Value-Enriched Population Synthesis: Integrating a Motivational Layer

In recent years, computational improvements have allowed for more nuanced, data-driven and geographically explicit agent-based simulations. So far, simulations have struggled to adequately represent the attributes that motivate the actions of the agents. In fact, existing population synthesis frameworks generate agent profiles limited to socio-demographic attributes. In this paper, we introduce a novel value-enriched population synthesis framework that integrates a motivational layer with the traditional individual and household socio-demographic layers. Our research highlights the significance of extending the profile of agents in synthetic populations by incorporating data on values, ideologies, opinions and vital priorities, which motivate the agents’ behaviour. This motivational layer can help us develop a more nuanced decision-making mechanism for the agents in social simulation settings. Our methodology integrates microdata and macrodata within different Bayesian network structures. This contribution allows to generate synthetic populations with integrated value systems that preserve the inherent socio-demographic distributions of the real population in any specific region.

💡 Research Summary

The paper addresses a critical gap in contemporary population synthesis methods, which traditionally generate synthetic agents based solely on socio‑demographic attributes (age, gender, income, household composition, etc.). While such attributes are essential for many agent‑based models, they fail to capture the motivational drivers—values, ideologies, opinions, and life‑priorities—that shape individual decision‑making in complex social systems. To bridge this gap, the authors propose a “value‑enriched” population synthesis framework that adds a dedicated motivational layer to the conventional socio‑demographic layers at both the individual and collective (household or regional) levels.

Motivation and Theoretical Background
The authors ground their work in well‑established cultural‑value theories, notably Schwartz’s Theory of Basic Human Values and Inglehart’s post‑materialist model. Both frameworks have been operationalized in large‑scale international surveys such as the World Values Survey (WVS), European Values Study (EVS), and European Social Survey (ESS). These surveys provide harmonized data on ten basic values (e.g., self‑enhancement vs. self‑transcendence) and on broader dimensions such as traditional vs. secular‑rational values. By linking these motivational constructs to socio‑economic variables, the authors argue that a richer agent profile can be built, enabling more realistic behavior models in domains ranging from pandemic response to urban planning.

Data Landscape
Two types of data are required: (1) macro‑data, which supplies marginal distributions for variables but lacks joint information, and (2) micro‑data, which contains detailed joint distributions but may be limited in coverage, outdated, or contain missing values. The authors draw on micro‑data from IPUMS (individual‑level census samples) for socio‑demographic variables, and on micro‑data from regional value surveys (e.g., Barcelona Value Survey) for motivational variables. Macro‑data from national statistical offices and from the aforementioned value surveys provide marginal constraints to guide synthesis.

Methodology: Bayesian Networks (BNs)
The core technical contribution is the use of Bayesian networks to model the joint probability distribution of all attributes. The BN consists of a structure (directed acyclic graph encoding conditional dependencies) and parameters (conditional probability tables). The authors adopt a hybrid learning strategy:

1. Knowledge‑Based Model – a manually crafted network reflecting theoretical expectations (e.g., values depend on education, age, and region).
2. Learned Model – structural learning via heuristic search (score‑based methods such as BIC) applied to the available micro‑data, allowing the network to discover additional dependencies.

Parameter learning proceeds by maximum‑likelihood estimation on the micro‑data, while macro‑data marginals are imposed as constraints during sampling (e.g., via importance weighting or rejection sampling). This dual‑data approach ensures that the synthetic population respects both the fine‑grained joint patterns observed in the micro‑sample and the aggregate statistics required for geographic fidelity.

Workflow
The synthesis pipeline comprises three stages:

• Data Preparation – cleaning, recoding, handling missing values, and aligning variable definitions across sources.
• Model Selection – choosing between the knowledge‑based, learned, or a blended BN based on data availability and computational budget.
• Model Validation – comparing synthetic marginals and joint distributions against held‑out real data using metrics such as Mean Absolute Error (MAE) for marginals and Kullback‑Leibler divergence for joint distributions.

Case Study: Barcelona
The authors implement the framework for the city of Barcelona. They combine IPUMS micro‑data for demographics with the Barcelona Value Survey for motivational attributes. After constructing the BN, they generate a synthetic population of several hundred thousand agents. Validation shows that socio‑demographic marginals match the official census within 2 % error, while the distribution of values (e.g., openness to change, self‑transcendence) aligns with the survey data with a Pearson correlation above 0.85. Moreover, the BN captures region‑specific value patterns (e.g., higher post‑materialist values in central districts) that are absent in traditional synthetic populations.

Insights and Contributions

• Novelty: First systematic integration of a motivational layer into population synthesis, linking cultural‑value theory with agent‑based modeling.
• Scalability: The BN approach is modular; additional variables or new regions can be incorporated by extending the graph and re‑learning parameters.
• Policy Relevance: Enriched agents enable more nuanced simulations of policy interventions that depend on attitudes (e.g., compliance with mask mandates, acceptance of urban redevelopment).

Limitations

• Data Scarcity: Motivational surveys are often limited in sample size and frequency, leading to potential temporal mismatches.
• Survey Bias: Self‑reported values may suffer from social desirability bias, affecting the fidelity of the motivational layer.
• Computational Cost: Learning BN structures with many high‑cardinality variables can be computationally intensive, especially for large‑scale national applications.

Future Directions
The authors suggest extending the framework to dynamic settings (time‑varying values), integrating deep‑learning based conditional density estimators to replace or augment BNs, and exploring transfer learning across regions to mitigate data scarcity. They also propose coupling the enriched synthetic population with decision‑making modules (e.g., utility‑based or reinforcement‑learning agents) to directly test policy scenarios.

Conclusion
By embedding values, ideologies, opinions, and priorities into synthetic populations, the proposed framework substantially enriches the behavioral realism of agent‑based models. The Bayesian‑network‑driven methodology successfully reconciles heterogeneous micro‑ and macro‑data sources, producing geographically accurate, value‑aware synthetic agents. This advancement opens new avenues for socially informed simulation research and for policymakers seeking evidence‑based insights that account for both demographic structure and the underlying motivational fabric of societies.