Dalla Lana School of Public Health

Presented by the Geoinformatics of Spatial and Environmental Health (Ge-iSEE) Lab …

The Geoinformatics of Spatial and Environmental Health (Ge-iSEE) Lab is delighted to host a seminar that brings together scholars and practitioners to strengthen interdisciplinary collaboration and advance knowledge in public health and environmental epidemiology.

About the Talk

Many scholars apply a single method, such as Bayesian Hierarchical Model (BHM) (Murray et al 1996; Global Nutrition Target Collaborators. 2024), Kriging (Gooveats 2005), or other alternatives, to all diseases and all regions. We refer to this practice as One-To-All (OTA). Others try multiple methods and select the one with the smallest cross validated error. We call this the All-To-One (ATO). The OTA approach is appropriate when the assumption of the estimator is align with the properties of the disease, but thus alignment is rarely tested or justified in existing literature. The ATO is suitable for a specific disease under its specific sample conditions, but it lacks a theoretical basis for generalization to other diseases.

To address the limitations of OTA and ATO, we propose using the Spatial Statistic Trinity (SST) as a framework for selecting appropriate models for disease mapping. Data always flow from the population to a sample and ultimately into an estimator. Therefore, the estimation error is determined by (1) the properties of the population, (2) the conditions of the sample, (3) the method of estimation, and their combination, collectively term the Spatial Statistic Trinity. Each components has multiple possibilities: populations may be independent and identical distribution (IID), or spatial autocorrelated (SAC), spatial stratified heterogeneity (SSH), or exhibit both SAC and SSH; sample may be simple random, systematic, snowball, convenience, purposive, self-selected, etc; estimators are numerous, including simple averages, BHM, Kriging, Sandwich, GM, DLNM, Machine Learning, and AI. The combinations of the triple are vast, whether scholars are aware of SST or not. A method for disease mapping is appropriate when the assumption of the model for disease mapping align with the property of the disease and the sample condition. For instance, Bayesian Hierarchical Models (BHM) may work well for monotonic populations but fail when modelling U-shaped like relationships such as those between cardiovascular disease and temperature. Kriging is appropriate when population is denominated by spatial autocorrelation and second order stationarity, but inefficient otherwise. I illustrate SST using an example of age-standardized breast cancer mortality, comparing predictions derived from Kriging (which assumes SAC) and Sandwich Estimators (which account for SSH). Our results show that adapting SST-based model selection significantly improves accuracy and robustness in disease mapping.

About the Speaker

Professor Jinfeng Wang is a Distinguished Professor of Spatial Statistics at the Chinese Academy of Sciences and a leading scholar in GIScience and spatial epidemiology. He is internationally known for developing spatial stratified heterogeneity theory, the Geodetector method, and advances in spatial sampling and disease mapping.

With a PhD from CAS and international training in Austria and the UK, he has led numerous national research projects and published over 240 SCI papers. He also develops widely used spatial analysis software and serves on editorial boards of major journals such as Spatial Statistics and IJGIS.