Joint Bayesian Nowcasting of Severe Acute Respiratory Illness and COVID-19 Positives in Brazil

Surveillance of severe acute respiratory illness (SARI) in Brazil provides an early warning system for respiratory outbreaks, including COVID-19, and statistical nowcasting methods are vital for correcting long and variable reporting delays that would otherwise hinder timely decision-making. Coherent joint prediction of SARI and COVID-positive SARI in Brazil could improve outbreak response planning and hospital resource management, but most existing nowcasting methods only target one outcome at a time. Furthermore, existing approaches usually focus solely on reconstructing recent incidence rather than forecasting future trends, which could support more proactive risk mitigation. Here, we propose a Bayesian hierarchical framework for joint nowcasting and short-term forecasting of two outcomes, where one is a strict subset of the other. Building on the generalized-Dirichlet-multinomial (GDM) method for correcting delayed reporting, a new beta-binomial component links SARI and COVID-positive counts, while separate conditional GDM components capture their distinct delay patterns. To allow for changes over time in the level of disease and delay distributions, flexible latent effects are included in all model components. Using national surveillance data from 2021 to 2024, we conduct a 20-date rolling prediction experiment across Brazil's 27 federative units. Compared to a well-established Bayesian nowcasting approach, our joint model achieves about one-third lower mean absolute error and continuous ranked probability score for contemporaneous nowcasts, with the largest gains in high-incidence regions. Meanwhile, energy scores indicate improved calibration for joint forecasts of total and COVID-positive SARI relative to comparable independent models.