Bayesian causal inference for environmental epidemiology data

点击翻译按钮获取中文摘要

The long term goal of the proposed research program is to develop Bayesian causal inference methods for environmental epidemiology data, with a specific emphasis on mediation analysis and longitudinal data. Bayesian causal inference is a thriving area of innovation in statistics, with important applications in environmental epidemiology. Environmental epidemiology is concerned with estimating the causal effects of environmental exposures (e.g. air pollution on human health), and to distinguish these effects from spurious associations due to confounding or other biases. Bayesian methods are well-suited to environmental epidemiology owing to their natural probability interpretation of uncertainty and their flexible handling of bias and missing data. The vision of the proposed research is to develop novel Bayesian causal inference methods that are motivated by important real-data applications and statistically valid.  The proposed methods will be available to data-analysts in environmental health sciences in particular, but also more generally in other domains including medicine, economics and the social sciences. The proposed research program is divided in three scientific objectives in Bayesian causal inference: 1) To extend my recent work on Bayesian methods for causal mediation analysis to account for missing data, with particular attention to environmental studies with missing biomarkers of environmental exposure, 2) To develop new Bayesian causal inference methods to explore sensitivity to bias from non-ignorable missing covariate and outcome data in longitudinal studies, and 3) To investigate novel methods for Bayesian quantile regression to detect heteroscedasticity and estimate heterogeneity of causal effects. The research will build on my recent progress, which includes developing new causal inference techniques that have been published in high-quality statistics journals and environmental epidemiology journals. A key component of the research will be to study the performance of new Bayesian methods compared to standard frequentist approaches, in terms of coverage probability, bias and average squared error. The methodological approach will include computer simulation experiments, analytical results based on simple models, and a study of performance when applied real datasets. Statistical codes using the software Stan will be made available in order to increase uptake of new methodologies. Trainees will be involved in all aspects of the research and gain knowledge in biostatistical methods applied to unique datasets.