A probabilistic model was developed which allows one to estimate sensitivity and specificity of diagnostic tests for coronary artery disease without reference to angiography. The feasibility of the model was evaluated first in a series of computer simulations, and the model was then applied to the assessment of ejection fraction in 933 patients without prior myocardial infarction who underwent exercise radionuclide ventriculography. In 196 patients who were referred to angiography, the conventional abnormal ejection fraction criterion--an absolute rise of less than 0.05 with exercise--had a sensitivity of 79% and a specificity of 68% when referenced to coronary angiography. In 737 patients who were not referred for angiography and who were analyzed instead by our probabilistic model, sensitivity was 63% (p = 0.004 compared to that in the 119 angiographically diseases patients) and specificity was 79% (p = 0.036 compared to that in the 77 angiographic normals). Both the higher sensitivity and lower specificity in the catheterized patients are consistent with a preferential referral of positive test responders to angiography, and of negative test responders away from angiography. The distortion of test sensitivity and specificity which results from this selection bias can be circumvented by substituting a probabilistic estimate of disease for conventional angiographic ascertainment.