Oct 16, 2013

Thriving au naturel amid science on steroids

Brian A. Nosek
University of Virginia, Center for Open Science

Jeffrey R. Spies
Center for Open Science

Photo of Brian Nosek Photo of Jeffrey Spies Last fall, the present first author taught a graduate class called “Improving (Our) Science” at the University of Virginia. The class reviewed evidence suggesting that scientific practices are not operating ideally and are damaging the reproducibility of publishing findings. For example, the power of an experimental design in null hypothesis significance testing is a function of the effect size being investigated and the size the sample to test it—power is greater when effects are larger and samples are bigger. In the authors’ field of psychology, for example, estimates suggest that the power of published studies to detect an average effect size is .50 or less (Cohen, 1962; Gigerenzer & Sedlmeier, 1989). Assuming that all of the published effects are true, approximately 50% of published studies would reveal positive results (i.e., p < .05 supporting the hypothesis). In reality, more than 90% of published results are positive (Sterling, 1959; Fanelli, 2010).

How is it possible that the average study has power to detect the average true effect 50% of the time or less and yet does so about 90% of the time? It isn’t. Then how does this occur? One obvious contributor is selective reporting. Positive effects are more likely than negative effects to be submitted and accepted for publication (Greenwald, 1975). The consequences include [1] the published literature is more likely to exaggerate the size of true effects because with low powered designs researchers must still leverage chance to obtain a large enough effect size to produce a positive result; and [2] the proportion of false positives – there isn’t actually an effect to detect – will be inflated beyond the nominal alpha level of 5% (Ioannidis, 2005).

The class discussed this and other scientific practices that may interfere with knowledge accumulation. Some of the relatively common ones are described in Table 1 along with some solutions that we, and others, identified. Problem. Solution. Easy. The class just fixed science. Now, class members can adopt the solutions as best available practices. Our scientific outputs will be more accurate, and significant effects will be more reproducible. Our science will be better.

Alex Schiller, a class member and graduate student, demurred. He agreed that the new practices would make science better, but disagreed that we should do them all. A better solution, he argued, is to take small steps: adopt one solution, wait for that to become standard scientific practice, and then adopt another solution.

We know that some of our practices are deficient, we know how to improve them, but Alex is arguing that we shouldn’t implement all the solutions? Alex’s lapse of judgment can be forgiven—he’s just a graduate student. However, his point isn’t a lapse. Faced with the reality of succeeding as a scientist, Alex is right.

Table 1

Scientific practices that increase irreproducibility of published findings, possible solutions, and barriers that prevent adoption of those solutions

Practice Problem Possible Solution Barrier to Solution
Run many low-powered studies rather than few high-powered studies Inflates false positive and false negative rates Run high-powered studies Non-significant effects are a threat to publishability; Risky to devote extra resources to high-powered tests that might not produce significant effects
Report significant effects and dismiss non-significant effects as methodologically flawed Using outcome to evaluate method is a logical error and can inflate false positive rate Report all effects with rationale for why some should be ignored; let reader decide Non-significant and mixed effects are a threat to publishability
Analyze during data collection, stop when significant result is obtained or continue until significant result is obtained Inflates false positive rate Define data stopping rule in advance Non-significant effects are a threat to publishability
Include multiple conditions or outcome variables, report the subset that showed significant effects Inflates false positive rate Report all conditions and outcome variables Non-significant and mixed effects are a threat to publishability
Try multiple analysis strategies, data exclusions, data transformations, report cleanest subset Inflates false positive rate Prespecify data analysis plan, or report all analysis strategies Non-significant and mixed effects are a threat to publishability
Report discoveries as if they had resulted from confirmatory tests Inflates false positive rate Pre-specify hypotheses; Report exploratory and confirmatory analyses separately Many findings are discoveries, but stories are nicer and scientists seem smarter if they had thought it in advance
Never do a direct replication Inflates false positive rate Conduct direct replications of important effects Incentives are focused on innovation, replications are boring; Original authors might feel embarrassed if their original finding is irreproducible

Note: For reviews of these practices and their effects see Ioannidis, 2005; Giner-Sorolla, 2012; Greenwald, 1975; John et al., 2012; Nosek et al., 2012; Rosenthal, 1979; Simmons et al., 2011; Young et al., 2008

The Context

In an ideal world, scientists use the best available practices to produce accurate, reproducible science. But, scientists don’t live in an ideal world. Alex is creating a career for himself. To succeed, he must publish. Papers are academic currency. They are Alex’s ticket to job security, fame, and fortune. Well, okay, maybe just job security. But, not everything is published, and some publications are valued more than others. Alex can maximize his publishing success by producing particular kinds of results. Positive effects, not negative effects (Fanelli, 2010; Sterling, 1959). Novel effects, not verifications of prior effects (Open Science Collaboration, 2012). Aesthetically appealing, clean results, not results with ambiguities or inconsistencies (Giner-Sorolla, 2012). Just look in the pages of Nature, or any other leading journal, they are filled with articles producing positive, novel, beautiful results. They are wonderful, exciting, and groundbreaking. Who wouldn’t want that?

We do want that, and science advances in leaps with groundbreaking results. The hard reality is that few results are actually groundbreaking. And, even for important research, the results are often far from beautiful. There are confusing contradictions, apparent exceptions, and things that just don’t make sense. To those in the laboratory, this is no surprise. Being at the frontiers of knowledge is hard. We don’t quite know what we are looking at. That’s why we are studying it. Or, as Einstein said, “If we knew what we were doing, it wouldn’t be called research”.   But, those outside the laboratory get a different impression. When the research becomes a published article, much of the muck goes away. The published articles are like the pictures of this commentary’s authors at the top of this page. Those pictures are about as good as we can look. You should see the discards. Those with insider access know, for example, that we each own three shirts with buttons and have highly variable shaving habits. Published articles present the best-dressed, clean-shaven versions of the actual work.

Just as with people, when you replicate effects yourself to see them in person, they may not be as beautiful as they appeared in print. The published version often looks much better than reality. The effect is hard to get, dependent on a multitude of unmentioned limiting conditions, or entirely irreproducible (Begley & Ellis, 2012; Prinz et al., 2011).

The Problem

It is not surprising that effects are presented in their best light. Career advancement depends on publishing success. More beautiful looking results are easier to publish and more likely to earn rewards (Giner-Sorolla, 2012). Individual incentives align for maximizing publishability, even at the expense of accuracy (Nosek et al., 2012).

Consider three hypothetical papers shown in Table 2. For all three, the researchers identified an important problem and had an idea for a novel solution. Paper A is a natural beauty. Two well planned studies showed effects supporting the idea. Paper B and Paper C were conducted with identical study designs. Paper B is natural, but not beautiful; Paper C is a manufactured beauty. Both Paper B and Paper C were based on 3 studies. One study for each showed clear support for the idea. A second study was a mixed success for Paper B, but “worked” for Paper C after increasing the sample size a bit and analyzing the data a few different ways. A third study did not work for either. Paper B reported the failure with an explanation for why the methodology might be to blame, rather than the idea being incorrect. The authors of Paper C generated the same methodological explanation, categorized the study as a pilot, and did not report it at all. Also, Paper C described the final sample sizes and analysis strategies, but did not mention that extra data was collected after initial analysis, or that alternative analysis strategies had been tried and dismissed.

Table 2

Summary of research practices for three hypothetical papers

Step Paper Paper B Paper C
Data collection Conducted two studies Conducted three studies Conducted three studies
Data analysis Analyzed data after completing data collection following a pre-specified analysis plan Analyzed data after completing data collection following a pre-specified analysis plan Analyzed during data collection and collected more data to get to significance in one case. Selected from multiple analysis strategies for all studies.
Result reporting Reported the results of the planned analyses for both studies Reported the results of the planned analyses for all studies Reported results of final analyses only. Did not report one study that did not reach significance.
Final paper Two studies demonstrating clear support for idea One study demonstrating clear support for idea, one mixed, one not at all Two studies demonstrating clear support for idea

Paper A is clearly better than Paper B. Paper A should be published in a more prestigious outlet and generate more attention and accolade. Paper C looks like Paper A, but in reality it is like Paper B. The actual evidence is more circumspect than the apparent evidence. Based on the report, however, no one can tell the difference between Paper A and Paper C.

Two possibilities would minimize the negative impact of publishing manufactured beauties like Paper C. First, if replication were standard practice, then manufactured effects would be identified rapidly. However, direct replication is very uncommon (Open Science Collaboration, 2012). Once an effect is the literature, there is little systematic ethic to self-correct. Rather than be weeded out, false effects persist or just slowly fade away. Second, scientists could just avoid doing the practices that lead to Paper C making this illustration an irrelevant hypothetical. Unfortunately, a growing body of evidence suggests that these practices occur, and some are even common (e.g., John et al., 2012).

To avoid the practices that produce Paper C, the scientist must be aware of and confront a conflict-of-interest—what is best for science versus what is best for me. Scientists have inordinate opportunity to pursue flexible decision-making in design and analysis, and there is minimal accountability for those practices. Further, humans’ prodigious motivated reasoning capacities provide a way to decide that the outcomes that look best for us also has the most compelling rationale (Kunda, 1990). So, we may convince ourselves that the best course of action for us was the best course of action period. It is very difficult to stop doing suspect practices when we have thoroughly convinced ourselves that we are not doing them.

The Solution

Alex needs to publish to succeed. The practices in Table 1 are to the scientist, what steroids are to the athlete. They amplify the likelihood of success in a competitive marketplace. If others are using, and Alex decides to rely on his natural performance, then he will disadvantage his career prospects. Alex wants to do the best science he can and be successful for doing it. In short, he is the same as every other scientist we know, ourselves included. Alex shouldn’t have to make a choice between doing the best science and being successful—these should be the same thing.

Is Alex stuck? Must he wait for institutional regulation, audits, and the science police to fix the system? In a regulatory world, practices are enforced and he need not worry that he’s committing career suicide by following them. Many scientists are wary of a strong regulatory environment in science, particularly for the possibility of stifling innovation. Some of the best ideas start with barely any evidence at all, and restrictive regulations on confidence in outputs could discourage taking risks on new ideas. Nonetheless, funders, governments, and other stakeholders are taking notice of the problematic incentive structures in science. If we don’t solve the problem ourselves, regulators may solve them for us.

Luckily, Alex has an alternative. The practices in Table 1 may be widespread, but the solutions are also well known and endorsed as good practice (Fuchs et al., 2012). That is, scientists easily understand the differences between Papers A, B, and C – if they have full access to how the findings were produced. As a consequence, the only way to be rewarded for natural achievements over manufactured ones is to make the process of obtaining the results transparent. Using the best available practices privately will improve science but hurt careers. Using the best available practices publicly will improve science while simultaneously improving the reputation of the scientist. With openness, success can be influenced by the results and by how they were obtained.  

Conclusion

The present incentives for publishing are focused on the one thing that we scientists are absolutely, positively not supposed to control - the results of the investigation. Scientists have complete control over the design, procedures, and execution of the study. The results are what they are.

A better science will emerge when the incentives for achievement align with the things that scientists can (and should) control with their wits, effort, and creativity. With results, beauty is contingent on what is known about their origin. Obfuscation of methodology can make ugly results appear beautiful. With methodology, if it looks beautiful, it is beautiful. The beauty of methodology is revealed by openness.

Most scientific results have warts. Evidence is halting, uncertain, incomplete, confusing, and messy. It is that way because scientists are working on hard problems. Exposing it will accelerate finding solutions to clean it up. Instead of trying to make results look beautiful when they are not, the inner beauty of science can be made apparent. Whatever the results, the inner beauty—strong design, brilliant reasoning, careful analysis—is what counts. With openness, we won’t stop aiming for A papers. But, when we get them, it will be clear that we earned them.

References

Begley, C. G., & Ellis, L. M. (2012). Raise standards for preclinical cancer research. Nature, 483, 531-533.

Cohen , J. (1962). The statistical power of abnormal-social psychological research: A review. Journal of Abnormal and Social Psychology, 65, 145-153.

Fanelli, D. (2010). "Positive" results increase down the hierarchy of the sciences. PLoS ONE, 5(4), e10068. doi:10.1371/journal.pone.0010068

Fuchs H., Jenny, M., & Fiedler, S. (2012). Psychologists are open to change, yet wary of rules. Perspectives on Psychological Science, 7, 634-637. doi:10.1177/1745691612459521

Sedlmeier, P., & Gigerenzer, G. (1989). Do studies of statistical power have an effect on the power of studies? Psychological Bulletin, 105, 309-316.

Giner-Sorolla, R. (2012). Science or art? How esthetic standards grease the way through the publication bottleneck but undermine science. Perspectives on Psychological Science.

Greenwald, A. G. (1975). Consequences of prejudice against the null hypothesis. Psychological Bulletin, 82, 1-20.

Ioannidis, J. P. A. (2005). Why most published research findings are false. PLoS Medicine, 2, e124.

John, L., Loewenstein, G., & Prelec, D. (2012). Measuring the prevalence of questionable research practices with incentives for truth-telling. Psychological Science, 23, 524-532. doi:10.1177/0956797611430953

Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108, 480-498. doi:10.1037/0033-2909.108.3.480

Nosek, B. A., Spies, J. R., & Motyl, M. (2012). Scientific utopia: II. Restructuring incentives and practices to promote truth over publishability. Perspectives on Psychological Science, 7,615-631. doi:10.1177/1745691612459058

Open Science Collaboration. (2012). An open, large-scale, collaborative effort to estimate the reproducibility of psychological science. Perspectives on Psychological Science, 7, 657-660. doi:10.1177/1745691612462588

Prinz, F., Schlange, T. & Asadullah, K. (2011). Believe it or not: how much can we rely on published data on potential drug targets? Nature Reviews Drug Discovery, 10, 712-713.

Rosenthal, R. (1979). The file drawer problem and tolerance for null results. Psychological Bulletin, 86, 638-641. doi:10.1037/0033-2909.86.3.638

Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011). False-positive psychology: Undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological Science, 22, 1359-1366.

Sterling, T. D. (1959). Publication decisions and their possible effects on inferences drawn from tests of significance - or vice versa. Journal of the American Statistical Association, 54, 30-34.

Young, N. S., Ioannidis, J. P. A., & Al-Ubaydli, O. (2008). Why current publication practices may distort science. PLoS Medicine, 5, 1418-1422.

Share on: TwitterFacebookGoogle+Email

Comments