Abstract
Data‐driven statistical methods are useful for examining the spatial organization of human brain function. Cluster analysis is one approach that aims to identify spatial classifications of temporal brain activity profiles. Numerous clustering algorithms are available, and no one method is optimal for all areas of application because an algorithm's performance depends on specific characteristics of the data. K‐means and fuzzy clustering are popular for neuroimaging analyses, and select hierarchical procedures also appear in the literature. It is unclear which clustering methods perform best for neuroimaging data. We conduct a simulation study, based on PET neuroimaging data, to evaluate the performances of several clustering algorithms, including a new procedure that builds on the __k__th nearest neighbor method. We also examine three stopping rules that assist in determining the optimal number of clusters. Five hierarchical clustering algorithms perform best in our study, some of which are new to neuroimaging analyses, with Ward's and the beta‐flexible methods exhibiting the strongest performances. Furthermore, Ward's and the beta‐flexible methods yield the best performances for noisy data, and the popular K‐means and fuzzy clustering procedures also perform reasonably well. The stopping rules also exhibit good performances for the top five clustering algorithms, and the pseudo‐T^2^ and pseudo‐F stopping rules are superior for noisy data. Based on our simulations for both noisy and unscaled PET neuroimaging data, we recommend the combined use of the pseudo‐F or pseudo‐T^2^ stopping rule along with either Ward's or the beta‐flexible clustering algorithm. Hum Brain Mapp 23:109–119, 2004. © 2004 Wiley‐Liss, Inc.