In this paper the concepts of term clusters and cluster coefficients are defined and used in the context of system identification. It is argued that if a certain type of term in a nonlinear model is spurious, the respective cluster coefficient is small compared with the coefficients of the other clusters represented in the model. Once the spurious clusters have been detected, the corresponding terms can be deleted from the set of candidate terms. The consequences of doing this are (i) a drastic reduction in the size of the set of candidate terms and, consequently, a substantial gain in computation time is achieved; (ii) the final estimated model is more likely to reproduce the dynamics of the original system; and (iii) the final model is more robust to overparametrization. Numerical examples are included to illustrate the new procedure.