The Reliability of Replications: A Study in Computational Reproductions

Abstract

The paper reports findings from a crowdsourced replication. Eighty-four replicator teams attempted to verify results reported in an original study by running the same models with the same data. The replication involved an experimental condition. A “transparent” group received the original study and code, and an “opaque” group received the same underlying study but with only a methods section and description of the regression coefficients without size or significance, and no code. The transparent group mostly verified the original study (95.5%), while the opaque group had less success (89.4%). Qualitative investigation of the replicators’ workflows reveals many causes of non-verification. Two categories of these causes are hypothesized, routine and non-routine. After correcting non-routine errors in the research process to ensure that the results reflect a level of quality that should be present in ‘real-world’ research, the rate of verification was 96.1% in the transparent group and 92.4% in the opaque group. Two conclusions follow: (1) Although high, the verification rate suggests that it would take a minimum of three replicators per study to achieve replication reliability of at least 95% confidence assuming ecological validity in this controlled setting, and (2) like any type of scientific research, replication is prone to errors that derive from routine and undeliberate actions in the research process. The latter suggests that idiosyncratic researcher variability might provide a key to understanding part of the “reliability crisis” in social and behavioral science and is a reminder of the importance of transparent and well documented workflows.

Replicability of simulation studies for the investigation of statistical methods: the RepliSims project

“…[ 37 ] were replicated by researchers who specialize in mediation analysis, and these were two of the more replicable studies. Previous replication initiatives found that replicator expertise was indeed related to the reliability of a replication [ 54 , 55 ].…”