So far, a large number of studies have quantified the effect of COVID-19 lockdown measures on air quality in different countries worldwide. However, few studies have compared the influence of different approaches on the estimation results. The present study aimed to utilize a random forest machine learning approach as well as a difference-to-difference approach to explore the effect of lockdown policy on nitrogen dioxide (NO 2 ) concentration during COVID-19 outbreak period in mainland China. Datasets from 2017 to 2019 were adopted to establish the random forest models, which were then applied to predict the NO 2 concentrations in 2020, representing a scenario without the lockdown effect. The results showed that random forest models achieved remarkable predictive accuracy for predicting NO 2 concentrations, with index of agreement values ranging between 0.34 and 0.76. Compared with the modelled NO 2 concentrations, on average, the observed NO 2 concentrations decreased by approximately 16 µg/m 3 in the lockdown period in 2020. The difference-to-difference approach tended to underestimate the influence of COVID-19 lockdown measures. Due to the improvement of NO 2 pollution, around 3722 non-accidental premature deaths were avoided in the studied population. The presented machine learning modelling framework has a great potential to be transferred to other short-term events with abrupt pollutant emission changes.