Read on Twitter
Beyond excited to share new PREPRINT I co-1st authored with @smarek0502 . Using #ABCDStudy data, we show that the effect sizes of brain-wide associations and sampling variability are a key element of replication failures in typical samples. 1/x https://www.biorxiv.org/content/10.1101/2020.08.21.257758v1
Within ABCD (N=3,928! post denoising), we found that the largest, replicable univariate association between individual differences in brain structure/function and behavioral phenotypes (e.g., cognition, mental health) was r=.14. What does a max r=.14 mean for smaller studies? 2/x
In simulated smaller studies, we show that these effect sizes, coupled with sampling variability, assured statistical errors (false neg, inflation, sign errors) that undermine reproducibility. However, statistical errors decreased as samples increased into the thousands. 3/x
We replicated the effect sizes and our resampling procedure using the single-site, adult human connectome project data. 4/x
Multivariate methods improved reproducibility of brain-wide associations (max out of sample, r=.34), but were likewise constrained at smaller samples, more typical for brain-wide association studies #BWAS. 5/x
Similar to #GWAS, we think this work suggests that larger datasets and consortia (N⪆2,000) are needed to usher in a new era of reproducible human brain-wide association studies - #BWAS. Of course, however, smaller studies will continue to be essential to the field #MSC. 6/x
Huge thank you to everyone that made this work possible, in particular @ndosenbach who worked tirelessly with @smarek0502 and I over the last 16 months, including ~1 million zoom meetings. #teamscience.
ICYMI: This work is a culmination of initial observations made in our DCN paper last year https://www.sciencedirect.com/science/article/pii/S1878929319302932
