily when analyzing large sets of multidimensional data, where it is incredibly difficult to separate meaningful data from random “noise”.

“Even statistical methods do not always cope with such volumes of data, let alone the human brain,” says Keith Baggerley, a statistician at the Oncology Center named after him. Md Anderson at the University of Texas.

Andrew King, a management specialist from Dartmouth College in Hanover, New Hampshire, said that thanks to the wide distribution of specialized software, it became easier for researchers to check huge data sets (they don’t have to fully understand the essence of the methods used) and get small p -values ​​(which may be useless for a specific study).

“It's like a sport here,” says Hal Pashler, a psychologist at the University of California at San Diego, “we are chasing the best result.”

Limited hypotheses

One of the pitfalls that await in the early stages of the study is the neglect of counter arguments and other explanations that contradict the initial hypothesis. “As a rule, if a person wants to confirm his point of view, he formulates the questions in such a way as to get a knowingly affirmative answer,” says Jonathan Baron, a psychologist at the University of Pennsylvania at Philadelphia.

Such situations are not uncommon in the courts. In 1999, in Britain, a woman named Sally Clark was convicted of the murder of her two infant sons. The verdict was made on the basis of statistical data, according to which the chance of death of two children from sudden infant death syndrome (SIDS) was only 1 in 73 million - this fact was accepted as incriminating evidence.

Mathematician Ray Hill later calculated that double death from SIDS occurs in about 1 family of 297,000, while double killing children by their parents is in about 1 family of 2.7 million. The 9: 1 ratio is against murder. In 2003, Sally Clark’s conviction was reversed on the basis of new evidence.

Texas sniper error

There is another trap that can be reached during data analysis. It is explained in an old American joke about a Texas sniper, an incompetent shooter, who first fired at the wall of the shed and only then painted a target centered around the largest cluster of bullet holes.

Psychologist Uri Simonson of the University of Pennsylvania gives an accurate explanation of this naivety in its definition of the term "p-hacking": "Conducting information manipulations until statistical significance is reached p <0.05". In 2012, a study of the behavior of more than two thousand US psychologists showed how widely p-hacking is.

Half of the subjects selectively reported only those studies that "succeeded"; 35% presented unexpectedly obtained data as if a similar outcome was assumed from the very beginning.

"Asymmetry" of attention

At the data verification stage, there is another trap: we do not verify the accuracy of the expected results and pay more attention to the “intuitive-incomprehensible”. We are not aware that the error may be hiding elsewhere.

This behavior is quite common. In 2004, a study was conducted on how employees of three leading molecular biological laboratories test the results of 165 different experiments.

In 88% of cases, when the result did not meet the expectations, scientists believed that errors were made during the experiment, and they even did not admit that the theory was incorrect. At the same time, “logical results” were practically not discussed.

Fairy tales just like that

In the process of analysis, the data are compiled and interpreted, and researchers often begin to give unscientific theoretical substantiations, that is, tell “fairy tales” (Just-So Stories) - this phenomenon was named after Rudyard Kipling’s book “Fair Tales Just Like That” (“Just So Stories” ), in which bizarre explanations are given to ordinary things (for example, from where the leopard has spots).

Another temptation for scientists is to provide a rationale for why the expected result was not obtained, that is, to justify it. Matthew Hankins, a statistician at King’s College in London, collected over 500 original phrases that researchers used to convince readers that they should pay attention to their insignificant results (see here ).

Among them are “balancing on the verge of the level of significance (> 0.1)”, “on the very limit of significance (p = 0.099)” and “the result is not quite significant, but very likely (> 0.05)”.

Solutions

Each of the above-described traps stimulates the process of identifying potentially important scientific dependencies, but here one has to learn to eliminate false results and dead-end hypotheses, to specifically slow down the pace of research.

There is one solution - you need to revive the old tradition and begin to openly consider all the competing hypotheses, and also, if possible, invent experiments that would test them. This will allow you not to close on any one theory.

Transparency

Open science has become another solution to the problem. Researchers share their methods, data, program code and results with each other, for example, through the Center for Open Science.

An even more radical idea is the introduction of “registered reports”, when researchers submit their research plans for review even before the experiment begins. If the plan is approved, a report on the obtained results of the experiment (regardless of their significance) is guaranteed to be published.

This measure is intended to reduce the influence of researchers and reviewers on the work. Today, more than 20 journals offer or plan to offer the possibility of publishing such reports.

Work with rivals

Another method is great for solving controversial issues - you can invite academic rivals to join the work. Working with competing hypotheses and theories, rivals quickly identify logical errors and eliminate them.

Blind data analysis method

Blind data analysis is another way to get rid of cognitive impairment. He came from the field of physics, but in other areas it is still little known. The idea is that the researchers do not know how close they are to the desired results, so they are less likely to influence the results of the experiment.

One way to implement this method is to write a special program that will create alternative data sets, for example, by adding random noise or shear. Researchers at any stage do not know what data they are working with. The truth is revealed only at the very last moment, when any intentional manipulation of the results of the analysis will be obvious.

Scientists who support this method call it superfluous, but necessary work, which supports the researcher's confidence that he will get unbiased results. Therefore, the blind data analysis method is sometimes called “intellectual hygiene”.

Despite the fact that scientists do not differ in many ways from other people and are subject to the same weaknesses, the methods that are now being introduced to eliminate “errors of consciousness” in science show their effectiveness. Interestingly, not least among these methods are techniques using software products: as it turns out, they can not only make life easier for the researcher, but also to a large extent guarantee his impartiality.

PS We recommend to look at our story about the development of a quantum communication system, and most recently we wrote about how students become advanced programmers.