The "replication crisis" refers to a problem in the sciences where findings from previous experiments don't hold up when studies are repeated. It is a particular issue for those in the behavioral sciences and experimental psychology; studies that should hold up, don't, and researchers have struggled with the vast task of both confirming what is known and trying to glean new and accurate information. The concerns ramped up in the 2010s, when repeated experiments to corroborate published psychological studies ended up failing to replicate, adding to concerns about how studies are conducted and published in the first place.
In recent work published in Frontiers in Psychology, researchers in psychological and brain sciences in Art & Sciences at Washington University in St. Louis propose a path to sort through this complex problem: Include data from previous studies in the framework of a replicated study. They recommend gathering data using an approach called Bayesian modeling, a type of inference theorized by mathematician Thomas Bayes some 300 years ago.
"Bayesian frameworks can help us think about science in a much more incremental and accumulative way," said neuroscientist Todd Braver, the William R. Stuckenberg Professor in Human Values and Moral Development and a professor of psychological and brain sciences in WashU Art & Sciences, as well as a professor of radiology and of neuroscience at WashU Medicine.
Braver is co-author of the research, along with Joshua Jackson, the Saul and Louise Rosenzweig Professor of Personality Science in Arts & Sciences, and graduate student Thomas Dudey.
"I tell my students that they are lucky to be learning statistics now, when the barriers to using Bayesian models are quite low," Jackson said. "These models provide a much more flexible, powerful and actually intuitive way to do experimental research and ask critical questions about our data."
Braver explained that most experimental studies adopt a simplistic approach to replication, treating an experiment like it was being performed for the first time. "The earlier study had its statistics and you had your statistics," and so scientists treated it like a binary decision: does the new study replicate the old results or not?
"Even though we as scientists are drawing on previously published work, we weren't using the estimates drawn from their data," Braver said.
With today's computational tools, researchers have better options. "One of our main goals was to persuade other researchers that current software tools actually make it quite easy to incorporate these models into data analysis workflows," said Dudey, first author of the new study.
In this work, the scientists revisited one of Braver's studies on cognitive control. They used a Bayesian approach to re-analyze two datasets, treating the first dataset as if it were the original or "prior" study and the second dataset as one that had been prepared to replicate, or re-do, the first one.
Their new analysis yielded some important insights, including that there is much to learn even in subtle differences in the data.
"For the consistent effects, we gained further confidence that an original finding was quite stable and reliable," Dudey said. "Even when the effect shifted across datasets though, the direction of the shift still provided important clues about which factors might require further investigation."
All of this makes Braver think a Bayesian approach is the right way to incorporate prior data. "We want to accumulate knowledge from everything that's come before," Braver said. "This framework gives you a continuous way of relating the current data to what we knew previously."
Dudey TA, Jackson JJ, Cooper SR and Braver TS (2026) Hierarchical Bayesian Regression for experimental psychology: a case study of cognitive control. Front. Psychol. 17:1643463. doi: 10.3389/fpsyg.2026.1643463
The author(s) declared that financial support was received for this work and/or its publication. This research was supported by National Institutes of Health (NIH) grant R37 MH066078 and T32 NS115672, Office of Naval Research grant MURI N00014-22-S-F0, and McDonnell Center for Systems Neuroscience.
