Robust Bayesian Hierarchical Models for Small-Area Estimation

Thesis Abstract

In many countries, reliable estimates of socio-economic indicators at small geographic levels are essential for targeted policy design, yet conventional area-level models often fail to capture heterogeneity and outlier behavior in sparse data settings. This study addresses the bias and instability inherent in small-area estimation by developing robust Bayesian hierarchical models that integrate unit-level and area-level information while down-weighting the influence of outliers through heavy-tailed priors and robust likelihoods. The aim is to deliver accurate, uncertainty-quantified small-area estimates for key indicators such as unemployment rate, poverty incidence, and educational attainment across 120 districts, using data aggregated from administrative records and a complementary survey frame. Specific objectives are (i) to formulate a unified Bayesian hierarchical framework that combines area-level Fay-Herriot components with unit-level nested random effects and robust error distributions; (ii) to compare performance against standard Gaussian and t-distribution-based hierarchies via simulation studies under varying outlier contamination and missingness patterns; (iii) to assess the impact of informative priors derived from auxiliary covariates, including migration flows and health indicators, on small-area estimates; (iv) to implement model averaging and cross-validation schemes to enhance predictive accuracy and calibrate posterior intervals; and (v) to translate probabilistic estimates into policy-ready decision metrics by deriving reliable risk-averse decision rules for resource allocation. The methodology adopts a mixed-methods design within a quantitative-analytic paradigm. The population comprises all districts within a national framework, while the primary dataset consists of administrative counts (n = 2,400,000 individuals) and a stratified random sample survey (n = 60,000 respondents) linked via geographic identifiers. Data collection instruments include administrative census extracts, household survey questionnaires adhering to international standard indicators, and covariate databases on education, health, and labor market characteristics. The analysis proceeds in three stages (1) specification and estimation of a robust Bayesian hierarchical model that extends the Fay-Herriot structure by incorporating unit-level random effects and a robust (e.g., Student-t or Laplace) error distribution to accommodate outliers; (2) implementation of spike-and-slab or Laplacian-type priors for shrinkage of small-area random effects where data are extremely sparse; (3) evaluation of model performance through posterior predictive checks, Watanabe-Akaike information criterion (WAIC), and leave-one-area-out cross-validation. Computation is carried out in a high-performance computing environment using Markov chain Monte Carlo methods with Hamiltonian Monte Carlo sampling (Stan) and variational Bayes as a scalability alternative. Model validation includes simulation scenarios with controlled contamination, missingness, and varying sample sizes to assess robustness and calibration of credible intervals. The study also employs causal-inspired diagnostics to examine the influence of auxiliary covariates and prior specifications on estimates. Key expected findings include (i) robustness gains in small-area estimates under outlier regimes, evidenced by reduced mean squared error and improved coverage probabilities relative to conventional Gaussian hierarchical models; (ii) pronounced improvements in districts with sparse data where unit-level information is limited, due to effective borrowing of strength through robust hierarchical structure; (iii) demonstrable gains from incorporating informative priors based on migration and health covariates, yielding more accurate unemployment and poverty risk assessments; (iv) reliable predictive intervals that maintain nominal coverage under data sparsity and contamination. The study contributes to knowledge by providing a transferable robust Bayesian framework for small-area estimation that blends methodological rigor with practical applicability, offering a generalizable blueprint for statistical agencies and researchers working with sparse or noisy data. The theoretical underpinnings draw on robust Bayesian inference, hierarchical modeling, and small-area estimation theory, with reference to the works of Fay and Herriot (1979) and robust Bayesian literature on heavy-tailed priors and outlier resistance. The main conclusion anticipates that robust Bayesian hierarchical models will deliver consistently superior small-area estimates in the presence of outliers and data sparsity, with improved decision support for resource allocation and policy targeting. Recommendations include adopting the proposed framework in national statistical programs, expanding covariate integration for further gains, and developing user-friendly software interfaces to facilitate routine production of small-area indicators with robust uncertainty quantification.