2/3 Imagine if NFL coaches were hired because they were friends with the White House. You’d end up with bad football teams pretty fast. Same deal with NIH IC Directors and science.

Yes! bsky.app/profile/mwco...

I think both of those explanations are plausible. I co-authored a paper on functional/effective connectivity in 2019 that may be helpful: Reid et al. (2019). "Advancing functional connectivity research from association to causation". Nature Neuroscience. www.colelab.org/pubs/Reid201...

Ideally, regularized partial correlation would have become the default back then. Instead, 90%+ of studies have continued to use pairwise correlations, especially with fMRI. I think one reason is that the advantages of the new approach hadn't been shown clearly, which is what we try to do here.

The graphical lasso functional connectivity approach that performed best in the paper can be implemented using the Activity Flow Toolbox: colelab.github.io/ActflowToolb...
And also using the paper's code release: github.com/ColeLab/Reli... [12/n]

Together, results demonstrated vast improvements in fMRI functional connectivity estimation using regularized partial correlation. Thanks to first author Kirsten Peterson, and coauthors Ruben Sanchez-Romero and
@ravimill.bsky.social!
doi.org/10.1162/IMAG... #neuroscience #neuroimaging [11/n]

And regularization improved prediction of individual differences in demographics (age) and behavior/cognition (general intelligence) relative to standard partial correlation. The glasso results were more interpretable than pairwise correlation (fewer false connections) 10/n

Also empirical, prediction of task-evoked activity (via activity flow modeling) was better with regularized partial correlation 9/n

As another empirical validation, regularized partial correlation was much less susceptible to motion artifacts than pairwise correlation. Percent connections linked to motion = Pairwise correlation FC: 56.4% vs. graphical lasso FC: 0.01% 8/n

First empirical validation: regularized partial correlation was much closer to structural connectivity, which doesn’t have the causal confounding problem (despite other issues) 7/n

This pattern of results was mirrored in empirical resting-state fMRI data across 4 validation measures. Regularization was key to estimating individual subject-level networks with reduced confounding. 6/n

In simulations, pairwise (standard) correlation led to many false connections, but so did partial correlation. Regularized partial correlation (glasso) better recovered the true network organization 5/n

We hypothesized that low reliability of partial correlation is due to overfitting to noise, with regularization (model simplification) improving reliability. 4/n

Pairwise correlations are known to be susceptible to false positives in theory. For example, region A causing activity in unconnected regions B and C (B<-A->C) can lead to a false B-C connection. Partial correlation can correct for this error, but not reliably 3/n

In brief: Improvements to pairwise (standard) correlation: 1) reduced false connections (confounding), 2) reduced sensitivity to in-scanner motion, 3) better correspondence to task-related activity, and 4) more interpretable links with individual differences in behavior 2/n

Maybe being anxious is a sign of being a good scientist? Accurate theories/hypotheses should stand up to many tests developed from a skeptical perspective. Starting from empirical constraints & modeling their interaction can help keep theories grounded, perhaps increasing the odds they'll be correct

Once you have a flow model that generates a phenomenon of interest, you can lesion each empirical constraint (e.g., each connection or task-evoked activation) to determine which contributed to generation of that phenomenon. Follow-up empirical stimulation or lesion work can further verify this.

I agree these are vexing problems that need more solutions. An approach that's working for us is to integrate empirical constraints into generative models. Connectivity and task response data used to form activity flow models (empirical neural networks). Details here: www.colelab.org/pubs/2024_Ac...