The false discovery rate (FDR, Benjamini and Hochberg 1995) is a powerful approach to multiple testing. However, the original approach developed by Benjamini and Hochberg (1995) applies only to independent tests. Yekutieli (2008) showed that a modification of the Benjamini-Hochberg (BH) approach can be used in the presence of dependent tests and labelled his procedure as separate subsets BH (ssBH). However, Yekutieli (2008) left the practical specification of the subsets of p values largely unresolved. In this paper we propose a modification of the ssBH procedure based on a selection of the subsets that guarantees that the dependence properties needed to control the FDR are satisfied. We label this new procedure as the separate pairs BH (spBH). An extensive Monte Carlo analysis is presented that compares the properties of the BH and spBH procedures.