Constrained Likelihood Inference in Instrumental Variable Regression with Invalid Instruments and Its Application to GWAS Summary Data

Abstract

There has been increasing interest in instrumental variables regression for causal inference. In genetics, transcriptome-wide association studies (TWAS), also known as PrediXcan, have recently emerged as a widely applied tool to discover causal/target genes by integrating an outcome GWAS dataset with another gene expression/ transcriptome GWAS (called eQTL) dataset; they can not only boost statistical power but also offer biological insights by identifying (putative) causal genes for a GWAS trait, e.g. low-density lipoprotein cholesterol (LDL). Statistically TWAS apply (two-sample) two-stage least squares (2SLS) with multiple correlated SNPs as instrumental variables (IVs) to predict/impute gene expression, in contrast to typical (two-sample) Mendelian randomization (MR) approaches using independent SNPs as IVs, which are expected to be lower-powered. However, some of the SNPs used may not be valid IVs as a result of their (horizontal) pleiotropic/direct effects on the trait not mediated through the gene of interest, leading to false conclusions by TWAS (or MR). We propose a general inferential method for possibly high-dimensional data to account for confounding and invalid IVs while selecting valid IVs simultaneously via two-stage constrained maximum likelihood; we develop a theory for the likelihood method subject to a truncated L1-constraint approximating the L0-constraint for asymptotically valid and efficient statistical inference on causal effects. We demonstrate both theoretically and numerically the superior performance of the proposed method over the standard 2SLS/TWAS and other methods. We apply the methods to identify causal genes for LDL by integrating GWAS summary data with eQTL data.