As genome-wide association studies diversify beyond European ancestry, Mendelian randomization (MR) is increasingly applied to infer causal relationships in under-represented populations. Yet the field lacks clear methodological guidance for handling ancestry-driven discordance in SNP effect sizes, raising concerns about instrument validity and bias when European-derived variants are used in non-European settings. Existing trans-ethnic MR approaches typically focus on genome-wide polygenic overlap to increase power but offer limited insight into how per-variant effect-size mismatch influences causal effect estimation in the target-ancestry.
We propose an alternative framing that treats ancestry mismatch as a weak-instrument problem. Here, reduced cross-ancestry effect concordance directly attenuates SNP–exposure associations in the target population, inducing predictable weak-instrument bias when European instruments are transported across ancestries. We developed a two-phase simulation framework that fixes the true causal effect in the target population (e.g., East Asian exposure and outcome) and evaluates how well existing MR methods recover this effect from EUR-derived SNPs under varying cross-ancestry genetic correlations, trait heritability, allele-frequency differences, LD discrepancies, and directional pleiotropy. We further compared alternative instrument strategies, including direct EUR-derived instruments, target-ancestry instruments, and hybrid approaches that replace EUR effect estimates with target-ancestry effect sizes.
In our simulations, inverse-variance weighted (IVW) estimates were highly sensitive to MAF differences, LD-clumping thresholds, and the extent of cross-ancestry effect-size discordance. In contrast, median-based estimators consistently recovered the target-ancestry causal effect, even when individual SNPs showed substantial heterogeneity across ancestries. Empirical analyses in East Asian cohorts demonstrated similar behaviour: For BMI on type 2 diabetes, the EAS-based causal estimate (OR=2.10 [95% CI: 1.59–2.76] per SD increase) closely matched the median-based estimate using EUR instruments (OR=2.12 [1.80–2.49]), whereas EUR-derived IVW estimates (OR 1.60–2.86) varied markedly with instrument selection. By conceptualising ancestry-dependent effect heterogeneity as a weak-instrument phenomenon, our findings offer practical guidance for designing reliable MR analyses in diverse populations.