A1728
Title: Jointly modeling zeros, regimes, and lags in epidemic count data via penalized spatio-temporal mixture models
Authors: Weiao Gan - City University of Hong Kong (China)
Zudi Lu - City University of Hong Kong (China) [presenting]
Ngai Hang Chan - City University of Hong Kong (Hong Kong)
Abstract: Daily epidemic count data frequently violate standard spatio-temporal Poisson assumptions, exhibiting excess zeros, multiple latent transmission regimes under similar histories, and dependence on high-dimensional temporal and spatial lags that induce collinearity and model-selection uncertainty. Existing methods typically address these issues in separate, stagewise procedures, which are computationally costly and ill-suited to mixture settings where predictor relevance may vary by regime. A unified observation-driven framework is proposed that jointly models zero inflation, latent regimes, and predictor selection within a single spatio-temporal conditional model. Excess zeros are handled via a structural-zero mechanism, latent regimes through a finite mixture with regime-specific lag structures, and complexity through dual penalization that simultaneously selects active regimes and regime-specific predictors. By conditioning on observed history, the model targets one-step-ahead prediction without introducing latent state dynamics. Estimation is achieved using a stabilized penalized EM algorithm with tractable weighted generalized linear model updates. The resulting approach is interpretable, computationally efficient, and well-suited to moderately high-dimensional epidemic surveillance data, as demonstrated in simulations and real-world data applications.
