Patients (672) admitted to a department of medicine during five consecutive months were followed by an investigator who identified 110 clinical manifestations which could have been considered adverse drug reactions. From these, 42 were excluded because they did not correspond to the definition of adverse reaction or they were inadequately documented. The remaining 68 cases were submitted to three independent observers who had to reply to a series of questions; from these replies five degrees of probability for the reaction itself were deduced. Reactions (54; 49% of the manifestations reported) were considered as certain or probable by at least two observers, but only 27 rections of these (25%) were attributed to the same drug by all three observers. There was a low level of agreement between any two observers (paired agreement ratio: 0.6 to 0.7) and little difference of agreement between any one observer and each of the others (personal agreement ratio: 0.6 to 0.7).