🤖 AI Summary
Conditional autoregressive (CAR) models, widely used in spatial modeling, suffer from inherent conditional construction that induces underestimation of marginal variances at boundary regions, severe heteroscedasticity, and geometry-dependent artifacts—compromising posterior inference reliability in applications such as disease mapping. To address these limitations, we propose the homoscedastic CAR (HCAR) distribution, which integrates a global variance constraint and an adjacency-structure-aware weight rescaling mechanism into the classical CAR framework, explicitly correcting edge effects and geometric bias. HCAR preserves spatial dependence modeling capability while ensuring uniform marginal variances across all regions, thereby enhancing posterior stability and interpretability. Empirical evaluation on multiple real-world disease mapping datasets demonstrates that HCAR effectively mitigates boundary anomalies, yields more plausible marginal distributions, and remains computationally compatible with standard Bayesian inference pipelines.
📝 Abstract
Conditional auto-regressive (CAR) distributions are widely used to induce spatial dependence in the geographic analysis of areal data. These distributions establish multivariate dependence networks by defining conditional relationships between neighboring units, resulting in positive dependence among nearby observations. Despite their practical convenience, the conditional nature of CAR distributions can lead to undesirable marginal properties, such as inherent heterogeneity assumptions that may significantly impact the posterior distributions.
In this paper, we highlight the variance issues associated with CAR distributions, particularly focusing on edge effects and artifacts related to the region's geometry. We show that edge effects may be more significant and widespread in the outcomes of disease mapping studies than previously anticipated. To address these homoscedasticity concerns, we introduce a new conditional autoregressive distribution designed to mitigate these problems. We demonstrate how this distribution effectively resolves the practical issues identified in earlier models.