From the finding that males are more variable than females on mathematics test scores, it is argued that one might expect a gender-by-item-difficulty interaction such that easy items are easier for females than for males and hard items are harder for females than for males. This prediction was tested in two studies using data from nine forms of a basic skills test in mathematics. The hypothesis was generally supported. For each form, there was a negative correlation between the item difficulty differences (males-females) estimated on males and females separately and the difficulty of the item estimated on the combined sample. Males tended to outperform females on the hardest items; females tended to outperform males on the easiest items. A possible explanation is offered for the finding, one which posits a shift in the ability measured as items become more difficult. Implications for the construction of mathematics tests and the conduct of meta-analyses are discussed.