A method for including socio-demographic factors in social contact matrices for compartment-based epidemic models

This paper presents a method to extend existing age-based social contact matrices by incorporating additional socio-demographic factors using population structure data and mixing assumptions, demonstrating that such stratification significantly alters epidemic dynamics and outcomes, particularly for minority groups.

The Gist

Imagine you are trying to predict how a rumor (or a virus) spreads through a massive, crowded party.

For a long time, scientists have tried to model this by dividing the partygoers into age groups. They assume that teenagers hang out mostly with other teenagers, and grandparents mostly with other grandparents. This helps, but it's like looking at the party through a blurry, black-and-white photo. You see the age groups, but you miss the other details that actually drive how the rumor spreads.

This paper introduces a new way to take that blurry, black-and-white photo and turn it into a high-definition, color image by adding socio-demographic factors (like ethnicity, income, or education) without needing to interview every single person at the party again.

Here is the breakdown of their method and findings using simple analogies:

1. The Problem: The "One-Size-Fits-All" Map

Think of the old way of modeling disease as using a single, flat map of a city. This map shows the roads (social contacts) between different neighborhoods (age groups). It's useful, but it assumes everyone in a neighborhood behaves exactly the same.

In reality, even within a single neighborhood, some people are "super-connectors" (they know everyone), while others are "hermits" (they stay home). If you ignore these differences, your map might tell you the virus will spread slowly, when in reality, it could explode through a specific group of people who hang out together more than the average.

2. The Solution: The "Layer Cake" Method

The authors created a mathematical "recipe" to slice that flat map into a layer cake.

• The Base Layer: You start with the existing map of age-based contacts (who talks to whom based on age).
• The New Slices: You want to add a new ingredient, like "ethnicity" or "income."
• The Challenge: You don't have a new map showing exactly how a 20-year-old Maori person talks to a 60-year-old European person. You only have the total population numbers and a few guesses about how much different groups prefer to mix with their own kind (assortativity) versus others.

The Recipe:
Instead of needing a survey of every single interaction, the authors say: "Let's assume the total number of handshakes stays the same, but we redistribute them based on two rules:"

1. Population Size: If a group is bigger, they naturally have more total handshakes.
2. Mixing Preference: If a group likes to stick to their own kind (high assortativity), they will shake fewer hands with outsiders and more with insiders.

Using just these two rules and the population numbers, their math "fills in the gaps" to create a detailed, multi-layered map of who is talking to whom.

3. The Experiment: Testing the New Map

To see if this new map works, the authors ran two types of simulations:

A. The "Hypothetical Party" (Made-up Data)
They created imaginary parties with different rules (e.g., one group is 90% of the crowd, another is 10%; one group talks twice as much as the other).

• The Result: When they ignored the extra groups, the party looked safe. But when they used their new "layer cake" map, they saw that the virus could actually spread much faster in specific minority groups.
• The Analogy: It's like realizing that while the average person at the party only drinks one soda, a specific group of friends is passing a giant pitcher around. If you only count the average, you miss the dehydration risk for that specific group.

B. The "Real-World Projection" (New Zealand Data)
They took real data from a European survey (POLYMOD) and applied it to New Zealand's population structure (Maori, Pacific, Asian, and European/Other).

• The Result: Even though they assumed some groups had lower contact rates, the age structure of those groups changed everything. For example, because the Asian population in their data was younger on average, the virus spread differently than if they had just looked at the "average" New Zealander.
• The Surprise: The new model predicted that the total number of infections might actually be lower than the old model, but the distribution would change. More older people would get sick, and fewer young people, shifting the burden of the disease.

4. Why This Matters (The "So What?")

The paper argues that if you use the old, blurry map, you might make bad decisions.

• Hidden Inequities: You might think the virus is under control because the "average" infection rate is low, while a specific minority group is actually being hit hard.
• Policy Impact: If a government plans a hospital response based on the old map, they might not have enough beds for the specific age group that ends up getting sick in the new model.

Summary

The authors didn't invent a new virus or a new cure. They invented a mathematical lens.

Think of it like upgrading from a black-and-white TV (age only) to a color TV with 3D sound (age + ethnicity/income). You don't need to film the whole movie again; you just use the existing footage and apply a filter that reveals the hidden details of who is really interacting with whom. This helps us see that the "average" experience is often a lie, and that the smallest groups at the party are often the ones most sensitive to how the virus spreads.

Technical Summary

Technical Summary: Extending Social Contact Matrices with Socio-Demographic Factors

Problem Statement

Compartment-based epidemic models rely on social contact matrices to define interaction rates between population groups. While these models commonly stratify populations by age, they frequently ignore other socio-demographic factors (e.g., ethnicity, deprivation, education) that significantly influence transmission dynamics. Relying solely on age-based stratification can obscure critical heterogeneities, leading to biased estimates of the basic reproduction number ($R_0$), final epidemic size, and health disparities.

Existing methods to incorporate multiple socio-demographic factors are limited. They typically require comprehensive social contact surveys that capture pairwise interaction rates across all combinations of factors (e.g., age $\times$ ethnicity). Such data are expensive, time-consuming to collect, and often suffer from perception biases. Furthermore, direct estimation of these high-dimensional matrices is often infeasible, leaving models unable to capture the compounding effects of intersecting socio-demographic identities.

Methodology

The authors propose a principled mathematical method to extend an existing age-based social contact matrix ($C_{ij}$) by incorporating an additional socio-demographic factor (denoted as $a$ and $b$ for groups within the new factor). This results in an extended contact matrix $C_{ia,jb}$ representing contacts between individuals in age group $i$ and socio-demographic group $a$ with those in age group $j$ and socio-demographic group $b$.

The method constructs this extended matrix using only three types of inputs:

1. The original age-based contact matrix ($C_{ij}$).
2. The population structure of the new socio-demographic factor (group sizes $N_{ia}$).
3. Assumptions regarding relative contact rates ($F_a$) and assortativity ($\epsilon$).

Core Constraints

The extended matrix must satisfy three conditions to ensure physical and epidemiological validity:

1. Aggregation Consistency: Summing contacts over the new socio-demographic groups must recover the original age-based matrix ($C_{ij}$). This ensures the extension redistributes existing contacts rather than creating or destroying them.
2. Symmetry: The matrix must satisfy the reciprocity condition ($N_{ia}C_{ia,jb} = N_{jb}C_{jb,ia}$), ensuring that if group A contacts group B, group B contacts group A.
3. Relative Interaction Condition: The ratio of total contact rates between two socio-demographic groups ($F_{a1}/F_{a2}$) must be consistent across all age groups.

Mixing Scenarios

The authors derive solutions for two mixing regimes and a linear combination of both:

• Proportionate Mixing: Contacts are distributed across socio-demographic groups in proportion to their population size and relative contact rates.
• Segregated (Assortative) Mixing: Contacts occur preferentially within the same socio-demographic group. The authors provide a specific formulation for "segregated mixing" where interactions are split based on the proportion of the socio-demographic group within each age group.
• Assortative Mixing: A general solution is defined as a linear combination of the proportionate and segregated matrices, controlled by an assortativity constant $\epsilon$ (where $\epsilon=0$ is purely proportionate and $\epsilon=1$ is purely segregated).

The method allows for the re-indexing of the four-dimensional array into a two-dimensional matrix, theoretically permitting the extension of matrices with arbitrary numbers of socio-demographic factors, though the authors note practical limitations regarding model complexity and the need for additional assumptions.

Key Contributions

1. A Data-Efficient Extension Method: The paper introduces a method to stratify contact matrices by additional socio-demographic factors without requiring granular pairwise survey data. It relies instead on population structure data and aggregate mixing assumptions.
2. Mathematical Formulation: The authors provide explicit formulas for constructing extended matrices under proportionate, segregated, and mixed assortative conditions, proving that these constructions satisfy the necessary symmetry and aggregation constraints.
3. Demonstration of Hidden Heterogeneity: Through numerical simulations, the study demonstrates how ignoring socio-demographic factors can mask significant variations in epidemic outcomes, particularly for minority groups.

Results

The authors validated the method using two approaches: hypothetical populations and a projection of the European POLYMOD contact survey onto Aotearoa New Zealand's age-ethnic structure.

Hypothetical Simulations

• Impact of Relative Contact Rates: Varying the relative contact rates between socio-demographic groups significantly altered the basic reproduction number ($R_0$) and attack rates. $R_0$ generally increased as the contact rate of one group deviated from the mean, driven by the fact that high-contact groups drive transmission more than low-contact groups reduce it.
• Minority Group Sensitivity: Epidemic outcomes for minority groups were found to be highly sensitive to variations in model parameters (contact rates and assortativity).
• Non-Identifiability: The study found that different combinations of assortativity and relative contact rates could produce similar aggregate attack rates, suggesting these parameters may be non-identifiable from aggregate data alone.

Aotearoa New Zealand Projection

Using the POLYMOD matrix projected onto New Zealand's ethnic structure (Māori, Pacific, Asian, European/Other):

• Shift in Dynamics: Incorporating ethnicity and relative contact rates increased the estimated $R_0$ but decreased the total projected number of infections by approximately 2% (roughly 100,000 people) compared to an age-only model.
• Age Structure Shift: The inclusion of ethnicity shifted the age structure of infections. The attack rate increased for older populations while decreasing for younger populations. This occurred because the younger ethnic groups (Māori and Pacific) had higher contact rates, but the older groups (European/Other) had a different age distribution that altered the overall transmission dynamics.
• Policy Implications: The results suggest that age-only models may overestimate total cases while underestimating the burden on specific age-ethnic subgroups, potentially leading to suboptimal policy decisions regarding resource allocation and intervention targeting.

Significance and Claims

The paper claims that its method provides a "principled way" to construct extended contact matrices when detailed empirical data is lacking. The primary significance lies in the ability to:

• Quantify Inequities: Enable models to explicitly account for socio-demographic heterogeneities that drive health disparities.
• Improve Policy Relevance: Provide more accurate estimates of $R_0$ and epidemic size for specific subgroups, which is critical for designing interventions that minimize inequalities.
• Reveal Obscured Variation: Demonstrate that aggregate models can obscure critical variations in disease outcomes, particularly for minority groups, which are often the most affected during epidemics.

The authors remain modest regarding the method's limitations, acknowledging that it imposes a specific structure on the contact matrix and assumes uniform assortativity and relative contact rates across age groups. They note that the method can produce negative matrix elements in extreme parameter scenarios, requiring verification of non-negativity before application. Furthermore, the method inherits general limitations of compartmental models, such as the assumption of homogeneous mixing within compartments and the lack of spatial resolution.