The Largest Normalized Residual Test (LNRT) has been widely utilized in commercial Power System State Estimation (PSSE) software for bad data identification. The LNRT has proved effective in dealing with single bad data as well as multiple non-interacting and multiple interacting but non-conforming bad data. However, it is known for a long time that when two bad data are both interacting and conforming, i.e. their errors are in agreement, the LNRT may fail to identify either one. Moreover, it has been shown recently that even two interacting and non-conforming bad data can cause the failure of the LNRT. Drawing on sensitivity analysis in linear regression, we develop normalized deleted residuals for suspected measurements so that the agreement in measurement errors are broken. Therefore, the LNRT for normalized deleted residuals will be able to identify the actual bad data point. Furthermore, in the case of AC PSSE, the method does not require calculation of a new hat matrix when a measurement is deleted from the data set. This makes the method computationally cost-effective. Simulation results for identifying different conforming and non-conforming interacting bad data proves that the proposed method can enhance the effectiveness of the LNRT.