After the implementation of the Medicare Prospective Payment System (PPS) in the USA, many European countries like France have introduced DRGs to curb hospital overspending. However, there has been some reluctance from hospital actors, especially because of the heterogeneous nature of DRG's. To analyse this situation, we propose a method based on distribution modelization of length of stays and costs within DRGs. For each DRG, the model is based on a mixture of Poisson and Weibull distributions identified as subgroups. The subgroups are characterized by their means and their proportions which are estimated by maximization of data likelihood. For a particular DRG, the proportion of long stay or high-cost patients can be explained by the introduction of clinical variables in the model. First the model was applied to the DRG "leukemia and lymphoma" (HCFA V.3), using 133 discharge abstract files from the Dijon public teaching hospital which were classified into this DRG in 1993. Among the studies parameters only acute leukemia, neutropenia < 500 PNN/mm3, high dose aplastic chemotherapy, central venous catheterization, parenteral nutrition, use of protected or laminar air flow room, septicemia, large spectrum intravenous antibiotherapy, and blood transfusion had a significant influence on the distribution of the patients in the long stay or costly subgroup. Second, for DRG "chronic bronchopneumopathies" (n = 220) the significant parameters were mechanical ventilation, antibiotherapy, post hospitalization medicalized care.