📊 statistics

A Mixed Model Approach for Estimating Regional Functional Connectivity from Voxel-level BOLD Signals

This paper proposes a novel mixed-effects modeling approach that explicitly accounts for intra-regional variability and measurement error to estimate regional functional connectivity, demonstrating through simulations and real-world data that it significantly outperforms the traditional Correlation of Averages method by reducing bias and improving test-retest reliability.

Original authors: Ruobin Liu, Chao Zhang, Chau Tran, Sophie Achard, Wendy Meiring, Alexander Petersen

Published 2026-06-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruobin Liu, Chao Zhang, Chau Tran, Sophie Achard, Wendy Meiring, Alexander Petersen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a massive, bustling city. In this city, there are thousands of neighborhoods (called voxels in brain scans) grouped into larger districts (called regions). When you rest, these neighborhoods are constantly chatting with each other, sending signals back and forth. Scientists want to map these conversations to understand how the city works, especially when things go wrong, like in diseases.

For a long time, the standard way to listen to these conversations was a bit like trying to understand a neighborhood by asking a single "spokesperson" to summarize what everyone said. This method, called Correlation of Averages (CA), takes all the tiny voices in a district, mashes them into one average voice, and then compares that average to the average voice of another district.

The authors of this paper argue that this "spokesperson" method is flawed. It's like trying to hear a whisper in a noisy room by only listening to the loudest person; you miss the nuance, the background chatter, and the specific details of who is talking to whom. Because of this, the old method often gets the connections wrong, either missing real friendships between districts or inventing fake ones.

The New Approach: Listening to Every Voice

The authors propose a new, smarter way to listen. Instead of averaging the voices, they use a Mixed Model. Think of this as a high-tech sound engineer who can isolate every single voice in the city, even the quiet ones, while filtering out the static and noise (like heartbeats or scanner glitches).

Here is how their new system works, broken down into simple steps:

1. The Two-Stage Detective Work
Because the city is so huge (thousands of neighborhoods), listening to everyone at once is computationally impossible for a standard computer. So, the authors built a two-step detective process:

Stage 1 (The Local Scout): First, they send a scout into each district individually. The scout learns the local rules: How noisy is this neighborhood? How much do the houses within this district talk to each other? They figure out the "local noise" and "local chatter" without worrying about other districts yet.
Stage 2 (The City-Wide Map): Once the scouts have their local reports, the main detective looks at pairs of districts. Using the local reports from Stage 1, they can now accurately measure how much District A is truly talking to District B, stripping away the local noise that used to confuse the old method.

2. The Magic Shortcut (Vecchia's Approximation)
Even with the two-stage plan, the math is still incredibly heavy. To solve this, they used a clever mathematical shortcut called Vecchia's approximation.

The Analogy: Imagine trying to predict the weather for a whole continent by looking at every single tree. It's too much data. Instead, you look at a tree, then only look at its 100 closest neighbors to make a prediction, ignoring the trees on the other side of the world. This shortcut allows the computer to do the heavy lifting quickly without losing much accuracy.

Why Does This Matter? (The Proof)

The authors didn't just invent this; they tested it rigorously.

The Simulation Test: They created fake brain data in a computer where they knew the "true" answers. When they compared their new method to the old "spokesperson" method, the new method was much more accurate. It didn't get tricked by noise, and it didn't invent fake connections. It found the real friendships between brain districts with much higher precision.
The Real-World Test: They applied this to real data from the Human Connectome Project, which involves scanning the same people's brains twice (a "test-retest"). If a method is good, it should give you the same map both times.
- The old method (CA) gave somewhat different maps the second time, like a blurry photo that changes slightly every time you take it.
- The new method gave much more consistent maps, like a sharp, high-definition photo that looks the same every time you snap it.

The Bottom Line

This paper introduces a new statistical tool that treats brain data with the respect it deserves: by listening to every single voxel instead of just averaging them out. By using a two-step process and a clever mathematical shortcut, they can build a much clearer, more reliable map of how different parts of the brain connect.

The result is a method that is less biased, more consistent, and statistically sound, offering neuroscientists a better way to see the true architecture of the human brain.

Technical Summary: A Mixed Model Approach for Estimating Regional Functional Connectivity from Voxel-level BOLD Signals

Problem Statement
Resting-state functional connectivity (FC) quantifies the synchrony between activity patterns of anatomically separated brain regions. In functional magnetic resonance imaging (fMRI), data are acquired as blood-oxygen-level-dependent (BOLD) signals across numerous spatial voxels. The prevailing method for estimating inter-regional connectivity is the Correlation of Averages (CA) estimator, which aggregates voxel-level signals within a region (typically via spatial averaging) and computes the Pearson correlation between these regional averages.

While computationally simple, the CA estimator suffers from demonstrable bias and a lack of statistically principled justification. As shown by Achard et al. (2023), the CA estimator converges to a value that depends not only on the true inter-regional correlation but also on nuisance parameters, including intra-regional spatial correlation, signal-to-noise ratios, and the number of voxels. Consequently, CA estimates can be severely biased, particularly when spatial correlations are high or signal strength is low. Furthermore, existing voxel-level modeling approaches (e.g., Kang et al., 2012; Castruccio et al., 2018) have primarily focused on task-based activation or group-level inference, often failing to clearly distinguish intra-regional from inter-regional variability or to provide robust individual-level network estimation.

Methodology
The authors propose a rigorous statistical framework based on linear mixed-effects models to estimate inter-regional functional connectivity directly from voxel-level BOLD signals.

Spatiotemporal Mixed-Effects Model:
The observed BOLD signal at voxel $v_{jl}$ in region $R_j$ and wavelet coefficient $m$ is modeled as:
$X_{jlm} = \mu_j + \eta_{jm} + \gamma_{jlm} + \epsilon_{jlm}$
- $\mu_j$ : Regional mean level.
- $\eta_{jm}$ : A zero-mean random effect representing the regional signal component, inducing inter-regional dependence via correlation matrix $R = \{\rho_{jj'}\}$ .
- $\gamma_{jlm}$ : A zero-mean local random effect capturing intra-regional spatial dependence, uncorrelated with $\eta_{jm}$ .
- $\epsilon_{jlm}$ : Uncorrelated measurement noise.
  The model operates in the wavelet domain to isolate scale-specific temporal dependence and reduce high-frequency noise. The covariance structure assumes separability between spatial and wavelet domains, utilizing Matérn and Gaussian kernels.
Two-Stage Estimation Procedure:
To address the computational intractability of full maximum likelihood estimation (MLE) for high-dimensional spatiotemporal data (where $N$ can exceed 20,000 voxels), a two-stage approach is developed:
- Stage 1: Restricted Maximum Likelihood (ReML) is applied to data from each region independently to estimate intra-regional covariance parameters (spatial correlation, noise variance). This step treats the regional random effects as fixed to isolate local structure.
- Stage 2: The inter-regional correlation parameters ( $\rho_{jj'}$ ) are estimated by conditioning on the Stage 1 estimates. For pairs of regions, the authors employ Vecchia's likelihood approximation (Vecchia, 1988; Guinness, 2018). This method approximates the joint likelihood as a product of conditional distributions based on a sparse set of nearest neighbors, reducing computational complexity from $O(N^3)$ to a tractable level while maintaining estimation accuracy.
Inference:
The asymptotic properties of the estimators are established under an expanding time-window regime (diverging number of wavelet coefficients). Approximate confidence intervals are constructed using the Fisher information matrix derived from the Vecchia approximation.

Key Contributions

Novel Estimator: The paper introduces a maximum likelihood-based estimator for inter-regional connectivity derived from a mixed-effects model, moving beyond the moment-based CA estimator.
Explicit Modeling of Variability: The framework explicitly separates and models sources of variability: inter-regional correlation, intra-regional spatial dependence, and measurement noise. This allows for unbiased estimation of connectivity parameters that are not confounded by regional signal aggregation artifacts.
Computational Scalability: This work presents the first application of Vecchia's likelihood approximation to functional connectivity estimation with voxel-level fMRI data, enabling the analysis of massive datasets that were previously computationally prohibitive for full likelihood methods.
Improved Reproducibility: The method is designed to produce more consistent individual-level brain networks, addressing the challenge of test-retest reliability in fMRI.

Results

Simulation Studies: Extensive simulations comparing the proposed ReML and Vecchia estimators against CA, the Average of Covariances (ACE), and EBLUE estimators demonstrate superior performance. The proposed estimators exhibit significantly lower bias and mean squared error, particularly in scenarios with high intra-regional spatial correlation or low signal-to-noise ratios. Crucially, the proposed method achieves nominal coverage for 95% confidence intervals, whereas CA-based intervals systematically fail to cover the true parameter.
Human Connectome Project (HCP) Application: The method was applied to test-retest data from 42 young adult subjects. Connectivity networks were constructed for both the proposed method and the CA estimator. Using the Concordance Correlation Coefficient (CCC) to measure agreement between repeated scans (Exam 0 vs. Exam 1), the proposed mixed model approach consistently yielded higher concordance than CA. This indicates that the proposed method produces more reliable and reproducible individual brain networks.

Significance
The paper argues that the widespread use of the CA estimator introduces systematic bias into functional connectivity studies, potentially affecting the validity of group comparisons and individual characterization. By providing a statistically principled, mixed-model approach that accounts for voxel-level dependencies and noise, the authors offer a method that yields more accurate and reliable estimates of brain network architecture. The demonstrated improvement in test-retest reproducibility suggests that this approach can lead to a more robust scientific understanding of individual variability in connectivity networks and their associations with health outcomes. The authors note that while the method is designed for fMRI, the underlying model could be applicable to other contexts involving spatially grouped time series.

The New Approach: Listening to Every Voice

Why Does This Matter? (The Proof)

The Bottom Line

More like this