A cautionary tale about control variables in IV estimation