but much improved Type 1 error rates in Scenario 3. This is because the invalid variants in this scenario are stronger than the valid variants, and so receive more weight in the weighted analyses. It is unclear that this would occur in applied practice, although it suggests that the simple median method is a worthwhile additional sensitivity analysis. The inverse‐standard error weighted median estimator has improved Type 1 error properties compared with the IVW median estimator particularly in Scenario 3 (although not uniformly), indicating the potential to use different choices of weights in median‐based methods to provide further sensitivity analyses.
