We previously evaluated radioassays for insulin autoantibodies (IAA), glutamate decarboxylase autoantibodies (GAA) and ICA512bdc autoantibodies (ICA512bdcAA) in the prediction of type I diabetes. Among first degree relatives, the five year risk of diabetes was 0% if no autoantibody was positive, 15% if only one was positive, 44% if two were positive and 100% if all three were positive. We measured IAA, GAA and ICA512bdcAA in 45 patients with new onset type I diabetes (sampled within 7 days of insulin therapy), 882 first degree relatives of patients with type I diabetes, and 217 control subjects. ICA512bdc is a construct of the ICA512/IA2 molecule (amino acid residues 256-979), including the intracellular domain. Based on receiver-operating characteristic plots, there was no significant difference between the three assays in their ability to discriminate between disease and control subjects. The upper limits of normal for the assays were determined as the 99th percentile of levels in the control subjects or higher. Using these cut-offs, 76% of new onset patients were positive for two or more autoantibodies, and 98% were positive for one or more. In comparison, none of 198 control subjects tested for all three assays were positive for more than one autoantibody. In relatives with a single autoantibody, or exactly two autoantibodies, there was no difference in diabetes-free survival according to which one of the autoantibodies was present (P = 0.70, logrank test), or which particular combination of autoantibodies was present (P = 0.56, logrank test) respectively. Our conclusions were as follows: the number of autoantibodies (counting IAA, GAA and ICA512bdcAA) is important in prediction, rather than the particular autoantibody specificities present. Among patients with new onset insulin-dependent diabetes, the absence of any of these autoantibodies justifies the consideration of non-autoimmune forms of diabetes in the differential diagnosis.