A total of 49 bladder cancer patients with tumor stages pT3b, pT4a and/or pelvic lymph node involvement without microscopic or macroscopic evidence of residual tumor was randomized into 2 comparative groups: the chemotherapy group was to receive 3 adjuvant cycles of methotrexate, vinblastine and cisplatin plus doxorubicin (M-VAC) or epirubicin (M-VEC) after radical cystectomy. The control group received no additional treatment. The protocol was activated in May 1987. Patient recruitment was concluded in December 1990 because an interim analysis of the 49 randomized patients revealed a significant prognostic advantage in favor of 26 patients randomized to the chemotherapy group compared to 23 in the control group (p = 0.0015, log rank test for relapse-free survival curves). Of the 26 patients randomized for adjuvant chemotherapy 18 were treated with M-VAC or M-VEC (3 cycles in 16 patients and 2 cycles in 2). Of the remaining 8 patients 7 refused chemotherapy before or during cycle 1 and 1 received chemotherapy without cisplatin because of impaired renal function. An update of the patients in August 1991 revealed a further increase in the prognostic difference between the 2 trial arms (p = 0.0012). Of 18 patients who received treatment with M-VAC or M-VEC only 3 have had tumor progression to date compared to 18 of 23 patients in the control group. Further statistical analysis of the data was performed on the basis of Cox's regression model, incorporating various criteria as explanatory variables, including patient sex and age, pT stage and number of involved lymph nodes. This multivariate analysis revealed a significant decrease in the risk of tumor recurrence (p = 0.0007, 2-sided) after adjuvant chemotherapy. The number of lymph nodes involved was also of prognostic significance (p = 0.0028, 1-sided). The results indicate that the survival time after radical cystectomy can be prolonged considerably by adjuvant polychemotherapy in cases of locally advanced bladder carcinoma. Fortunately, all of these conclusions are not affected by switching from an intent-to-treat analysis to an analysis of the therapy actually performed. The p values obtained from the latter are 0.0005 (log rank test) and 0.0001 (Cox model with the same set of additional regressors).