【摘要】Multiple endpoints are often naturally clustered based on their scienti c functions. Tests that compare these clustered outcomes between independent groups may lose eciency if the cluster structures are not properly accounted for. We propose a cluster-adjusted multivariate test procedure for the comparison and demonstrate its gain of eciency over a number of popular test procedures that ignore the clusters. Data from a dietary intervention trial and a case-control study on neural tube defects are used to exemplify the methods