Ziyu Liu

The COVID-19 pandemic has reshaped global infrastructure, highlighting the importance of effective infectious disease management. Identifying when and where infection trends change abnormally can aid strategic planning; yet, existing change point detection methods struggle due to the non-linear nature of infection trends, spatial and temporal dependencies, regional demographic and healthcare variations, and differing preventive measures. To address this issue, we propose a Bayesian method that can detect candidate regional disease-related change periods while overcoming these challenges. Specifically, we develop a Bayesian function-on-function regression model that learns from infection trends across multiple regions by incorporating both time-invariant features and the historical effect of time-dependent functional covariates. Temporal dependence in the covariate effects is captured through neighborhood-based spike-and-slab priors, whose latent binary inclusion indicators are, in turn, modeled by Ising priors. A Gibbs sampling framework is derived to approximate the joint posterior distribution of the model parameters. We compared the performance of the proposed framework against two widely used change-point detection methods, BCP and Segmented. In our simulation studies, BLIDE achieves an F1-score of $1.000$ under high signal-to-noise conditions and maintains an F1-score above $0.95$ even when noise dominates the trends, substantially outperforming BCP (F1-scores $0.454$ and $0.131$, respectively) and Segmented (F1-scores below $0.05$ across all scenarios).