A causal machine-learning framework for studying policy impact on air pollution: a case study in COVID-19 lockdowns

Area of Science:

• Environmental Epidemiology
• Statistical Modeling
• Air Quality Analysis

Background:

• Policy interventions and natural experiments require rigorous statistical analysis to determine causal impacts on air pollution.
• Confounding factors often complicate the accurate assessment of environmental policy effects.
• COVID-19 lockdowns serve as a relevant case study for evaluating methods to analyze environmental changes.

Purpose of the Study:

• To present a comprehensive framework for estimating and validating causal changes in air pollution time series.
• To propose flexible machine learning-based comparative interrupted time series (CITS) models.
• To introduce a diagnostic criterion for empirical validation against false effects.

Main Methods:

• Utilizing flexible machine learning-based comparative interrupted time series (CITS) models.
• Outlining assumptions for causal effect identification, highlighting advantages of machine learning models over common methods.
• Proposing a diagnostic criterion for validating causal effects, particularly in pre-intervention periods.

Main Results:

• Machine learning approaches demonstrated superior performance in guarding against false effects compared to conventional methods.
• The framework was applied to analyze the impact of COVID-19 lockdowns on atmospheric nitrogen dioxide (NO2) levels in the eastern United States.
• Significant decreases in NO2 levels were observed in Boston, New York, Baltimore, and Washington, DC during the pandemic lockdowns.

Conclusions:

• The proposed validation framework is crucial for selecting appropriate methods for air pollution time series analysis.
• Machine learning-based CITS models are effective tools for studying causal changes in air pollution.
• The study highlights the environmental impact of COVID-19 lockdowns on air quality in major US urban centers.