Bayesian Modeling for Aggregated Relational Data: A Unified Perspective

1. Problem Statement

Aggregated Relational Data (ARD) is a widely used method for studying social networks, particularly when full network data is unavailable or too costly to collect. Instead of asking individuals about specific connections, ARD asks respondents to report the number of people they know within specific subpopulations (e.g., "How many people named 'Christina' do you know?").

Despite its utility in estimating hidden population sizes (e.g., drug users, homeless populations) and personal network sizes, the field faces several critical challenges:

• Fragmented Implementation: Existing Bayesian ARD models are implemented in disparate ways, often requiring custom, non-standardized code (e.g., custom Gibbs-Metropolis algorithms), making reproducibility and comparison difficult.
• Identifiability and Scaling: ARD models suffer from inherent identifiability issues where personal degree ($d_i$) and subpopulation size ($N_k$) are confounded. They can only be recovered up to a multiplicative scaling factor. While post-hoc rescaling is common, it is often treated as an ad-hoc step rather than an integrated part of the inference.
• Model Selection and Comparison: There is a lack of systematic, unified tools for comparing competing ARD models. Standard Bayesian model selection tools (like Leave-One-Out Cross-Validation via PSIS) often fail for ARD due to the hierarchical nature of the data and individual-level parameters.
• Complexity vs. Computation: There is a trade-off between model complexity (capturing non-random mixing, homophily, and latent structures) and computational feasibility, which has not been systematically evaluated across different model classes.

2. Methodology

The authors propose a unified framework for Bayesian ARD modeling using Stan, a state-of-the-art probabilistic programming language utilizing Hamiltonian Monte Carlo (HMC).

A. Unified Model Collection

The paper provides a coherent collection of Stan implementations for a spectrum of ARD models, ranging from simple to complex:

1. Null Models:
   • Erdős-Rényi: Assumes random ties with a common degree for all individuals.
   • Handcock & Jones (Null): Allows for individual variation in degree (gregariousness) but assumes random mixing between individuals and subpopulations.
2. Models with Non-Random Mixing:
   • Overdispersed Model (Zheng et al., 2006): Introduces an overdispersion parameter to account for extra variability in tie formation (barrier effects), modeled via a Negative Binomial distribution.
   • Latent Space Model (McCormick & Zheng, 2015): Positions individuals and subpopulations in a latent social space (hypersphere) to capture structural dependencies based on proximity.
3. Covariate-Based Models:
   • Discrete Mixing: Uses demographic categories (e.g., gender, age) to model mixing probabilities.
   • Kernel Smoothing: Extends mixing to continuous covariates (e.g., age) using smooth kernels.
4. Network Scale-Up Models (NSUM): Includes extensions of the classic NSUM method, such as the Barrier Effects model, which accounts for recall and transmission biases.

B. Technical Innovations

• Within-Iteration Rescaling: A key methodological contribution is the integration of rescaling procedures directly into the MCMC sampling loop (via a transformed parameters block in Stan). Instead of fitting unconstrained models and rescaling post-hoc, the model parameters are constrained during sampling to match known subpopulation sizes. This significantly improves chain convergence and eliminates a post-processing step.
• Exact K-Fold Cross-Validation: The authors demonstrate that standard approximate Leave-One-Out (LOO) methods (using Pareto Smoothed Importance Sampling, PSIS) often fail for ARD models because removing a single observation drastically changes the posterior of individual-level parameters (e.g., personal degree $d_i$). They propose and implement exact K-fold cross-validation by holding out random entries of the ARD matrix rather than entire rows, allowing for valid model comparison in many cases.
• Synthetic Data Generation: To evaluate these models, the authors generate synthetic datasets based on:
  1. A true Latent Space model (McCormick & Zheng, 2015).
  2. A replication of the classic McCarty et al. (2001) dataset using the Barrier Effects model (Maltiel et al., 2015).
     This allows for "ground truth" evaluation of parameter recovery (degrees and subpopulation sizes).

3. Key Contributions

1. Unified Stan Implementation: The paper provides a ready-to-use suite of Stan code for major Bayesian ARD models, lowering the barrier to entry for researchers and ensuring reproducibility.
2. Systematic Model Comparison: Through synthetic experiments, the authors compare how different models recover true personal network sizes and hidden subpopulation sizes.
3. Diagnostic Framework: The paper evaluates the efficacy of Posterior Predictive Checks (PPCs) and Cross-Validation for ARD. It highlights that while PPCs can detect gross model misspecification (e.g., failure to capture overdispersion), they often struggle to distinguish between two plausible, complex models.
4. Identification of LOO Failure: The work explicitly documents why standard PSIS-LOO fails for ARD (due to hierarchical individual parameters) and provides a practical alternative (entry-wise K-fold CV) for models without entry-level parameters.

4. Results

• Parameter Recovery:
  • Degree Estimation: The Varying Degree model and the Overdispersed model generally provided the most accurate estimates of personal network sizes across different data generating processes. The Latent Space model, while theoretically sound, was difficult to fit and often overestimated unknown subpopulation sizes.
  • Subpopulation Size: The "True" model (the one used to generate the data) generally performed best, but simpler models (like Overdispersed) often performed nearly as well, suggesting that extreme complexity may not always yield better practical estimates.
• Convergence: The within-iteration rescaling procedure was shown to be critical. Models fitted without this constraint exhibited poor convergence diagnostics ($\hat{R} > 1.1$), whereas the rescaled models converged rapidly ($\hat{R} \approx 1.0$).
• Model Checking:
  • PPCs: Posterior predictive checks successfully identified when a model failed to capture overdispersion (e.g., the Erdős-Rényi model failed to match the variance in the data). However, for complex models, PPCs often showed similar agreement with the data, making it hard to select a "winner."
  • Cross-Validation: Exact K-fold CV successfully identified the true data-generating model in the synthetic Latent Space experiment, outperforming PPCs in distinguishing the correct model structure. However, CV is computationally expensive and currently inapplicable to models with entry-level parameters (like some NSUM variants).

5. Significance and Future Directions

This paper serves as a practical guide and reference for statisticians and practitioners working with ARD.

• Practical Guidance: It clarifies the trade-offs between model complexity and computational cost, advising researchers to start with simpler models (like Overdispersed) and only move to complex latent space models if necessary and if data supports it.
• Reproducibility: By providing open-source Stan code, the authors democratize access to advanced Bayesian ARD modeling, moving the field away from custom, hard-to-reproduce algorithms.
• Future Research: The authors identify a critical gap: the lack of a general, robust model selection procedure that works for all ARD models (especially those with complex hierarchical structures). They call for new statistical techniques to handle model selection in the presence of individual-level parameters and to better incorporate covariate information.

In summary, the paper unifies the fragmented landscape of Bayesian ARD modeling, provides robust computational tools for fitting and checking these models, and offers evidence-based guidance on selecting the appropriate model complexity for specific research goals (e.g., estimating hidden population sizes vs. understanding social structure).