In total, 103 cases were randomly selected from the Groningen, Nottingham and South Verona registers. Six raters (two in each of the register areas) were involved in coding these cases according to the ICD–9. In general, interrater agreement was satisfactory when codes were grouped into a limited number of categories. Nevertheless, considerable variation in agreement rates was found. We distinguished three steps in the diagnostic process. The selected logistic model showed that reliability is significantly affected in each step, but only substantial in the first where clinicians formulate a diagnosis.