Data from 166 general hospitals in New York State (1981) is used to estimate a quadratic and logarithmic long-run cost function. Both equations fit the data very well but give very different results. The quadratic appears in confirm the commonly-held view of a shallow U-shaped average cost curve, whereas the log function indicates significant economies of scale: a total and average cost elasticity of 0.9 and -0.10, respectively (using beds or patient days to measure output). Ramsey's RESET test is used to discriminate between the two models and the quadratic is clearly rejected as a misspecification. Scale economies thus exist even where the usual quadratic suggests otherwise.