Statistical Inference for Novel Study Designs

点击翻译按钮获取中文摘要

Modern research studies call for modern statistical methodologies. A study design is the set of methods and procedures used in collecting data for a research problem. To investigate causal relationships, modern researchers in biomedical and social sciences use increasingly complex study designs to reduce confounding bias, increase statistical efficiency, and/or to meet logistical or ethical constraints. Research data collected in these novel study designs are often being analysed by model-based approaches, which assume the data are generated from certain statistical models. As a consequence, the results of such studies can be very sensitive to the modelling assumptions.Recent developments have provided a suite of methods to elucidate causal inference using experimental and observational data. They have been successfully applied to understand classical study designs such as completely randomised experiments, cross-sectional cohort studies, and instrumental variables. However, there is still a significant gap between the theory of causal inference and the practice of modern research studies involving complex processes of intervention and data collection. Many valuable datasets are still being analysed by inferior, error-prone methods, which is a fundamental issue for the ongoing replication crisis in several scientific disciplines.This project will investigate statistical inference in two novel study designs with exponentially increasing popularity. The first design is within-family Mendelian randomisation. Mendelian randomisation is a popular method in epidemiology and genetic that uses genetic variants as instrumental variables. Intuitively, genetic variants are randomised during conception and should be independent of any confounders. However, this is not true without controlling for family effects. This project will provide a rigorous justification of how the causal inference in Mendelian randomisation can be based precisely on the randomisation in meiosis (the division of germ cells) and propose new methods that use the family structure in the data. The new methods will be applied to large genetic datasets to examine recent findings in health research.The second design investigated in this project is the stepped wedge trials. In this design, the trial participants gradually cross over from the control condition to the treatment, and the cross-over times are randomised. Currently, stepped wedge trials are routinely analysed by linear mixed-effect models, but recent works have demonstrated that this is extremely sensitive to the modelling assumptions. This project will develop new methods that are exactly based on randomisation. This will greatly enhance the robustness of stepped wedge trials.