Abstract:
|
Transcriptome-wide association studies (TWAS, or PrediXcan) have been increasingly used to identify causal genes by integrating GWAS with eQTL data. The basic methodology underlying TWAS (and Mendelian randomization, MR) is the (two-sample) two-stage least squares (2SLS) instrumental variables regression for causal inference, which imposes strong assumptions on the SNPs to be valid instrumental variables (IVs). These assumptions are most likely to be violated in practice, e.g. due to widespread horizontal pleiotropy of the SNPs. We first consider some simple and powerful methods to detect invalid IVs/SNPs, then propose more robust and more powerful methods than existing ones for causal inference in the presence of invalid IVs.
|