Evaluating statistical methods to predict indoor black carbon in an urban birth cohort

Most air pollution epidemiology studies rely on outdoor exposure data from various sources, such as reference monitors, low-cost monitors, models, or Earth observations. However, people spend 90 % of their time indoors, with 70 % of that time spent at home, which may result in misclassification of air pollution exposure when using data reflecting ambient concentrations. In this study, we evaluated methods to predict residential indoor black carbon (BC) from outdoor BC, PM2.5, and housing characteristics to support future efforts in estimating personal air pollution exposure. Households from the Healthy Start cohort in Denver, CO hosted paired indoor/outdoor low-cost air samplers for one-week periods during spring 2018, summer 2018, and winter 2019. Participants completed questionnaires about housing characteristics like building type, flooring, and heating and cooling methods. Filters were analyzed for BC using transmissometry. Ridge, LASSO and ordinary least squares regression (OLS) techniques were used to build predictive models of indoor BC given the available set of covariates. Leave-one-out cross-validation was used to assess the predictive accuracy of each model. We hypothesized that Ridge and LASSO will obtain improved predictive performance over the OLS model due to regularization. A total of 27 households participated, with 39 paired measurements available after data cleaning. All winter data were excluded due to high variability and incomplete sampling times for outdoor measurements. Performance issues suggested insufficient weatherproofing of monitors for low temperatures. The Ridge regression showed the best predictive performance. The final inference model included outdoor PM2.5, hard floors, and the presence of pets in the home, accounting for approximately 28 % of the variability in indoor BC concentrations measured in participant homes. In the absence of indoor monitoring, household characteristics like flooring and the presence of pets can help predict indoor levels of BC.