We conducted preregistered replications of 28 classic and contemporary published findings, with protocols that were peer reviewed in advance, to examine variation in effect magnitudes across samples and settings. Each protocol was administered to approximately half of 125 samples that comprised 15,305 participants from 36 countries and territories. Using the conventional criterion of statistical significance ( p < .05), we found that 15 (54%) of the replications provided evidence of a statistically significant effect in the same direction as the original finding. With a strict significance criterion ( p < .0001), 14 (50%) of the replications still provided such evidence, a reflection of the extremely high-powered design. Seven (25%) of the replications yielded effect sizes larger than the original ones, and 21 (75%) yielded effect sizes smaller than the original ones. The median comparable Cohen’s ds were 0.60 for the original findings and 0.15 for the replications. The effect sizes were small (< 0.20) in 16 of the replications (57%), and 9 effects (32%) were in the direction opposite the direction of the original effect. Across settings, the Q statistic indicated significant heterogeneity in 11 (39%) of the replication effects, and most of those were among the findings with the largest overall effect sizes; only 1 effect that was near zero in the aggregate showed significant heterogeneity according to this measure. Only 1 effect had a tau value greater than .20, an indication of moderate heterogeneity. Eight others had tau values near or slightly above .10, an indication of slight heterogeneity. Moderation tests indicated that very little heterogeneity was attributable to the order in which the tasks were performed or whether the tasks were administered in lab versus online. Exploratory comparisons revealed little heterogeneity between Western, educated, industrialized, rich, and democratic (WEIRD) cultures and less WEIRD cultures (i.e., cultures with relatively high and low WEIRDness scores, respectively). Cumulatively, variability in the observed effect sizes was attributable more to the effect being studied than to the sample or setting in which it was studied.

This article was originally submitted for publication to the Editor of Advances in Methods and Practices in Psychological Science (AMPPS) in 2015. When the submitted manuscript was subsequently posted online (Silberzahn et al., 2015), it received some media attention, and two of the authors were invited to write a brief commentary in Nature advocating for greater crowdsourcing of data analysis by scientists. This commentary, arguing that crowdsourced research "can balance discussions, validate findings and better inform policy" (Silberzahn & Uhlmann, 2015, p. 189), included a new figure that displayed the analytic teams' effectsize estimates and cited the submitted manuscript as the source of the findings, with a link to the preprint. However, the authors forgot to add a citation of the Nature commentary to the final published version of the AMPPS article or to note that the main findings had been previously publicized via the commentary, the online preprint, research presentations at conferences and universities, and media reports by other people. The authors regret the oversight.

We conducted preregistered replications of 28 classic and contemporary published findings with protocols that were peer reviewed in advance to examine variation in effect magnitudes across sample and setting. Each protocol was administered to approximately half of 125 samples and 15,305 total participants from 36 countries and territories. Using conventional statistical significance (p &lt; .05), fifteen (54%) of the replications provided evidence in the same direction and statistically significant as the original finding. With a strict significance criterion (p &lt; .0001), fourteen (50%) provide such evidence reflecting the extremely high powered design. Seven (25%) of the replications had effect sizes larger than the original finding and 21 (75%) had effect sizes smaller than the original finding. The median comparable Cohen’s d effect sizes for original findings was 0.60 and for replications was 0.15. Sixteen replications (57%) had small effect sizes (&lt; .20) and 9 (32%) were in the opposite direction from the original finding. Across settings, 11 (39%) showed significant heterogeneity using the Q statistic and most of those were among the findings eliciting the largest overall effect sizes; only one effect that was near zero in the aggregate showed significant heterogeneity. Only one effect showed a Tau &gt; 0.20 indicating moderate heterogeneity. Nine others had a Tau near or slightly above 0.10 indicating slight heterogeneity. In moderation tests, very little heterogeneity was attributable to task order, administration in lab versus online, and exploratory WEIRD versus less WEIRD culture comparisons. Cumulatively, variability in observed effect sizes was more attributable to the effect being studied than the sample or setting in which it was studied.

Over the last ten years, Oosterhof and Todorov's valence-dominance model has emerged as the most prominent account of how people evaluate faces on social dimensions. In this model, two dimensions (valence and dominance) underpin social judgments of faces. Because this model has primarily been developed and tested in Western regions, it is unclear whether these findings apply to other regions. We addressed this question by replicating Oosterhof and Todorov's methodology across 11 world regions, 41 countries, and 11,570 participants. When we used Oosterhof and Todorov's original analysis strategy, the valence-dominance model generalized across regions. When we used an alternative methodology to allow for correlated dimensions we observed much less generalization. Collectively, these results suggest that, while the valence-dominance model generalizes very well across regions when dimensions are forced to be orthogonal, regional differences are revealed when we use different extraction methods, correlate and rotate the dimension reduction solution.

Twenty-nine teams involving 61 analysts used the same dataset to address the same research question: whether soccer referees are more likely to give red cards to dark skin toned players than light skin toned players. Analytic approaches varied widely across teams, and estimated effect sizes ranged from 0.89 to 2.93 in odds ratio units, with a median of 1.31. Twenty teams (69%) found a statistically significant positive effect and nine teams (31%) observed a non-significant relationship. Overall 29 different analyses used 21 unique combinations of covariates. We found that neither analysts' prior beliefs about the effect, nor their level of expertise, nor peer-reviewed quality of analysis readily explained variation in analysis outcomes. This suggests that significant variation in analysis of complex data may be difficult to avoid, even by experts with honest intentions. Crowdsourcing data analysis, a strategy by which numerous research teams are recruited to simultaneously investigate the same research question, makes transparent how defensible, yet subjective analytic choices influence research results.