Comparisons of Web, telephone and face-to-face surveys have conflated differences in sample composition (who participates in each mode) with mode effects (how the same person would respond to questions administered using different modes). Using an experimental design that eliminates differences in sample composition, we estimate mode differences between a self-administered Web survey and a face-to-face interview. The design recruited a diverse set of participants at a central location, secured agreement to participation, and only then randomized interview mode for each participant. To improve efficiency, respondents were randomized in blocks, further controlling differences between the two treatment groups. The experiment demonstrates that there are large mode differences between selfadministered Web interviews and face-to-face interviews. The face-to-face interview causes respondents to moderate attitudes, acquiesce, or give socially appropriate answers to tough questions. The effects are especially pronounced for respondents with low cognitive skills, suggesting that low observed attitude constraint in this group may be an artifact of survey mode.