OBJECTIVE: To assess the test-retest reliability (repeatability) of Borg's 6-20 rating of perceived exertion (RPE) scale using a more appropriate statistical technique than has been employed in previous investigations. The RPE scale is used widely in exercise science and sports medicine to monitor and/or prescribe levels of exercise intensity. The "95% limits of agreement" technique has recently been advocated as a better means of assessing within-subject (trial to trial) agreement than traditional indicators such as Pearson and intraclass correlation coefficients. METHODS: Sixteen male athletes (mean (SD) age 23.6 (5.1) years) completed two identical multistage (incremental) treadmill running protocols over a period of two to five days. RPEs were requested and recorded during the final 15 seconds of each three minute stage. All subjects successfully completed at least four stages in each trial, allowing the reliability of RPE responses to be examined at each stage. RESULTS: The 95% limits of agreement (bias +/- 1.96 x SDdiff) were found to widen as exercise intensity increased: 0.88 (2.02) RPE units (stage 1), 0.25 (2.53) RPE units (stage 2), -0.13 (2.86) RPE units (stage 3), and -0.13 (2.94) RPE units (stage 4). Pearson correlations (0.81, 0.72, 0.65, and 0.60) and intraclass correlations (0.82, 0.80, 0.77, and 0.75) decreased as exercise intensity increased. CONCLUSIONS: These findings question the test-retest reliability of the RPE scale when used to monitor subjective estimates of exercise intensity in progressive (or graded) exercise tests.