Finally read this paper in detail. Overall very compelling case for abandoning small individual/group differences studies with neuroimaging data. https://twitter.com/smarek0502/status/1297154759556763649

Results are reminiscent of @nicebread303 "corridor of stability" work https://www.nicebread.de/at-what-sample-size-do-correlations-stabilize/ and @GaelVaroquaux "cross validation failure" https://hal.inria.fr/hal-01545002  - but nice to see these phenomena demonstrated so strikingly at this scale in real data

Suggestions:
1) Would be nice to plot positive predictive value alongside error rates in Figure 2

2) I'm concerned that rigorous motion exclusion could result in restricted range and/or Berkson's paradox. Would be nice to see more detail about how distributions of behavioral/demographic variables are affected by exclusion.

Definitely shows the immense value of large shared datasets - big kudos to the ABCD and HCP teams for making this possible through their data sharing efforts!