The translog form, if fitted to data exhibiting monotonically decreasing average costs, generally will imply spurious U-shaped average costs. The reason is that the log-log transform of any cost function characterized by monotonically declining average costs is strictly convex in the logarithm of size except in limited cases. Simulations suggest that the severity of the bias is exacerbated by a realistically skewed sample distribution. Similar properties of alternative forms also are considered. Estimates fitted to three samples of U.S. banks bear out the potential importance of these concerns.