Identification of Effective Connectivity Subregions

Clark Glymour; Joseph D. Ramsey; Kun Zhang; Ruben Sanchez-Romero

REVIEW 4 major objections 5 minor 1 cited by

Identification of Effective Connectivity Subregions

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read ROI averaging hides which voxels drive inter-regional communication, and two causal search algorithms recover those connectivity subregions directly from voxel-level resting-state fMRI.

desk verdict A simple, clearly-presented idea for defining voxel-level connectivity subregions, with real-data consistency results that are suggestive but undercut by concatenation artifacts and missing null controls. read the letter →

arxiv 1908.03264 v1 pith:T3GPBPLS submitted 2019-08-08 q-bio.NC cs.LGeess.IVq-bio.QM

Ruben Sanchez-Romero , Joseph D. Ramsey , Kun Zhang , Clark Glymour This is my paper · ORCID

classification q-bio.NCcs.LGeess.IVq-bio.QM

keywords effectiveconnectivityvoxel-levelsubregionsFASKalgorithmTwo-Stepresting-statefMRImedialtemporallobecausalgraphicalmodels

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Standard resting-state fMRI connectivity analysis averages voxel time series inside regions of interest, and this paper argues that the averaging discards the signal that distinguishes which voxels actually drive communication between regions. The paper's proposal is to infer a directed graph at the voxel level over all voxels in a set of ROIs, then define a connectivity subregion for each member of a pair of ROIs as the set of voxels in that ROI that have an adjacency to the other ROI. Applied to eighty longitudinal resting-state scans of one person's medial temporal lobe, two independent causal-search procedures, FASK and Two-Step, recover similar subregions that are spatially consistent across scanning sessions and across hemispheres. The intended payoffs are anatomical: subregions can coexist inside one ROI, they tend to sit near contact surfaces of adjacent regions, and their directed edges are roughly balanced, suggesting reciprocal communication. The methodological point is that correlation and partial-correlation analyses cannot distinguish direct signaling from common causes, indirect paths, or conditioning on common effects, while the two algorithms used here are designed to do so.

What carries the argument

The machinery that carries the argument is a pair of graphical search algorithms that exploit the non-Gaussianity of BOLD signals instead of second-order correlations. FASK first runs FAS-stable, an order-independent conditional-independence search using BIC scores, to identify undirected adjacencies without mistaking common causes or indirect paths for direct connections; it then orients each adjacency as a two-way or one-way edge by comparing correlations conditioned on positive values of each variable, with a heuristic extra-adjacency rule for cases where a true 2-cycle cancels to near-zero correlation. Two-Step writes the linear causal system as $X=(I-B)^{-1}E$, uses adaptive lasso in a first step to restrict the free parameters of $(I-B)$, and then uses sparse ICA to estimate the connectivity matrix $B$ so that the residual components $E$ are as independent and non-Gaussian as possible. The two procedures are the source of the voxel-level adjacencies; the Connectivity Subregions Algorithm is simply the extraction rule that turns those adjacencies into subregions $S_P$ and $S_Q$ for each pair of ROIs.

What would settle it

Run the same FASK and Two-Step searches on each of the ten sessions separately instead of on their concatenation; if the recovered subregions do not reappear in the single-session graphs, or if their spatial consistency across sessions is no better than a null distribution obtained by permuting voxel coordinates within each ROI, then the reported connectivity subregions are artifacts of the concatenation or of voxel-level autocorrelation rather than stable signaling.

Watch

Extended reading notes

Core claim

The central claim is that effective-connectivity subregions are identifiable from voxel-level BOLD time series without any anatomical prior about where within an ROI the communicating voxels lie. For a pair of ROIs $P$ and $Q$ with exclusive voxel sets $V_P$ and $V_Q$, a voxel $v_P\in V_P$ belongs to the connectivity subregion $S_P$ exactly when the adjacency $v_P-v_Q$ appears in the voxel-level adjacency graph for some $v_Q\in V_Q$, and $S_Q$ is defined symmetrically. To get those adjacencies, the authors first estimate a full directed, possibly cyclic graph over all 570 (left) or 530 (right) voxels spanning entorhinal cortex, CA32DG, CA1, and subiculum, using FASK or Two-Step on each of eight concatenated ten-session datasets. Across the eight temporally separated datasets, the undirected voxel graphs agree at Jaccard values around 0.7 for both algorithms, most subregions agree at 0.77–0.94, and the direction counts between subregion pairs are nearly symmetric. The authors conclude that connectivity subregions are a recoverable feature of resting-state data, that they can differ and overlap within a single ROI, and that most medial-temporal-lobe interregional signaling is reciprocal.

Load-bearing premise

The load-bearing premise is that the 4610 BOLD time points obtained by concatenating ten separate sessions can be treated as independent, identically distributed samples for causal search, even though each session has its own baseline, drift, and motion and the BOLD signal is autocorrelated.

Editorial extensions

If this is right

A single ROI can contain several overlapping connectivity subregions, each tied to a different partner region, so ROI-level connectivity is not a homogeneous property of the region.
Adjacency structure at the voxel level is reproducible across sessions (Jaccard around 0.7), while edge orientations are much less so (around 0.3), so connectivity-subregion identification should be treated as stable but direction claims as provisional.
Medial temporal lobe communication inferred this way is largely reciprocal, with comparable numbers of directed edges in both directions between subregions.
If a single scanning session can support the same search, as the paper's simulations suggest, connectivity subregions could be mapped without the expense of longitudinal data.
Including all voxels in the search, not just the pair of interest, protects the inferred subgraphs against false adjacencies mediated by common causes or intermediate voxels in other regions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next test is to compare recovered subregions against anatomical tracers or known axonal termination zones in an animal model; the paper's observation that subregions hug contact surfaces is a prediction that neuroanatomy could confirm.
Because session-to-session orientation estimates are unstable, direction-of-influence conclusions in single-subject resting-state studies may need pooling across sessions or subjects, rather than thresholding each graph independently.
The concatenation of ten sessions into one 4610-point sample treats the scans as exchangeable; checking whether the same subregions survive when each session is analyzed separately would directly test that assumption.
The same subregion-extraction rule could be applied to time-varying or task-based connectivity graphs, asking whether the voxels driving inter-regional communication shift with cognitive state.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the paper gives a simple definition of a 'connectivity subregion' — voxels in ROI P that share an estimated adjacency with voxels in ROI Q — and applies it to high-resolution MTL data from one person across 80 sessions. The new contribution is the adjacency-based subregion concept, not the search algorithms, which are the authors' own FASK and Two-Step. That distinction matters: the paper is essentially an application note with a clean formalization.

The good parts: The definition is transparent and easy to reuse. The authors explicitly say Jaccard is a consistency measure, not an informativeness measure, which is the right caveat. They also concede that single-scan stability 'seems advisable,' which is an honest acknowledgment of a weakness. Both algorithms are open source, so the work is reproducible. For two of the five ROI pairs (CA32DG–CA1 and CA1–SUB), the subregion Jaccard indices are high (around 0.85–0.90) and consistent across the two methods, so the central idea has some empirical support.

The soft spots are real and load-bearing. The eight datasets are made by concatenating ten temporally ordered sessions into one 4610-point time series. FASK's FAS-stable step and Two-Step's ICA step assume exchangeable, stationary samples. BOLD data are autocorrelated, and each session has its own baseline, drift, and motion. Appending sessions creates artificial discontinuities and mean shifts that can act as common causes, producing exactly the kinds of adjacencies the algorithms are designed to find — except these are artifacts. Because those artifacts would be present in every concatenated dataset, the across-dataset Jaccard consistency could be driven by repetition of the same artifacts, not by true signaling. The paper provides no null baseline (e.g., phase-randomized data or session-shuffled voxels) to show that the observed Jaccard values exceed what an artifact-only process would produce. The right-hemisphere results are described as 'similar' but no quantitative cross-hemisphere overlap is given, so the cross-hemisphere claim is asserted rather than tested. The low consistency for ENT-related subregions is acknowledged in the text, which is honest, but it also weakens the claim that the method works generally.

The authors' own Discussion suggests running on a single scan would be advisable; I agree. A revision that re-runs the analysis on individual sessions (or with session as a covariate), includes surrogate nulls, and quantifies mean graph density would make the empirical case much stronger.

Bottom line: the paper is worth a serious referee, mainly because the subregion definition is genuinely useful and the application is clear. But as it stands, the empirical support is conditional on the concatenation assumptions, and the claims outrun the evidence. If I were the editor, I'd send it out, with the expectation that the authors address the iid violation and the missing null.

Referee Report

4 major / 5 minor

Summary. The paper proposes a procedure to identify 'connectivity subregions' within anatomically defined ROIs by applying two causal search algorithms, FASK and Two-Step, to voxel-level BOLD time series from a single individual's medial temporal lobe. For each pair of ROIs, subregions are defined as the sets of voxels that are adjacent to at least one voxel in the other ROI in the inferred directed graph. Using eight datasets formed by concatenating ten resting-state sessions each, the authors report Jaccard consistency of the inferred subgraphs and subregions across datasets, with high consistency for CA32DG-CA1 and CA1-SUB and lower consistency for entorhinal-related pairs. They conclude that the methods recover stable, spatially localized subregions, including across hemispheres, and suggest that most MTL inter-regional signaling is reciprocal.

Significance. If the assumptions underlying FASK and Two-Step hold, this work offers a principled alternative to seed-based correlation or partial-correlation parcellation, potentially recovering intra-ROI functional differentiation from effective-connectivity information. The paper is clearly written, uses publicly available longitudinal data, and builds on previously published, code-released algorithms. However, the empirical demonstration rests on a concatenation scheme that violates the algorithms' iid assumptions, and the consistency metrics lack null baselines and direct cross-method/cross-hemisphere comparisons. As a proof-of-concept, the idea is promising but the current evidence is insufficient to support the strong claims in the abstract.

major comments (4)

[Data and Voxel level graph results (pp. 6-8)] The concatenation of ten temporally ordered sessions into a single 4610-point time series violates the iid/exchangeability assumptions required by FAS-stable's BIC conditional independence tests and Two-Step's ICA step. BOLD time series are autocorrelated, and each session has its own baseline, drift, and motion profile; appending sessions creates artificial mean shifts and boundary discontinuities. Any such session-level factor is a common cause of many voxels, so it can create the exact adjacency structure FASK and Two-Step are designed to avoid, but as artifact rather than as true signaling. The Discussion's concession that 'investigation of stability over multiple scans of the same subject seems advisable' (p. 15) conflicts with the concatenation design. Please provide evidence that the inferred subregions are stable when analyses are performed on single sessions or when session effects are explicitly modeled (e.g., nuisance regressors, session-specific means), or temper the causal claims accordingly.
[Tables 1-4 and Jaccard discussion (pp. 7-11)] All consistency claims are based solely on Jaccard indices without any null baseline. The paper correctly notes that an overfitting complete graph yields Jaccard = 1, so high values such as 0.89 for S_CA32DG and 0.90 for S_CA1 in Table 4 do not by themselves indicate meaningful stability. Please report expected Jaccard values under a null model matched for graph density (or permutation-based nulls) and, if possible, confidence intervals, so the reader can assess whether the observed consistency exceeds chance.
[Abstract and Connectivity subregions section (pp. 11-13)] The abstract and text claim that FASK and Two-Step 'recovered similar subsets of voxels' and that subregions show 'spatial consistency across different scanning sessions and across hemispheres,' but no quantitative comparison is provided between the two methods for the same data, nor between the left and right hemispheres. Tables 3-4 only report within-method consistency across the eight concatenated datasets. Please add quantitative cross-method (FASK vs Two-Step) and cross-hemisphere similarity measures (e.g., Jaccard or Dice coefficients) for the identified subregions and report them in the main text or supplementary material.
[FASK and Two-Step Methods (pp. 3-5)] Both algorithms' correctness rests on the absence of unmeasured confounders: FAS-stable is a PC variant that assumes causal sufficiency, and Two-Step explicitly ignores the M/C terms ('In our application, we ignore the possible unobserved confounders C,' p. 4). Resting-state fMRI is subject to well-known physiological and scanner confounds. If latent common causes are present, the inferred adjacencies—and hence the subregions S_P—may reflect shared nuisance signals rather than effective connectivity. Please justify this assumption for the present data or provide robustness analyses (e.g., with global signal regression or inclusion of physiological regressors).

minor comments (5)

[Connectivity Subregions Algorithm (p. 7)] There is a typographical error in step 4: the notation 'v_P ∈ V_P' is repeated; the second instance should be 'v_Q ∈ V_Q'.
[References (p. 16)] The in-text citation 'Pearl, Glymour and Jewel' should be 'Pearl, Glymour and Jewell', and the reference list entry should be checked for consistency.
[References (p. 17)] The citation 'Jaccard et al., 1912' should be 'Jaccard, 1912', since the work is single-authored.
[Data (p. 6)] Please clarify whether the 57 volumes removed 'due to drifting' are removed from each session individually, and how the concatenation handles the session boundaries (e.g., any overlap, gap, or detrending).
[Voxel level graph results (p. 7)] The parameters for FASK (c = 1, alpha = 1e-7, Delta = 0.3) and Two-Step (sigma = 75 log(n), lambda = 20, t = 0.05) are set 'to match the sparsity obtained with FASK'; no sensitivity analysis is provided. Since the consistency results may depend on these choices, a brief robustness check or a reference to prior simulation calibration would strengthen the report.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the subregion definition is operational and the consistency results are fresh empirical outputs.

full rationale

The paper's derivation chain is: (1) run FASK or Two-Step on each concatenated voxel dataset to obtain a voxel-level directed graph; (2) drop edge orientations to form an adjacency graph; (3) for each ROI pair, define connectivity subregions as the sets of voxels incident to inter-region adjacencies in that graph. This is an explicit operational definition rather than a prediction derived from a fitted quantity: the S_P sets are read directly from the inferred graphs, and nothing in the algorithm is fitted to the subregion output to make the output appear. The free parameters (c, alpha, Delta, sigma, lambda, t) are stated before the runs, and the two algorithms were published previously with simulation support in Sanchez-Romero et al. (2019); citing that work is normal method attribution, not a load-bearing self-citation chain, and it is not used as a uniqueness theorem to forbid alternatives. The paper's empirical claims are the Jaccard consistency values across eight concatenated datasets and the qualitative similarity between the two methods' recovered subregions; these are new measurements made on the data, not consequences of the subregion definition. The main weakness—treating concatenated, autocorrelated BOLD sessions as exchangeable samples—is an assumption about the data and algorithm validity, not a circularity in the derivation. Therefore no step reduces by construction to its own inputs, and the analysis is self-contained with respect to circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The result depends on the causal discovery assumptions of FASK and Two-Step inherited from Sanchez-Romero et al. (2019): linearity, non-Gaussian noise, faithfulness, and, for Two-Step, no unmeasured confounding. In addition, the concatenation of multiple fMRI sessions requires stationarity and approximate iid sampling, which is not guaranteed. Six user-set tuning parameters are free choices that materially affect the inferred graphs.

free parameters (6)

FASK penalty c = 1
BIC penalty discount for FAS-stable conditional independence tests; affects the adjacency set.
FASK alpha = 1e-7
Significance threshold for 2-cycle detection correlation-difference tests.
FASK Delta = 0.3
Extra adjacency heuristic threshold; value taken from prior simulations, not from this dataset.
Two-Step sigma = 75*log(4610), about 633
Adaptive lasso penalty in step one; set to match FASK sparsity, a data-dependent choice.
Two-Step lambda = 20
Sparse ICA penalty for (I-B) estimation; user-adjusted.
Two-Step zeroing threshold t = 0.05
Threshold below which estimated B coefficients are set to zero.

assumptions (5)

domain assumption The BOLD signals are generated by a linear causal process with non-Gaussian noise terms.
Required for FASK's correctness; stated in Methods: 'The correctness of the output of FASK assumes that the noise terms in the causal process are skewed.'
domain assumption There are no unmeasured confounders affecting the voxel variables.
Two-Step's model includes C but the application sets C=0: 'In our application, we ignore the possible unobserved confounders C.'
domain assumption Concatenated sessions form an iid sample from a stationary process.
Data section concatenates ten sessions to 4610 datapoints; FAS-stable and ICA assume exchangeable independent samples.
domain assumption The true voxel-level graph is faithful to the observed distribution.
Needed for FAS-stable adjacency search to recover the true graph, inherited from Spirtes et al. (2000).
domain assumption Linear model X = BX + MC + E with sparse B and independent components E.
Assumed by Two-Step and sparse ICA; stated in the Two-Step section.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Identification of Effective Connectivity Subregions." pith.science (2026). https://pith.science/paper/T3GPBPLS

@misc{pith2026190803264,
  author       = {Pith},
  title        = {Pith review of: Identification of Effective Connectivity Subregions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3GPBPLS}},
  note         = {Machine review of arXiv:1908.03264}
}

read the original abstract

Standard fMRI connectivity analyses depend on aggregating the time series of individual voxels within regions of interest (ROIs). In certain cases, this spatial aggregation implies a loss of valuable functional and anatomical information about smaller subsets of voxels that drive the ROI level connectivity. We use two recently published graphical search methods to identify subsets of voxels that are highly responsible for the connectivity between larger ROIs. To illustrate the procedure, we apply both methods to longitudinal high-resolution resting state fMRI data from regions in the medial temporal lobe from a single individual. Both methods recovered similar subsets of voxels within larger ROIs of entorhinal cortex and hippocampus subfields that also show spatial consistency across different scanning sessions and across hemispheres. In contrast to standard functional connectivity methods, both algorithms applied here are robust against false positive connections produced by common causes and indirect paths (in contrast to Pearson's correlation) and common effect conditioning (in contrast to partial correlation based approaches). These algorithms allow for identification of subregions of voxels driving the connectivity between regions of interest, recovering valuable anatomical and functional information that is lost when ROIs are aggregated. Both methods are specially suited for voxelwise connectivity research, given their running times and scalability to big data problems.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Causal Transfer in Medical Image Analysis
cs.CV 2026-03 accept novelty 5.0 of 10

Causal Transfer Learning unifies structural causal models, invariant risk minimisation and counterfactuals with transfer learning to produce domain-robust medical image models.

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages · cited by 1 Pith paper

[1]

Arnold Anteraper, S., Guell, X., D’Mello, A., Joshi, N., Whitfield-Gabrieli, S., & Joshi, G. (2019). Disrupted cerebrocerebellar intrinsic functional connectivity in young adults with high-functioning Autism Spectrum Disorder: A data-driven, whole-brain, high-temporal resolution functional magnetic resonance imaging study. Brain Connectivity, 9(1), 48...

work page 2019
[2]

Kelly, C., Toro, R., Di Martino, A., Cox, C

New Phytologist, 11(2), 37–50. Kelly, C., Toro, R., Di Martino, A., Cox, C. L., Bellec, P., Castellanos, F. X., & Milham, M. P. (2012). A convergent functional architecture of the insula emerges across imaging modalities. NeuroImage, 61(4), 1129–1142. Kerr, K. M., Agster, K. L., Furtak, S. C., & Burwell, R. D. (2007). Functional neuroanatomy of th...

work page 2012
[1779]

Zhang, K., Peng, H., Chan, L., & Hyvärinen, A. (2009). ICA with sparse connections: Revisited. International Conference on Independent Component Analysis and Signal Separation, 195–202. Springer. Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical Association, 101(476), 1418–1429. 18 Identification of Eff...

work page 2009
[8885]

R., Bornstein, A

Preston, A. R., Bornstein, A. M., Hutchinson, J. B., Gaare, M. E., Glover, G. H., & Wagner, A. D. (2010). High-resolution fMRI of content-sensitive subsequent memory responses in human medial temporal lobe. Journal of Cognitive Neuroscience, 22(1), 156–173. 17 Sanchez-Romero, R., Ramsey, J. D., Zhang, K., Glymour, M. R. K., Huang, B., & Glymour, C. (2...

work page 2010

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Arnold Anteraper, S., Guell, X., D’Mello, A., Joshi, N., Whitfield-Gabrieli, S., & Joshi, G. (2019). Disrupted cerebrocerebellar intrinsic functional connectivity in young adults with high-functioning Autism Spectrum Disorder: A data-driven, whole-brain, high-temporal resolution functional magnetic resonance imaging study. Brain Connectivity, 9(1), 48...

work page 2019

[2] [2]

Kelly, C., Toro, R., Di Martino, A., Cox, C

New Phytologist, 11(2), 37–50. Kelly, C., Toro, R., Di Martino, A., Cox, C. L., Bellec, P., Castellanos, F. X., & Milham, M. P. (2012). A convergent functional architecture of the insula emerges across imaging modalities. NeuroImage, 61(4), 1129–1142. Kerr, K. M., Agster, K. L., Furtak, S. C., & Burwell, R. D. (2007). Functional neuroanatomy of th...

work page 2012

[3] [1779]

Zhang, K., Peng, H., Chan, L., & Hyvärinen, A. (2009). ICA with sparse connections: Revisited. International Conference on Independent Component Analysis and Signal Separation, 195–202. Springer. Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical Association, 101(476), 1418–1429. 18 Identification of Eff...

work page 2009

[4] [8885]

R., Bornstein, A

Preston, A. R., Bornstein, A. M., Hutchinson, J. B., Gaare, M. E., Glover, G. H., & Wagner, A. D. (2010). High-resolution fMRI of content-sensitive subsequent memory responses in human medial temporal lobe. Journal of Cognitive Neuroscience, 22(1), 156–173. 17 Sanchez-Romero, R., Ramsey, J. D., Zhang, K., Glymour, M. R. K., Huang, B., & Glymour, C. (2...

work page 2010