Dong, Q. (STAT) – Prevalence Mapping and Model Evaluation for Binary and Categorical Outcomes from Household Surveys

Reliable subnational estimates of health and demographic indicators from household surveys are crucial for understanding public health progress, especially in low- and middle-income countries (LMICs), where vital registration data are limited, and household surveys serve as the primary source of information. We develop a principled workflow for prevalence mapping using household survey data, addressing all stages of analysis. We then propose a generalizable model comparison framework for prevalence mapping in the absence of ground truth. Specifically, we develop cross-validation schemes for small area estimation (SAE), preserving stratification and clustering structures from the complex survey design. We introduce error metrics that enable the evaluation of models based on their accuracy in estimating finite-population prevalence at the target level of inference. Through simulation studies, we demonstrate that conventional practices such as leave-one-area-out (LOAO) cross-validation or comparisons based on aggregated estimates can lead to misleading conclusions. In contrast, our proposed frameworks can more accurately rank SAE models. Finally, we extend the prevalence mapping framework to accommodate categorical and ordinal outcomes. For area-level models, we apply continuation-ratio logit transformations to design-based direct estimates and model the transformed outcomes using a multivariate Fay–Herriot model. We explore modeling the dependence structures between spatial regions and outcome categories through structured random effects. We also explore extending similar random effect construction to unit-level models, where we directly model cluster-level categorical response.

Event Host: Qianyu Dong, Ph.D. Student, Statistical Science

Advisor: Zehang Richard Li