The ability of four experienced clinicians to predict short-term outcome (serum creatinine level at 1 year) and long-term outcome (renal insufficiency) was evaluated in 87 patients with lupus nephritis. The correlational agreement and the accuracy of their predictions were contrasted with the actual outcomes observed and with statistically generated prognostic regression models. In contrast to previously published data, all four clinicians predicted both short-term outcomes (P < 0.001) and long-term outcomes (P < 0.02) well. The clinicians' predictions approximated that of a statistically generated computer model for both agreement and accuracy for renal function at 1 year. The four clinicians identified nearly identical clinical variables as important in determining prognosis. Provision of biopsy data to the clinicians improved short-term and long-term prediction slightly. The value of the statistical models was 'validated' by demonstrating that three of the four clinical variables identified by the models, but not by the clinicians, could enhance clinical prediction (P < 0.05). In addition, the extent of tubulo-interstitial involvement on biopsy, a predictor that has recently received increased attention, could improve the long-term clinical predictions of all four clinicians (P < 0.05).