Residual exposure stratification subtracts the genetically predicted component of an exposure before forming exposurestrata. Under a suitable additive first-stage model, this avoids directly conditioning on the component caused by the instrument and can reduce the resulting collider bias.