PURPOSE: We explored health differences between population groups who describe their health as excellent, very good, good, fair, or poor. METHODS: We used data from a population-based survey which included self-rated health (SRH) and three global measures of health: the SF36 general health score (computed from the 4 items other than SRH), the EQ-5D health utility, and a visual analogue health thermometer. We compared health characteristics of respondents across the five health ratings. RESULTS: Survey respondents (N = 1.844, 49.2 % response) rated their health as excellent (12.2 %), very good (39.1 %), good (41.9 %), fair (6.0 %), or poor (0.9 %). The means of global health assessments were not equidistant across these five groups, for example, means of the health thermometer were 95.8 (SRH excellent), 88.8 (SRH very good), 76.6 (SRH good), 49.7 (SRH fair), and 33.5 (SRH poor, p < 0.001). Recoding the SRH to reflect these mean values substantially improved the variance explained by the SRH, for example, the linear r (2) increased from 0.50 to 0.56 for the health thermometer if the SRH was coded as poor = 1, fair = 2, good = 3.7, very good = 4.5, and excellent = 5. Furthermore, transitions between response options were not explained by the same health-related characteristics of the respondents. CONCLUSIONS: The adjectival SRH is not an evenly spaced interval scale. However, it can be turned into an interval variable if the ratings are recoded in proportion to the underlying construct of health. Possible improvements include the addition of a rating option between good and fair or the use of a numerical scale instead of the classic adjectival scale.