We examine dynamic choice behavior in a natural experiment with large stakes and a demographically divers sample. The television game show Deal Or No Deal offers a rich paradigm to examine the latent decision processes that people use to make choices under uncertainty when they face future options linked to current choices. We have four major findings. First, we show that popular utility functions that assume constant relative or absolute risk aversion and expected utility theory defined over the prizes cannot characterize these choices, which exhibit increasing relative risk aversion over prizes ranging from a penny to nearly half a million U.S. dollars. Second, the argument of the utility function under expected utility theory reflects the integration of game show prizes with regular income. These decision makers do not segregate the income from the lotteries they face on the game show from the income that they bring to the game show. Allowing for this integration of income and game show prizes leads to choice behavior consistent with constant relative risk aversion. Third, we examine th e effects of allowing contestants to make choices characterized by non-standard decision models. We find evidence of some probability weighting, but no loss aversion. We also find evidence that contestants make decisions as if using more than one latent criteria, mixing traditional utility evaluations, probability weighting, and aspiration levels. Fourth, we design and implement laboratory experiments patterned after the natural experiment, to gauge how qualitatively reliable the lab inferences are in the same type of dynamic choice task. We find that choices in the lab are dramatically different in one respect – subjects in those tasks do segregate the income from their prizes from their extra-lab income, in contrast to game show contestants who integrate the two.