Gang Chen

I had the privilege of seeking help from Dr. Fox intermittently between 2008 and 2015 via r-help and email regarding multivariate linear modeling and the R package 'car'. His insight, generosity, and willingness to help the community are greatly missed.

Thanks for the great question! Text can only go so far. Hopefully the attached image does a better job than my words. The key is integrating spatial relatedness among neighboring voxels directly into the hierarchy (the green part), which is what sets it apart from the usual mass-univariate approach.

Memory has a strange way of letting us live multiple lives at once. Thank you for sharing this. Wishing you steadiness in the present.

Hyperbolic trade-off: The importance of balancing trial and subject sample sizes in neuroimaging Here we investigate the crucial role of trials in task-based neuroimaging from the perspectives of statistical efficiency and condition-level generali…

Thanks! A takeaway from our earlier work is that the role of trials isn’t quite what people assume. If the goal is group-level inference, trial count still matters, but it can be traded off with subject number. It’s the combination of both that matters, not either one alone.
doi.org/10.1016/j.ne...

Hyperbolic trade-off: The importance of balancing trial and subject sample sizes in neuroimaging Here we investigate the crucial role of trials in task-based neuroimaging from the perspectives of statistical efficiency and condition-level generali…

If your ultimate inference target is the group level, then what matters is the joint contribution of trial number per condition and participant sample size, not either one in isolation. We explored this point in detail here:
www.sciencedirect.com/science/arti...

Next up: bringing this to everyday analysis.

AFNI’s new program SIMBA is in development and aims to make full whole-brain voxel-level hierarchical modeling accessible to users, hopefully within the next few months.

SIMBA: Scalable Image Modeling using a Bayesian Approach, A Consistent Framework for Including Spatial Dependencies in fMRI Studies Bayesian spatial modeling provides a flexible framework for whole-brain fMRI analysis by explicitly incorporating spatial dependencies, overcoming the limitations of traditional massive univariate app...

And the next step? Full voxel-level modeling.

Recent numerical advances cracked the scalability barrier. Voxel-level hierarchical modeling is now feasible, revealing just how punishing traditional multiple-comparison adjustments really are.
arxiv.org/abs/2511.12825

Whole-brain hierarchical modeling used to feel impossible under the Bayesian framework. It’s become within reach.

@mandymejia.bsky.social’s group demonstrated computational feasibility on the cortical surface and showed major gains in inferential efficiency.
www.sciencedirect.com/science/arti...

What if multiple comparisons weren’t an afterthought?

Hierarchical modeling at the group level bakes the adjustment into the model. Even early demos, despite brutal computational demands, already showed clear gains when applied to a set of regions.
link.springer.com/content/pdf/...

Is the “standard workflow” holding back fMRI analysis?

Mass-univariate analysis is still the bread-and-butter: intuitive, fast… and chronically overfitted. Add harsh multiple-comparison penalties, and we patch the workflow with statistical band-aids. No wonder the stringency debates never die.

New AFNI Academy playlist!

This tutorial presents afni_proc.py's quality control HTML for single subject FMRI.

The APQC HTML has systematic views of data and useful derived quantities. Users can instantly rate, comment and query the fully processed subject data.

www.youtube.com/watch?v=hD9z...

Is Representational Similarity Analysis Reliable? A Comparison with Regression Representational Similarity Analysis (RSA) is a popular method for analyzing neuroimaging and behavioral data. Here we evaluate the accuracy and reliability of RSA in the context of model selection, a...

However, splitting the RSA computation into two steps may lead to information loss. A single-step approach using regression or hierarchical modeling appears to improve precision, reliability and interpretability in estimating representational similarity. arxiv.org/abs/2511.00395

Representational Similarity Analysis (RSA) is a popular method in cognitive neuroscience for comparing representational patterns across conditions. It follows a "correlation-of-correlations" logic: compute (dis)similarities within each representational space, then correlate them across spaces.

Thanks to Zhengchen Cai, @kordinglab.bsky.social, Tom Liu, Josh Faskowitz, @fmri-today.bsky.social, Bharat Biswal, and @afni-pt.bsky.social for fueling this ride and helping turn it into a commentary.

Has resting-state fMRI leaned too much on inductive, data-driven modeling? It can reveal patterns, but also spurious results and weak explanations, the classic "tail wagging the dog." The real challenge is restoring theory-driven, deductive modeling to guide the science.

...but leaning solely on correlation carries hazards: omnipresent noise, over-interpretation, and a canyon separating correlation from true neural mechanisms. And when correlations start masquerading as causes? Welcome to the land of chaos, confusion, and boobytraps.

Data only shows associations. Turning those into claims about mechanism or causation? That requires a Rosetta Stone of prior knowledge + theory. Resting-state fMRI is purely observational; correlation is its currency. From this, plenty of "theoretical toys" about brain function can be built...

Blind data cleaning, automated pipelines and dichotomized results may give the illusion of standardization, rigor and reproducibility, but they risk turning science into ritual over inquiry. When mechanisms are obscure, don’t pretend they’re fixed; perhaps embrace variability and think creatively?

Well, whenever you take a break from being a task guy… don’t you technically become a rest guy?

Quite interesting! Are we veering into an ontological vs epistemological distinction here? Conceptually, brain activity can be decomposed into task-independent and task-induced components, but practically, the boundary between them is often blurred and difficult to disentangle in real data.

So does it boil down to this: trading one flavor of contamination (task engagement) for another (microsleep roulette)?

Has neuroimaging reached the glorious era where a magical residualization spell can summon the latent resting-state signal from the ashes of task-induced disruption? I’d love to see such an incantation, especially if it comes with a modeling wand.

**FMRI/neuroimaging folks**

Quick reminder @ the next AFNI Bootcamp: May 28-30, 2025. Learn through interactive data analysis!

Day 1-2: data viz, single subject analysis and QC.
Day 3: statistics, results reporting and group analysis.

Details, registration and schedule:
afni.nimh.nih.gov/bootcamp

We are pleased to announce the next AFNI Bootcamp, May 28-30, 2025.

First 2 days: data visualization, single subject analysis and QC. 3rd day: statistics, results reporting and group analysis.

Please see here for details, registration link and preliminary schedule:
afni.nimh.nih.gov/bootcamp

Science doesn’t grow in a vacuum; it thrives on shared ideas and fresh perspectives. Thanks to #sans2025 for building bridges and connecting the dots, and to @elisabaek.bsky.social & @jfguassimoreira.bsky.social for creating the opportunity!

Thanks for the kind shoutout! It was a pleasure rambling about statistics, science, and their rocky relationship. Grateful the audience didn’t throw tomatoes or shoes. I'm taking that as strong evidence of tolerance for variability and uncertainty. Am I allowed to skip the p-value for that evidence?

For those who think more data just means more headcount, here’s a quirky twist: the number of data points per individual actually matters--a lot. If you're into a bit of rigor, this article highlights a factor that’s often overlooked. Thanks for the shoutout!
www.sciencedirect.com/science/arti...

The mind craves binaries: good or bad, true or false, on or off. It’s tidy. It’s comforting. But the world rarely plays along. Reality tends to unfold in gradients, not in absolutes. And so does statistical evidence. Data analysis doesn’t speak in black and white, but in shades of uncertainty.

Do you like genes and estimating heritability? Then @gangchen6.bsky.social and D. Moraczewski have important news for you.

Conventional estimation methods ignore measurement error, leading to a bias. Don't worry: hierarchical modeling to the rescue!

www.frontiersin.org/journals/gen...

The p-value arms race has reached a new milestone -- 10⁻²⁶². At this quantum level of super precision, statistical modeling in quantitative genetics is on the verge of breaking the uncertainty principle.