Mortality during hospital care of AIDS-associated Pneumocystis carinii pneumonia (PCP) varies as much as 3-fold among reporting institutions. Prognostic factors for death during an episode of PCP have been identified that may be useful risk adjustors for quality-of-care studies. We illustrate a risk adjustment approach to evaluating institutional performance for PCP using a pilot data set from two Southern California hospitals differing widely in crude PCP death rates (61 vs 27%). Using admission AaDO2, hemoglobin and age in a logistic prediction model for hospital death, we found that outcomes in 90% of cases could be accurately classified. Nearly all of the "excess" mortality of the poor outcome hospital could be explained by greater pulmonary severity on admission. We discuss four conceptual issues in design of AIDS quality-of-care studies: confounding by therapeutic intention, defining relevant treatment components, determining the range of co-morbidity, and truncation of the episode of care.