Table 1.

Comparisons among different task-based fMRI analysis methods.

Methods	Purposes	Strengths	Limitations
General linear model	Estimating to what extent each known predictor contributes to the variability observed in the voxel's BOLD signal time course	(i) Mathematically simple, easily interpreted, and readily available in standard packages (e.g., the SPM software) (ii) Flexible to incorporate multiple quantitative and qualitative independent variables, such as low- frequency drifts and head motion	(i) Relies on assumptions such as appropriate repressors in the matrix and normality of the fMRI noise which are difficult to check
Psychophysiological interaction	“Searching” for regions that correlate differently with a particular region under certain experimental context	(i) Can explore the connectivity of the source area to the rest of the brain and how it interacts with the psychological variables	(i) Only involves one region of interest in one model (ii) Limited in the extent to which you can infer a causal relationship
Structural equation model	Estimating the degree to which the activity between different brain regions is connected and how this connectivity is affected by an experimental variable	(i) Can examine interactions of several regions of interest simultaneously and offer estimations of causal relationships (ii) Predetermined connections are based on prior anatomical or functional knowledge	(i) Causality is predetermined, and this might overlook several aspects of neural activity (ii) Assumes the interactions are linear (i.e., structural equation models are not time-series) (iii) Lacks temporal information
Dynamic causal model	Estimating and making inferences about the coupling among brain regions and how this coupling is affected by changes in experimental context at the neuronal level	(i) Biologically more accurate and realistic than other methods because DCM models interactions at the neuronal rather than the hemodynamic level and complex connectivity patterns between regions can be arbitrarily postulated	(i) Prespecified models are needed (ii) Requires much longer time to estimate parameters than SEM (iii) Neurodynamics in each region are characterized by a single state variable (“neuronal activity”)
Granger causality model	Measuring the predictability of one neural time-series from another	(i) No a priori specification of a model is needed. Thus this model can complement the hypothesis-driven methods and help to form directed graph models of regions and their interactions	(i) The causal relationship may be caused by the differences in hemodynamic latencies in different parts of the brain if long repetition times (TR) are used
Multivoxel pattern analysis	Applying pattern- classification algorithms to demonstrate the relationship between measures of brain activity and a perceptual state and provide an information-theoretic framework for the isolation of regions that uniquely represent a behavior	(i) Simultaneously examines the disparate signals carried within a set of voxels rather than examining individual voxels in parallel (ii) Can decode more complex information due to improved sensitivity and use of spatial information	(i) The possibility of overfitting increases as the classifier becomes more complex, which may result in poor performance in tests of generalization