Cross-country labor productivity differences are larger in agriculture than in non-agriculture. We propose a new explanation for these patterns in which the self-selection of heterogeneous workers determines sector productivity. We formalize our theory in a generalequilibrium Roy model in which preferences feature a subsistence food requirement. In the model, subsistence requirements induce workers that are relatively unproductive at agricultural work to nonetheless select into the agriculture sector in poor countries. When parameterized, the model predicts that productivity differences are roughly twice as large in agriculture as non-agriculture even when countries differ by an economy-wide efficiency term that affects both sectors uniformly. (JEL J24, J31, J43, O11, O13, O40) Cross-country labor productivity differences are much larger in agriculture than in the non-agricultural sector (Caselli 2005; Restuccia, Yang, and Zhu 2008). Because developing countries have most of their workers in agriculture, their low productivity in agriculture accounts for nearly all of their low productivity in the aggregate. This implies that understanding why productivity differences in agriculture are so large compared to those of the non-agricultural sector is at the heart of understanding world income inequality. 1 In this article we propose a new explanation for these productivity patterns in which the self-selection of heterogeneous workers determines sector productivity. We start from the well-known idea that in poor countries, where economy-wide efficiency is low, most people must work in the agricultural sector in order to satisfy subsistence consumption needs. This is what Schultz (1953) famously called the "food problem." Our insight is that precisely because the majority of workers in 1 Versions of this argument have been made by Caselli (2005); Restuccia, Yang, and Zhu (). For helpful comments we thank the three anonymous referees,