From July 1980 to November 1985, 109 patients with acute myelogenous and lymphoblastic leukemia who had reached a complete remission (CR) following induction treatment were assigned to a study comparing marrow transplantation with chemotherapy as a postremission treatment. Sixty-nine patients did not have a human leukocyte antigen (HLA)-identical donor, and therefore served as chemotherapy controls; 40 patients had HLA-identical donors, and therefore were assigned to the transplant arm. Of these, 23 were transplanted in first remission and 17 were not. Ten of these 17 were subsequently transplanted in relapse. Initially, only patients with poor prognosis determined by a predictive model were entered into the study. Subsequently, patients with moderately poor prognosis were admitted. Comparing the chemotherapy group with the patients transplanted in first CR, significant control of leukemia relapse in transplanted patients was seen in the subgroup with acute myelogenous leukemia (AML) (P less than .1), and acute lymphoblastic leukemia (ALL) (P less than .01), in the poor (P = .01) and intermediate subgroup (P = .01), and in the good-prognostic groups (P = .05). The survival was affected significantly in only the poor and intermediate subgroups. The use of predictive models might help to select patients for whom bone marrow transplantation is appropriate in first remission and those for whom bone marrow transplantation can be left as the initial treatment of relapse. Predictive models could further be helpful in comparing studies performed at different transplant centers.