Do We Adequately Control for Unmeasured Confounders When Estimating the Short-term Effect of Air Pollution on Mortality?

Numerous time series studies have quantified the potential association between daily variations in air pollution and daily variations in non-accidental deaths. In order to account for the presence of unmeasured confounders, a smooth function of time trend is typically used as a proxy for these variables. We shed light on the validity of the results obtained by using this approach. Specifically, we use data from the National Morbidity, Mortality and Air Pollution Study database, and carry out a carefully designed simulation study. Our findings suggest that the use of a smooth function of time trend cannot fully account for the presence of unmeasured confounders, especially when their impact is strong relatively to the effect of air pollution, and when several unobservables are not included in the model.