Estimating software development effort is frequently based on assessing the effort of one task relative to that of another. The author presents empirical results that show how relative estimation can result in biased assessments of similarity and overly optimistic effort estimates. Specifically, tasks tend to be assessed as more similar than they actually are. Furthermore, the similarity of two tasks depends on the direction of the comparison, and a comparison based on difference in work hours is distinct from one based on a ratio. From these results and other evidence, the author suggests ways to improve the accuracy of relative estimation.