A0610
Title: Spatial prediction and risk mapping: A generalized linear model approach with applications to disease modeling
Authors: Divya Kappara - Indian Institute of Technology Bombay (India) [presenting]
Siuli Mukhopadhyay Siuli Mukhopadhyay - IIT Bombay (India)
Shubham Niphadkar - INDIAN INSTITUTE OF TECHNOLOGY BOMBAY (India)
Abstract: Epidemiological data typically appear as over-dispersed count observations sampled from limited spatial locations. This sparsity motivates the use of spatial statistical methods to predict disease burden in unsampled regions while handling non-normality. Identifying high-risk regions based on response or covariate thresholds is crucial in disease modeling. A related approach is used in engineering optimization and experimental design, where response surfaces are fitted to data by evaluating the objective function at selected points. Further contour estimation, often based on threshold values of interest, identifies regions satisfying specific constraints or performance criteria. Building upon this framework in the context of disease modeling, a generalized linear model approach is proposed for predicting spatially referenced count data. The response variable is assumed to be conditioned on a weakly stationary latent spatial process accounting for both overdispersion and spatial correlation structure. The model estimates are then used to obtain predictions at new locations of interest, along with measures of prediction uncertainty, and identify high-risk contours of interest. The prediction uncertainty is quantified by means of resampling methods for spatial data under a GLM setup. The proposed method is illustrated through empirical and real data analysis, and the effect of varying sampling designs is also studied.
