Résumé : In an effort to control escalating health expenditures, especially in hospitals, many countries are planning or experimenting with prospective budgeting systems. Belgium is no exception and has recently introduced, with some success, limited fixed charges per hospital admission and/or per hospitalisation day for laboratory tests and radiographic investigations. More recently, the focus has shifted to hospital drug expenditures, which have shown high growth rates over the past few years. Until now, such expenditures have been reimbursed on a fee-for-service system, often with limited out-of-pocket charges for hospitalised patients. In order to curb the growth of drug expenditures, it is appropriate to investigate whether the financing of hospital drugs through a prospective budgeting system could be a feasible solution. Therefore, we constructed a database of over 270,000 admissions from a sample of 23 Belgian general and teaching (university) hospitals for the year 1991. Data were obtained from the official Minimum Basic Data Set or Resume Clinique Minimum, which contains summarised clinical and administrative information, plus detailed expenditures (including medications) for each hospital stay. This information allowed us to categorise each stay into an appropriate diagnosis-related group (DRG). Our first descriptive analysis identified a number of major variables that influenced patients' drug expenditures: all-patient DRG (APDRG), age, disease severity, length of stay in an intensive care unit, emergency admission, death during hospitalisation, and hospital type (teaching or general). A covariance analysis was then performed on all hospital stays combined, and separately on surgical and medical stays. The results indicated that these variables taken together account for between 56.5 and 76.3% of drug expenditures in medical and surgical stays, respectively, with the major variance explained by differences in APDRG category. However, when the data were disaggregated according to major diagnosis category, a large degree of heterogeneity in the explained variance was observed. In patients with drug use- and alcohol-related disorders, 5.2% of drug billings/expenditures were attributable to the APDRG, and the corresponding figure in patients undergoing circulatory system surgery was 84%. This means that, if DRGs are used to define a global prospective drug budget for a hospital, using the hospital's historical case mix as a weighting factor, we should pay particular attention to the hospital profile because the predictive power of such a system could be relatively low in some hospitals. Consequently, we need to construct larger confidence intervals for hospitals in which historical drug expenditures have low predictive power, or search for additional explanatory variables for expenditures in these hospitals.