Scaling Spatially Varying Distributed Lag Nonlinear Models for Nationwide Small-Area Data: A Divide-and-Conquer Approach

Abstract

Lagged and nonlinear exposure-response relationships are frequently modelled using distributed lag non-linear models (DLNMs). However, these relationships can vary substantially across locations due to differences in climate, population structure, adaptation and socio-economic characteristics, among others. Spatially-varying DLNMs can capture this heterogeneity, but their computational demands limit their application to fine-scale nationwide datasets with thousands of small areas. In this paper, we adapt a 'divide-and-conquer' approach to spatially varying DLNMs. The method partitions the spatial domain into smaller regions, expands each region with a k-order neighbourhood to reduce boundary effects, fits separate models within each partition using Laplacian P-splines, and combines the resulting estimates using conditional predictive ordinate-based mixture weights. We evaluated the approach in a simulation study based on municipalities in Sicily and applied it to daily summer temperature and mortality data from 7122 municipalities in mainland Italy between 2011 and 2021. Simulation studies show good performance, with relative-risk estimates comparable to those from a global model and relatively robust to the choice and number of partitioning units. Computational efficiency can be substantially improved by using parallel computations. The Italian application reveals significant spatial heterogeneity in the temperature-mortality relationship, with clear regional differences. The proposed framework makes spatially varying DLNMs feasible for large, fine-scale spatiotemporal datasets while retaining location-specific exposure-lag-response estimation.