This paper investigates an AR (Autoregressive)-filtered version of several conventional diagnostic tests for cross sectional dependence in mixed panels, including the adjusted LM test, the CD test by Pesaran (2008, 2015) and the Schott test by Schott (2005). We show that the revised tests asymptotically follow the standard normal distribution. The distinctive feature of these new tests is its simplicity to implement even though the exact time series property of each component of a panel is unknown or unobservable or the panel is large with mixed time series properties. Simulations show that the AR-filtered version of the CD test perform the best over other testing procedures in term of the finite sample performance and the computation time, especially for those cases with large cross-sectional dimension N (100 and 500). We also provide a new perspective to the linkages between asset returns of banking and non-banking firms using our novel AR-filtered CD statistics as it could behave like an early warning indicator of systemic risk and crises.