Serial correlation (SC) of noise inflates T-statistics in simultaneous multi-slice excitation (SMS) fMRI studies with short repetition times (TR<2s). The SC can be corrected using noise pre-whitening methods based on the high-order autoregressive (AR) model. This study aims to determine the optimal order selection (OS) method of AR model to achieve the best SC correction accuracy. By evaluating the false positive characteristics in rest/null datasets, our study showed that the corrected Akaike information criterion (AICc) has the best performance among the OS criteria. We recommend use the AR model with AICc to correct the SC in SMS fMRI experiments.