= Solution
Three useful extensions are:
* Replicate the bidirectional <Mendelian randomization> in independent, larger cohorts and ancestries, using nonoverlapping discovery and outcome samples to improve precision and reduce winner's-curse and population-specific artifacts.
* Apply pleiotropy-robust analyses such as weighted-median, mode-based, and MR-Egger estimates, inspect heterogeneity and leave-one-variant-out results, and use biologically distinct instrument sets. Agreement despite different sensitivities to invalid instruments strengthens the exclusion restriction.
* Triangulate with a design based on an independent source of causal variation, such as compulsory-schooling-law changes or another natural experiment, and with longitudinal refractive measurements. Agreement across designs with different biases is stronger evidence than enlarging one genetic analysis alone.
Back to article page
