|
Abstract:
|
While new discoveries from neuroscience and computational psychiatry merit great excitement, critical barriers impede their large-scale implementation. Specifically, underreported issues with biobehavioral measurement often negate promising research in ways that are rarely detected. These issues can prevent new findings from emerging, and also lead to nonreplication of promising prior findings. They are often unique to mental health, but they are common within the field and can easily reduce statistical power in simple bivariate results by over 60%, reduce effect sizes by over 70%, and increase sample sizes required by more than 10-fold. Furthermore, they prevent new computational methods from aggregating small effects into dependable knowledge, regardless of the amount of “big data” present. As a result, these issues threaten to waste individual and federal investment in mental health progress. In this presentation, “data pollution” will be defined as a new framework that unifies myriad measurement error sources, with the goal of generating novel ways to address problems with replicability and validity in both machine learning and traditional clinical research approaches.
|
