In this paper, we investigate the phase information for classification between clench speed and clench force motor imagery for BCI applications. The multivariate extensions of empirical mode decomposition (MEMD) are used to decompose EEG data into intrinsic mode functions (IMFs). Then, the phase information is got by transforming IMFs into analytic signal using Hiblert transforms. Six feature types are compared in the paper for channel C3, Cz and C4: the amplitude of IMFs, the power of IMFs, the amplitude of the corresponding analytic signal, the instantaneous phase of the analytic signal, the instantaneous frequency of the analytic signal and the phase-locking value (PLV) between two channels. The support vector machine with 5-fold crossvalidation is used to classify clench speed motor imagery from clench force motor imagery. The results show that for some subjects the instantaneous phase can get the best results, while PLV never performs best compared with other features. The minimum classification error rate of 0.25 is reached in our research.