Varieties of Voodoo
Ed Vul
Department of Brain and Cognitive Sciences
Massachusetts Institute of Technology
fMRI experiments provide an abundance of data, of which only a small,
a priori unknown, subset is relevant. As such, experimenters are
faced with a challenging statistical problem: How to estimate both the
location of the signal as well as the strength of the signal? (e.g.,
Which bit of the brain is correlated with neuroticism, and what is the
strength of the correlation?) I will argue that current analysis
techniques lead to overestimates of signal strength (e.g., correlation
magnitude), that have gone largely unnoticed by researchers. I will
advocate several simple analysis strategies that avoid this
overestimation, produce unbiased estimates of signal (correlation or
otherwise), and afford quantitative population-level inferences.
Back to workshop main page.