Mixed Membership Models for Multilevel Functional Data

1. Problem Statement

The paper addresses the challenge of analyzing multilevel functional data, specifically in the context of electroencephalography (EEG) studies involving children with Autism Spectrum Disorder (ASD) and typically developing (TD) controls.

• Data Structure: The data consists of $N$ subjects, where each subject contributes $J$ functions (e.g., EEG spectral densities across 25 channels) observed over a time/frequency domain $T$. This creates a product space of measurable functions $L(T)$.
• Heterogeneity: There is significant heterogeneity both within subjects (differences between channels) and between subjects (different spectral features).
• Limitation of Existing Methods: Traditional clustering (hard clustering) forces observations into a single class, which fails to capture the reality that subjects often exhibit a continuous mixture of underlying physiological patterns (e.g., a mix of periodic alpha waves and aperiodic noise). Furthermore, standard functional clustering often yields poorly separated, hard-to-interpret cluster means.
• Goal: To develop a Mixed Membership Model that allows each observation to partially belong to multiple "pure" latent classes simultaneously, while accounting for the multilevel structure (subject $\to$ channel) and ensuring model identifiability.

2. Methodology

The authors propose a hierarchical Bayesian framework that extends mixed membership models to multilevel functional data using a Multivariate Karhunen-Loève (K-L) decomposition.

A. Model Formulation

The observed functional data $Y_{ij}(t)$ (subject $i$, channel $j$) is modeled as a realization of a stochastic process. The core assumption is that the data is a convex combination of $K$ latent functional features (pure membership classes).

1. Sampling Model: The finite-dimensional margins are modeled as a mixture of dependent Gaussian Processes (GPs):
   $$Y_{ij}(t_i) \mid \Theta, X \sim \mathcal{N}\left( \sum_{k=1}^K Z_{ijk} \left[ S(t_i)^T \nu_k + \sum_{m=1}^M \chi_{ijm} S(t_i)^T \phi_{km} \right], \sigma_j^2 I_{n_i} \right)$$

   • $Z_{ijk}$: Mixing proportions on the unit simplex (probability subject $i$, channel $j$ belongs to feature $k$).
   • $\nu_k$: Feature-specific mean functions (basis coefficients).
   • $\phi_{km}$: Pseudo-eigenfunctions capturing covariance structure.
   • $\chi_{ijm}$: Pseudo-scores (integrated out to derive the marginal distribution).

2. Marginal Distribution: After integrating out pseudo-scores, the model becomes an additive model where the mean is a subject-by-channel convex combination of feature means, and the covariance is a weighted sum of feature covariance functions.

3. Scalability: To ensure scalability and flexibility, the covariance surface is approximated using $M$ scaled pseudo-eigenfunctions. Crucially, orthonormality is not imposed on the parameters $S(t_i)^T \phi_{km}$. This allows the model to operate outside the Stiefel manifold, facilitating better Markov Chain Monte Carlo (MCMC) mixing.

B. Priors and Identifiability

A major challenge in mixed membership models is identifiability (distinguishing between different mixtures).

• Covariance Regularization: The authors use a Multiplicative Gamma Process (MGP) shrinkage prior on the pseudo-eigenfunctions ($\phi$). This induces adaptive regularization, where the variance of the coefficients stochastically decreases as the index $m$ increases, effectively truncating the series.
• Repulsive Prior for Identifiability: To ensure the mixing proportions $Z$ are identifiable, the authors propose a hierarchically repulsive point-process prior on the unit simplex.
  • Subject-level centroids $\pi_i$ are drawn from a repulsive process that discourages centroids from clustering too closely (using an exponential repulsion term based on Euclidean distance).
  • Channel-level mixing probabilities $Z_{ij}$ are drawn from a Dirichlet distribution centered at the subject-level repulsive centroid $\pi_i$.
  • This geometric constraint ensures the "sufficiently scattered condition," covering the simplex effectively to separate latent classes.

C. Inference

Posterior inference is performed using standard MCMC strategies. The model is designed to be computationally tractable, avoiding complex manifold constraints.

3. Key Contributions

1. Extension to Multilevel Functional Data: The paper generalizes mixed membership models from standard data to multilevel functional data (subject $\to$ channel $\to$ function), capturing both within-subject and between-subject variability.
2. Hierarchical Repulsive Prior: A novel contribution is the use of a repulsive point-process prior on the simplex to solve the identifiability problem inherent in flexible mixed membership models, ensuring distinct latent classes.
3. Flexible Covariance Modeling: By relaxing the orthonormality constraint on the K-L expansion, the authors create a model that is easier to sample and more flexible than those constrained to Stiefel manifolds.
4. Application to ASD EEG: The framework is successfully applied to real-world EEG data to quantify the spectrum of neurodevelopmental differences between ASD and TD children.

4. Results (Application to ASD EEG)

The model was applied to resting-state EEG data from 58 ASD and 39 TD children (ages 2–12) in the alpha band (6–14 Hz).

• Model Selection: Using the "elbow method" (as AIC/BIC tended to overestimate $K$), the optimal number of latent features was determined to be $K=2$.
• Latent Features Identified:
  • Feature 1: An aperiodic, $1/f$ (pink) noise-like mean with no alpha peak.
  • Feature 2: A periodic component with a clear alpha peak, typical of older neurotypical subjects.
• Findings:
  • TD Children: Showed high probability of loading heavily on Feature 2 (strong alpha peak).
  • ASD Children: Exhibited high heterogeneity, showing continuous mixtures of both the aperiodic noise and the periodic alpha peak.
  • Spatial Analysis: The channel-level mixing probabilities revealed that the differences between TD and ASD groups were most pronounced in the temporal regions of the brain.
• Interpretation: The mixed membership approach successfully quantified the "spectrum" of ASD, showing that ASD is not a binary absence of alpha peaks but a heterogeneous mixture of periodic and aperiodic spectral patterns.

5. Significance and Discussion

• Scientific Impact: The study provides a novel quantitative perspective on neurodevelopmental disorders, moving beyond binary classification to a continuous spectrum of functional brain states. It confirms existing literature on the attenuation of alpha peaks in ASD while offering a more granular, probabilistic view.
• Methodological Impact: The proposed framework offers a scalable, flexible tool for analyzing complex multilevel functional data in neuroscience and other fields.
• Limitations & Future Work:
  • The authors note that information criteria (AIC/BIC) struggled to select the correct $K$, whereas the elbow method worked better.
  • Nonparametric alternatives (e.g., Indian Buffet Process) exist but lack well-understood inference properties for changing dimensions in this specific context.
  • Further theoretical work is needed to establish posterior convergence rates for overfitted mixed membership models in this multilevel setting.

In summary, this paper presents a robust statistical framework that bridges the gap between functional data analysis and mixed membership clustering, providing a powerful tool for uncovering latent heterogeneity in complex biological data.