Association indices were originally developed to describe species co-occurrences, but have been used increasingly to measure associations between individuals. However, no statistical method has been published that allows one to test the extent to which the observed association index values differ from those of a randomly associating population. Here, we describe an adaptation of a test developed by Manly (1995, Ecology, 76, 1109-1115), which uses the observed association data as a basis for a computer-generated randomization. The observed pattern of association is tested against a randomly created one while retaining important features of the original data, for example group size and sighting frequency. We applied this new method to test four data sets of associations from two populations of Hector's dolphin, Cephalorhynchus hectori, using the Half-Weight Index (HWI) as an example of a measure of association. The test demonstrated that populations with similar median HWI values showed clear differences in association patterns, that is, some were associating nonrandomly whereas others were not. These results highlight the benefits of using this new testing method in order to validate the analysis of association indices. Copyright 1998 The Association for the Study of Animal Behaviour.