Selection of Valid Instruments in Mendelian Randomization

• Activity Number: 669
• Type: Topic Contributed
• Date/Time: Thursday, August 8, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Hyunseung Kang*+ and Tony Cai and Dylan S Small
• Companies: Wharton School, University of Pennsylvania and University of Pennsylvania and University of Pennsylvania
• Keywords: Mendelian Randomization ; Instrumental Variables ; Genetic Epidemiology ; L1 Penalization ; Variable Selection

Abstract

With the exponential growth of genomic data, Mendelian randomization (MR) has recently gained attraction in genetic epidemiology as a method to elucidate a causal relationship using genes as instruments in an instrumental variables (IV) regression. The basic idea of MR is to extract variation in an exposure that is due to a Mendelian gene which is independent of confounders and use this confounder-free variation to estimate the effect of the exposure on the outcome. However, one of the major challenges in MR, like in many IV regressions, is choosing valid instruments; a valid instrument needs to be (i) independent of confounders, (ii) have no direct effect on the outcome, (iii) and is correlated to the exposure. Pleiotropy and linkage disequilibrium, common in MR, are conditions which can make genes invalid instruments

In this work, we present results concerning the selection of valid instruments. We provide conditions for identification in the presence of invalid instruments and we propose an estimation strategy based on penalized L1 methods popular in high dimensional inference. We demonstrate the method on simulated data.