In this article, we accomplish two things. First, we show that despite empirical psychologists’ nominal endorsement of a low rate of false-positive findings (≤ .05), flexibility in data collection, analysis, and reporting dramatically increases actual false-positive rates. In many cases, a researcher is more likely to falsely find evidence that an effect exists than to correctly find evidence that it does not. We present computer simulations and a pair of actual experiments that demonstrate how unacceptably easy it is to accumulate (and report) statistically significant evidence for a false hypothesis. Second, we suggest a simple, low-cost, and straightforwardly effective disclosure-based solution to this problem. The solution involves six concrete requirements for authors and four guidelines for reviewers, all of which impose a minimal burden on the publication process.
This article introduces a new approach for evaluating replication results. It combines effect-size estimation with hypothesis testing, assessing the extent to which the replication results are consistent with an effect size big enough to have been detectable in the original study. The approach is demonstrated by examining replications of three well-known findings. Its benefits include the following: (a) differentiating “unsuccessful” replication attempts (i.e., studies yielding p > .05) that are too noisy from those that actively indicate the effect is undetectably different from zero, (b) “protecting” true findings from underpowered replications, and (c) arriving at intuitively compelling inferences in general and for the revisited replications in particular.
Journals tend to publish only statistically significant evidence, creating a scientific record that markedly overstates the size of effects. We provide a new tool that corrects for this bias without requiring access to nonsignificant results. It capitalizes on the fact that the distribution of significant p values, p-curve, is a function of the true underlying effect. Researchers armed only with sample sizes and test results of the published findings can correct for publication bias. We validate the technique with simulations and by reanalyzing data from the Many-Labs Replication project. We demonstrate that p-curve can arrive at conclusions opposite that of existing tools by reanalyzing the meta-analysis of the “choice overload” literature.
In 2010–2012, a few largely coincidental events led experimental psychologists to realize that their approach to collecting, analyzing, and reporting data made it too easy to publish false-positive findings. This sparked a period of methodological reflection that we review here and call Psychology's Renaissance. We begin by describing how psychologists’ concerns with publication bias shifted from worrying about file-drawered studies to worrying about p-hacked analyses. We then review the methodological changes that psychologists have proposed and, in some cases, embraced. In describing how the renaissance has unfolded, we attempt to describe different points of view fairly but not neutrally, so as to identify the most promising paths forward. In so doing, we champion disclosure and preregistration, express skepticism about most statistical solutions to publication bias, take positions on the analysis and interpretation of replication failures, and contend that meta-analytical thinking increases the prevalence of false positives. Our general thesis is that the scientific practices of experimental psychologists have improved dramatically.
28 November 2018 in Advances in Methods and Practices in Psychological Science
Many psychological theories predict U-shaped relationships: The effect of x is positive for low values of x, but negative for high values, or vice versa. Despite implying merely a change of sign, hypotheses about U-shaped functions are tested almost exclusively via quadratic regressions, an approach that imposes an arbitrary functional-form assumption that in some scenarios can lead to a 100% rate of false positives (e.g., the incorrect conclusion that y = log(x) is U shaped). Estimating two regression lines, one for low and one for high values of x, allows testing for a sign change without a functional-form assumption. I introduce the Robin Hood algorithm as a way to set the break point between the lines. This algorithm delivers higher power to detect U shapes than all the other break-point-setting alternatives I compared with it. The article includes simulations demonstrating the performance of the two-lines test and reanalyses of published results using this test. An app for running the two-lines test is available at http://webstimate.org/twolines.