In this paper, we show that conventional heteroskedasticity and autocorrelation robust inference procedures based on the reduced form provide tests and confidence intervals for structural parameters that are valid when instruments are strongly or weakly correlated to the endogenous variables.