An experiment by Tversky and Kahneman (1981) illustrates that people's tendency to evaluate risky decisions separately can lead them to choose combinations of choices that are first-order stochastically dominated by other available combinations. We investigate the generality of this effect both theoretically and experimentally. We show that for any decisionmaker who does not have constant-absolute-risk-averse preferences and who evaluates her decisions one by one, there exists a simple pair of independent binary decisions where the decisionmaker will make a dominated combination of choices. We also characterize, as a function of a person's preferences, the amount of money that she can lose due to a single mistake of this kind. The theory is accompanied by both a real-stakes laboratory experiment and a large-sample survey from the general U.S. population. Replicating Tversky and Kahneman's original experiment where decisionmakers with prototypical prospect-theory preferences will choose a dominated combination, we find that 28% of the participants do so. In the survey we ask the respondents about several hypothetical large-stakes choices, and find higher proportions of dominated choice combinations. A statistical model that estimates preferences from the survey results is best fit by assuming people have utility functions that are close to prospect-theory value functions and that about 83% of people bracket narrowly. None of these results varies strongly with the personal characteristics of participants. We also demonstrate directly that dominated choices are driven by narrow bracketing: when we eliminate the possibility of narrow bracketing by using a combined presentation of the decisions, the dominated choices are eliminated in the laboratory experiment and are greatly reduced in the survey.