When a large number of genes are significant in correlating microarray gene expression data with patient prognosis, clustering of significant genes may be effective not only for further dimension reduction but also for identifying co-regulated genes that belong to the same molecular pathway related to disease biology and aggressiveness. Moreover, a reduced feature, such as the average expression across samples for a cluster of significant genes, can play an important role in reducing variance in prediction analysis. We propose a simple procedure to select gene clusters that have strong marginal association with survival outcome from a large pool of candidate hierarchical clusters of significant genes. Selected gene clusters can have better predictive capability than the other gene clusters and singleton genes. Application of such clustering to the data set from a clinical study for patients with multiple myeloma and associated microarrays is given. Copyright © 2007 John Wiley & Sons, Ltd.