Scientific Lectures //
Thinking Latent: Construct-Measure Confusion as a Primary Source of Inferential Error in fMRI
Tal Yarkoni, Ph.D., Director, Psychoinformatics Lab, Research Associate Professor, Department of Psychology, University of Texas at Austin
Presented: March 30, 2018
ABSTRACT: In this talk, I argue that many conceptual and statistical problems in functional neuroimaging research (and social and biomedical science more generally) can be traced to our strong collective tendency to confuse operational measures for the latent constructs they are intended to index. I demonstrate how a failure to account for measurement error and/or model misspecification routinely lead fMRI researchers to draw inappropriate inferences--for example, concluding that brain activity predicts some outcome of interest when "controlling for" potential confounds, or implying that one's findings generalize across a much broader universe of measurements than the statistical model actually supports. Drawing on concrete examples from recent work, I suggest that many of these problems could be overcome by (a) adopting standard mixed-effects models widely used in other literatures; (b) placing greater emphasis on naturalistic experimental designs and cross-study replication; and (c) de-emphasizing traditional efforts to "explain" how the brain works in favor of more pragmatic prediction-focused approaches.
BIO: Tal Yarkoni is a Research Associate Professor of Psychology at the University of Texas at Austin. His research program focuses on developing new methods for the analysis, synthesis, and interpretation of neuroimaging and behavioral data.