Different methods of eliciting the marginal propensity to consume give very different distributions. Mean MPCs range from below 0.1, indicating life-cycle consumers, to over 0.5, consistent with consumers being hand-to-mouth. We conducted a randomized survey experiment to test if this difference arises because of question wording: we compare using a direct question and a filtered question. Survey wording has large effects on (1) the mean MPC, (2) the extensive margin, and (3) how MPCs vary with payment size, spending horizon, and liquidity. MPCs elicited using a filtered question are much closer to results from using a covariance restriction approach.