Read on Twitter
Important, sobering study, likely to be widely misinterpeted. But if the strongest reproducible brain-behavior correlation is r = 0.19, we are measuring the brain, behavior, or both wrong. https://twitter.com/smarek0502/status/1297154759556763649
I found it extremely interesting that the strongest effective correlation nearly doubled when using a multivariate technique. Maybe some hints there, i.e. we need to build more comprehensive profiles, not macroscopic, overtly reductive ones.
I remember a discussion about reward prediction error in a thousand-n study. Highly significant prediction of psychiatric features, but a tiny correlation similar to what is reported here. How could this mechanism predict so little variance?
Well, people are complicated, dynamical systems. We undergo massive homeostatic, hormonal, and arousal based changes on 4,6, 12, and 25 hour cycles. A single 8 minute resting state scan and some behavioral tests on a computer likely only capture a tiny tip of the wave.
I recall a chapter in Olaf Sporn’s Networks of The Brain, surmising the the connectome changes its topography constantly, across micro, meso, and macro temporal and spatial scales. Maybe lying in a scanner for a few minutes out of all the those changes can’t reveal that much.
If we want to profile people, i.e. to develop sensitive and specific fingerprints that predict psychopathology and idiosyncratic behavior, we probably need to measure them multiple times, in multiple states, especially under different physiological and affective loads.
Maybe this paper can make us think more deeply not only about the limitations of the fMRI scanner, but also about the limitations of the boring, unrealistic games we make people play on computers because we believe that they isolate specific cognitive variables.
Maybe they do, but it turns out, those ‘specific cognitive variables’ have very little to do with being a person! So much for the ‘tasking state’.
After all, a psychiatric symptom or personality trait is itself a complex mixture of social, psychological, embodied, and environmental factors. Trying to predict those solely from a single fMRI session and some boring, reductive computer games is probably futile...
Or maybe not. Maybe we should just use MEG instead :D
In sum, “Heaven and earth contain more than what is dreamt of in thy data features, Horatio!”.