The problem of replicability in science

from xkcd

by Massimo Pigliucci

In recent months much has been written about the apparent fact that a surprising, indeed disturbing, number of scientific findings cannot be replicated, or when replicated the effect size turns out to be much smaller than previously thought.

Arguably, the recent streak of articles on this topic began with one penned by David Freedman in The Atlantic, and provocatively entitled “Lies, Damned Lies, and Medical Science.” In it, the major character was John Ioannidis, the author of some influential meta-studies about the low degree of replicability and high number of technical flaws in a significant portion of published papers in the biomedical literature. As Freedman put it in The Atlantic: “80 percent of non-randomized studies (by far the most common type) turn out to be wrong, as do 25 percent of supposedly gold-standard randomized trials, and as much as 10 percent of the platinum-standard large randomized trials.” Ioannidis himself was quoted uttering some sobering words for the medical community (and the public at large): “Science is a noble endeavor, but it’s also a low-yield endeavor. I’m not sure that more than a very small percentage of medical research is ever likely to lead to major improvements in clinical outcomes and quality of life. We should be very comfortable with that fact.” Ouch.

Julia and I actually addressed this topic during a Rationally Speaking podcast, featuring as guest our friend Steve Novella, of Skeptics’ Guide to the Universe and Science-Based Medicine fame. But while Steve did quibble with the tone of the Atlantic article, he agreed that Ioannidis’ results are well known and accepted by the medical research community. Steve did point out that it should not be surprising that results get better and better as one moves toward more stringent protocols like large randomized trials, but it seems to me that one should be surprised (actually, appalled) by the fact that even there the percentage of flawed studies is high — not to mention the fact that most studies are in fact neither large nor properly randomized.

The second big recent blow to public perception of the reliability of scientific results is an article published in The New Yorker by Jonah Lehrer, entitled “The truth wears off.” Lehrer also mentions Ioannidis, but the bulk of his essay is about findings in psychiatry, psychology and evolutionary biology (and even in research on the paranormal!). In these disciplines there are now several documented cases of results that were initially spectacularly positive — for instance the effects of second generation antipsychotic drugs, or the hypothesized relationship between a male’s body symmetry and the quality of his genes — that turned out to be increasingly difficult to replicate over time, with the original effect sizes being cut down dramatically, or even disappearing altogether. Again, the implications are disturbing. As Lehrer concludes at the end of his article: “Such anomalies demonstrate the slipperiness of empiricism. Although many scientific ideas generate conflicting results and suffer from falling effect sizes, they continue to get cited in the textbooks and drive standard medical practice. Why? Because these ideas seem true. Because they make sense. Because we can’t bear to let them go. And this is why the decline effect is so troubling.”

None of this should actually be particularly surprising to any practicing scientist. If you have spent a significant time of your life in labs and reading the technical literature, you will appreciate the difficulties posed by empirical research, not to mention a number of issues such as the fact that few scientists ever actually bother to replicate someone else’s results, for the simple reason that there is no Nobel (or even funded grant, or tenured position) waiting for the guy who arrived second. Still, in the midst of this I was directed by a tweet by my colleague Neil deGrasse Tyson (who has also appeared on the RS podcast, though in a different context) to a recent ABC News article penned by John Allen Paulos, which meant to explain the decline effect in science.

Paulos’ article is indeed concise and on the mark (though several of the explanations he proposes were already brought up in both the Atlantic and New Yorker essays), but it doesn’t really make things much better. For instance, Paulos suggests that one explanation for the decline effect is the well known statistical phenomenon of the regression toward the mean. This phenomenon is responsible, among other things, for a fair number of superstitions: you’ve probably heard of some athletes’ and other celebrities’ fear of being featured on the cover of a magazine after a particularly impressive series of accomplishments, because this brings “bad luck,” meaning that the following year one will not be able to repeat the performance at the same level. This is actually true, not because of magical reasons, but simply as a result of the regression to the mean: extraordinary performances are the result of a large number of factors that have to line up just right for the spectacular result to be achieved. The statistical chances of such an alignment to repeat itself are low, so inevitably next year’s performance will likely be below par. Paulos correctly argues that this also explains some of the decline effect of scientific results: the first discovery might have been the result of a number of factors that are unlikely to repeat themselves in exactly the same way, thus reducing the effect size when the study is replicated.

But that’s not all. Another major determinant of the unreliability of scientific results mentioned by Paulos is the well know problem of publication bias: crudely put, science journals (particularly the high-profile ones, like Nature and Science) are interested only in positive, spectacular, “sexy” results. Which creates a powerful filter against negative, or marginally significant results. What you see in science journals, in other words, isn’t a statistically representative sample of scientific results, but a highly biased one, in favor of positive outcomes. No wonder that when people try to repeat the feat they often come up empty handed.

A third cause for the problem, not mentioned by Paulos but addressed in the New Yorker article, is the selective reporting of results by scientists themselves. This is essentially the same phenomenon as the publication bias, except that this time it is scientists themselves, not editors and reviewers, who don’t bother to submit for publication results that are either negative or not strongly conclusive. Again, the outcome is that what we see in the literature isn’t all the science that we ought to see. And it’s no good to argue that it is the “best” science, because the quality of scientific research is measured by the appropriateness of the experimental protocols (including the use of large samples) and of the data analyses — not by whether the results happen to confirm the scientist’s favorite theory.

The conclusion of all this is not, of course, that we should throw the baby (science) out with the bath water (bad or unreliable results). But scientists should also be under no illusion that these are rare anomalies that do not affect scientific research at large. Too much emphasis is being put on the “publish or perish” culture of modern academia, with the result that graduate students are explicitly instructed to go for the SPU’s — Smallest Publishable Units — when they have to decide how much of their work to submit to a journal. That way they maximize the number of their publications, which maximizes the chances of landing a postdoc position, and then a tenure track one, and then of getting grants funded, and finally of getting tenure. The result is that, according to statistics published by Nature, it turns out that about ⅓ of published studies is never cited (not to mention replicated!).

John Platt, in a classical article published in Science in 1964 (the year I was born), famously wrote: “Scientists these days tend to keep up the polite fiction that all science is equal. Except for the work of the misguided opponent whose arguments we happen to be refuting at the time, we speak as though every scientist’s field and methods of study are as good as every other scientist’s, and perhaps a little better. This keeps us all cordial when it comes to recommending each other for government grants. ... We speak piously of taking measurements and making small studies that will ‘add another brick to the temple of science.’ Most such bricks lie around the brickyard.”

Most damning of all, however, is the potential effect that all of this may have on science’s already dubious reputation with the general public (think evolution-creation, vaccine-autism, or climate change). This sentiment was expressed cogently by Ioannidis, again in the Atlantic article: “If we don’t tell the public about these problems, then we’re no better than non-scientists who falsely claim they can heal. If the drugs don’t work and we’re not sure how to treat something, why should we claim differently? Some fear that there may be less funding because we stop claiming we can prove we have miraculous treatments. But if we can’t really provide those miracles, how long will we be able to fool the public anyway? The scientific enterprise is probably the most fantastic achievement in human history, but that doesn’t mean we have a right to overstate what we’re accomplishing.” Good questions, and good point.