In this article multilevel analysis (MLA) is used in order to analyse a large data set on general practitioners' (GPs') referrals to specialists in internal medicine. The hypothesis is tested that patients' morbidity is the main determinant of GPs' referral rates. From a Dutch survey among 161 GPs, referrals to specialists in internal medicine were selected. The results corroborate the main hypothesis: approximately half of the variation in the overall chance of being referred is associated with patient characteristics and approximately 45% with patients' morbidity measured in terms of the GPs' diagnosis. Referral chances differ mainly with diagnosis, which implies that the variation in the referral rates reflects a variation in case mix rather than differences in the quality of care or efficiency. The only doctor or practice characteristics that affect referral chances are the number of instruments available in a practice (more instruments, fewer referrals) and GPs' list sizes (larger lists, more referrals). I n the UK and The Netherlands, where patients do not have direct access to specialized types of care, the variation in general practitioners' (GPs') referral rates to consultants has been studied for many years. This interest in variation is related to 2 concerns. Firstly, referrals to secondary care trigger the use of expensive resources and, secondly, variation in referral rates might reflect differences in the quality of care, in the sense that some doctors may refer too few patients or that others refer too many.'' The variation in the number of referrals in relation to the number of enlisted patients or the number of doctorpatient encounters has been analysed in terms of doctor, patient or practice characteristics.3 This approach has not been very fruitful explaining variation.4'5 Wilkin and Smith5 therefore, argued that more sophisticated models of referral decisions should be developed. It is argued by other authors too1'3'6'7 that referral is a complex process, involving the interaction of both social and psychological factors. For that reason it is seen as more appropriate to explore the process of decision making, e.g. through indepth interviews,8 or by examining small numbers of actual consultations or referrals.3'6 Without questioning this it should be pointed out that a lack of this type of model may not have been the only reason why previous studies have failed to come up with explanations of variation in referral rates. For one thing, appropriate data sets for analysing the variation in referral rates are rare. One needs morbidity-specific information