Magnetic resonance angiography (MRA) for determining the degree of carotid stenosis prior to carotid endarterectomy is attractive because it does not have the high morbidity associated with conventional intra-arterial angiography. We assessed the interobserver variability in the estimation of the degree of stenosis amongst observers of different experience. In a prospective study, consecutive patients with transient ischaemic attacks and symptomatic tight carotid stenosis shown by Doppler ultrasound underwent conventional intra-arterial angiography and 2-D and 3-D time-of-flight MRA of the carotid bifurcations. The films of the processed MRA images were reviewed blind to other clinical and imaging data by eight observers of different levels of experience and coded for the presence and degree of stenosis. The stenosis on intra-arterial angiography was used as the reference standard. There was considerable variability between observers for estimation of the degree of stenosis on MRA. The observers' accuracy ranged from 41% (student) to 79% (experienced). From these estimations, excluding those of the student, it could be seen that up to 23% of patients who would have had an endarterectomy based on MRA should not have done so according to angiography, and up to 33% of patients who should have had an endarterectomy according to angiography would have inappropriately missed having an endarterectomy based on MRA results. Observer reliability of MRA processed images is not good, even in experienced hands. Using the source images and views of the circle of Willis might improve the accuracy, but a further study is required to assess this.