There has been controversy about whether confounding by indication for treatment--that is, owing to physicians' conscious efforts to base treatment decisions on patients' pretreatment prognoses--makes nonrandomized, observational comparisons of treatments invalid. Some now believe evidence from studies of practice variation means that physicians' treatment decisions have little relationship to patients' prognostic clinical characteristics. They therefore believe that patients who receive different treatments should vary little in their baseline prognoses, and multivariable statistical methods should easily be able to adjust for any resultant confounding, even when analyses are restricted to administrative rather than clinical data. The objective of this study is to determine whether adjusting for variables found in administrative data sets produces the same results as does adjusting for clinical variables. Data were reanalyzed from a previously enrolled prospective sequential cohort of 227 hospitalized patients with suspected bacteremia who had blood cultures. The treatment under study was aminoglycoside therapy given empirically, that is, before blood culture results were known. The outcome of interest was death during hospitalization. Univariable analyses suggest that empiric aminoglycoside therapy had a positive association with mortality, by univarible logistic regression, odds ratio (OR = 3.1 (95% confidence interval = [1.6, 5.8]). Few administrative variables had univariable associations with aminoglycoside use or death. Multivariable analyses that controlled for them still suggest that aminoglycosides increased mortality; for example, in one model, adjusted OR = 3.2 (1.6, 6.5). Many clinical variables were strongly associated with aminoglycoside use or death. Analyses that controlled for them suggested that empiric aminoglycosides did not increase mortality; for example, in one model, adjusted OR = 1.2 (0.55, 2.7.) Results of adjustment for confounding using administrative data disagreed with the results of adjustment using clinical data. It is concluded that nonrandomized, observational outcome studies that fail to control for prognostic differences between patients receiving different treatments may not always be valid.