📊 statistics

Bayesian Modeling and Prediction of Generalized Contact Matrices

This paper introduces a Bayesian modeling framework that leverages tensor structures and contingency table theory to infer generalized, multi-dimensional contact matrices from incomplete data, offering a statistically stable solution for infectious disease epidemiology that is implemented in an open-source Python package.

Original authors: Shozen Dan, David A. van Dyk, Zhi Ling, Swapnil Mishra, Oliver Ratmann

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shozen Dan, David A. van Dyk, Zhi Ling, Swapnil Mishra, Oliver Ratmann

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 trying to understand how a contagious virus spreads through a crowd. To do this, you need a map of who talks to whom. In the world of epidemiology, this map is called a contact matrix.

Traditionally, these maps have been very simple: they only show how often people of different ages talk to each other. It's like a grid where one side lists "Children" and the other lists "Adults," and the numbers tell you how many times they interact.

However, the authors of this paper argue that age isn't the only thing that matters. A virus might spread differently between a wealthy CEO and a factory worker, or between people of different races, regardless of their age. The paper introduces a new, more powerful tool called g-mix (generalized mixing) to create a much more detailed map that includes these other factors.

Here is a breakdown of their work using simple analogies:

1. The Problem: The "Blind Spot" in the Data

Imagine you are at a party. You know exactly who you are (your age, your job, your income), but you don't know the details of the people you talk to. You might say, "I talked to a 30-year-old," but you don't know if that person is a doctor, a student, or a retiree.

In real-world surveys, people report their own details but rarely know the details of everyone they met. This creates a "missing data" problem. The old maps (age-only) work fine, but when you try to add more details (like income or race), the data becomes too sparse and full of holes to draw a clear picture.

2. The Solution: The "Smart Puzzle" (g-mix)

The authors built a Bayesian modeling framework (a type of smart statistical puzzle solver) called g-mix. Think of it as a detective who can fill in the missing pieces of a puzzle using logic and patterns, even when some pieces are missing.

The "Reciprocity" Rule: The core of their logic is a simple truth: If Person A talks to Person B, then Person B must have talked to Person A. The number of interactions must be equal from both sides. The authors use this "mirror rule" to double-check their work and fill in the gaps. If the data says "Men talked to Women 100 times," the model ensures the math says "Women talked to Men 100 times" as well.
The "Smooth Surface" Analogy: Imagine the contact data as a bumpy, 3D landscape. The authors use a technique called splines (like a flexible ruler) to smooth out the bumps. This helps them estimate contact rates for specific groups even if they didn't have enough survey data for that exact group, by borrowing information from neighboring groups.

3. Handling the "Missing Pieces"

When the survey data is incomplete (e.g., we know the respondent's race but not the contact's), the model uses a clever trick borrowed from contingency tables (a type of statistical grid).

Think of it like a budget. If you know your total income and your total spending, you can figure out the limits of what you spent on specific categories, even if you lost the receipts. The model calculates the tightest possible bounds for the missing information. It doesn't just guess; it says, "We know the answer is definitely between X and Y," and then uses probability to find the most likely spot within that range.

4. Testing the Tool

The authors tested their new "g-mix" tool against older methods using computer simulations:

The Test: They created fake populations with known contact patterns and then tried to "re-discover" those patterns using only survey data.
The Result: The g-mix model was more accurate and faster than the old methods, especially when the data was sparse or when they tried to include multiple factors (like age + gender + income) at once. It handled the "missing pieces" much better than previous tools.

5. Real-World Examples

The team applied their model to two real datasets:

Germany (COVIMOD): They tracked how contact patterns changed during the COVID-19 pandemic. They found that during strict lockdowns, social inequalities (differences in contact patterns based on job type) disappeared—everyone stayed home equally. But when restrictions were lifted, the differences returned: manual workers had to keep working and interacting, while office workers could stay home.
United States (BICS): They analyzed data from Berkeley to see which factors mattered most. They found that adding race/ethnicity and sex to the age-based model significantly improved the accuracy of the map, proving that these factors are crucial for understanding how diseases spread.

Summary

In short, this paper presents a new, smarter way to draw the "social maps" that epidemiologists use to predict disease spread. Instead of just looking at age, the g-mix model can layer on other details like income, race, and job type. It uses the "mirror rule" of reciprocity and statistical smoothing to fill in the blanks where data is missing, providing a clearer, more accurate picture of how different groups of people interact. This helps scientists understand why diseases might hit some communities harder than others.

Technical Summary: Bayesian Modeling and Prediction of Generalized Contact Matrices

Problem Statement
Social contact matrices are fundamental to infectious disease epidemiology, quantifying close-range human interactions that drive transmission. Traditionally, these matrices are stratified solely by age. However, pathogens such as coronaviruses, influenza, and tuberculosis are significantly influenced by demographic and socioeconomic factors beyond age, including gender, race, ethnicity, and socioeconomic status (SES). Existing methods for inferring contact matrices often lack a principled mechanism to handle these additional features, particularly when data is sparse or when feature information is missing for reported contacts (the "partial-data" setting). Furthermore, enforcing structural constraints, such as reciprocity (the principle that the total number of contacts from group A to B must equal those from B to A), becomes computationally and statistically challenging in high-dimensional, multi-feature spaces.

Methodology
The authors propose the generalized Bayesian social mixing (g-mix) model, a framework designed to infer contact matrices stratified by multiple features while satisfying fundamental structural assumptions.

Model Formulation: The model treats contact counts as realizations of a Negative Binomial distribution. It decomposes the expected contact intensity into a baseline age-specific contact rate ( $\gamma_{a,b}$ ) and a set of rate modifiers ( $\delta^{k,\ell}_{a,b}$ ) that account for heterogeneity across strata (e.g., gender, SES).
Structural Constraints: The framework enforces three critical constraints:
1. Age Reciprocity: $\gamma_{a,b} = \gamma_{b,a}$ .
2. Modifier Reciprocity: $\delta^{k,\ell}_{a,b} = \delta^{\ell,k}_{b,a}$ .
3. Marginal Consistency: The modifiers must sum to one when weighted by population proportions, ensuring the model aligns with the underlying population structure.
Parameterization and Constraints: To satisfy these constraints computationally, the authors map unconstrained latent parameters ( $\omega$ ) to the simplex space using a softmax transformation. This approach allows the model to handle the reciprocity constraints via permutation matrices and ensures marginal consistency automatically. For multiple features, the model utilizes Kronecker sums of component tensors, assuming an additive log-linear structure on the latent parameters to manage the exponential growth of the parameter space.
Smoothing and Priors: The model employs Bayesian P-splines (penalized splines) with second-order intrinsic Gaussian Markov random field (IGMRF) priors. This allows the model to borrow strength across adjacent age groups, treating contact patterns as smooth surfaces.
Handling Missing Data (Partial Stratification): In real-world surveys, respondents often know their own attributes but not those of their contacts. The authors link this missing-data problem to the theory of contingency tables. By deriving the marginal contact intensities from the partial data and applying reciprocity, they establish sharp bounds (Fréchet–Hoeffding bounds) for the unobserved "attributable fractions" (the proportion of contacts between specific strata). These bounds are incorporated into a truncated Dirichlet prior to predict the full contact matrix.
Inference: The model is fitted using Stochastic Variational Inference (SVI) implemented in the Python library NumPyro, optimizing the Evidence Lower Bound (ELBO).

Key Contributions

Generalized Framework: The g-mix model extends contact matrix inference beyond age to include arbitrary categorical features (e.g., SES, race) while maintaining mathematical consistency.
Structural Enforcement: Unlike previous approaches that enforce reciprocity as a post-hoc correction, g-mix intrinsically enforces reciprocity and marginal consistency through its parameterization, reducing the effective parameter space and improving statistical efficiency.
Missing Data Solution: The paper provides a rigorous Bayesian approach to predicting full contact matrices from partial data (where contact attributes are unobserved), utilizing feasible mixing bounds derived from contingency table theory to constrain predictions.
Feature Selection: The framework enables principled model selection (via Leave-One-Out Cross-Validation and Expected Log Predictive Density) to identify which features significantly explain heterogeneity in contact patterns.
Open-Source Implementation: The authors released the cntmosaic Python package to facilitate adoption.

Results

Simulation Studies: The g-mix model was benchmarked against extended versions of socialmixr (bootstrapping), Prem et al.'s IGMRF approach, and van de Kassteele et al.'s method. In scenarios with increasing numbers of features (from age-only to multi-variable stratification), g-mix consistently achieved the lowest Mean Absolute Percentage Error (MAPE) and the best interval scores. While socialmixr was fastest, it suffered from high error rates and failure probabilities in sparse data regimes due to zero-division issues in bootstrapping.
Partial Data Prediction: When contact features were withheld (simulating real-world survey limitations), the model successfully predicted the full contact matrix. While Mean Absolute Percentage Error increased compared to the complete-data scenario, the posterior 95% coverage rates remained well-calibrated (e.g., ~94-98% in neutral/assortative scenarios), indicating robust uncertainty quantification.
Case Study 1 (Germany - COVIMOD): Applied to non-household contact data during the COVID-19 pandemic, the model revealed that strict non-pharmaceutical interventions (lockdowns) temporarily suppressed socioeconomic inequalities in contact patterns, acting as an "epidemiological equalizer." When restrictions were relaxed, inequalities re-emerged, with manual workers maintaining higher contact rates than managerial or retired groups.
Case Study 2 (USA - BICS): Feature selection analysis identified that sex and race/ethnicity, in addition to age, were statistically significant predictors of contact heterogeneity. The model detected that contact patterns varied significantly by race/ethnicity as non-pharmaceutical interventions were eased, with specific groups (e.g., "Non-Hispanic Asian and Other") showing consistently fewer contacts compared to others.

Significance and Claims
The paper claims that the g-mix framework moves the field closer to understanding how contact patterns vary across dimensions of social stratification beyond age. By providing a principled method to infer generalized contact matrices from existing survey data (even with missing contact attributes), the framework supports more accurate estimation of reproduction numbers ( $R_0$ ), better monitoring of intervention impacts, and improved design of vaccine allocation strategies. The authors emphasize that their approach allows epidemiologists to capture additional heterogeneity without requiring new, expensive data collection efforts, provided that population size data is available. They acknowledge limitations, such as the exclusion of survey sampling weights and the challenge of accurately predicting complete matrices from partial data in highly assortative scenarios, but posit that the framework offers a significant step forward in modeling infectious disease drivers.