Jointly modelling multiple ancestral populations using GWAS summary data improves causal inference

Population diversity contributing towards genome wide association studies is increasing, which has invited the expansion of causal inference through Mendelian randomization (MR) to multiple populations. Here we show that multi-ancestry MR can easily lead to causal effect estimates being inconsistent between ancestral groups if not performed appropriately. We introduce an analytical framework to manage these issues. It also handles disparities in statistical power for instrument identification between ancestries, and leverages instances of population-specific instrument effects to model horizontal pleiotropy, helping to examine a core MR assumption. We apply our framework to a series of cardiometabolic and behavioural exposures and outcome trait pairs combining European and East Asian samples, which indicate that SNP-exposure, exposure-outcome and horizontal pleiotropic effects tend to be relatively consistent across these ancestral groups, though with some notable apparent discrepancies which are likely driven by gene by gene or environment interactions. We also examine the relationship between LDL cholesterol and stroke across a broader set of ancestral groups which again, after accounting for biasing mechanisms, illustrates overall consistency of effects across ancestral groups. Our results suggest differential health outcomes are more likely driven by differential distributions of risk factors than mechanisms that would change susceptibility to risk factors.