Which Brain-Rhythm Lead-Lag Links Survive Across People?

When one brain signal reliably moves a few milliseconds before another, neuroscientists often read that as one region driving the other. Two things make that reading unsafe. First, an unrecorded third region can drive both at different delays and produce a lead that means nothing. Second, the standard step of splitting a signal into frequency bands such as alpha (8–12 Hz) or gamma (above 30 Hz) runs the data through a filter, and the filter can itself shift timing. We propose a simple test that targets both problems. A lead-lag link within a frequency band counts as a candidate causal link only if the lagged prediction it implies stays the same across subjects and across reasonable choices of filter. Links that pass are kept; links whose strength or sign changes from person to person, or from filter to filter, are discarded as products of confounding or of the filter.

This gives cross-subject variability a use. Small neuroimaging studies usually treat it as noise that drains statistical power. Under the logic of invariant causal prediction, the same variability helps separate causes from correlates: a true causal mechanism should look the same in every environment that leaves it intact, while a spurious association usually will not. We will build this test for band-resolved scalp EEG. We will measure, on simulations with known ground truth, how much it reduces false directed links. We will also measure how the reliability of the surviving links scales with the number of subjects, using the 123-subject FACED dataset.

Core Claim

Band-specific lagged dependencies that are invariant across subjects and across filter designs recover true directed interactions more precisely than pooled or pairwise lag methods. They also reproduce better between disjoint groups of subjects. We expect this to hold in a specific regime: when latent drivers or filtering artifacts vary across environments. In the regime where a confounder's delays are identical in everyone, the method should offer no help, and we will show that failure explicitly rather than hide it.

The claim rests on a conditional-invariance statement in the style of Causal Inference by using Invariant Prediction. Let $x^{(b)}_j(t)$ be the band-$b$ signal (or its amplitude envelope) at channel $j$. Let $S$ be a candidate set of parent channels with lags up to $L$. Let $e$ index an environment, meaning a subject or a filter variant. A set $S$ is accepted for target $j$ in band $b$ if

$$ P\big(x^{(b)}_j(t)\,\big|\,\{x^{(b)}_k(t-\ell)\}_{k\in S,\,1\le\ell\le L},\ x^{(b)}_j(t-1{:}L)\big)\ \text{is the same for every } e, $$

and the inferred parents of $j$ are the intersection of all accepted sets.

Relation to Prior Work

Several papers motivate this work by showing that lag alone is not enough. Fine-scale cross-correlograms, which plot how often one signal fires at each time offset from another, link millisecond synchrony to slower correlations (Bair et al.). Jitter correction subtracts the slow baseline to isolate precise offsets (Jia et al. 2020). Both remain pairwise, and both are blind to an upstream driver with unequal delays. Precedence findings in EEG, such as prestimulus phase predicting perception (Busch et al.), are exactly the kind of result that VanRullen et al. caution could come from unobserved upstream or recurrent drivers.

Other papers constrain the design. Network modelling methods for FMRI found that lag-based methods detect fewer than 20% of true connections on simulated BOLD, the blood-oxygen fMRI signal, because blood-flow delays swamp neural ones. The same study found that partial correlation and regularized inverse covariance detect more than 90% of edges but sit near 50% on direction. We take two lessons from this. We work on EEG, where lags are neural. We also use inverse covariance for what it does well, finding which pairs are connected, and leave direction to the invariance test. The Case for Time in Causal DAGs shows that aggregating over time creates dependencies that can point backward in time. Band-pass filtering is a continuous form of such aggregation, which is why we treat filter choice as an environment rather than a fixed preprocessing step.

Our technique comes from the invariance literature. Peters et al. supply invariant causal prediction, with error-rate guarantees, but for independent samples and with a search that is exponential in the number of variables. Weichwald & Peters argue conceptually that invariance across subjects can substitute for interventions in neuroscience, but report no experiments. Differentiable Invariant Causal Discovery makes invariance differentiable, but only for static graphs.

The closest temporal methods do the opposite of what we propose. They fit one broadband model and assume no hidden confounders. Neural Granger Causality uses sparse per-channel networks. Shapley Regularized Neural Granger Causality explicitly assumes causal sufficiency, meaning no unobserved confounders, and ignores zero-lag effects. CausalFormer is an interpretable transformer. None of the three splits signals into bands or tests across subjects.

On the EEG side, GNN4EEG uses band-wise features but builds graph edges from distance or from correlations without direction. LaBraM predicts spectra at scale but produces no directed graph. Neither addresses the questions we pose.

Research Program

The method has three stages.

The first stage finds candidate connections. For each band, we decompose the signal with causal (one-sided) filters and with zero-phase filters. Comparing the two isolates distortions introduced by filtering. We then estimate a sparse inverse-covariance skeleton pooled over subjects, which keeps each channel's candidate parent set small enough for the exponential search to be tractable.

The second stage is the invariance test. For each target channel, band, and candidate set, we fit a low-order lagged regression within each environment, either linear or a small additive nonlinear model. We then test whether the coefficients and residual distribution are equal across environments. Environments are defined two ways: as individual subjects, and as filter variants. The filter variants are band edges shifted by ±1–2 Hz, alpha bands centered on each individual's own alpha peak, and different filter orders. Shifted bands address the arbitrary-boundary problem; varying the filter order addresses spectral leakage. Because the signals are autocorrelated, they violate the independence assumption of invariant causal prediction. We handle this by prewhitening residuals and calibrating p-values with a block bootstrap. Scalp volume conduction, the smearing of one source's signal across nearby electrodes at zero lag, is handled with a surface-Laplacian montage and by excluding zero-lag terms. We will report this choice as a limitation rather than a solution.

The third stage is a sample-size analysis. We subsample the number of subjects (environments) and measure two things as functions of N: the precision and recall of the surviving links, and how often the same links are found in two disjoint halves of the subjects. This yields the practical number this field lacks: how many subjects a band-resolved directed-connectivity claim needs before it reproduces.

Baselines

The naive reference point is jitter-corrected pairwise cross-correlation within each band, the procedure of Jia et al. 2020. Peak significance is assessed against surrogates and corrected for false discovery rate across pairs, bands, and lags. It is the method most experimentalists would actually use, so the method must beat it on direction precision without losing more than a small amount of recall.

The strongest existing temporal causal-discovery competitor is CausalFormer, run from the released code (lingbai-kong/CausalFormer). It reports an F1 of 0.66 on NetSim, against 0.61 for CUTS and 0.59 for TCDF. On NetSim, Learning domain-specific causal discovery from time series reports AUPRC of about 0.9 for a supervised learned detector. However, that detector needs ground-truth perturbation data that does not exist for human EEG, so we treat it as an upper reference rather than a competitor. We will also run Shapley Regularized Neural Granger Causality and the cMLP model from Neural Granger Causality, because they are the sparse lagged models whose causal-sufficiency assumption our method relaxes. Each baseline will be fit on subject-pooled data. To match our use of bands, each will also be fit separately per band. All methods get the same channels, the same lag budget, the same nested-cross-validation tuning budget, and comparable GPU time. On our EEG simulator there is no reported result found on Lacuna for any of these methods, so all numbers there will be produced by us.

For edge detection on fMRI, the reference is inverse covariance, with reported sensitivity above 90% and direction accuracy near 50% in Network modelling methods for FMRI. The best reported direction accuracy is about 65%, for Patel's τ. On real EEG, the downstream reference is RGNN in GNN4EEG, at 31.3% accuracy for 9-class cross-subject classification on FACED.

Two ablations isolate the key ingredient. The first keeps everything except the invariance test, replacing it with pooled lagged regression plus significance testing. The second keeps the invariance test but uses only subjects as environments, dropping filter variants. Beating the baselines means higher direction precision at matched recall in simulation, and higher between-half reproducibility on FACED. We would accept a modest recall loss, which is the expected cost of conservatism.

Why not just…?

Why not just use jitter-corrected pairwise cross-correlograms with group statistics? Pairwise tests cannot separate a direct link from a two-step chain, and an upstream driver with unequal delays produces a stable, significant lead between two unconnected channels. The simulation includes both structures. We expect the pairwise method to report these false links with confidence, and pooling more subjects only makes the false lead more significant.

Why not just fit a multivariate neural Granger model on pooled data? Conditioning on all recorded channels does fix the chain problem. If the confounder is recorded, this alternative is a real threat, and the first ablation tests exactly that case. The argument for invariance applies only to unrecorded drivers whose influence differs across subjects. Shapley Regularized Neural Granger Causality assumes such drivers away. If the pooled model matches our precision in the hidden-driver simulations, the invariance step is unnecessary, and we will say so.

Why not just use inverse covariance and skip direction? For detecting which pairs are connected it is excellent: above 90% sensitivity in Network modelling methods for FMRI, which is why we adopt it as stage one. But it is symmetric and sits at chance on direction, and direction is the scientific question.

Why not just fine-tune a foundation model like LaBraM for cross-subject robustness? If the goal were classification accuracy, this might well win. LaBraM reports 0.6616 balanced accuracy on TUEV against 0.5281 for BIOT. We do not claim to beat it on accuracy. Our claim concerns recovering directed links. The FACED classification test is only a secondary check that invariant edges carry transferable structure, and a null result there would not undercut the main claim.

Proposed Experiments

The first experiment uses a ground-truth EEG simulator. We will build coupled band-limited oscillators. Alpha links will run top-down and gamma links bottom-up, so that broadband averaging corrupts the edges. The simulator also includes hidden drivers and a realistic channel-mixing matrix. Across simulated subjects we vary the hidden drivers' delays and gains, and in a control condition we hold them fixed. Metrics are direction precision, recall, AUPRC, and the rate of reversed edges, swept over 10 to 120 subjects. The decisive check is a crossover. Invariance should sharply cut false links when hidden-driver delays vary across subjects and should give no benefit when they are fixed. A benefit in both conditions would suggest the test is simply conservative, not causal. Separately, the filter-variant environments should remove edges reversed by zero-phase filtering that survive the subject-only ablation.

The second experiment is a negative control on NetSim. Following Network modelling methods for FMRI, lag-based direction should fail on blood-flow signals. If our method nonetheless claims many confident directions there, its calibration is broken.

The third experiment uses FACED with 123 subjects. Real EEG has no ground truth, so we measure reproducibility instead. We split subjects into disjoint halves many times and compute, for each N, the overlap of the band-resolved directed edge sets found in each half. We compare against pairwise cross-correlation and pooled CausalFormer. The prediction is that invariant edges reproduce better at every N and level off at a smaller N. As the secondary check, we use these edges as the graph inside RGNN under the GNN4EEG nested cross-validation protocol, and compare against its 31.3% cross-subject accuracy and against correlation-based edges.

If invariant edges neither reproduce better nor transfer better, then cross-subject invariance does not act as a causal filter for EEG lags at this scale, and that is itself a result the field needs.