Chris
Oldnall
Estimating causal effects in the instrumental variable framework, without ex-ante knowledge of valid instruments
Abstract.
In the field of Biology, the use of instrumental variables has taken precedence for estimating causal relationships between exposures and outcomes, in the form of Mendelian randomisation. This framework assumes the instrument is a genetic variant, which is essentially inherited at birth and remains unchanged throughout one's life. The framework however also stipulates that the instrument is only valid if it acts on the outcome only through the exposure of interest — a breach of this situation makes the estimation null, and in the biological sphere is referred to as pleiotropy. Many attempts have been made to identify ‘pleiotropic’ instruments and remove them to perform valid estimation, however with no mathematically testable criteria, this procedure is not robust. This work instead looks at the approach by Sun et al.¹ in developing a solution which leverages potentially invalid instruments with valid ones in the form of G-estimators with unconditional moment conditions. This approach proves promising, but is limited by scalability leading to trade-offs between biological accuracy and computational ability. This talk demonstrates the severity of pleiotropic rich systems, and will discuss the computational limitations and trade-offs that one has to make to employ G-estimation in the Mendelian randomisation application.