Statistical and Machine Learning Models for the Analysis of Longitudinal Air Pollution Categorical Outcomes: An Illustration Using EPA Data from the USA

Tracking air quality matters for public health because it monitors invisible pollutants that cause serious respiratory and cardiovascular diseases, among other adverse health effects. Air quality is measured by identifying the levels of key pollutants and fine particulate matter present in the atmosphere, among other components. Inefficient energy use increases emissions of CO2 and other greenhouse gases, which affect air quality. This ultimately affects sensitive groups, e.g., children, older adults, and people with asthma or heart problems, who face the highest risks from poor air quality. Fine particulate matter, such as PM2.5 and PM10, is a major concern when air quality is unhealthy. PM2.5 is air contamination from tiny particles that negatively impacts the environment and human health at high levels. Breathing in harmful levels will increase the risk of various health problems, e.g., heart disease, asthma, and low birth weight. Unhealthy levels can also reduce visibility and cause the air to appear hazy. Therefore, it is crucial to identify how pollutant levels degrade air quality. For these reasons, monitoring stations worldwide track air quality and store data repeatedly over time at many sites, along with other risk factors, producing a massive amount of data (big data). Modeling such data and predicting the risk of PM2.5 trajectories is of enormous interest to researchers and policymakers. This paper proposes a modeling framework that uses a Markov chain and a statistical model, along with six machine learning algorithms, to measure how covariates affect trajectory risk. We illustrate the proposed model using PM2.5 outdoor air pollution data from the United States from 2000 to 2020. Using predicted marginal, conditional, and joint probabilities, we compare trajectories from all models for stations with a specific covariate pattern. We report prediction accuracy for logistic regression and six machine learning models, NN, DT, RF, SVM, KNN, and NB, using 10-fold cross-validation and Leave-One-Out Cross-Validation (LOOCV). Overall, the datasets achieved 49% to 89% aggregated prediction accuracy across all left-out test evaluations for LOOCV and out-of-fold validation (test) sets. The proposed methodology will enable prediction of trajectory risk from longitudinal, correlated responses and covariates.