Suppose I have an instrument Z for some endogenous variable X. Z does a much better job predicting X for some X values than others. E.g. major better predicts occupation for nurses than for management consultants. What problems does this introduce? Is there a go-to reference for this issue?