Statistics Seminar - Mengjie Bian - : Bootstrap Inference in Mendelian Randomization
Speaker: Mengjie Bian (McMaster University)
Title: :Bootstrap Inference in Mendelian Randomization
Abstract: Mendelian randomization studies (MR) have been become increasingly common in health-related research. Using genetic variants as instrument variables (IV), Mendelian randomization provides a way to investigate the causal relationship between exposure and outcome. Here we consider two-stage least squares to estimate the causal effect of exposure on the outcome. It has been known in the fields of economics that the two-stage least squares estimator can perform poorly in finite sample when the instruments are weak. However, it is poorly understood in the epidemiology field. We consider the bias when the instruments are weakly correlated with risk factor. Techniques for obtaining the confidence intervals are presented, especially using bootstrapping method. We particularly consider two types of bootstrap methods, which are "raw bootstrap"(resampling the original data) and model-based bootstrap (resampling residuals). We also consider Wald and bootstrap-based methods for hypothesis testing.