iiinova — where your digital self keeps evolving

Simon Vandekar

@vandyatvandy.bsky.social

Biostatistician with neuroimaging data. Loves dogs, mushrooms, data.

Simon Vandekar @vandyatvandy.bsky.social · 01/12/2025

@mntj21.bsky.social and Ishaan use the method to show that structural and functional brain imaging variables do not improve predictive performance of psychopathology factor scores in reprobrainchart.github.io. w/ @aarona-b.bsky.social

Brain imaging data improve prediction accuracy for age, but not psychopathology factor scores.

000

Simon Vandekar @vandyatvandy.bsky.social · 01/12/2025

Out-of-sample machine learning prediction accuracy underestimates the true maximum achievable predictive accuracy (MAPA) for brain phenotype associations. @mntj21.bsky.social used double ML to improve estimates for ML accuracy. w/ @edwardhkennedy.bsky.social www.biorxiv.org/content/10.1...

Double machine learning prediction accuracy is less biased for estimating the maximum achievable prediction accuracy (MAPA) for brain-phenotype associations than conventional Pearson's correlation.

140

Simon Vandekar @vandyatvandy.bsky.social · 22/01/2025

Don't throw out your data though! If you model separate between- and within-subject effects, you'd see that they're different! Brain-behavior associations across individuals are different than changes within individuals. You can model those separately in your own data without any fancy stats!

111

Simon Vandekar @vandyatvandy.bsky.social · 22/01/2025

Tidbit that you can implement in your own analyses: If you collect longitudinal data you improve efficiency by throwing out half your data! WTF?! If you use baseline ("1st") versus a common longitudinal analysis ("All") effect size decreases! Read on! @kaidikang.bsky.social @meharpist.bsky.social

111