Eleanor Sanderson1, and Jack Bowden1
1MRC Integrative Epidemiology Unit, Univerity of Bristol.
One of the key assumptions of any Mendelian randomisation (MR) analysis is that the exposures are strongly predicted by the set of SNPs used as instruments. In multivariable MR (MVMR) this assumption requires that each exposure is strongly predicted by the SNPs included conditional on the predicted value of the other exposures in the model. Many of the traits considered in the NMR data here are associated with highly overlapping groups of SNPs and therefore it is particuarly important with data of this type to consider whether these exposures can be reliably predicted by the set of SNPs if multiple traits from this data are going to be included as exposures in an MVMR analysis.
Here I consider whether multiple traits in this dataset can be predicted at the same time. If a group of exposures can all be strongly predicted by the set of SNPs then, assuming the other instrumental variable assumptions are satisfied, it will be possible to estimate the direct effect of each exposure on an outcome. However, if some or all of the exposures are weakly predicted then any MR analysis including those exposures will be subject to weak instrument bias.