In this paper, a novel statistical tool, stochastic context-free models (SCFMs), is introduced to model and analyze brain voxel activation in fMRI time series. SCFMs characterize the dynamic process where Blood-oxygen-level dependent (BOLD) responses are assumed to be driven by brain voxel activation in pre-designed experiments. Classical state space methods such as hidden Markov models(HMMs) make strong Markov assumptions on states behaviors. Whereas, in SCFMs, more powerful context-free grammar rules are used to model such behaviors in accordance to paradigm design. The methodologies of evaluation, inference, and decoding based on SCFMs are presented. Experimental results using both HMMs and SCFMs show that the later models can better capture the completeness of the target activation patterns, and encapsulate more hierarchical information in the resulting probabilistic parsing tree.