Using Explainable Machine Learning to Interpret the Effects of Policies on Air Pollution: COVID-19 Lockdown in London

Abstract

Activity changes during the COVID-19 lockdown present an opportunity to understand the effects that prospective emission control and air quality management policies might have on reducing air pollution. Using a regression discontinuity design for causal analysis, we show that the first UK national lockdown led to unprecedented decreases in road traffic, by up to 65%, yet incommensurate and heterogeneous responses in air pollution in London. At different locations, changes in air pollution attributable to the lockdown ranged from −50% to 0% for nitrogen dioxide (NO2), 0% to +4% for ozone (O3), and −5% to +0% for particulate matter with an aerodynamic diameter less than 10 μm (PM10), and there was no response for PM2.5. Using explainable machine learning to interpret the outputs of a predictive model, we show that the degree to which NO2 pollution was reduced in an area was correlated with spatial features (including road freight traffic and proximity to a major airport and the city center), and that existing inequalities in air pollution exposure were exacerbated: pollution reductions were greater in places with more affluent residents and better access to public transport services.