The aim of the study reported here was to estimate the sensitivity and specificity of the gamma interferon (IFN-gamma) ELISA for paratuberculosis in sheep using receiver-operating characteristic analysis. Bacteriologic culture of tissues was used to define the reference positive population (n = 33). Two reference negative populations were used: culture-negative sheep from infected flocks (n = 77), and sheep from noninfected flocks (n = 358). We also evaluated the accuracy of 2 Mycobacterium avium subspecies paratuberculosis (MAP) antigen preparations, a whole-cell sonicate (MpS) and a johnin purified-protein derivative (PPDj). The source of the reference negative sheep used in the analysis affected overall accuracy of the IFN-gamma ELISA. The area under the curve was 0.683 (95% confidence interval 0.574-0.787), using culture-negative sheep from infected flocks, was 0.831 (0.764-0.889), using sheep from noninfected flocks for the MpS, and was 0.809 (0.726-0.881) and 0.897 (0.862-0.925) for the PPDj, respectively. Using the MpS, the cut point that classified the most sheep correctly was an optical density reading of 0.20, for sensitivity of 40.7% (19.4-57.6) and specificity of 88.7% (77.0-95.7) or 97.6% (93.04-99.5), depending on the reference negative population used. Using the PPDj, the cut point that classified the most sheep correctly was 0.25 for sensitivity of 66.7% (47.2-82.7) and specificity of 93.5% (85.5-97.9) or 98.3% (96.4-99.4), respectively. The PPDj was more accurate at identifying MAP-infected sheep than was the MpS (P = 0.034).