Causal Inference Methods in Air Pollution Research
PAPADOGEORGOU-DISSERTATION-2018.pdf (7.516Mb)(embargoed until: 2020-05-01)
MetadataShow full item record
CitationPapadogeorgou, Georgia. 2018. Causal Inference Methods in Air Pollution Research. Doctoral dissertation, Harvard University, Graduate School of Arts & Sciences.
AbstractWhile the air pollution concentrations in the United States continue to decrease, one important and politically charged question remains: Is long-term exposure to low levels of air pollution still harmful? Several approaches have been developed to estimate the relationship between exposure to particulate matter with diameter at most 2.5 micrometers (PM2.5) and various health outcomes. However, none of these approaches account for the fact that different variables might act as confounders of the exposure response relationship at different exposure levels.
In chapter 1, we developed a Bayesian methodology for the estimation of the causal exposure-response curve for exposure to PM2.5 on cardiovascular hospitalizations. This method allows for flexible estimation of the shape of the exposure-response relationship, and for differential confounding adjustment at different levels of the exposure. Moreover, it provides a principled way to identify the confounding importance of different predictors at different exposure levels.
Over the last few decades, there have been various regulations in the Unites States aiming to reduce emissions from power plants with the ultimate goal of reducing ambient air pollution concentrations and pollution-related hospitalizations. However, the effectiveness of these regulations has not been adequately studied. Since nitric oxide and nitrogen dioxides (NOx) are important precursors of ozone formation, we focused on the comparative effectiveness of a class of NOx emission reduction technologies against alternatives on ambient ozone concentrations.
In chapter 2, we developed causal inference methodology rooted in propensity score matching to adjust for unobserved spatial confounding, such as unmeasured weather and atmospheric conditions. We showed that unobserved confounding by spatial variables is likely to be present, and that incorporating spatial proximity in the matching of treated units to control units returns effect estimates that are more in line with subject-matter knowledge.
In chapter 3, we addressed the issue of interference in the studies of air pollution regulations. The movement of emissions and air pollution leads to interference, since interventions that take place at one power plant can affect air pollution levels in the area surrounding other power plants. In more detail, assuming that the power plants can be clustered in groups within which there is interference but not across them, we defined new estimands for causal inference with interfering units that correspond to quantities of interest under realistic treatment allocation programs. Consistent estimators and asymptotic results were derived and were employed to quantify the comparative effectiveness of NOx emission control technologies.
Citable link to this pagehttp://nrs.harvard.edu/urn-3:HUL.InstRepos:41128507
- FAS Theses and Dissertations