A1298
Title: Distributed instrumental variable analysis in three UK studies
Authors: Yanchun Bao - University of Essex (United Kingdom) [presenting]
Hongsheng Dai - University of Essex (United Kingdom)
Wei Liang - Xiamen University (China)
Abstract: In observational data analysis, the instrumental variable analysis is a popular approach to obtain the causal effects of a risk factor (exposure) X on an outcome (response) Y. An instrument variable G is correlated with the exposure X but not correlated to the confounder U or directly correlated to the response Y. This is also called Mendelian randomization when genetic variants are used as the instrument variable to examine the causal effect of a modifiable exposure on a particular disease in an observational study. Merging results from different data sources is challenging for such studies based on instrument variables. One reason is that the data privacy barriers do not allow data from different studies to be transferred and stored for centralized analysis. Distributed analyses have to be implemented. However, the heterogeneity of the instrument variable G indifferent studies makes it very challenging to combine results from different data sources into a final conclusion using a naive meta-analysis approach or divide-and-conquer approach. A novel distributed analysis with instrumental genetic variables is developed to overcome the heterogeneity of the instrument variables in different studies. Simulation and data applied in three UK studies have been used to illustrate the proposed method.
