Bias, Measurement Error, and Double-Dipping: When Can GNN Convolutions Help Brain Connectome Prediction?

Tommaso Castellani, Jiaqi Li, Muriah D. Wheelock, Rezwana Rosen Razzaque, Helmut Laufs, Hong Chen, Claire DonnatarXiv 2026 · arXiv

TL;DR

While Graph Neural Networks (GNNs) are a go-to for analyzing brain connectivity, they frequently fail to outperform simple linear baselines. This paper provides a rigorous statistical explanation for this phenomenon, revealing that GNNs only provide a benefit in a narrow "Goldilocks" regime where the noise reduction from graph smoothing outweighs the signal distortion and estimation errors—a condition rarely met in standard fMRI pipelines.

What Problem Does This Paper Address?

In neuroscience, researchers use fMRI data to build "connectomes"—matrices representing the functional connectivity between different brain regions (ROIs). GNNs are theoretically ideal for this because they can incorporate the graph structure of the brain. However, recent benchmarks have shown a frustrating trend: simple, graph-agnostic models like SVMs or even basic linear regressions often match or beat state-of-the-art GNNs.

The authors tackle the fundamental question: Why does message passing often fail in brain connectome analysis? They argue that the issue stems from the fact that we never observe the "true" connectome; we only have a noisy estimate derived from a finite time series. This creates an errors-in-variables (EIV) problem where the noise in the data corrupts both the features and the graph structure itself.

Methods

The authors formulate connectome prediction as a statistical estimation problem. They decompose the impact of graph convolution into three distinct mechanisms:

  1. Smoothing-induced Bias and Variance: Averaging features over a graph reduces noise (variance) but can distort the actual predictive signal (bias).
  2. Graph Estimation Error: Because the graph is estimated from noisy data, "smoothing" often happens over the wrong edges.
  3. Double-Dipping: In standard pipelines, the same noisy matrix is used to define both the node features and the graph edges, coupling the noise in a way that can be detrimental.

To analyze these effects, they use a latent community model tailored to neuroscience, where ROIs are assumed to belong to functional communities. This allows them to mathematically define a "favorable regime" where GNNs should theoretically help.

Figure 2: Brain connectome GNN pipeline and statistical challenges. Top: the standard analytical pipeline, from ROI time series to prediction byi. Bottom: statistical challenges induced by the pipeline. The errors-in-variables component (red) is analyzed in this paper; the remaining challenges (gray, dashed) are left to future work.

Datasets

The study validates its theory using both synthetic simulations and three major real-world fMRI datasets:

  • ABIDE (Autism Brain Imaging Data Exchange): Used for age regression and autism diagnosis classification.
  • ADHD-200: Used for diagnosis classification and predicting inattentive symptom scores.
  • Frankfurt Dataset: Concurrent EEG-fMRI recordings used to classify sleep vs. wake states.

Key Insights and Ideas

  • The Smoothing Trade-off: Graph convolution acts as a low-pass filter. It only helps if the "signal" you are trying to predict is "community-smooth" (i.e., similar across ROIs in the same functional group). If the predictive signal is localized to specific edges or contrasts, smoothing actually destroys the information you need.
  • The "Oracle" Gap: Simulations show that if you had the perfect graph (the "Oracle"), GNNs would help significantly. However, as soon as you estimate the graph from the same noisy data used for features, most of that gain vanishes due to estimation error.
  • Double-Dipping is a Distraction: While "double-dipping" (using the same data for graph and features) is often blamed for issues, the authors find its effect is relatively small compared to the massive "bottleneck" of general graph estimation error.
  • Recording Length Matters: GNNs are most likely to help when recordings are very short (high noise). As the recording length ($T$) increases, the raw features become clean enough that the bias introduced by smoothing is no longer worth the small reduction in variance.
Table 1: Graph constructions and prediction conclusions. Comparisons concern unregularized population linear prediction under the stated assumptions.

Key Results

  • Baseline Dominance: On real-world datasets, standard GNNs (GCN, GAT, GraphSAGE) consistently fail to outperform simple linear models and Random Forests.
  • Narrow Favorable Regime: In simulations, the authors show that GNNs only win when the recording length is short and the signal is perfectly aligned with the community structure.
  • Validation of Theory: By using "proxies" for oracle graphs on real data (e.g., using the full-length recording to build a graph for a short-window prediction), they confirmed that the theoretical "Oracle" gain exists but is lost in practical pipelines.
Figure 1: Performance across regression and classi- fication tasks on 3 datasets. Graph-based methods fail to outperform graph-agnostic baselines, both linear (Linear Model) and non-linear (Random For- est, SVM, MLP). Mean ± 1 SD over 10 seeds. See Appendix C for details.

Limitations and Future Work

The theoretical framework primarily focuses on linear models in an asymptotic setting (infinite subjects). While the authors show empirically that non-linearities like ReLU and Attention don't change the overall picture, a full non-linear theory is still missing.

The authors suggest that future work should move away from simple thresholding for graph construction. Since the "noisy graph" is the primary bottleneck, developing more robust ways to estimate the underlying latent structure of the brain could finally allow GNNs to live up to their potential in connectomics.

Summary updated

Abstract

Graph Neural Networks (GNNs) are an important tool for fMRI prediction, where empirical functional connectivity matrices serve both as features and for graph construction. However, recent studies and our own experiments show that these methods often fail to outperform simpler graph-agnostic baselines. We propose a statistical explanation for this failure by formulating connectome prediction as an errors-in-variables (EIV) problem: the true connectome is latent, while the observed covariance matrix is a noisy finite-time-series estimate. In a linear asymptotic setting, we decompose the effect of graph convolution into three mechanisms: smoothing-induced bias and variance reduction, graph estimation error, and a double-dipping effect arising when the same noisy matrix determines both graph and features. Under a latent community model adapted to neuroscience, we identify a narrow favorable regime, set by the brain community structure, the predictive signal, the graph construction and the recording length, in which message passing can improve prediction. Simulations validate this decomposition, while real-data experiments illustrate that connectome GNN pipelines often fall outside this regime. Our results provide a first step towards explaining when graph convolutions can help brain connectome prediction, and why they often do not.
BibTeX
@misc{castellani2026bias,
  author = {Tommaso Castellani and Jiaqi Li and Muriah D. Wheelock and Rezwana Rosen Razzaque and Helmut Laufs and Hong Chen and Claire Donnat},
  title = {{Bias, Measurement Error, and Double-Dipping: When Can GNN Convolutions Help Brain Connectome Prediction?}},
  year = {2026},
  eprint = {2611.10364},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10364},
}
Figure