HLA and disease: relative-risk regression methods and multiple testing considerations.

This paper discusses certain statistical issues arising in an analysis to relate HLA-B antigens to the incidence of graft-versus-host disease in a clinical study. The Cox regression method is used to develop a global test of the hypothesis of no association and to produce estimated relative-risk factors corresponding to the presence of each particular allele. Multiple-testing considerations are central to any attempt to identify pairs of alleles having significantly different relative-risk factors. A number of simultaneous-testing approaches are considered including cumulative P-value plotting and a direct simulation of the null distribution of the maximum pairwise ratio.