Both fundamental and practical differences between two common approaches to fault detection and isolation are examined. One approach is based on causal state-variable or parity-relation models developed from theory or identified from plant test data. The faults are then detected and isolated using structural or directional residuals from these models. The multivariate statistical process control approaches are based on noncausal models built from historical process data using multivariate latent variable methods such as PCA and PLS. The faults are then detected by referencing future data against these covariance models, and isolation is attempted through examining contributions to the breakdown of the covariance structure. There are major differences between these approaches arising mainly from the different types of models and data utilized to build them. Each of them has clear, but complementary, strengths and weaknesses. These are discussed using simulated data from a CSTR process.