Synthetic population generation provides detailed datasets when real data are unavailable, supporting research and policy analyses that require representative populations, with applications in transportation and many other fields. However, most existing methods focus on single-level data (e.g., individuals) or single-period snapshots, limiting their ability to capture hierarchical structures (e.g., households with members) and the temporal evolution of real populations. To address these limitations, this thesis develops a unified simulation framework that advances synthetic population generation along two key dimensions: hierarchy and time.
The first part addresses the challenge of creating internally consistent synthetic households. We formulate household synthesis as a joint sampling problem over household- and individual-level attributes, and propose a practical extension of the Gibbs sampling approach that accounts for different household types. By ordering individuals by decreasing age, the method eliminates arbitrary role assignments and sequential grouping steps common in existing approaches, thereby reducing the possibility of generating illogical households and enabling more diverse and realistic household representations.
Building on this, the second part extends the framework to an adaptive, time-evolving population generation process. Rather than generating static snapshots that quickly become outdated as populations evolve, we combine dynamic projection with periodic Gibbs resampling. This allows populations to evolve through simulated demographic events while realigning with new data as they become available. The adaptive mechanism reduces computational costs by avoiding full regeneration, improves representativeness under data scarcity or bias, and enhances robustness to external shocks that disrupt demographic trends.
Finally, we address the problem of synthetic panel data generation, enabling individual trajectories to be tracked across multiple time periods. The core of our method is the formulation of universal, time-independent variables encoding life course trajectories, from which time-specific populations can be derived. These variables can be sampled directly from demographic models for data-free generation or estimated via a tailored maximum likelihood approach when cross-sectional data are available. The resulting synthetic panels combine model-based regularities with data-driven refinement, producing longitudinal datasets that maintain internal consistency while reflecting observed demographic patterns.
Taken together, the contributions of this thesis shift research efforts from repetitive regeneration toward adaptive enrichment, developing methods that evolve synthetic populations by integrating new information over time. By combining the strengths of model-based and data-driven paradigms, the proposed framework enables the generation, projection, and continuous updating of synthetic populations by integrating data
