Children aged three to six were shown drawings of three objects that were identical except for differences in size in one or both dimensions. Each child was asked to determine if one object was the tallest. A missing feature model was not supported, since a child was often correct for some items and incorrect for others. Errors were partially predictable by certain properties of the objects, such as preponderance of curved or straight lines. In a second experiment, the same children were presented with a different set of pictures that included some from the first experiment, now identified with new names (e.g., "three houses" became "three arrows"). Identification of tallness followed the same pattern as in the first experiment, with many children providing different responses for identical but differently labeled items. The older children made only half as many reversal errors as the younger; both groups erred systematically, but according to different criteria. This research supports the idea that concepts such as "tall" develop not by the progressively restrictive accrual of features, but by a progression from idiosyncratic, object-bound attribution to an eventually universalized class-bound attribution. (JB)