Estimating the Influence of Wind on Air Pollution Using a Causal Inference Pipeline

Changes in wind patterns can substantially alter the air pollution level of a city. However, it is challenging to estimate a causal effect from observed data. Since wind patterns are not randomly distributed over time and are related to other weather parameters influencing air pollution, researchers must adjust for these confounding factors. As an alternative to current practices, we implement a causal inference pipeline to embed an observational study within an hypothetical randomized experiment. We illustrate this new approach for air pollution studies using 4018 daily observations from Paris, France, over the 2008-2018 period. Using the Neyman-Rubin potential outcomes framework, we first define treatment of interest as the comparison of air pollutant concentrations when winds are blowing from the North-East (824 units) with concentrations when wind come from other directions (3,194 units). We then implement a matching algorithm to approximate a pair randomized experiment and find 119 matched pairs. By selecting units that are comparable in regards to various confounders, matching allows us to adjust nonparametrically for observed confounders while avoiding model extrapolation to treated days without similar control days. Once the balance of treated and control groups was deemed satisfactory, we estimate the average differences in air pollutant concentrations and their sampling variability using Neymanian inference, a mode of inference specifically designed for randomized experiments. We find that North-East winds increase PM10 concentrations by 4.8 μg/m³ (95% CI: 2.6, 6.9). As in any observational studies, an unobserved confounder could bias our results. We therefore carry out a quantitative bias analysis which reveals that an unobserved variable 2 times more common among treated units could make our data compatible with small negative effects up to very large effects (95% CI: -2.3, 10). Our causal inference approach highlights the importance of checking covariates balance and bias from unmeasured confounders in air pollution studies.