The fragility of experimental findings

Because of the importance of the topic, I am putting up only one thing here today: A long excerpt of an article that gives many examples of initially exciting findings that fade away as people try to repeat them. Because of my career in academic research, all of the processes described below are very familar to me but they are very little known to the general public.

One thing that amuses me is that many of the findings decribed below that proved false in the long run were really strong-looking findings -- infinitely stronger than the pathetic little relative ratios usually reported in diet-related research. Skepticism is the only well-informed response to all new medical and diet-related findings

On September 18, 2007, a few dozen neuroscientists, psychiatrists, and drug-company executives gathered in a hotel conference room in Brussels to hear some startling news. It had to do with a class of drugs known as atypical or second-generation antipsychotics, which came on the market in the early nineties. The drugs, sold under brand names such as Abilify, Seroquel, and Zyprexa, had been tested on schizophrenics in several large clinical trials, all of which had demonstrated a dramatic decrease in the subjects’ psychiatric symptoms. As a result, second-generation antipsychotics had become one of the fastest-growing and most profitable pharmaceutical classes. By 2001, Eli Lilly’s Zyprexa was generating more revenue than Prozac. It remains the company’s top-selling drug.

But the data presented at the Brussels meeting made it clear that something strange was happening: the therapeutic power of the drugs appeared to be steadily waning. A recent study showed an effect that was less than half of that documented in the first trials, in the early nineteen-nineties. Many researchers began to argue that the expensive pharmaceuticals weren’t any better than first-generation antipsychotics, which have been in use since the fifties. “In fact, sometimes they now look even worse,” John Davis, a professor of psychiatry at the University of Illinois at Chicago, told me.

Before the effectiveness of a drug can be confirmed, it must be tested and tested again. Different scientists in different labs need to repeat the protocols and publish their results. The test of replicability, as it’s known, is the foundation of modern research. Replicability is how the community enforces itself. It’s a safeguard for the creep of subjectivity. Most of the time, scientists know what results they want, and that can influence the results they get. The premise of replicability is that the scientific community can correct for these flaws.

But now all sorts of well-established, multiply confirmed findings have started to look increasingly uncertain. It’s as if our facts were losing their truth: claims that have been enshrined in textbooks are suddenly unprovable. This phenomenon doesn’t yet have an official name, but it’s occurring across a wide range of fields, from psychology to ecology. In the field of medicine, the phenomenon seems extremely widespread, affecting not only antipsychotics but also therapies ranging from cardiac stents to Vitamin E and antidepressants: Davis has a forthcoming analysis demonstrating that the efficacy of antidepressants has gone down as much as threefold in recent decades.

For many scientists, the effect is especially troubling because of what it exposes about the scientific process. If replication is what separates the rigor of science from the squishiness of pseudoscience, where do we put all these rigorously validated findings that can no longer be proved? Which results should we believe? Francis Bacon, the early-modern philosopher and pioneer of the scientific method, once declared that experiments were essential, because they allowed us to “put nature to the question.” But it appears that nature often gives us different answers.

Jonathan Schooler was a young graduate student at the University of Washington in the nineteen-eighties when he discovered a surprising new fact about language and memory. At the time, it was widely believed that the act of describing our memories improved them. But, in a series of clever experiments, Schooler demonstrated that subjects shown a face and asked to describe it were much less likely to recognize the face when shown it later than those who had simply looked at it. Schooler called the phenomenon “verbal overshadowing.”

The study turned him into an academic star. Since its initial publication, in 1990, it has been cited more than four hundred times. Before long, Schooler had extended the model to a variety of other tasks, such as remembering the taste of a wine, identifying the best strawberry jam, and solving difficult creative puzzles. In each instance, asking people to put their perceptions into words led to dramatic decreases in performance.

But while Schooler was publishing these results in highly reputable journals, a secret worry gnawed at him: it was proving difficult to replicate his earlier findings. “I’d often still see an effect, but the effect just wouldn’t be as strong,” he told me. “It was as if verbal overshadowing, my big new idea, was getting weaker.” At first, he assumed that he’d made an error in experimental design or a statistical miscalculation. But he couldn’t find anything wrong with his research. He then concluded that his initial batch of research subjects must have been unusually susceptible to verbal overshadowing. (John Davis, similarly, has speculated that part of the drop-off in the effectiveness of antipsychotics can be attributed to using subjects who suffer from milder forms of psychosis which are less likely to show dramatic improvement.) “It wasn’t a very satisfying explanation,” Schooler says. “One of my mentors told me that my real mistake was trying to replicate my work. He told me doing that was just setting myself up for disappointment.”

Schooler tried to put the problem out of his mind; his colleagues assured him that such things happened all the time. Over the next few years, he found new research questions, got married and had kids. But his replication problem kept on getting worse. His first attempt at replicating the 1990 study, in 1995, resulted in an effect that was thirty per cent smaller. The next year, the size of the effect shrank another thirty per cent. When other labs repeated Schooler’s experiments, they got a similar spread of data, with a distinct downward trend. “This was profoundly frustrating,” he says. “It was as if nature gave me this great result and then tried to take it back.” ....

The most likely explanation for the decline is an obvious one: regression to the mean. As the experiment is repeated, that is, an early statistical fluke gets cancelled out. The extrasensory powers of Schooler’s subjects didn’t decline—they were simply an illusion that vanished over time. And yet Schooler has noticed that many of the data sets that end up declining seem statistically solid—that is, they contain enough data that any regression to the mean shouldn’t be dramatic. “These are the results that pass all the tests,” he says. “The odds of them being random are typically quite remote, like one in a million. This means that the decline effect should almost never happen. But it happens all the time! Hell, it’s happened to me multiple times.”

In 2001, Michael Jennions, a biologist at the Australian National University, set out to analyze “temporal trends” across a wide range of subjects in ecology and evolutionary biology. He looked at hundreds of papers and forty-four meta-analyses (that is, statistical syntheses of related studies), and discovered a consistent decline effect over time, as many of the theories seemed to fade into irrelevance. In fact, even when numerous variables were controlled for—Jennions knew, for instance, that the same author might publish several critical papers, which could distort his analysis—there was still a significant decrease in the validity of the hypothesis, often within a year of publication. Jennions admits that his findings are troubling, but expresses a reluctance to talk about them publicly. “This is a very sensitive issue for scientists,” he says. “You know, we’re supposed to be dealing with hard facts, the stuff that’s supposed to stand the test of time. But when you see these trends you become a little more skeptical of things.”

What happened? Leigh Simmons, a biologist at the University of Western Australia, suggested one explanation when he told me about his initial enthusiasm for the theory: “I was really excited by fluctuating asymmetry. The early studies made the effect look very robust.” He decided to conduct a few experiments of his own, investigating symmetry in male horned beetles. “Unfortunately, I couldn’t find the effect,” he said. “But the worst part was that when I submitted these null results I had difficulty getting them published. The journals only wanted confirming data. It was too exciting an idea to disprove, at least back then.” For Simmons, the steep rise and slow fall of fluctuating asymmetry is a clear example of a scientific paradigm, one of those intellectual fads that both guide and constrain research: after a new paradigm is proposed, the peer-review process is tilted toward positive results. But then, after a few years, the academic incentives shift—the paradigm has become entrenched—so that the most notable results are now those that disprove the theory.

Jennions, similarly, argues that the decline effect is largely a product of publication bias, or the tendency of scientists and scientific journals to prefer positive data over null results, which is what happens when no effect is found. The bias was first identified by the statistician Theodore Sterling, in 1959, after he noticed that ninety-seven per cent of all published psychological studies with statistically significant data found the effect they were looking for. A “significant” result is defined as any data point that would be produced by chance less than five per cent of the time. This ubiquitous test was invented in 1922 by the English mathematician Ronald Fisher, who picked five per cent as the boundary line, somewhat arbitrarily, because it made pencil and slide-rule calculations easier. Sterling saw that if ninety-seven per cent of psychology studies were proving their hypotheses, either psychologists were extraordinarily lucky or they published only the outcomes of successful experiments. In recent years, publication bias has mostly been seen as a problem for clinical trials, since pharmaceutical companies are less interested in publishing results that aren’t favorable. But it’s becoming increasingly clear that publication bias also produces major distortions in fields without large corporate incentives, such as psychology and ecology.

While publication bias almost certainly plays a role in the decline effect, it remains an incomplete explanation. For one thing, it fails to account for the initial prevalence of positive results among studies that never even get submitted to journals. It also fails to explain the experience of people like Schooler, who have been unable to replicate their initial data despite their best efforts. Richard Palmer, a biologist at the University of Alberta, who has studied the problems surrounding fluctuating asymmetry, suspects that an equally significant issue is the selective reporting of results—the data that scientists choose to document in the first place. Palmer’s most convincing evidence relies on a statistical tool known as a funnel graph. When a large number of studies have been done on a single subject, the data should follow a pattern: studies with a large sample size should all cluster around a common value—the true result—whereas those with a smaller sample size should exhibit a random scattering, since they’re subject to greater sampling error. This pattern gives the graph its name, since the distribution resembles a funnel.

The funnel graph visually captures the distortions of selective reporting. For instance, after Palmer plotted every study of fluctuating asymmetry, he noticed that the distribution of results with smaller sample sizes wasn’t random at all but instead skewed heavily toward positive results. Palmer has since documented a similar problem in several other contested subject areas. “Once I realized that selective reporting is everywhere in science, I got quite depressed,” Palmer told me. “As a researcher, you’re always aware that there might be some nonrandom patterns, but I had no idea how widespread it is.” In a recent review article, Palmer summarized the impact of selective reporting on his field: “We cannot escape the troubling conclusion that some—perhaps many—cherished generalities are at best exaggerated in their biological significance and at worst a collective illusion nurtured by strong a-priori beliefs often repeated.”

Palmer emphasizes that selective reporting is not the same as scientific fraud. Rather, the problem seems to be one of subtle omissions and unconscious misperceptions, as researchers struggle to make sense of their results. Stephen Jay Gould referred to this as the “shoehorning” process. “A lot of scientific measurement is really hard,” Simmons told me. “If you’re talking about fluctuating asymmetry, then it’s a matter of minuscule differences between the right and left sides of an animal. It’s millimetres of a tail feather. And so maybe a researcher knows that he’s measuring a good male”—an animal that has successfully mated—“and he knows that it’s supposed to be symmetrical. Well, that act of measurement is going to be vulnerable to all sorts of perception biases. That’s not a cynical statement. That’s just the way human beings work.”

According to Ioannidis, the main problem is that too many researchers engage in what he calls “significance chasing,” or finding ways to interpret the data so that it passes the statistical test of significance—the ninety-five-per-cent boundary invented by Ronald Fisher. “The scientists are so eager to pass this magical test that they start playing around with the numbers, trying to find anything that seems worthy,” Ioannidis says. In recent years, Ioannidis has become increasingly blunt about the pervasiveness of the problem. One of his most cited papers has a deliberately provocative title: “Why Most Published Research Findings Are False.”

The problem of selective reporting is rooted in a fundamental cognitive flaw, which is that we like proving ourselves right and hate being wrong. “It feels good to validate a hypothesis,” Ioannidis said. “It feels even better when you’ve got a financial interest in the idea or your career depends upon it. And that’s why, even after a claim has been systematically disproven”—he cites, for instance, the early work on hormone replacement therapy, or claims involving various vitamins—“you still see some stubborn researchers citing the first few studies that show a strong effect. They really want to believe that it’s true.”

That’s why Schooler argues that scientists need to become more rigorous about data collection before they publish. “We’re wasting too much time chasing after bad studies and underpowered experiments,” he says. The current “obsession” with replicability distracts from the real problem, which is faulty design. He notes that nobody even tries to replicate most science papers—there are simply too many. (According to Nature, a third of all studies never even get cited, let alone repeated.) “I’ve learned the hard way to be exceedingly careful,” Schooler says. “Every researcher should have to spell out, in advance, how many subjects they’re going to use, and what exactly they’re testing, and what constitutes a sufficient level of proof. We have the tools to be much more transparent about our experiments.”

In a forthcoming paper, Schooler recommends the establishment of an open-source database, in which researchers are required to outline their planned investigations and document all their results. “I think this would provide a huge increase in access to scientific work and give us a much better way to judge the quality of an experiment,” Schooler says. “It would help us finally deal with all these issues that the decline effect is exposing.”

Although such reforms would mitigate the dangers of publication bias and selective reporting, they still wouldn’t erase the decline effect. This is largely because scientific research will always be shadowed by a force that can’t be curbed, only contained: sheer randomness. Although little research has been done on the experimental dangers of chance and happenstance, the research that exists isn’t encouraging.

In the late nineteen-nineties, John Crabbe, a neuroscientist at the Oregon Health and Science University, conducted an experiment that showed how unknowable chance events can skew tests of replicability. He performed a series of experiments on mouse behavior in three different science labs: in Albany, New York; Edmonton, Alberta; and Portland, Oregon. Before he conducted the experiments, he tried to standardize every variable he could think of. The same strains of mice were used in each lab, shipped on the same day from the same supplier. The animals were raised in the same kind of enclosure, with the same brand of sawdust bedding. They had been exposed to the same amount of incandescent light, were living with the same number of littermates, and were fed the exact same type of chow pellets. When the mice were handled, it was with the same kind of surgical glove, and when they were tested it was on the same equipment, at the same time in the morning.

The premise of this test of replicability, of course, is that each of the labs should have generated the same pattern of results. “If any set of experiments should have passed the test, it should have been ours,” Crabbe says. “But that’s not the way it turned out.” In one experiment, Crabbe injected a particular strain of mouse with cocaine. In Portland the mice given the drug moved, on average, six hundred centimetres more than they normally did; in Albany they moved seven hundred and one additional centimetres. But in the Edmonton lab they moved more than five thousand additional centimetres. Similar deviations were observed in a test of anxiety. Furthermore, these inconsistencies didn’t follow any detectable pattern. In Portland one strain of mouse proved most anxious, while in Albany another strain won that distinction.

The disturbing implication of the Crabbe study is that a lot of extraordinary scientific data are nothing but noise. The hyperactivity of those coked-up Edmonton mice wasn’t an interesting new fact—it was a meaningless outlier, a by-product of invisible variables we don’t understand. The problem, of course, is that such dramatic findings are also the most likely to get published in prestigious journals, since the data are both statistically significant and entirely unexpected. Grants get written, follow-up studies are conducted. The end result is a scientific accident that can take years to unravel.

More here