A new bayesian ecological inference model with covariates

Ecological inference models are essential tools for estimating transition matrices from aggregate data, particularly in fields like political science and epidemiology. Building on the multinomial-Dirichlet framework developed by King, Rosen, Tanner, and Jiang (2001), this paper proposes a refined ec...

Descripción completa

Detalles Bibliográficos
Autor: Candela Rubio, David
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/447219
Acceso en línea:https://hdl.handle.net/2117/447219
Access Level:acceso abierto
Palabra clave:Bayesian statistical decision theory
Inference
Ecology
Model jeràrquic bayesià
Model bayesià
Inferència ecològica
Dades electorals
Eleccions
Louisiana
Nova Zelanda
Bayesian hierarchical model
Bayesian model
Ecological inference
Election data
Elections
New Zealand
Estadística bayesiana
Inferència
Ecologia
Àrees temàtiques de la UPC::Matemàtiques i estadística::Estadística matemàtica::Inferència estadística
Descripción
Sumario:Ecological inference models are essential tools for estimating transition matrices from aggregate data, particularly in fields like political science and epidemiology. Building on the multinomial-Dirichlet framework developed by King, Rosen, Tanner, and Jiang (2001), this paper proposes a refined ecological inference model that introduces an alternative covariate formulation. Our approach maintains the flexibility and relaxed assumptions of the original model while addressing key structural and computational limitations, such as numerical stability and efficiency. We further investigate the role of random effects by implementing two model variants, with and without random effects, to evaluate their impact on interpretability, computational speed, convergence, and robustness to unobserved heterogeneity. Using real-world voter registration data from Louisiana and simulated data with New Zealand elections as reference, we benchmark the proposed model against established Bayesian and numerical ecological inference methods, including the original covariate model. Our results demonstrate that the refined model consistently matches or exceeds the accuracy of existing approaches while reducing computational time by up to 80%, and achieving greater numerical stability. Notably, excluding random effects can lead to order-of-magnitude speed improvements without significantly compromising accuracy at the global level. We also highlight the importance of covariate relevance, finding that only covariates strongly correlated with the latent transition structure substantially improve estimation. These findings clarify when model complexity is justified and when simpler models suffice. In summary, this study contributes to the field of ecological inference by introducing a robust, data-supported, and computationally streamlined method for integrating covariates into the estimation of transition matrices. We explore the wider relevance of our approach across different domains and suggest future research avenues, such as developing faster numerical techniques and refining prior assumptions for covariate modelling.