For optimal case mix adjustment, it is necessary to build risk or severity models to fit the disease and its outcome. For cardiac disease, for example, several risk models predict the outcome of mortality. These risk models include variables that, by themselves in multivariable models, predict death. If a variable is associated with death independent of treatment or quality of care, it has to be adjusted for and, therefore, included in the model. Similar models have to be developed and tested for end-stage renal disease. A complete risk model would consist of age, sex, the severity of the primary disease of interest (here, end-stage renal disease), not just the presence but also the severity of all co-morbid disease, and, finally, a measurement of functional status or quality of life. This last measurements is associated with outcome beyond and independent of the above noted severity components. These factors (or "severity" or "patient mix" dimensions) then have to be developed in relation to a specific outcome or response variable. Cholesterol and hypertension are risk factors for long-term events and mortality, but not for 2-year symptoms and quality of life. Thus, the model has to be defined by the type of outcome and by the time interval over which the variable acts. For end-stage renal disease, at least one of the major types of outcome will be quality of life. Our case mix measures for hospital co-morbidity and for office practice chronic disease are defined by and intended to predict functional status in the short run, because that is the outcome of interest for dollars spent and for most of the care rendered to patients with chronic disease before the very end stages. Creating a valid case mix or severity measurement aimed at or defined by quality of life thus involves the creation of new models, the testing of those models, the revision based on the empirical testing, and, finally, generalizing to other sites beyond which the data were collected. Useful case mix measurements will be parsimonious, feasible, reliable, validated against the outcome of interest, shown to work in diverse settings, and acceptable to users, especially clinicians.