On smoothing time series with low average counts

Generalized Additive Models have been widely adopted for studies of acute effects of particulate matter on mortality and morbidity. Monitoringof pollutants and health outcomes increased worldwide and investigators thus increasingly relied on automatic selection methods that exist ofsummary statistics such as AIC and PACF. Methodological studies have used simulations to compare selection methods and their impact on largescale multi-city analyses and concluded that aggressive smoothing is to be preferred. For smaller groups, these effects can be visualised with simpleresidual plots. Data from Belgian cities is used to illustrate the effect of over-smoothing on time series with low average counts.