We analysed data from 936 newly-diagnosed patients with advanced, aggressive non-Hodgkin's lymphoma (NHL) treated in three randomised European Organisation for Research and Treatment of Cancer (EORTC) trials performed between 1980 and 1999 (median follow-up of 8.7 (0.2-20.4) years). The CHOP-like regimen CHVmP/BV (cyclophosphamide, doxorubicin, teniposide and prednisone with bleomycin and vincristine at mid-interval), was compared with CHVmP (CHVmP/BV without bleomycin and vincristine), ProMACE-MOPP (methotrexate, doxorubicin, cyclophosphamide, etoposide, mechlorethamide, vincristine, procarbazine and prednisone) and CHVmp/BV with additional, autologous stem-cell transplantation, respectively. Overall, treatment with CHVmP/BV resulted in a better long-term outcome with 63% complete responses being observed and an overall survival (OS) of 59 and 43% at 5 and 10 years, respectively. Remarkably, OS after CHVmP/BV improved across the trials, even after stratifying for the International Prognostic Index (IPI). This finding could not be directly related to better salvage treatments during the last decade. Selection bias appears to be responsible: stepwise corrections for small differences in inclusion criteria eliminated the difference in OS, especially when histological subgroups were studied. This systemic review underlines the difficulties encountered in retrospective sub-set analyses and the biases that can be introduced when recent studies are compared with older ones.