The methodological challenges associated with conducting research on air pollution mixtures are well-known: correlated co-pollutants result in unstable effect estimates and large standard errors, hindering the assignment of causality to any one exposure. There is still relatively little research in the growing multi-pollutant literature that is focused on the mixture itself as the unit of analysis. In this dissertation, I implement a statistical method from the causal inference literature to estimate the effects of ambient air pollution, as single pollutants and in a two-pollutant mixture.

I analyze the effects of single-pollutant and multi-pollutant summertime ambient air pollution exposures on pulmonary function in a cohort of children with asthma living in Fresno, California. I employ a technique from the causal inference literature, the Population Intervention Model (PIM), to describe the health effects that would result from several hypothetical interventions that involve lowering concentrations of ambient air pollution. By describing the health effects of the ambient air pollutants in these terms, this approach estimates results that are relevant to real-world policy questions. Furthermore, this analytical approach permits the calculation of air pollution health effects that correspond to multiple pollutants dynamically changing within a mixture, as ambient air pollution is actually experienced by people. I interpret each of these health effects according to whether it reflects a realistic, or even a possible, exposure scenario during the study period and in the region where data were collected. I achieve this through an examination of the individual and joint distributions of the pollutants under study.

This dissertation contains several analyses, corresponding to single- and multi-pollutant exposure regimens. In the first analysis, I analyze the effects of ambient summertime NO2 on FEF25-75 in a single-pollutant approach that demonstrates the methodological approach. All analyses use central-site exposure data, assigning all subjects on a given study day the same air pollution exposure values. Ambient PM10-2.5 is analyzed throughout as a summertime pollutant of secondary interest, both in a single-pollutant PM10-2.5 analysis, and in a mixture analysis. For the multi-pollutant mixture analysis, I extend the Population Intervention Model framework demonstrated in the single-pollutant analyses to a two-pollutant summer analysis of ambient NO2 and PM10-2.5, estimating health effects associated with an intervention that dynamically alters levels of one or both pollutants. In this two-pollutant analysis, I estimate the effects of lowering levels of one co-pollutant while "controlling for" the other (i.e., holding it at observed levels), as well as the effects of a joint intervention that decreases levels of both pollutants.

The Background chapter presents a brief history of air pollution epidemiology and policy, and reviews the epidemiologic and statistical research upon which this dissertation builds. The Methods chapter describes the data collection protocol of the Fresno Asthmatic Children's Environment Study (FACES), the theoretical basis for the chosen methodological approach, and the details of the statistical methods employed in these analyses. In the Results section, I describe the characteristics of the FACES study sample, provide tabular and graphical descriptions of the distribution of ambient air pollution in the study, and present the results of the single- and multi-pollutant PIM analyses. In the Discussion section, I provide interpretation of the effects estimated in these various analyses, and refer back to the single- and multi-pollutant exposure distributions to situate the various health effects in appropriate context, and to speculate on potential sources of bias.

All health effects calculated in these analyses were estimated relatively imprecisely; however, comparison of the magnitude and direction of the risk differences across analyses demonstrates patterns and provides information about the respiratory effects of the pollutants analyzed in this study. Furthermore, consideration of the individual and joint distributions of the two exposures yields key insight that guides the interpretation of these findings, especially as relates to parameter identifiability. In this analysis, there is ample evidence that the types of air pollution profiles described by two interventions are not realistic given the observed data, and furthermore that there is not support in the data to estimate health effects for these interventions. These parameters were defined to be comparable to standard practice in the multi-pollutant literature. The finding that they were not identifiable in the FACES data argues against giving weight to these specific findings, and also raises broader questions about parameters of this type: large, isolated single-pollutant concentration changes in a multi-pollutant exposure regimen. The work presented here emphasizes that such parameters should be scrutinized for positivity and data support before commencing analysis, regardless of the analytical approach chosen.