📊 statistics

Bayesian Multi-Group Functional Factor Models with Parameter-Expanded Cumulative Shrinkage Priors

This paper proposes a Bayesian multi-group functional factor analysis framework that utilizes B-spline bases and a parameter-expanded cumulative shrinkage prior to jointly model group-specific and shared latent structures in functional data, automatically determining the number of active factors while effectively distinguishing between common and distinct variations, as demonstrated through simulations and an application to EEG data.

Original authors: Xuanye Dai, Anna Gottard, Michele Guindani, Marina Vannucci

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xuanye Dai, Anna Gottard, Michele Guindani, Marina Vannucci

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 you are a music producer trying to understand a massive library of recordings. You have two distinct groups of musicians: Group A (let's say, jazz enthusiasts) and Group B (rock fans). You have hundreds of recordings from each group, and every recording is a long, continuous melody (a "trajectory") that changes over time.

Your goal is to figure out two things:

What musical elements do both groups share? (Maybe they both use a standard drum beat or a specific chord progression).
What makes each group unique? (Maybe the jazz group improvises wildly, while the rock group has a specific guitar riff).

This is exactly the problem the paper solves, but instead of music, they are looking at EEG brain waves (electrical signals from the brain) from alcoholic and healthy subjects.

Here is a simple breakdown of their solution, the Bayesian Multi-Group Functional Factor Model, using some creative analogies.

1. The Problem: Too Much Noise, Too Many Details

Brain waves are like a chaotic ocean. They are continuous, wiggly lines with thousands of data points. If you try to analyze every single wiggle, it's impossible to see the big picture. It's like trying to understand a symphony by looking at every single vibration of every single string.

Scientists need to compress this data. They need to find the "main themes" (factors) that explain the music without recording every single note.

2. The Solution: The "Shared vs. Unique" Orchestra

The authors built a statistical model that acts like a smart music editor. It breaks down the brain waves into three parts:

The Group Average (The Conductor): First, it calculates the "average" brain wave for the alcoholics and the "average" for the healthy people. This is the baseline.
The Shared Themes (The Common Rhythm): It looks for patterns that appear in both groups. In the EEG study, they found a "shared factor" that looked like a standard brain response (a specific bump in the wave around 100ms and another around 200ms). This is like a drum beat that both jazz and rock bands use. It represents the brain's basic, universal reaction to seeing a picture.
The Unique Solos (The Group-Specific Riffs): Then, it looks for patterns that only appear in one group.
- For the alcoholic group, the model found extra "solos" (factors) that showed delayed or messy reactions. It's like the jazz band playing a riff that is slightly out of sync or more chaotic than usual.
- For the healthy group, the model found different "solos" that showed sharp, clear, and synchronized reactions.

3. The Secret Sauce: The "Automatic Shrinkage" Filter

The hardest part of this job is guessing how many themes (factors) exist. Do the jazz band have 2 solos? 5? 10? If you guess wrong, your model is either too simple (missing the music) or too complex (hearing ghosts).

The authors used a clever trick called a "Parameter-Expanded Cumulative Shrinkage Prior."

The Analogy: Imagine you have a row of 100 volume knobs on a mixing board. You don't know which ones are actually playing music and which ones are just static.

Most models would ask you to guess how many knobs to turn up.
This new model has a "Smart Shrinkage Filter." It starts by turning all knobs up slightly. Then, it automatically turns the volume down on the knobs that aren't doing anything important.
Crucially, it turns the volume down more and more as you go down the line. It's like a filter that says, "The first few knobs are probably important, but by the time we get to knob #50, if it's not loud, it's definitely silence."
This allows the computer to automatically decide how many factors are real and how many are just noise, without the human needing to guess.

4. The Result: What Did They Find?

When they applied this to real brain data:

The Shared Part: They found the brain's standard "Hello" signal (a quick reaction to seeing an image). Both groups had this.
The Unique Part: They found that the alcoholic group had "blurry" or "delayed" extra signals. Their brains took longer to process the image, and the reaction was less focused. The healthy group had sharp, quick, and focused extra signals.

Why This Matters

Think of this model as a smart translator.

Old methods might have tried to analyze the two groups separately, missing the fact that they share a common language (the shared brain response).
Or, they might have tried to average them together, washing out the unique differences (the specific struggles of the alcoholic group).

This new model acts like a bilingual translator that understands the common language shared by everyone, while also highlighting the unique dialects spoken by each specific group. It helps doctors and scientists see exactly where and how the brain activity differs, which is a huge step forward for understanding conditions like alcoholism.

In a nutshell: They built a mathematical tool that automatically finds the "common beats" and "unique solos" in messy brain wave data, helping us see the hidden differences between healthy and alcoholic brains with much greater clarity.

1. Problem Statement

Functional data analysis (FDA) involves trajectories observed over continuous domains (e.g., time, space). While standard methods like Functional Principal Component Analysis (fPCA) and Bayesian latent factor models exist for single groups, they struggle with multi-group functional data.

The Challenge: In many scientific applications (e.g., neuroscience comparing healthy vs. diseased subjects), data arises from multiple groups. Researchers need to disentangle shared latent structures (common patterns across all groups) from group-specific variations (unique patterns within specific groups).
Limitations of Existing Methods:
- Standard fPCA treats groups independently or aggregates them, failing to explicitly separate shared vs. specific components.
- Existing multi-group latent factor models for discrete data cannot be directly applied to dense, ordered functional grids.
- Current Bayesian functional factor models often rely on priors (like the Multiplicative Gamma Process) that are sensitive to hyperparameter tuning and do not automatically select the optimal number of factors.

2. Methodology: Bayesian Multi-Group Functional Factor Model (BMGFFM)

The authors propose a hierarchical Bayesian framework that decomposes functional trajectories into group-specific means, shared latent components, and group-specific latent components.

A. Model Formulation

For subject $i$ in group $s$ , the observed curve $y_{is}(t)$ is modeled as:
$y_{is}(t) = \mu_s(t) + \sum_{l=1}^{L} \eta_{ils} \lambda_l(t) + \sum_{k=1}^{K_s} \rho_{iks} \phi_{sk}(t) + \epsilon_{is}(t)$

$\mu_s(t)$ : Group-specific mean function.
$\lambda_l(t)$ : Shared latent components (common to all $S$ groups).
$\phi_{sk}(t)$ : Group-specific latent components (unique to group $s$ ).
$\eta_{ils}, \rho_{iks}$ : Latent factor scores (Gaussian priors).
Basis Expansion: All smooth functions are represented as linear combinations of a common set of cubic B-spline basis functions ( $R$ bases), achieving a low-rank representation.
Penalization: Penalized B-splines are used to enforce smoothness on the mean coefficients via a roughness penalty matrix.

B. Prior Construction: Parameter-Expanded Cumulative Shrinkage Process (CUSP)

The core innovation is the use of a Parameter-Expanded Cumulative Shrinkage Process (CUSP) prior on the factor loading matrices ( $\Lambda$ for shared, $\Phi_s$ for group-specific).

Goal: To automatically determine the number of active factors ( $L^*$ and $K_s^*$ ) by shrinking redundant factors to zero.
Mechanism:
- The loading matrix is parameterized as $\Lambda = \Xi \times \text{diag}(\gamma_1, \dots, \gamma_L)$ .
- $\Xi$ : An auxiliary matrix with entries drawn from a mixture of point masses at $\pm 1$ (improving MCMC mixing compared to standard normals).
- $\gamma_l$ : Global scale parameters controlled by a cumulative shrinkage process.
- Shrinkage Logic: A stick-breaking construction generates probabilities $\pi_l$ that increase with the index $l$ . As $l$ increases, the probability that the $l$ -th factor is "inactive" (shrunk to near zero) increases.
- Parameter Expansion: Auxiliary parameters are introduced to handle posterior dependencies, improving the convergence and mixing of the MCMC sampler.

C. Posterior Inference & Identifiability

Algorithm: A Gibbs sampler is used to draw from the posterior distribution.
Identifiability Issue: Factor models suffer from rotational invariance, sign switching, and label switching. In multi-group settings, decomposing covariance into shared and specific parts is not uniquely identifiable at the loading matrix level.
Solution: The authors propose a post-processing strategy operating in covariance space rather than parameter space:
1. Selection: Identify the modal factor configuration (most frequent $L^*, K_s^*$ ) across MCMC iterations.
2. Alignment: Apply a Rotation-Sign-Permutation (RSP) algorithm (using Varimax rotation) to align posterior draws.
3. Covariance Decomposition: Instead of interpreting raw loadings, they compute the posterior mean covariance operators for shared and group-specific components. They then perform spectral decomposition on these covariance matrices to derive unique, interpretable loading subspaces.

3. Key Contributions

Novel Framework: First Bayesian multi-group functional factor model that explicitly separates shared and group-specific latent structures using a low-rank B-spline decomposition.
Automatic Dimension Selection: Introduction of the Parameter-Expanded CUSP prior for functional data, which induces increasing shrinkage on higher-order factors, allowing the model to automatically select the number of active shared and group-specific factors without manual tuning.
Robust Inference: A rigorous post-processing pipeline that resolves non-identifiability issues (rotation, sign, permutation) by working in the covariance space, ensuring stable and interpretable estimates.
Superior Performance: Demonstrated through simulations that the model outperforms competitors (BFPCA, SFFM, filt-fPCA) in scenarios with mixed shared and specific structures.

4. Results

Simulation Studies

Scenarios: Tested under three conditions: (A) Mixed shared/specific factors, (B) Only shared factors, (C) Only group-specific factors.
Factor Recovery: The model accurately recovered the true number of factors ( $L$ and $K_s$ ) in most scenarios, particularly when both shared and specific structures existed.
Estimation Accuracy:
- MSE: BMGFFM achieved the lowest Mean Squared Error (MSE) in Scenario A (mixed structures). In pure shared or pure specific scenarios, it performed comparably to or slightly worse than specialized single-group methods (BFPCA), but remained robust.
- RV Coefficient: High similarity (RV > 0.9) between true and estimated covariance-derived loadings in mixed scenarios.
Sensitivity: The model was robust to hyperparameter choices, though the shrinkage threshold ( $v_0$ ) significantly impacted the selection of active factors.

Real Data Application: EEG Analysis

Data: Event-Related Potentials (ERPs) from 77 alcoholic and 45 healthy control subjects (COGA dataset).
Findings:
- Factor Selection: The model identified 1 shared factor, 4 alcoholic-specific factors, and 5 control-specific factors.
- Shared Structure: The single shared factor captured canonical ERP components (N100 and P200), representing fundamental sensory/perceptual processes common to all humans.
- Group-Specific Structure:
  - Alcoholic Group: Showed distinct, delayed, and blurred neural responses (e.g., pronounced positive fluctuations at 240–347ms), suggesting altered cognitive processing.
  - Control Group: Exhibited sharper, more synchronized peaks, indicating efficient stimulus evaluation.
- Visualization: The model successfully reconstructed individual curves with tight credible intervals, clearly distinguishing the neural dynamics between the two groups.

5. Significance

This paper provides a powerful tool for multi-group functional data analysis in fields like neuroscience, genomics, and econometrics.

Scientific Insight: It moves beyond simple group comparisons to reveal what is common and what is unique in complex trajectories.
Methodological Advancement: By combining B-splines with a parameter-expanded shrinkage prior, it solves the dual problems of dimensionality reduction and automatic model selection in a unified Bayesian framework.
Clinical Relevance: The application to EEG data demonstrates the model's ability to detect subtle, group-specific neural biomarkers (e.g., delayed ERP responses in alcoholism) that might be obscured by standard averaging or single-group analyses.

The proposed BMGFFM effectively balances flexibility and parsimony, offering a robust solution for understanding complex, heterogeneous functional data structures.