The pitfalls of incidence-based time series regression for inferring the effects of weather on infectious diseases.

A central question in environmental epidemiology is how the weather affects infectious diseases. Time-series regression (TSR) on population level case-counts data is widely used to estimate weather effects; however, this design may be biased due to the complexities of infectious disease dynamics, including nonlinear feedback, various types of noise, and latent, dynamic variables such as population immunity. Here, we assess the reliability of incidence-based TSR through a controlled simulation study across four different climates and fifty scenarios representing different pathogens. For each scenario, we simulated 10 years of weekly case-count data using a transmission model that included real-world weather data on temperature and relative humidity. We then examined whether the ground-truth weather effects could be recovered from simulated data using negative binomial generalized additive models, a flexible class of TSR models commonly used in empirical applications. We find that these models frequently fail to yield accurate and precise estimates of weather effects, even under favorable conditions such as no process noise and low observation noise (overdispersion). Hence, our results suggest that more causally elaborated formulations of TSR models, or more mechanistic approaches like transmission models, are needed for statistical inference of weather effects from population data.