Abstract
Failures to replicate evidence of new discoveries have forced scientists to ask whether this unreliability is due to suboptimal implementation of methods or whether presumptively optimal methods are not, in fact, optimal. This paper reports an investigation by four coordinated laboratories of the prospective replicability of 16 novel experimental findings using rigour-enhancing practices: confirmatory tests, large sample sizes, preregistration and methodological transparency. In contrast to past systematic replication efforts that reported replication rates averaging 50%, replication attempts here produced the expected effects with significance testing (P < 0.05) in 86% of attempts, slightly exceeding the maximum expected replicability based on observed effect sizes and sample sizes. When one lab attempted to replicate an effect discovered by another lab, the effect size in the replications was 97% that in the original study. This high replication rate justifies confidence in rigour-enhancing methods to increase the replicability of new discoveries.
Similar content being viewed by others
Main
Science progressively learns about the world through the discovery of replicable findings1,2. Efforts to systematically replicate studies across various scientific fields have reported seemingly disappointing replication rates ranging from 30% to 70%, with effect sizes (ESs) about half the size of the original findings3,4,5,6,7,8,9,10,11,12,13. This has been a cause for concern among many—but not all14,15.
Seemingly low replicability can be the consequence of false positives or exaggerated ESs among the original studies, resulting from low statistical power, measurement problems, errors, P-hacking and regression to the mean due to selective reporting favouring original positive results15,16. Low replicability can also be the consequence of false negatives or suppressed ESs among replication studies, resulting from sampling error, low statistical power, heterogeneity of the phenomena, different analytic strategies, questionable research practices and/or lack of fidelity to the original protocols17,18,19,20,21,22. Finally, failures to replicate or declining ESs can be the consequence of unknown or unarticulated moderating influences and boundary conditions that differ between the original and replication studies, indicating shortcomings in the theoretical and methodological specifications23.
Collectively, these factors comprise likely explanations for why replications are less successful and produce weaker ESs over time24,25,26. In this Article, we report the results of a prospective replication study examining whether low replicability and declining effects are inevitable when using proposed rigour-enhancing practices.
Four laboratories conducting discovery-oriented social-behavioural research participated in a prospective replication study (Supplementary Information section 1). Over five years, the labs conducted their typical research, examining topics covering psychology, marketing, advertising, political science, communication, and judgement and decision-making (Table 1). Each lab engaged in pilot testing of new effects based on their laboratory’s business-as-usual practices. These practices could involve collecting data with different sample providers and with any sample size the lab saw fit. All pilots were required to have their materials, procedure, hypotheses, analysis plan and exclusions preregistered prior to data collection. Ostensible discoveries were then nominated for confirmation and replication by the discovering labs. The main criterion for moving from piloting and exploration into the confirmation and replication protocol was that the lab believed they had discovered a new effect that was statistically distinguishable from zero during the piloting phase. Each of the four labs submitted four new candidate discoveries for a self-confirmatory test and four replications, for a total of 16 confirmatory tests and 64 replications. In the self-confirmatory test, the discovering lab conducted a preregistered study with a large sample (N ≳ 1,500) and shared a report of the methodology. Regardless of the outcome of the self-confirmatory test, in the replication phase, all four labs conducted independent preregistered replications using the written methodology and any specialized study materials shared by the discovering lab (for example, videos constructed for delivering interventions). Ordinarily, we would promote strong communication between labs to maximize sharing of tacit knowledge about the methodology, but in this case, to maintain the independence of each replication, we opted to discourage communication with the discovering lab outside of the documented protocols except for critical methodology clarifications (Supplementary Information section 2). The replicating labs used equally large sample sizes (all N ≳ 1,500), and each lab used a different sample provider.
Preregistration, reporting all outcomes, large sample sizes, transparent archiving, sharing of materials and commitment to high-fidelity replication procedures should minimize irreplicability or declining ESs stemming from questionable research practices, selective reporting, low-powered research or poorly implemented replication procedures. Such optimizing might promote higher replicability than previously reported in the literature. If—despite these rigour-enhancing practices—low replicability rates or declining effects are observed, such rates or declines could be intrinsic to social-behavioural scientific investigation24,25,27,28,29,30.
Each of the 16 ostensible discoveries were obtained through pilot and exploratory research conducted independently in each laboratory. Not every pilot study the labs conducted was put forward for confirmation and replication. Like all exploratory research, labs sometimes found errors, did not find signals of potential effects or just lost interest in pursuing it further. The labs introduced 4 provisional discoveries each, resulting in 16 self-confirmatory tests and 64 replications (3 independent and 1 self-replication for each), testing replicability and decline. All confirmatory tests, replications and analyses were preregistered both in the individual studies (Supplementary Information section 3 and Supplementary Table 2) and for this meta-project (https://osf.io/6t9vm).
Results
Replicability rate
Self-confirmatory tests
Of the 16 discoveries put forward for replication, 81% (13/16) produced statistically significant results during the self-confirmation phase Effect Size (\({\bar{d}}=0.27\); t(15) = 5.61; P < 0.001; 95% confidence interval (CI), 0.17 to 0.37; estimated between-study heterogeneity s.d., 0.18). The average ES of the self-confirmatory tests was smaller than the estimated average ES of the published psychological literature (\({\bar{d}}=0.43\))31, even when considering only the 13 statistically significant findings (simple average \({\bar{d}}=0.32\)). No lab produced self-confirmatory tests with larger average ESs than the other labs (robust approximate Hotelling’s T 2(6.01), 0.60; P = 0.64).
Replications
One way of assessing replicability is to examine whether each replication rejects the null hypothesis at P < 0.05 in the expected direction6. Including all 16 self-confirmatory tests, 55/64 (86%) replications were successful. The average ES of the replications was \({\bar{d}}=0.26\) (t(15) = 7.61; P < 0.001; 95% CI, 0.19 to 0.33), similar to the ES observed in the self-confirmatory tests (\({\bar{d}}=0.27\); t(15) = 5.61; P < 0.001; 95% CI, 0.17 to 0.37) yet larger than replication ESs observed in prior attempts to systematically replicate the established social-behavioural literature (\({\bar{d}}=0.16\))32.
Focusing only on the 13 statistically significant self-confirmatory tests, 47/52 (90%) of the subsequent replications yielded statistically significant effects in the hypothesized direction. The average ES of the replications was \({\bar{d}}=0.32\), similar to the ES observed in the 13 self-confirmatory tests (\({\bar{d}}=0.32\)).
For the three statistically non-significant self-confirmatory tests, 8/12 (67%) of the subsequent replications produced statistically significant effects in the hypothesized direction: 1/4 for one (d = 0.03), 3/4 for the second (d = 0.09) and 4/4 for the third (d = 0.15). These replication ESs were slightly larger than the ES observed in the original self-confirmatory tests (\({\bar{d}}=0.01\); Fig. 1).
An alternative index of replicability examines the consistency of ESs generated by the initial self-confirmatory test and its subsequent replications (within-study heterogeneity; \({\hat{{\rm{\tau }}}}_{{\rm{within}}}\)). On the basis of a multilevel meta-analysis, little variation in ESs was observed beyond what would be expected by sampling variation alone (\({\hat{{\rm{\tau }}}}_{{\rm{within}}}=0.06\); profile likelihood 95% CI, 0.04 to 0.08; P < 0.001). This indicates that the self-replications and independent replications would not have perfectly replicated the ESs of the self-confirmatory tests, even if all samples were large enough to make sampling error negligible. The degree of variation was, however, smaller than the variation in ESs across the 16 discoveries (\({\hat{{\rm{\tau }}}}_{{\rm{between}}}=0.14\); profile likelihood 95% CI, 0.1 to 0.2; P < 0.001). This suggests that the variation due to sampling and procedural differences across labs was much smaller than the variation due to the phenomena being investigated7,8.
In an exploratory model, differences between self-confirmatory tests and self-replication ESs were found to be fully attributable to sampling error, and average ESs in independent replications strongly correlated with ESs from self-confirmatory tests and self-replications (r = 0.83; profile likelihood 95% CI, 0.52 to 0.95; P < 0.001). There was a small amount of heterogeneity across the independent replications (\({\hat{{\rm{\tau }}}}_{{\rm{within}}}=0.05\); 95% CI, 0.03 to 0.08; P < 0.001). This suggests that conducting replications in new samples and with independently implemented methods increased variability in the observed ESs but did not systematically reduce the observed ESs.
On the basis of a power analysis of the 13 self-confirmatory tests with statistically significant results, the average replication power was 0.96 with a median approaching 1 and an average power in replication studies of specific discoveries ranging from 0.62 to approaching 1. The observed replication rate of 90% is slightly smaller than expected on the basis of these power estimates. Considering all self-confirmatory tests (including statistically non-significant ones), the average power across attempted replications was 0.80 with a median of 0.99 and a range of 0 to approaching 1. The observed replication rate of 86% in the replications for all 16 discoveries was somewhat larger than expected on the basis of these power estimates. Overall, replication rates were consistent with power estimates calculated using the ESs observed in self-confirmatory tests.
Declines in ESs across replications
Within-study heterogeneity across replications was estimated to be s.d. = 0.06, suggesting little heterogeneity overall, despite 75% of the replications being conducted independently using different sample providers. There was modest evidence that one lab produced slightly smaller ESs in replications than did one other lab, controlling for the average size of effects in the initial self-confirmatory tests from each lab (robust approximate Hotelling’s T 2(12.31), 3.51; P = 0.05).
When comparing the self-confirmatory tests to their self-replication attempts, no significant evidence of declining effects appeared. No originating lab’s self-confirmatory test ES was significantly different from its self-replication attempt (Fig. 2). On average, self-replications had the same ES as the self-confirmatory tests (\({{\underline{d}}}_{{\rm{difference}}}={-}0.00\); t(15) = −0.15; P = 0.88; 95% CI, −0.03 to 0.03; between-study heterogeneity \(\hat{{\rm{\tau }}}=0.00\); profile likelihood 95% CI, 0.00 to 0.06; P = 1.00). In three cases, however, there was a statistically significant difference between the ES in the self-confirmatory test and the average ES across independent replications. As both positive and negative discrepancies occurred, independent replication ESs were the same size as those of self-confirmatory tests, on average (\({\underline{d}}_{\mathrm{difference}} = -0.01\); t(15) = −0.21; P = 0.84; 95% CI, −0.06 to 0.05). The discrepancies were heterogeneous across studies, with an estimated between-study s.d. of \(\hat{{\rm{\tau }}}=0.09\) (profile likelihood 95% CI, 0.06 to 0.16; P < 0.001). On the basis of this degree of heterogeneity, independent replications of new discoveries would be predicted to differ from the ES of the self-confirmatory test by as much as s.d. = 0.2 (95% prediction interval, −0.20 to 0.21).
The results of half of the self-confirmatory tests and replications were blinded from data analysis until all replications were completed, to test whether awareness of outcomes influenced replication success and/or the potential declining of ESs over replications. Whether the studies’ results were blinded did not moderate the results (Unstandardized Regression Coefficient, b = −0.01; t(14) = −0.26; P = 0.80; 95% CI, −0.07 to 0.06 for differences between self-confirmatory tests and self-replications; b = 0.07; t(14) = 1.34; P = 0.20; 95% CI, −0.04 to 0.19 for differences between self-confirmatory tests and independent replications). Finally, when we tested ESs sequentially over time, no evidence for a decline in ESs from the self-confirmatory test through the final replication was observed (b = −0.002; t(73) = −0.38; P = 0.71; 95% CI, −0.02 to 0.01; Fig. 3). These results did not significantly change when we removed the fixed effect for each lab (for all additional and robustness checks, see Supplementary Information section 4). In addition, the null was not rejected when we compared the slopes of change in ESs among ‘blind’ versus ‘not blind’ studies (b = 0.02; t(73) = 1.75; P = 0.10; 95% CI, −0.01 to 0.05).
An uninteresting reason for high replicability would be if the discoveries, although novel, are obviously true. Trivial findings might be particularly easy to replicate. To assess this, we conducted two additional studies (Supplementary Information section 5; the protocol (no. 156-19-0689) was deemed exempt by the Office of Research on Human Subjects (Institutional Review Board) at the University of California, Santa Barbara). In the first study, 72 researchers reviewed a synopsis of most of the research designs and predicted the direction of each finding. On average, raters correctly predicted the direction and significance of the self-confirmatory tests 42% of the time, incorrectly predicted null results 38% of the time and incorrectly predicted the direction of the findings 20% of the time. In the second study, 1,180 laypeople reviewed synopses of the research designs from this study, which showed high replicability, and from a prior study of published findings from the same fields with similar methodologies that showed low replicability6,8,9. The synopses were generated by independent researchers with experience in designs using synopses. On multiple preregistered criteria, the participants were no better at predicting the outcomes of the highly replicable discoveries presented here (meanpresent studies = 41% correct prediction) than at predicting the other less replicable findings from the prior investigation (meancomparison studies = 42%; \(\hat{\varDelta }\) = −1.65; 90% CI, −3.4 to 0.1; t(1,172) = −6.32; P < 0.001 for a preregistered equivalence test of the null hypothesis that the difference between present studies and comparison studies would exceed 5 percentage points, H0: Δ ≥ 5.00). Notably, the average accuracy rate of researchers in the first study was nearly identical to the average accuracy among laypeople in the second. Additionally, the accuracy of predictions for specific findings was significantly associated with the absolute magnitude of the average ESs from independent replications (b = 2.79; z = 2.95; P = 0.003; 95% CI, 0.94 to 4.66 for the findings in the present study; b = 0.66; z = 3.05; P = 0.002; 95% CI, 0.24 to 1.09 for the comparison findings); absolute ES explained 35% of the variance in predictability rates. These findings indicate that the sample of discoveries used here were not of a prima facie different type of content that would yield high replication rates. Nor were the content or hypotheses more obvious or predictable than similar findings with low replication rates.
Discussion
Rather than beginning with published findings and attempting to replicate them in a retrospective replication investigation, we implemented a prospective investigation33. By discovering new experimental effects, we were able to apply a meta-scientific34 lens to the entire process from discovery to a confirmatory test, and through a sequence of replication attempts. By subjecting ostensible discoveries to large-sample, preregistered confirmatory tests (13/16; 81% supported at P < 0.05; d = 0.27), we ensured that the primary findings for replication were free from P-hacking and questionable research practices, unlikely to be artefacts of low statistical power, and fully documented. This was true not only of the original discoveries of new effects but also in their replications.
These 16 discoveries, four from each of four independent labs, were then subjected to independent sequential replication (55/64; 86% supported at P < 0.05; d = 0.26). Considering replications of only the 13 statistically significant self-confirmatory tests, the observed replicability rate was 90%. Considering the power to detect ESs from all 16 self-confirmatory tests, the replication rate could not be any higher. The replication ESs were the same size on average as those of the self-confirmatory tests when conducted by the discovering labs and 97% the ESs found by independent labs. Prior replication efforts in the social-behavioural sciences report replication success rates of about 50% on average, producing ESs less than half of the originally reported ESs3,4,5,6,7,8,9,10,11,12,13,32. The present findings establish a benchmark showing that high replicability in the social-behavioural sciences is achievable in both statistical inference and ES estimation, when the original studies and replications are conducted using ostensibly best practices and replications are conducted with complete fidelity.
We investigated whether low replicability and declining ESs should be expected from the social-behavioural sciences because of the complexity of the phenomena, hidden moderators18 and other factors that might be intrinsic to the phenomena being studied or to the replication process24,25. Instead, we found a high replicability rate. The present results are reassuring about the effectiveness of what we think of as best practices in scientific investigations. When novel findings were transparently subjected to preregistered, large-sample confirmatory tests—and when replications involved similar materials and were implemented with a commitment to faithfulness to testing the same hypothesis with fidelity to the original procedure—the observed rate of replication was high. Furthermore, we saw no statistically significant evidence of declining ESs over replications, either when holding materials, procedures and sample source constant (except for sampling error) or when materials, procedures and sample sources varied but were faithful to the original studies.
We further showed, with one survey of scientists and one survey of the lay public, that our studies were not a priori more predictable than similar other studies with a known replication rate. Our high replication rate thus cannot be attributed to our studies being different in especially replicable topics, designs or hypotheses. It would also be possible to imagine that we observed higher replicability than other investigations because of the qualities of the researchers involved in this project, such as being better at imagining and discovering new, replicable phenomena. Although we could be motivated to believe this possibility, the principal investigators in this project all have direct experience with their own published findings failing to replicate. Also, in this and other research, the participating labs have established practices of making risky predictions, most of which fail to materialize into reliable phenomena. If there is an investigator influence on the observed findings, we believe that it is aligned with our interpretation of the present evidence as being due to the adoption of rigour-enhancing practices as lab norms rather than individual exceptionality35.
It is likely that we observed high replicability because of the rigour-enhancing methodological standards adopted in both the original research leading to discovery and the replication studies. First, rather than using exploratory discoveries as the basis for claiming a finding, all discoveries were subjected to preregistered self-confirmatory tests. This eliminated inflation of false positives and ESs by pre-commitment to research designs and analysis plans36. Second, once a discovery was submitted for a self-confirmatory test, we committed to reporting the outcomes. This eliminated publication bias, which is particularly pernicious when selective reporting of study findings systematically ignores null results37,38,39. Third, all self-confirmatory tests and replications were conducted with large sample sizes (N ≳ 1,500), resulting in relatively precise estimates. Fourth, each lab was part of the process of both discovering and replicating findings. This may have motivated teams to be especially careful in both characterizing their methods and carrying out their replications. Fifth, if there were essential specialized materials for the experimental design, the discovering lab made them available as supplementary materials. Sharing original materials should increase understanding of and adherence to critical features of original experimental methodologies. We expect that all these features contributed to improving replicability to varying degrees. Future investigations could manipulate these features to learn more about their causal contributions to replicability.
Even when using rigour-enhancing processes, independent replications conducted by other laboratories produced effect sizes that differed from self-confirmatory effect sizes by larger margins than would be expected due to sampling error alone, producing both larger and smaller ESs with a heterogeneity on the order of s.d. = 0.2. The added heterogeneity may be due to ambiguities or imprecise descriptions in the materials provided to replicating labs, or due to sample heterogeneities introduced by using varying sample providers. This finding highlights the value of multi-lab replication processes even when the original studies follow rigorous practices.
An important question is the constraints on the generalizability of these findings to other research. Our 16 novel findings in social-behavioural sciences each involved two between-subject conditions that could be administered online. All samples for the self-confirmatory tests and replications were drawn from online panels of American adults. These 16 findings do not characterize a representative sample of any methodology or discipline, although they do represent common methodologies, samples and research questions from the social-behavioural sciences. Due to the limited number of participating labs, lab-level variation in the replicability of findings was incalculable; to the extent that labs vary in how they select potential replication targets, the replication rates observed in the present study may not generalize to a broader population of research groups, although—as discussed above—the discovered effects did not differ from similar findings when asking laypeople and a group of researchers to predict them.
The most obvious areas for further testing of the replicability ceiling include more complex experimental or observational designs, in-person behavioural designs that have intricate staging or unusual protocols, sampling strategies that are more inclusive of the world’s population, more heterogeneous conditions including variations in procedure and time between investigations, and research domains beyond the social-behavioural sciences. It is possible that our computer-administered protocols are inherently more replicable, though prior replication efforts of similar methodologies suggest that this is not the case8,9,12, and our follow-up investigation found that the present findings were no more predictable a priori than other findings with similar methodologies that frequently failed to replicate (Supplementary Information section 5). Systematic investigation will be fruitful for understanding the boundary conditions for achieving high replicability.
Low replicability or declining ESs in social-behavioural research are not inevitable. We did not observe declining effects due to idiosyncrasies of different laboratory practices or different sampling conditions. Whereas prior research demonstrates that such declines can occur, the present research demonstrates that they do not necessarily occur. The encouraging evidence here should empower scientists with confidence that what we believe to be rigour-enhancing practices may indeed efficiently generate reliable insights. With the adoption of rigorous research practices such as confirmatory testing, large samples, preregistration, strong documentation and fidelity of replication, high replicability is achievable—perhaps even likely.
Methods
All research complies with all relevant ethical regulations; all studies were approved by the local lab’s Institutional Review Board. The four participating labs conducted pilot and exploratory research in the social sciences pursuing their own typical practices and research interests independently of the other labs. The labs were encouraged to investigate any aspect of social-behavioural science, with the requirements that the discoveries submitted for self-confirmatory testing and replication be novel and non-obvious and not involve deception.
The labs submitted promising discoveries for self-confirmatory tests and replication if they met the following inclusion criteria: a two-group between-subjects manipulation with one focal dependent variable, with methods administered via computer online to adults within a single 20-minute study session.
Prior to conducting the self-confirmatory test, the discovering labs preregistered the study design and analysis plan, including materials, protocol, plans for data cleaning and exclusion, and specification of the analysis model. Once a self-confirmatory test was preregistered, the lab wrote a methods section to share with the other labs. These methods sections had to include everything that the discovering lab believed would be required for an independent lab to conduct an effective replication. This was done to capture the naturalistic conditions when a researcher reads a methods section and conducts a replication based on it.
Following preregistration, no changes could be made to the methods or procedures, and all labs were committed to replicating the protocol regardless of the outcome of the self-confirmatory test. The discovering lab conducted its self-confirmatory test with about 1,500 participants, and then the project coordinator initiated the replication process with the other labs. The labs were assigned the order to conduct replications in a Latin square design to equate lab-specific effects across the order of replications (Supplementary Information section 6).
Sharing study descriptions
After a lab identified an ostensible discovery for a self-confirmatory test, they distributed a description of the methodological details that they believed would be required for an independent lab to run a replication. When the replicating labs considered the instructions to be ambiguous on a meaningful part of the design (71% of studies), the replicating labs sought clarifications about methodology from the discovering lab. Usually these were trivial clarifications or confirmations, but not always (Supplementary Information section 2).
Replications were done sequentially following the same protocol as the self-confirmatory tests, including preregistration. Variation from the 1,500 participants per study was due to idiosyncrasies in how the panels and labs managed participant flow and the application of preregistered exclusion criteria. In most cases, the panels allowed more participants to complete the questionnaire.
The discovering labs could specify required exclusion criteria, such as attention checks. The replicating labs could also choose to preregister and implement exclusions for attention checks following their own laboratory’s best practices. This was done to capture the natural way researchers conduct replications using their own view of best practices. To maintain the ecological validity of labs conducting research in their own style, and to maximize the independence of each replication, all sharing of materials was managed by a project coordinator to prevent unintended communication of designs or results.
Main studies
Sixteen new discoveries of social-behavioural phenomena were submitted to self-confirmatory testing and replication, four from each of the participating laboratories. Table 1 catalogues the new discoveries with a brief name, a one-sentence summary of the finding and a citation to the research. Supplementary Table 3 provides links to comprehensive information for each self-confirmatory test and replication, including the preregistration with the design and analysis plan, research materials, data, analysis code, analysis output and written reports of the methods and results.
Participants
The population of interest for the self-confirmatory tests and replications was adults living in the United States who could read and write in English. The participants were members of panels that had been recruited through non-probability sampling methods to complete online questionnaires in return for small amounts of money or redeemable ‘points’40,41. Labs contracted with different sample providers to provide participants (Stanford University: Toluna, SSI and Dynata; University of California, Santa Barbara: CriticalMix; University of California, Berkeley: Luth; University of Virginia: SoapBox Sample and Lightspeed GMI). We used different sample providers to minimize potential overlap in sampling, although we cannot be sure that some participants are not part of multiple panels and also repeated our studies as part of different panels. These samples were taken from the providers’ online, opt-in, non-probability panels. The sample providers were instructed to provide American adults drawn in a stratified way with unequal probabilities of selection from the panels so that the people who completed each survey would resemble the nation’s adult population (according to the most recently available Current Population Survey, conducted by the US Census Bureau) in terms of gender, age, education, ethnicity (Hispanic versus not), race (allowing each respondent to select more than one race), region and income. This method produced samples designed to look similar to probability samples on the matched characteristics, but the samples may still have differed in unknown ways on unmatched characteristics. The sample providers may have varied in their success at achieving representativeness. A potential lack of adherence to that sampling plan was non-consequential for the conducted studies. For none of the discoveries were the findings presumed to be limited to a subsample of adults, although there may have been a priori or post facto hypothesizing about moderation by demographic variables. For the pilot and exploratory studies, the labs used whatever samples they wished (for example, panel, MTurk or participants visiting the laboratory).
Blinding and sample-splitting manipulations
Two planned manipulations of secondary interest were included to explore potential reasons for variation in the replicability rate or its decline over time. One involved randomly assigning participant recruitment for each data collection of 1,500 participants into a first and second wave of 750 to investigate declines in ES across a single data collection. We assign less confidence to this manipulation, however, as not all panels may have consistently followed our strict protocols for this random assignment (see Supplementary Information section 7 for all additional procedures that the labs and sample providers were instructed to follow). The second manipulation randomly assigned 8 of the 16 new discoveries (2 from each team) to blind the results of the primary outcome variable from the self-confirmatory tests and replications for all team members until all replications for that finding had been completed. For the other 8 discoveries, the data were analysed and reported to the other teams as the results became available. This was to determine whether explicitly blinding research findings would moderate replicability rates and/or declining ESs across replications24,25.
Confirmatory analysis
Meta-analysis
In all analyses, meta-analytic models estimated with restricted maximum likelihood were used, as implemented in the metafor package (version 4.2-0) for R version 4.2.2 (ref. 42,43). For single-level models, Knapp–Hartung corrections for standard errors were used. For multilevel models, cluster-robust variance estimation with small-sample corrections was used to account for the limited number of independent studies40. Preregistration of the overall analysis plan is available at https://osf.io/6t9vm.
We summarized the overall distribution of effects using a multilevel meta-analysis model, including fixed effects to distinguish replications from self-confirmatory tests, with random effects for each unique discovery and each unique ES nested within discovery27. The study-level variance component describes heterogeneity in the phenomena investigated in different studies and labs. The ES-level variance component describes heterogeneity across replications of the same phenomena.
Confirmation versus self-replication and independent replications
A random-effects meta-analysis was estimated to analyse the differences between the self-confirmatory test and the replication of the same discovery by the same lab. A negative average change would be evidence of declining replication ES, even when conducted by the same investigators.
Comparing self-confirmatory tests to replication results from other labs allows for assessment of the impact of between-lab differences in replicability success. Again, a random-effects meta-analysis was used to analyse differences between the ES in the self-confirmatory test and the average ES estimate in the three independent replications. Negative average differences would be evidence of declining replication ESs in cross-lab replication. The random-effects model provides an estimate of heterogeneity in the differences between self-confirmatory tests and replications beyond what would be expected by sampling error alone. Positive heterogeneity would indicate that ESs from self-confirmatory tests could not be exactly replicated by independent labs.
Slope across replications
According to one theory, declines in ESs over time are caused by a study being repeatedly run25. If accurate, the more studies run between the self-confirmatory test and the self-replication, the greater the decline. To examine temporal decline effects across all replications, we aggregated ES estimates from each self-confirmatory test with each of the replications and conducted a meta-analytic growth curve. The model also included random effects for each self-confirmatory test or replication attempt of each study that were allowed to covary within study according to an auto-regressive structure. The ESs were recoded for this analysis so that all effects were positive and a slope to non-significance or weakening ES would be negative in sign.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
The data for each of the individual studies can be found by following the OSF links presented in supplementary information (section 3 and Supplementary Table 2). The data for the overall analysis presented here can be found at https://osf.io/bnq5j/.
Code availability
The statistical code for each of the individual studies can be found by following the OSF links presented in supplementary information (section 3 and Supplementary Table 2). The statistical code for the overall analysis presented here can be found at https://osf.io/bnq5j/.
Change history
11 December 2023
Editor’s Note: Readers are alerted that this paper is subject to criticisms that are being considered by the editors. The criticisms relate to the preregistration and corresponding reporting of hypotheses, predictions, and analyses. We are investigating the concerns raised and an editorial response will follow the resolution of these issues.
24 September 2024
This article has been retracted. Please see the Retraction Notice for more detail: https://doi.org/10.1038/s41562-024-01997-3
References
Putnam, H. Mathematics, Matter and Method: Volume 1, Philosophical Papers (CUP Archive, 1975).
Schmidt, S. Shall we really do it again? The powerful concept of replication is neglected in the social sciences. Rev. Gen. Psychol. 13, 90–100 (2009).
Begley, C. G. & Ellis, L. M. Drug development: raise standards for preclinical cancer research. Nature 483, 531–533 (2012).
Camerer, C. F. et al. Evaluating replicability of laboratory experiments in economics. Science 351, 1433–1436 (2016).
Camerer, C. F. et al. Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015. Nat. Hum. Behav. 2, 637–644 (2018).
Open Science Collaboration.Estimating the reproducibility of psychological science. Science 349, aac4716 (2015).
Errington, T. M. et al. Investigating the replicability of preclinical cancer biology. eLife 10, e71601 (2021).
Klein, R. A. et al. Many Labs 2: investigating variation in replicability across samples and settings. Adv. Methods Pract. Psychol. Sci. 1, 443–490 (2018).
Ebersole, C. R. et al. Many Labs 3: evaluating participant pool quality across the academic semester via replication. J. Exp. Soc. Psychol. 67, 68–82 (2016).
Wagenmakers, E. J. et al. Registered replication report: Strack, Martin, & Stepper (1988). Perspect. Psychol. Sci. 11, 917–928 (2016).
O’Donnell, M. et al. Registered replication report: Dijksterhuis and van Knippenberg (1998). Perspect. Psychol. Sci. 13, 268–294 (2018).
Piovesan, M., Døssing, F. S. & Wengström, E. R. Registered replication report: Rand, Greene & Nowak (2012). Perspect. Psychol. Sci. 12, 527–542 (2017).
Verschuere, B. et al. Registered replication report on Mazar, Amir, and Ariely (2008). Adv. Methods Pract. Psychol. Sci. 1, 299–317 (2018).
Gilbert, D. T., King, G., Pettigrew, S. & Wilson, T. D. Comment on ‘Estimating the reproducibility of psychological science’. Science 351, 1037 (2016).
Wilson, B. M., Harris, C. R. & Wixted, J. T. Science is not a signal detection problem. Proc. Natl Acad. Sci. USA 117, 5559–5567 (2020).
Reich, E. S. Timing glitches dog neutrino claim. Nat. News 483, 17 (2012).
Simmons, J. P., Nelson, L. D. & Simonsohn, U. False-positive psychology: undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychol. Sci. 22, 1359–1366 (2011).
Van Bavel, J. J., Mende-Siedlecki, P., Brady, W. J. & Reinero, D. A. Contextual sensitivity in scientific reproducibility. Proc. Natl Acad. Sci. USA 113, 6454–6459 (2016).
Hedges, L. V. & Schauer, J. M. Consistency of effects is important in replication: rejoinder to Mathur and VanderWeele (2019). Psychol. Methods 24, 576–577 (2019).
Glymour, C., Scheines, R. & Spirtes, P. Discovering Causal Structure: Artificial Intelligence, Philosophy of Science, and Statistical Modeling (Academic Press, 2014).
Protzko, J. Null-hacking, a lurking problem. Preprint at PsyArXiv https://doi.org/10.31234/osf.io/9y3mp (2018).
Bryan, C. J., Yeager, D. S. & O'Brien, J. Replicator degrees of freedom allow publication of misleading failures to replicate. Proc. Natl Acad. Sci USA 116, 25535–25545 (2019).
Nosek, B. A. & Errington, T. M. What is replication? PLoS Biol. https://doi.org/10.1371/journal.pbio.3000691 (2020).
Schooler, J. Unpublished results hide the decline effect: some effects diminish when tests are repeated. Nature 470, 437–438 (2011).
Protzko, J. & Schooler, J. W. Decline effects: types, mechanisms, and personal reflections. In Psychological Science Under Scrutiny: Recent Challenges and Proposed Solutions (eds Lilienfeld, S. O. & Waldman, I. D.) 85–107 (Wiley Blackwell, 2017).
Yeager, D. S., Krosnick, J. A., Visser, P. S., Holbrook, A. L. & Tahk, A. M. Moderation of classic social psychological effects by demographics in the US adult population: new opportunities for theoretical advancement. J. Pers. Soc. Psychol. 117, e84–e99 (2019).
Iso-Ahola, S. E. Reproducibility in psychological science: when do psychological phenomena exist? Front. Psychol. 8, 879 (2017).
Iso-Ahola, S. E. Replication and the establishment of scientific truth. Front. Psychol. 11, 2183 (2020).
Strack, F. & Stroebe, W. What have we learned? What can we learn? Behav. Brain Sci. 41, e151 (2018).
Stroebe, W. & Strack, F. The alleged crisis and the illusion of exact replication. Perspect. Psychol. Sci. 9, 59–71 (2014).
Richard, F. D., Bond, C. F. Jr & Stokes-Zoota, J. J. One hundred years of social psychology quantitatively described. Rev. Gen. Psychol. 7, 331–363 (2003).
Kvarven, A., Strømland, E. & Johannesson, M. Comparing meta-analyses and preregistered multiple-laboratory replication projects. Nat. Hum. Behav. https://doi.org/10.1038/s41562-019-0787-z (2019).
Schweinsberg, M. et al. The pipeline project: pre-publication independent replications of a single laboratory′s research pipeline. J. Exp. Soc. Psychol. 66, 55–67 (2016).
Schooler, J. W. Metascience could rescue the ‘replication crisis’. Nature 515, 9 (2014).
Protzko, J. & Schooler, J. W. No relationship between researcher impact and replication effect: an analysis of five studies with 100 replications. PeerJ 8, e8014 (2020).
Wagenmakers, E. J., Wetzels, R., Borsboom, D., van der Maas, H. L. & Kievit, R. A. An agenda for purely confirmatory research. Perspect. Psychol. Sci. 7, 632–638 (2012).
Greenwald, A. G. Consequences of prejudice against the null hypothesis. Psychol. Bull. 82, 1–20 (1975).
Rosenthal, R. The file drawer problem and tolerance for null results. Psychol. Bull. 86, 638–641 (1979).
Ernst, A. F., Hoekstra, R., Wagenmakers, E. J., Gelman, A. & van Ravenzwaaij, D. Do researchers anchor their beliefs on the outcome of an initial study? Testing the time-reversal heuristic. Exp. Psychol. 65, 158–169 (2018).
Tipton, E. & Pustejovsky, J. E. Small-sample adjustments for tests of moderators and model fit using robust variance estimation in meta-regression. J. Educ. Behav. Stat. 40, 604–634 (2015).
Konstantopoulos, S. Fixed effects and variance components estimation in three‐level meta‐analysis. Res. Synth. Methods 2, 61–76 (2011).
Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1–48 (2010).
R Core Team. R: a language and environment for statistical computing. R Foundation for Statistical Computing https://www.R-project.org/ (2022).
Protzko, J., Buttrick, N., Ebersole, C. R., Lundmark, S. & Schooler, J. Not all explanations are equal: only explanations invoking a change from the true self mitigate responsibility. Preprint at PsyArXiv https://doi.org/10.31234/osf.io/sx7cn (2020).
Protzko, J., Zedelius, C. M. & Schooler, J. W. Rushing to appear virtuous: time pressure increases socially desirable responding. Psychol. Sci. 30, 1584–1591 (2019).
Evers, E. R. K., O′Donnell, M. & Inbar, Y. Arbitrary fairness in reward and punishments. J. Exp. Psychol. Gen. 152, 956–967 (2023).
Laukkonen, R. E. et al. Irrelevant insights make worldviews ring true. Sci. Rep. 12, 2075 (2022).
Inbar, Y. & Evers, E. Worse is bad: divergent inferences from logically equivalent comparisons. J. Exp. Psychol. Gen. 151, 665–675 (2022).
Acknowledgements
We thank B. Fetzer and M. Kidwell for their help with the project. We also thank B. Ouimette, P. Noel, B. Schembari, M. Gross, B. Kaveladze, R. Laukkonen, N. Buchen, R. Narr, M. Austin, J. Rodriguez, K. Kurosawa, H. Harrington, M. Basili and D. Dienhart for their help in completing the individual studies comprising this project. We thank S. Hoogeveen, A. Sarafoglou, A. Dreber Almenberg, and E. J. Wagenmakers for help in creating the study descriptions. This work was sponsored by grants from the Fetzer Franklin Fund of the John E. Fetzer Memorial Trust to each of the participating laboratories at the University of California, Santa Barbara; the University of Virginia; the University of California, Berkeley; and Stanford University. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript, aside from the participation of J.W.
Author information
Authors and Affiliations
Contributions
Conceptualization: J.P., J.K., L.N., B.A.N., J.W. and J.W.S. Methodology: J.P., J.K., M.B., S.L., J.E.P., J.W. and J.W.S. Validation: J.P., J.K., L.N., B.A.N., J.A., M.B., N.B., M.D., C.R.E., S.L., B.M., M.O., H.P. and S.S.R. Formal analysis: J.P., M.D. and J.E.P. Investigation: J.P., J.K., L.N., B.A.N., J.A., M.B., N.B., M.D., C.R.E., S.L., B.M., M.O., H.P., S.S.R. and J.W.S. Resources: J.P., L.N., B.A.N., J.A., M.B., N.B., M.D., C.R.E., S.L., B.M., M.O., H.P. and S.S.R. Software: S.L. Data curation: J.P., J.K., M.B., S.L. and J.E.P. Writing—original draft: J.P. and J.W.S. Writing—review and editing: J.P., J.K., L.N., B.A.N., J.A., M.B., N.B., M.D., C.R.E., S.L., B.M., M.O., H.P., J.E.P., J.W. and J.W.S. Project administration: J.P., J.W. and J.W.S. Supervision: J.K., L.N., B.A.N., M.B., J.W. and J.W.S. Funding acquisition: J.K., L.N., B.A.N., J.W. and J.W.S. Visualization: B.A.N. and J.E.P. Author names are presented in the following order: project manager (J.P.), individual principal investigators in alphabetical order (J.K., L.N. and B.A.N.), authors in alphabetical order (J.A., M.B., N.B., M.D., C.R.E., S.L., B.M., M.O., H.P., J.E.P., S.S.R. and J.W.), and project principal investigator and senior author (J.W.S.).
Corresponding author
Ethics declarations
Competing interests
B.A.N. is the executive director of the non-profit Center for Open Science with a mission to “increase openness, integrity, and reproducibility of research”. J.W. was the scientific director of the Fetzer Franklin Fund that sponsored this research, and B.A.N. and J.W.S were also on the scientific advisory board. J.W. made substantive contributions to the design and execution of this research, but as a funder did not have controlling interest in the decision to publish or not. All other authors declare no competing interests.
Peer review
Peer review information
Nature Human Behaviour thanks Malcolm Macleod and Malte Elson for their contribution to the peer review of this work. Peer reviewer reports are available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article has been retracted. Please see the retraction notice for more detail: https://doi.org/10.1038/s41562-024-01997-3
Supplementary information
Supplementary Information
Supplementary materials, including Supplementary Figs. 1–5 and Tables 1–3.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Protzko, J., Krosnick, J., Nelson, L. et al. RETRACTED ARTICLE: High replicability of newly discovered social-behavioural findings is achievable. Nat Hum Behav 8, 311–319 (2024). https://doi.org/10.1038/s41562-023-01749-9
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1038/s41562-023-01749-9
This article is cited by
-
Claims about scientific rigour require rigour
Nature Human Behaviour (2024)
-
What reproducibility crisis? New research protocol yields ultra-high replication rate
Nature (2023)