DeutschClear Cookie - decide language by browser settings
Deffner, Veronika (2016): Exposure modeling and exposure measurement error correction in health outcome models with longitudinal data structure: Exposure to particulate matter. Dissertation, LMU München: Faculty of Mathematics, Computer Science and Statistics



The human organism is permanently exposed to various environmental factors, which influence its performance, e.g. climatic conditions, nutrition intake behavior or air quality. This thesis focuses on the human exposure to particulate matter. The “Augsburger Umweltstudie” (2007–2008) was conducted at the KORA study center in Augsburg and aimed at the investigation of the association between particulate matter and human health. An accompanying validation study was conducted in 2011 in order to collect information about the errors in the measurement of particulate matter. The complexity of these errors arises from different data sources: 1.) The usage of population-specific exposure measurements, e.g. from one or several fixed site measurement stations, instead of personal measurements involves a Berkson type error. The deviations between personal and populations–specific measurements are driven by the microenvironment of the person and the climatic conditions. In the first part of the work, a two–level exposure model is developed for the association between the fixed–site and the personal exposure measurements of the validation study including the selection of relevant covariates and the appropriate consideration of categorical covariates in the analysis of longitudinal data. 2.) The mobile devices used for personal exposure measurements exhibit classical measurement error, which is partially device–specific and autocorrelated. 3.) In order to use as much information as possible missing personal exposure measurements are filled in with population–specific exposure measurements resulting in a mixture of Berkson and classical measurement error. The second part of the work aims at the development and application of methods to include knowledge about Berkson, classical and mixture error into regression models of the health outcome. Therefore, the method–of–moments is extended to longitudinal data and to the different types of errors with individual–specific and autocorrelated structures. Validation studies and expert knowledge provide information about the size of the measurement error, but prior knowledge is often afflicted with uncertainty. Approaches for the adequate inclusion of prior knowledge about the measurement errors in the Bayesian health outcome model are evaluated in the third part of the thesis. The role of prior knowledge in regression models with an error–prone covariate differs from conventional Bayesian regression models and is strongly affected by the interaction between the parameters in the model. The thesis is closed with the application of the developed method–of–moments and Bayesian approach to the Augsburger Umweltstudie by integrating information from the validation studies.