Which Contemporaneous fMRI Edges Are Hidden Common Drivers?

Whole-brain causal discovery on resting-state fMRI now runs at the scale of the Human Connectome Project (HCP). These methods take brain-activity time series from about a hundred brain regions and output a directed graph of who drives whom. They all rest on one assumption, causal sufficiency: every common cause of two measured regions is itself measured. That assumption is almost certainly false in the brain. Subcortical nuclei sit outside cortical parcellations, breathing and heartbeat are not modeled, and neighbouring areas are lumped together. When it fails, two regions that merely share an unmeasured driver get reported as one region causing the other. The best-known whole-brain result on HCP data reports that about 70% of its recovered edges are contemporaneous, meaning they connect signals at the same time point. Same-time dependence is exactly the pattern an unmeasured common driver produces. So we do not currently know how much of the published "causal connectome" is direct influence and how much is shared input.

We propose a two-stage method that scales to 116 regions and can label an edge as "confounded" rather than forcing a direction onto it. We pair it with a test on real data that needs no ground truth. The test hides measured regions on purpose and checks whether the method's answer on the remaining regions agrees with what its own full-brain answer implies. The method is a reweighting of existing tools. The real contribution is the measurement: an estimate, with a falsifiable validity check, of what fraction of HCP contemporaneous edges are bidirected (confounded) rather than directed.

Core Claim

The claim has two parts. First, separate the lagged and contemporaneous parts of the problem. Regress each region's signal on the full past of all regions, the move CaLLTiF already makes. What is left are "innovations": the part of each signal not predictable from the past. A latent-confounder-aware search over these innovations then needs only about 116 variables and has very many samples. Second, HCP supplies a quality check that toy benchmarks lack. HCP has a subcortical block of 16 regions that can be measured or hidden at will, and two scanning days per subject. These allow a check of marginalization consistency: a correct latent-aware method run on a subset of regions should reproduce the latent projection of its full-set output. Latent projection is the standard rule for turning a graph with hidden nodes into one over the remaining nodes, where hidden common parents become bidirected edges.

Methods that assume sufficiency must fail this check whenever a hidden region has two or more measured children. Such methods have no symbol for a bidirected edge, so they must write either a directed edge or nothing. Measured regions with a shared hidden parent remain dependent no matter what else is conditioned on, which rules out "nothing" as a correct output. The check uses the identifiability result in Learning Linear Non-Gaussian Causal Models in the Presence of Latent Variables, which shows that a latent with only one child cannot be told apart from that child's noise. We therefore score only hidden regions with at least two measured children, where a difference is detectable in principle.

Positioning Against Prior Work

The work that motivates this proposal, and also its main competition, is the two HCP studies. Whole-Brain Causal Discovery Using fMRI (CaLLTiF) conditions on all lagged variables and then orients same-time edges. On 200 HCP subjects with 116 regions, it reports roughly 70% contemporaneous edges and a flow from attention and default-mode networks toward sensorimotor networks. Causal Discovery for fMRI data: Challenges, Solutions, and a Case Study (GANGO) finds the graph skeleton with a greedy score search and orients edges using skewness and kurtosis. On 442 subjects with 360 cortical regions, it reports about 2.25% edge density. GANGO's authors name unmeasured subcortical and autonomic inputs as an open source of bias. Neither method can output a bidirected edge. We keep CaLLTiF's lag conditioning and replace the sufficiency-based contemporaneous step.

The method itself comes from the latent-confounder literature. Iterative Causal Discovery in the Possible Presence of Latent Confounders and Selection Bias (ICD) returns a partial ancestral graph (PAG), a graph whose edge marks can encode "possibly confounded". It ties the size of each conditioning set to graph distance, which keeps the number of tests bounded, and it reports a 5.6x runtime advantage over FCI on 35 nodes. An implementation is available at IntelLabs/causality-lab. We run ICD on the innovations with a capped conditioning radius. Local Causal Discovery Without Causal Sufficiency (LatentLCD) detects bidirected edges locally around one target. We use it as a per-edge cross-check on the edges that matter most scientifically.

Other work constrains the design. Higher-order cumulant methods, BANG and ReLVLiNGAM, need very large samples, and both were tested on at most about 15 variables. ParceLiNGAM has only been run on a 10-node simulated fMRI network. We therefore do not use any of them as the global search. Instead, pooled HCP innovations give on the order of a million time points. That is enough to apply cumulant-based orientation to individual edges that ICD leaves undecided, which turns their main weakness, sample size, into a non-issue at local scale.

Towards Robust and Specific Causal Discovery from fMRI adds an observed confounder indicator. It handles only confounders that indicator captures, and we treat it as a baseline to beat rather than a method to reuse.

Research Program

Stage one fits each subject's lagged model by regressing every region on all regions at lags 1 to $L$. The lag order $L$ is chosen by the same multiple-comparison logic CaLLTiF uses. This stage keeps that method's lagged edges. Stage two pools innovations within subject across the four runs, which gives 4,800 time points. On these innovations it runs ICD with kernel or partial-correlation tests and a conditioning radius of 2 or 3. The output is a per-subject PAG over contemporaneous structure. A group graph is formed by stability selection across subjects. Stage three revisits ICD's undecided edge marks. For those edges only, it applies a pairwise cumulant rank test on group-pooled innovations, and LatentLCD supplies a second local opinion on selected edges.

The scientific analyses then ask three questions. What fraction of CaLLTiF's contemporaneous edges does our method mark as bidirected? Are those edges concentrated among cortical pairs that share a thalamic or basal-ganglia parent in the full 116-region graph? Does the attention/default-mode-to-sensorimotor flow survive once confounded edges are removed?

Baselines

The naive reference point is the sparse partial-correlation network: the inverse covariance matrix of the same regions, thresholded to match each method's edge density. It costs nothing to compute and is the standard sufficiency-free statistical summary. Any causal method that cannot beat it on our metrics is not adding anything.

The strongest existing method on this exact question is CaLLTiF (Whole-Brain Causal Discovery Using fMRI). It is the only published method validated at 116 HCP regions and on the 91-node macaque simulation. We run it from released descriptions on identical subjects, parcellation and preprocessing, with the same $\alpha$-selection rule, and we give it the same compute budget. The file reports only qualitative results for it: higher F1 than PCMCI, MVGC, DYNOTEARS and VARLiNGAM on the Sanchez-Romero and macaque simulations, and about 70% contemporaneous edges on HCP. No reported numerical F1 for CaLLTiF was found on Lacuna, so we will regenerate its numbers in our simulations.

GANGO is the second strong baseline. We run it at 116 regions, and at 360 regions where compute permits. For the latent-aware side, the comparison point is plain ICD run on raw time series without the lagged residualization stage. Its prior numbers come from smaller settings: 5.6x faster than FCI at 35 nodes. On simulated confounded fMRI, the only reported reference point is ParceLiNGAM's ordering precision of 0.54–0.60 against 0.24–0.31 for pairwise LvLiNGAM on the 10-node confounded simulation from Dynamic Causal Modeling (DCM, a standard biophysical fMRI generative model). We will rerun ParceLiNGAM there as a sanity anchor.

The key ablation swaps ICD for PC inside our own pipeline. PC is the classic constraint-based search that assumes sufficiency. With the lag stage, tests and budget all held fixed, this swap isolates the effect of allowing bidirected edges.

The method counts as beating CaLLTiF if it does all of the following. On simulations with hidden nodes, its arrowhead precision and bidirected-edge F1 must be significantly higher while its skeleton F1 does not drop. On real data, its marginalization-consistency score must be significantly higher, and its day-to-day reproducibility must not be worse.

Why not just…?

Why not just regress out nuisance signals and run CaLLTiF? This means removing motion, global signal and other proxies, the logic of Towards Robust and Specific Causal Discovery from fMRI. This is a real competitor, so we include "global-signal regression plus CaLLTiF" as a baseline. Regression can only remove confounders that the proxies capture. Our hide-the-subcortex test creates confounders that are, by construction, not in any nuisance regressor: real thalamic signals. If nuisance regression alone passes the marginalization-consistency test as well as our method does, we will report that the simpler fix suffices.

Why not just use partial correlation? It is undirected and also assumes sufficiency. A shared hidden driver leaves a nonzero partial correlation that looks exactly like a direct link. It cannot express "confounded", so it also fails marginalization consistency by construction. It still serves as our density-matched floor.

Why not just run FCI or ICD directly on the raw time series? The search over lagged copies of 116 regions multiplies the number of variables, and prior latent-aware methods have been shown only up to 35 nodes. The "ICD without residualization" baseline measures directly whether the lag stage is needed. If raw ICD matches us at feasible cost, the residualization step is unnecessary and we drop it.

Why not just trust sufficiency-based methods, since CaLLTiF is consistent across subjects? Consistency across subjects does not establish correctness. A shared subcortical driver is itself consistent across subjects. The hide-region experiment tests this objection directly. If CaLLTiF's subset outputs already match the latent projection of its full outputs, sufficiency is empirically harmless here, and we would report that as a result.

Proposed Experiments and Decisive Checks

In simulation, we use the 91-node macaque connectome and the Sanchez-Romero topologies. We hide 10–20% of nodes, chosen from those with at least two observed children. We also include a nonlinear BOLD forward model, so that non-Gaussian structure comes partly from confounder mixing. This setting probes the tension GANGO raises about whether non-Gaussian features are real causal asymmetry or artifact. Metrics are skeleton F1, arrowhead precision, and bidirected-edge precision and recall against the true latent projection.

On HCP S1200, we use the same 200-subject, 116-region setup as CaLLTiF, with a replication sample of different subjects. The decisive check hides the 16 subcortical regions. The prediction is specific. CaLLTiF and PC-in-our-pipeline will add directed cortical edges between pairs that share a subcortical parent in the full graph. Our method will mark most of these pairs bidirected, and its subset output will match the latent projection of its full-graph output on a significantly larger fraction of edges. As a control, we hide random sets of cortical regions of equal size. Reproducibility is measured as the edge overlap between graphs estimated separately from day-1 and day-2 runs.

The headline quantity is the fraction of CaLLTiF's contemporaneous edges that our method marks bidirected, with bootstrap intervals. Results near zero would vindicate prior whole-brain graphs. A large fraction would mean that much of the published directed flow should be reinterpreted as shared input.